Microscale three-dimensional wind field retrieval method based on three-dimensional laser wind measurement radar

Abstract

The present application relates to the technical field of wind field inversion, in particular to a micro-scale three-dimensional wind field inversion method based on three-dimensional laser wind measurement radar, comprising obtaining discrete wind field data through multiple radars and performing space-time registration under a unified coordinate system to obtain spatial coordinates and resolution difference characteristics; positioning the missing boundary based on the characteristic gradient and marking it as a mask area; extracting local data patches of fixed dimensions and dividing and processing them into visible non-mask data and to-be-reconstructed mask data; performing three-dimensional relative position coding calculation to obtain patch feature tensors and generating mask position tensors in combination with the mask area; inputting the patch feature tensors into a pre-trained three-dimensional space generation model to extract hidden layer features and input them into a decoder to perform reconstruction mapping constrained by physical rules and generate reconstructed three-dimensional wind fields. The present application generates a dynamic mask by extracting a resolution difference gradient and introduces a fluid mechanics continuity equation as a loss constraint to realize continuous wind field inversion.

Description

Technical Field

[0001] This invention relates to the field of wind field inversion technology, specifically to a microscale three-dimensional wind field inversion method based on three-dimensional laser wind radar. Background Technology

[0002] Three-dimensional laser wind radar, due to its high spatiotemporal resolution, high precision, and excellent detection capabilities in clear-sky atmospheres, is playing an increasingly important role in wind field detection, aviation meteorological support, micro-site selection of wind farms, and wake analysis. The core technology for three-dimensional wind field inversion is to transform one-dimensional discrete radial wind speed data acquired by radar into a three-dimensional continuous wind field vector. However, due to the limited detection range of a single radar, practical applications often require the overlapping and stitching of data from multiple radars, or the processing of wind field transitions from large-scale background to micro-scale local areas (such as wake boundaries). Existing processing methods, when faced with overlapping detection boundaries, radar sector scanning edges, and multi-scale transition regions, often resort to grid-based forced interpolation due to extremely mismatched observation resolutions. This rigid interpolation method disrupts the physical continuity of the flow field and lacks the ability to explore the topological interaction patterns of tangential and radial spatial directions. As a result, the three-dimensional flow field loses its topological integrity at the boundaries, easily leading to abrupt breaks in the wind field vector direction and large-area data cliffs.

[0003] To address this, a microscale three-dimensional wind field inversion method based on three-dimensional laser wind radar is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide a microscale three-dimensional wind field inversion method based on three-dimensional laser wind radar to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A microscale three-dimensional wind field inversion method based on three-dimensional laser wind radar includes: Discrete wind field data is acquired by multiple three-dimensional laser wind radars, including radial wind speed and corresponding radar line-of-sight direction vector; the discrete wind field data is spatiotemporally registered in a unified coordinate system to obtain spatial coordinate and resolution difference features; based on the resolution difference features, the defect boundary interval is located and the defect boundary interval is marked as a mask region. Based on the spatial coordinates, local data blocks of fixed dimensions are extracted from the discrete wind field data, and the local data blocks are divided into visible non-masked data and masked data to be reconstructed according to the mask region. Based on the spatial coordinates and the radar line-of-sight direction vector, a three-dimensional relative position encoding calculation is performed on the local data blocks to obtain a block feature tensor containing geometric projection information, and a mask position tensor is generated by combining the mask region. The extracted feature tensor is input into a pre-trained 3D spatial generation model for hidden layer feature extraction. The extracted hidden layer features are then merged with the mask position tensor and input into the decoder of the 3D spatial generation model to perform a reconstruction mapping constrained by physical rules, thereby generating the reconstructed 3D wind field corresponding to the mask region.

[0006] Preferably, the discrete wind field data consists of radial wind speed point sets and their corresponding radar line-of-sight vectors obtained by multiple three-dimensional laser wind measuring radars in a spherical coordinate system during original scanning. The radar line-of-sight vectors include spatial azimuth and elevation information. The multiple three-dimensional laser wind measuring radars are deployed at different spatial locations within the target detection area, with their scanning centers pointing towards the target detection area, so that the scanning fields of multiple radars form a three-dimensional spatial overlap area within the target detection area. Each radar performs coordinated scanning of azimuth and elevation angles through a dual-axis optical scanning mirror to obtain radial wind speed point sets in different line-of-sight directions within the target detection area.

[0007] Preferably, the unified coordinate system is a three-dimensional Cartesian coordinate system; the spatiotemporal registration includes spatial registration and temporal registration. The spatial registration process involves mapping the radial wind speed point sets of each radar in the original spherical coordinate system to the three-dimensional Cartesian coordinate system using the coordinate transformation formula from spherical coordinate system to three-dimensional Cartesian coordinate system to generate spatial coordinates; the temporal registration process involves performing time synchronization interpolation on the scanning data of each radar according to their respective timestamp information to align the data points of each radar to the same time reference; the resolution difference feature acquisition process involves calculating the local spatial resolution index of the discrete wind field data of each radar in the three-dimensional spatial overlap area formed by the scanning fields of multiple three-dimensional laser wind measuring radars. The local spatial resolution index includes at least one of scanning point density, adjacent point spacing, and angle sampling interval. The resolution difference feature is generated by comparing the local spatial resolution indexes of different radars point by point.

[0008] Preferably, the mask region acquisition process involves calculating the spatial gradient within the three-dimensional spatial region corresponding to the discrete wind field data based on the resolution difference feature, obtaining the resolution difference gradient value at each spatial location; marking the defect boundary points based on the resolution difference gradient values, and forming the defect boundary interval by the defect boundary points; assigning mask identifiers to all spatial locations within the defect boundary interval in a three-dimensional Cartesian coordinate system to generate the mask region; the mask identifier is a binary identifier, with the mask identifier corresponding to the spatial location within the defect boundary interval set to a first identifier value, and the mask identifier corresponding to the spatial location outside the defect boundary interval set to a second identifier value, wherein the first identifier value and the second identifier value are 0 and 1, respectively.

[0009] Preferably, the local data block is a three-dimensional data sub-block containing a fixed number of grid nodes extracted from the discrete wind field data after three-dimensional spatial meshing in a three-dimensional Cartesian coordinate system based on the spatial coordinates according to a preset voxel size and block step size; the three-dimensional data sub-block has a preset number of grid dimensions along the x-axis, y-axis and z-axis directions, and the number of grid dimensions in each dimension is consistent in the local data block of the fixed dimension; the local data block is divided into visible non-masked data and masked data to be reconstructed according to the mask region.

[0010] Preferably, the process of dividing the visible non-masked data and the masked data to be reconstructed is to match and determine each spatial position in the local data block based on the mask identifier. The determination process is to divide the data at the spatial position with the first identifier value into the masked data to be reconstructed and set its value to zero; divide the data at the spatial position with the second identifier value into the visible non-masked data, retain its original value, and construct an input tensor with complete spatial dimensions.

[0011] Preferably, the three-dimensional relative position encoding is based on the spatial coordinates of each grid node within the local data block. The relative coordinates of each grid node are calculated with the center position of each block as the origin and normalized. Then, the normalized relative coordinates in the x-axis, y-axis, and z-axis directions are encoded respectively using a sinusoidal position encoding function. The encoding results of each axis are spliced ​​and fused along the channel dimension to generate a three-dimensional position encoding tensor that matches the spatial dimension of the local data block. The block feature tensor is formed by splicing the radial wind speed and radar line-of-sight vector corresponding to the local data block with the three-dimensional position encoding tensor along the channel dimension.

[0012] Preferably, the 3D spatial generation model is an encoder-decoder network constructed based on a 3D convolutional neural network. The encoder is composed of stacked 3D convolutional downsampling layers, used to extract multi-scale hidden layer features from the input segmented feature tensor. The decoder is composed of stacked 3D transposed convolutional upsampling layers, which fuse multi-scale features with the encoder through cross-layer skip connections. This decoder merges the hidden layer features with the mask position tensor and performs conditional decoding mapping to generate the reconstructed 3D wind field corresponding to the mask region. The loss function of the 3D spatial generation model during the training phase includes a data reconstruction loss term and a physical constraint loss term. The physical constraint loss term is constructed based on the fluid dynamics continuity equation, using the calculated 3D spatial divergence of the output reconstructed 3D wind field as a penalty term to constrain the output reconstructed 3D wind field of the 3D spatial generation model to satisfy the fluid mass conservation theorem.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention extracts local spatial resolution difference features in the overlapping area of ​​multiple three-dimensional laser wind measurement radars, and accurately locates the missing boundary intervals based on the spatial gradient of these features to generate a dynamic mask. This changes the traditional inversion method's processing logic of blindly relying on global forced interpolation when facing multi-scale transition regions. This mechanism adaptively identifies and accurately isolates distorted boundaries with extremely mismatched observation resolutions, thereby effectively avoiding the amplification of boundary errors caused by direct splicing of heterogeneous observation data, and realizing reasonable connection and smooth transition of multi-source radar discrete data in spatial scale.

[0014] 2. This invention constructs a spatially complete input tensor by performing masking and zeroing on local data blocks containing missing boundaries, and deeply integrates the radar line-of-sight vector with the three-dimensional relative position code, overcoming the defect of traditional rigid interpolation methods that destroy the physical continuity of the flow field. This scheme not only completely preserves the original spatial physical topology of the three-dimensional mesh, but also enables the three-dimensional spatial generation model to fully extract and restore the tangential and radial interaction law of radial wind speed in three-dimensional geometric space, fundamentally ensuring the topological integrity of the three-dimensional spatial flow field at the edge of the radar scanning sector and the boundary of multiple radar overlap.

[0015] 3. This invention introduces the fluid dynamics continuity equation as a physical constraint during the training phase of the 3D spatial generation model, and adds the 3D spatial divergence of the output reconstructed 3D wind field as a penalty term to the loss function. This solves the problem that traditional pure data-driven or conventional mesh completion algorithms are prone to producing abrupt changes in wind field vector direction and data cliffs that do not conform to physical common sense. The physical information neural network architecture forces the decoding and reconstruction process to strictly follow the fluid mass conservation law, ensuring that the generated 3D microscale wind field does not diverge or become distorted in terms of underlying physical laws, and obtains a 3D continuous vector field that conforms to the laws of real fluid dynamics. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the process for microscale three-dimensional wind field inversion based on three-dimensional laser wind radar; Figure 2 A schematic diagram of the spatiotemporal registration and feature acquisition process; Figure 3 This is a schematic diagram of the logic flow for generating a 3D model and its physical constraints. Detailed Implementation

[0017] 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.

[0018] Please see Figure 1 This invention provides a microscale three-dimensional wind field inversion method based on three-dimensional laser wind radar, and the technical solution is as follows: A microscale three-dimensional wind field inversion method based on three-dimensional laser wind radar includes: Discrete wind field data is acquired by multiple three-dimensional laser wind radars, including radial wind speed and corresponding radar line-of-sight direction vector; the discrete wind field data is spatiotemporally registered in a unified coordinate system to obtain spatial coordinate and resolution difference features; based on the resolution difference features, the defect boundary interval is located and the defect boundary interval is marked as a mask region. Based on the spatial coordinates, local data blocks of fixed dimensions are extracted from the discrete wind field data, and the local data blocks are divided into visible non-masked data and masked data to be reconstructed according to the mask region. The masked data to be reconstructed is then processed by setting the mask to zero. Based on the spatial coordinates and the radar line-of-sight direction vector, the local data blocks are subjected to three-dimensional relative position encoding calculation to obtain a block feature tensor containing geometric projection information, and a mask position tensor is generated by combining the mask region; the block feature tensor is extracted and input into a pre-trained three-dimensional space generation model for hidden layer feature extraction, and the extracted hidden layer features are merged with the mask position tensor and input into the decoder of the three-dimensional space generation model to perform a reconstruction mapping constrained by physical rules, thereby generating the reconstructed three-dimensional wind field corresponding to the mask region.

[0019] Example 1: Discrete wind field data is acquired by multiple three-dimensional laser wind radars, including radial wind speed and the corresponding radar line-of-sight vector. The discrete wind field data consists of radial wind speed point sets and their corresponding radar line-of-sight vectors obtained by multiple three-dimensional laser wind measuring radars in spherical coordinates. The radar line-of-sight vectors include spatial azimuth and elevation information. The multiple three-dimensional laser wind measuring radars are deployed at different spatial locations in the target detection area, with their scanning centers pointing towards the target detection area, so that the scanning fields of multiple radars form a three-dimensional spatial overlap area within the target detection area. Each radar performs coordinated scanning of azimuth and elevation angles through a dual-axis optical scanning mirror to obtain radial wind speed point sets in different line-of-sight directions in the target detection area.

[0020] Specifically, multiple unobstructed stations are selected around and within the target detection area. The three-dimensional spatial coordinates of each radar location and the device's north reference angle are measured and calibrated using a global navigation satellite system and a total station. The configuration of the multiple three-dimensional laser wind-measuring radars is at least two units; in preferred embodiments, two or three radars are typically used for network detection. After deployment, the hardware optical orientation of each radar is adjusted so that the scanning center points to the center of the target detection area. The maximum detection range and scanning angle parameters of each radar are set and constrained by the internal main control computer to ensure that the laser beams emitted by multiple radars physically intersect and cover each other within the target detection area. This creates a three-dimensional spatial overlap zone within the target detection area, ensuring that any air mass within the target detection area can be simultaneously detected by at least two radars at different locations, detecting backscattered signals from different angles. The dual-axis optical scanning mirror includes a horizontal rotation mechanism and a pitch deflection mechanism. During the acquisition process, the servo control unit inside the radar synchronously sends drive pulses to the azimuth servo motor controlling the horizontal rotation mechanism and the pitch servo motor controlling the pitch deflection mechanism according to a preset three-dimensional volume scanning strategy. The azimuth servo motor drives the optical mirror to change the emission azimuth of the laser beam at a set angular velocity, completing the spatial azimuth angle scan. At the same time, the pitch servo motor drives the mirror to change the pitch state of the laser emission at a set step angle, completing the pitch angle scan.

[0021] The laser pulse emitted by the laser is incident on the atmosphere after passing through a dual-axis optical scanning mirror. It is then scattered by wind-driven aerosol particles, generating a backscattered echo signal with a polarization frequency shift, which is captured by the radar receiving telescope. The radar's digital signal processing board performs analog-to-digital conversion and fast Fourier transform on the received mixed signal to extract the center value of the Doppler frequency shift caused by the aerosol particle motion. Based on the Doppler effect formula, the velocity of atmospheric aerosols at the corresponding distance from the laser emission point along the laser emission direction is calculated as follows: ; in, Represents radial wind speed; This represents the operating wavelength emitted by the laser. As the dual-axis optical scanning mirror continuously scans within the overlapping area of ​​three-dimensional space, the radar cyclically performs extraction and calculation at each distance from the gate and angle step point to calculate the radial wind speed at each measurement point. An independent local spherical coordinate system is established with the optical transmission center of each radar as the origin. While calculating the radial wind speed at each point, the photoelectric encoder data on the dual-axis optical scanning mirror is read in real time and simultaneously latched. The pulse count value of the horizontal encoder is converted into the spatial azimuth angle of the current laser beam, and the pulse count value of the elevation encoder is converted into the elevation angle of the current laser beam. Based on the spatial azimuth and elevation angle information, the unit radar line-of-sight direction vector for each measurement point is calculated using the following formula: ; in, Represents the radar line-of-sight direction vector. Represents spatial azimuth. The pitch angle represents the radial distance, the spatial azimuth angle, the radar line-of-sight direction vector calculated from the pitch angle, and the radial wind speed obtained at the same time are time-stamped and spatially aligned to generate discrete wind field data. By networking multiple three-dimensional laser wind-measuring radars and forming a three-dimensional spatial overlap zone in the target area, the traditional assumption that single-radar wind measurement relies on a large-scale uniform wind field is broken. This enables multi-angle independent physical observation of the same air mass in a complex microscale flow field, thus providing sufficient boundary conditions to overcome the bottleneck of only being able to obtain radial wind speed and to solve the three-dimensional wind field vector.

[0022] See Figure 2 Spatiotemporal registration of the discrete wind field data is performed under a unified coordinate system to obtain spatial coordinate and resolution difference features; The unified coordinate system is a three-dimensional Cartesian coordinate system; the spatiotemporal registration includes spatial registration and temporal registration. The spatial registration process involves mapping the radial wind speed point sets of each radar in the original spherical coordinate system to the three-dimensional Cartesian coordinate system using the coordinate transformation formula from spherical coordinate system to three-dimensional Cartesian coordinate system to generate spatial coordinates; the temporal registration process involves performing time synchronization interpolation on the scanning data of each radar according to their respective timestamp information to align the data points of each radar to the same time reference; the resolution difference feature acquisition process involves calculating the local spatial resolution index of the discrete wind field data of each radar in the three-dimensional spatial overlap area formed by the scanning fields of multiple three-dimensional laser wind measuring radars. The local spatial resolution index includes at least one of scanning point density, adjacent point spacing, and angle sampling interval. The resolution difference feature is generated by comparing the local spatial resolution indexes of different radars point by point.

[0023] Specifically, the bottom center physical calibration point of the target detection area is set as the origin of the three-dimensional Cartesian coordinate system, with due east as the positive horizontal axis, due north as the positive vertical axis, and the vertically upward direction as the positive vertical axis. The spatial registration process is as follows: using geodetic equipment to obtain the preset absolute installation center coordinates of each radar in the three-dimensional Cartesian coordinate system, the radial wind speed point set of each radar in the original spherical coordinate system is mapped point by point to the three-dimensional Cartesian coordinate system through the coordinate transformation formula from the spherical coordinate system to the three-dimensional Cartesian coordinate system, and the three-dimensional spatial coordinates of each wind field data point are calculated. The coordinate transformation formula is: ; ; ; in, , and Represents the three-dimensional spatial coordinates generated after mapping. and Representing the The absolute installation center coordinates of the radar Represents the radial distance of the data points. Represents spatial azimuth; Represents pitch angle; The time registration process involves performing time synchronization interpolation on the scanning data of each radar according to its respective timestamp information, aligning the data points of each radar to the same time reference. In this embodiment, the same time reference is preset to an equally spaced integer-second standard time sequence, with a preset time alignment step size of 1 second. Using each integer-second standard time node as a reference, for any valid data point in a unified coordinate system, three adjacent radial wind speed points carrying physical timestamp information are searched and extracted from the discrete wind field data of a single radar, denoted as (…). ), ( )and( ),in, Represents a timestamp. This represents the corresponding radial wind speed; This embodiment presupposes the use of a second-order Lagrange polynomial interpolation algorithm to construct a mathematical model, which will be used to determine the current whole-second standard time node. Substituting the values ​​into the mathematical formula and performing a smooth fit, the scalar value of the wind speed aligned to that node is calculated. : ; Eliminate the asynchronous time difference of electromechanical scanning, so that all acquisition points of each radar are strictly aligned to the same preset time reference in the time dimension.

[0024] The three-dimensional spatial overlap area under a unified coordinate system is pre-defined as a three-dimensional voxel grid with a side length of 10m. Within each independent 10m three-dimensional voxel grid, the data of each radar is quantitatively calculated. In this embodiment, the local spatial resolution index is pre-defined to include three items: scan point density, adjacent point spacing, and angle sampling interval. The pre-defined calculation method for scan point density is to directly count the total number of valid data points of a single radar falling within the current voxel grid. The pre-defined calculation method for adjacent point spacing is to extract the nearest neighbor three-dimensional Euclidean distance between all point clouds of the radar within the current grid and calculate the arithmetic mean. The pre-defined calculation method for angle sampling interval is to calculate the average absolute value of the difference in spatial azimuth angle and the difference in elevation angle between adjacent detection beams within the grid. After the calculation of each indicator is completed, the local spatial resolution indicators corresponding to each of the multiple radars in the same three-dimensional voxel grid are extracted and compared point by point. The preset comparison method in this embodiment is to calculate the absolute value of the difference and the ratio deviation between the corresponding indicators of different radars, and combine the absolute value of the difference and the ratio deviation in all voxel grids in the whole space into a three-dimensional tensor matrix, and output the resolution difference feature in this matrix form. By establishing a three-dimensional Cartesian coordinate system and performing spatial mapping, the geometric perspective differences of multi-source radar observations are eliminated, achieving unified alignment of discrete data from multiple radars in physical space. In terms of time registration, a second-order Lagrange polynomial interpolation algorithm is introduced to eliminate the asynchronous time difference of electromechanical scanning, effectively preserving the dynamic characteristics of the nonlinear smooth evolution of the wind field and avoiding the physical distortion caused by traditional linear interpolation. In addition, by multidimensionally quantizing the local spatial resolution index and generating a three-dimensional feature tensor matrix, the uneven sampling density of the overlapping area caused by detection distance and beam divergence is objectively characterized.

[0025] Based on the resolution difference features, the defect boundary interval is located and the defect boundary interval is marked as a mask region; The mask region acquisition process involves calculating the spatial gradient within the three-dimensional spatial region corresponding to the discrete wind field data based on the resolution difference characteristics, obtaining the resolution difference gradient value at each spatial location; marking the defect boundary points based on the resolution difference gradient values, and forming the defect boundary interval by the defect boundary points; assigning mask identifiers to all spatial locations within the defect boundary interval in a three-dimensional Cartesian coordinate system to generate the mask region; the mask identifier is a binary identifier, with the mask identifier corresponding to the spatial location within the defect boundary interval set to a first identifier value, and the mask identifier corresponding to the spatial location outside the defect boundary interval set to a second identifier value, wherein the first identifier value and the second identifier value are 0 and 1, respectively. Specifically, the spatial gradient calculation employs a three-dimensional Sobel operator. In this embodiment, the three-dimensional Sobel operator uses... A three-dimensional discrete convolution kernel is used; in the three directions of the horizontal, vertical, and axial axes of the three-dimensional Cartesian coordinate system, the three-dimensional discrete convolution is performed to differentiate the resolution difference features represented as a three-dimensional tensor matrix, respectively, to calculate the partial derivative values ​​of the three-dimensional voxel grid at each spatial location in the three spatial dimensions. In this embodiment, the square root of the sum of the squares of the partial derivatives in the three directions is assumed to be the synthesized gradient magnitude, and the specific calculation formula is as follows: ; in, This represents the gradient value of the resolution difference at that spatial location. , and These represent the partial derivative values ​​of the resolution difference feature at that spatial location along the horizontal, vertical, and triangular axes of the three-dimensional Cartesian coordinate system, respectively. A gradient judgment threshold is set based on the background noise level and inherent measurement error of multiple 3D laser wind measuring radars to filter out minor resolution fluctuations caused by normal measurement errors of the radar system. The resolution difference gradient value at each spatial location is compared with the gradient judgment threshold. When the resolution difference gradient value at a certain spatial location is greater than or equal to the gradient judgment threshold, the coordinate attribute of that spatial location is marked as a missing boundary point, thereby extracting a boundary point cloud set composed of multiple missing boundary points within the 3D spatial region. The missing boundary points enclose the missing boundary interval. In specific implementation, a 3D concave hull algorithm (preferably the Alpha-shape algorithm) is used to reconstruct the geometric surface of all missing boundary points in the boundary point cloud set. The missing boundary points located at the edges are extracted as vertices, constructing a closed 3D polyhedron that closely follows the true topological shape of the missing region. The internal 3D spatial geometry enclosed by this 3D polyhedron forms the missing boundary interval. A ray intersection determination algorithm is used to traverse each spatial location within the 3D spatial region mesh: a virtual ray is emitted from the currently traversed spatial location in any specified direction, and the number of geometric intersections between the virtual ray and the outer surface of the defect boundary interval (i.e., each plane of the 3D polyhedron bounding box) is calculated. If the calculated number of intersections is odd, the current spatial location is determined to be within the defect boundary interval, and the mask identifier corresponding to the spatial location within the defect boundary interval is set to a first identifier value; if the calculated number of intersections is even or zero, the current spatial location is determined to be outside the defect boundary interval, and the mask identifier corresponding to the spatial location outside the defect boundary interval is set to a second identifier value. In this embodiment, the first identifier value and the second identifier value are set to 0 and 1, respectively. That is, through the above-mentioned ray intersection determination algorithm, spatial positions determined to be located within the defect boundary interval are uniformly assigned a mask identifier of 0, and spatial positions determined to be located outside the defect boundary interval are uniformly assigned a mask identifier of 1; through the above-mentioned spatial position traversal and binary assignment calculation, a three-dimensional binary tensor matrix that is completely consistent with the grid dimension and coordinates of the three-dimensional spatial region is finally output. The three-dimensional binary tensor matrix and the spatial range corresponding to all nodes containing the first identifier value of 0 together generate the mask region; By combining the three-dimensional Sobel operator with a judgment threshold set based on background noise, the abrupt boundary of wind field resolution is effectively extracted, and the interference of normal measurement errors of the system is reduced. The three-dimensional concave hull algorithm is used to construct a closed polyhedron, which better adapts to the irregular real defect morphology caused by complex terrain or occlusion, and reduces the misjudgment of effective observation data compared with the traditional convex hull algorithm. The three-dimensional binary mask tensor matrix is ​​generated by combining the ray cross judgment algorithm, which realizes accurate spatial positioning and isolation of the defect area.

[0026] Based on the spatial coordinates, local data chunks of fixed dimensions are extracted from the discrete wind field data; The local data block is a three-dimensional data sub-block containing a fixed number of grid nodes extracted from the discrete wind field data after three-dimensional spatial meshing in a three-dimensional Cartesian coordinate system based on the spatial coordinates and according to a preset voxel size and block step size. The three-dimensional data sub-block has a preset number of grid dimensions along the x-axis, y-axis and z-axis, and the number of grid dimensions in each dimension is consistent in the local data block of the fixed dimension. The local data block is divided into visible non-masked data and masked data to be reconstructed according to the mask region.

[0027] The process of dividing the visible non-masked data and the masked data to be reconstructed is to match and determine each spatial position in the local data block based on the mask identifier. The determination process is to divide the data at the spatial position with the first identifier value into the masked data to be reconstructed and set its value to zero; and divide the data at the spatial position with the second identifier value into the visible non-masked data, retain its original value, and construct an input tensor with complete spatial dimensions. Specifically, based on the spatial coordinates, a three-dimensional spatial mesh is generated in a three-dimensional Cartesian coordinate system according to a preset voxel size and dicing step size. The preset voxel size is essentially the minimum discrete physical control volume characterizing the spatial resolution of the reconstructed three-dimensional wind field. Its specific value is determined by a comprehensive evaluation based on the lower limit of the hardware physical distance resolution of the multiple three-dimensional laser wind measuring radars, the characteristic physical scale of the microscale flow field within the target detection area, and the computational cost of the three-dimensional spatial generation model. In this embodiment, the voxel size is 10 meters in all three directions (horizontal, vertical, and longitudinal). The preset dicing step size is the distance of 16 mesh nodes (i.e., a sliding step size of 160 meters in actual physical space, to ensure 50% data overlap between adjacent dicings, thereby maintaining the continuity of spatial features). The specific meshing process is as follows: obtain the minimum spatial boundary coordinates of the detection area corresponding to the discrete wind field data, and use the following coordinate transformation formula to convert the continuous spatial coordinates into discrete three-dimensional mesh index nodes: ; in, , () represents spatial coordinates. , () is the minimum coordinate reference point for the boundary of the detection area. For voxel size, This represents the floor function. The mapped 3D mesh index corresponding to this spatial location; A three-dimensional sliding traversal is performed on the global spatial grid according to the stated block size to extract three-dimensional data sub-blocks containing a fixed number of grid nodes from the discrete wind field data, which are then used as the local data blocks. In this embodiment, the grid dimensions of the three-dimensional data sub-blocks are preset to be 32 along the horizontal, vertical, and longitudinal axes, and the number of grid dimensions in each dimension remains consistent across all extracted local data blocks of fixed dimensions. That is, each local data block exhibits a fixed node size. A three-dimensional data cube matrix; based on the pre-generated mask region, the local data block is divided into visible non-masked data and mask data to be reconstructed.

[0028] The process of dividing the visible non-masked data into the masked data to be reconstructed is as follows: Using tensor bit-by-bit alignment, based on the mask identifiers at each spatial location within the masked region, each grid node in the local data block is matched and determined. Specifically, the determination and execution process is as follows: The mask identifiers corresponding to the current grid node within the local data block are read one by one. If the matching determination indicates that the current grid node possesses a first identifier value (in this embodiment, the first identifier value is preset to 0, representing the area within the defect boundary), then the data at that spatial location is divided into masked data to be reconstructed, and the mask to be reconstructed is zeroed out, i.e., the radial wind speed values ​​within it are forcibly erased and uniformly replaced with 0. If the matching determination indicates that the current grid node possesses a second identifier value (in this embodiment, the second identifier value is preset to 1, representing the area outside the defect boundary), then the data at that spatial location is divided into visible non-masked data, completely preserving the radial wind speed values ​​obtained from the original radar scan. By processing all data within the local data block After completing the matching and zeroing / preservation processes at each spatial location, a spatially complete input tensor is constructed. This input tensor strictly maintains the original fixed three-dimensional mesh dimension. Its internal node data consists of visible non-masked data retaining the true physical observation values, and masked data to be reconstructed after normalization and erasure. This achieves data isolation of missing regions without disrupting the original three-dimensional topology of the wind field, enabling the constructed input tensor to directly meet the standardized tensor dimension input requirements of subsequent three-dimensional spatial generation models.

[0029] Based on the spatial coordinates and the radar line-of-sight direction vector, perform three-dimensional relative position encoding calculation on the local data block to obtain the block feature tensor containing geometric projection information, and generate the mask position tensor by combining the mask region; The three-dimensional relative position encoding is based on the spatial coordinates of each grid node within the local data block. The relative coordinates of each grid node are calculated and normalized using the center position of each block as the origin. Then, the normalized relative coordinates in the x, y, and z axes are encoded using a sinusoidal position encoding function. The encoding results of each axis are then concatenated and fused along the channel dimension to generate a three-dimensional position encoding tensor that matches the spatial dimension of the local data block. The block feature tensor is formed by concatenating the radial wind speed and radar line-of-sight vector corresponding to the local data block with the three-dimensional position encoding tensor along the channel dimension.

[0030] Specifically, for any local data block containing a preset number of grid dimensions, all grid nodes contained within it are traversed, and the spatial coordinates of each grid node are read in real time.

[0031] Calculate the center position coordinates of the current local data block and use them as the origin of the relative coordinate system. Specifically, the calculation method is as follows: extract the maximum and minimum spatial coordinates of the current local data block in the three-dimensional Cartesian coordinate system along the horizontal, vertical, and triangular axes, calculate the arithmetic mean of the maximum and minimum spatial coordinates in each axis, and combine the arithmetic mean of each direction as the center position coordinates.

[0032] Subtract the corresponding axial value of the center position coordinate from the spatial coordinates of each grid node along the horizontal, vertical, and longitudinal axes respectively to calculate the relative coordinates of each grid node along the horizontal, vertical, and longitudinal axes.

[0033] After obtaining the relative coordinates of each grid node, normalization is performed on the relative coordinates. The normalization denominator is set to the absolute value of the maximum relative distance on one side of the local data block in physical space. By dividing the relative coordinates of each axis by the normalization denominator, the numerical range of the relative coordinates of each grid node in the horizontal, vertical, and longitudinal directions is linearly mapped to obtain the normalized relative coordinates.

[0034] The normalized relative coordinates along the horizontal, vertical, and axial directions are encoded using a preset sinusoidal position encoding function. The number of position encoding channels for each independent axis is set to [value missing]. For any even-numbered channel index and odd channel index (in The value ranges from 0 to The calculation formula for the sinusoidal position coding function is as follows: (where the integer is an integer). ; ; During encoding, the normalized relative coordinates of the horizontal axis are used as input parameters. Substituting into the above formula, the feature values ​​of all channels are calculated iteratively, and combined to construct the positional encoding feature vector in the horizontal axis direction; similarly, the normalized relative coordinates of the vertical axis and the relative coordinates of the vertical axis are used as input parameters respectively. Substituting into the above formula, the position encoding feature vectors in the vertical axis direction and the vertical axis direction can be calculated independently.

[0035] After encoding all grid nodes along each axis, the encoding results in the horizontal, vertical, and longitudinal directions are concatenated and fused along the channel dimension. After concatenation and fusion, each grid node will obtain a composite relative position feature vector that integrates features from the three directions. By traversing all grid nodes of the entire local data block and performing this concatenation operation, a three-dimensional position encoding tensor whose spatial grid dimension perfectly matches that of the local data block is generated. Extract the radial wind speed scalar data of each grid node within the current local data block after masking and zeroing or preservation; extract the three-dimensional radar line-of-sight direction vector of the corresponding grid node, which is composed of three unit components decomposed along the axis of the three-dimensional Cartesian coordinate system; read the corresponding three-dimensional position encoding tensor. Align the above three types of data according to the node position and sequentially perform matrix concatenation along the channel dimension; without changing the original spatial grid structure, output a block feature tensor containing geometric projection and relative position information, and with the channel dimension fully expanded; By calculating and normalizing relative coordinates with the center of local data blocks as the origin, the spatial position dependence caused by global absolute coordinates is eliminated, giving the 3D spatial generation model excellent spatial translation invariance and cross-regional generalization ability. Furthermore, a sinusoidal position encoding function containing multi-frequency band mapping is introduced to perform high-frequency feature mapping on continuous relative coordinates, which effectively overcomes the defect of conventional neural networks in directly capturing high-frequency spatial details from low-dimensional continuous coordinates (i.e., the spectral bias problem), enabling the model to perceive the topological changes of microscale wind fields at the defect boundary.

[0036] See Figure 3 The segmented feature tensor is extracted and input into a pre-trained 3D space generation model for hidden layer feature extraction. The extracted hidden layer features are then merged with the mask position tensor and input into the decoder of the 3D space generation model to perform a reconstruction mapping constrained by physical rules, thereby generating the reconstructed 3D wind field corresponding to the mask region.

[0037] The 3D spatial generation model is an encoder-decoder network built on a 3D convolutional neural network. The encoder consists of stacked 3D convolutional downsampling layers, used to extract multi-scale hidden layer features from the input segmented feature tensor. The decoder consists of stacked 3D transposed convolutional upsampling layers, which fuse multi-scale features with the encoder through cross-layer skip connections. This decoder merges the hidden layer features with the mask position tensor and performs conditional decoding mapping to generate the reconstructed 3D wind field corresponding to the mask region. The loss function of the 3D spatial generation model during the training phase includes a data reconstruction loss term and a physical constraint loss term. The physical constraint loss term is constructed based on the fluid dynamics continuity equation, using the calculated 3D spatial divergence of the output reconstructed 3D wind field as a penalty term to constrain the output reconstructed 3D wind field of the 3D spatial generation model to satisfy the fluid mass conservation theorem. Specifically, the 3D spatial generation model is an encoder-decoder network built based on a 3D convolutional neural network. In this embodiment, the encoder is presumably composed of three stacked 3D convolutional downsampling layers, each layer sequentially including 3D convolution operations, batch normalization processing, and nonlinear activation processing. The presumed kernel size for the 3D convolution operations is... Spatial downsampling is performed with a stride of 2. As the layers deepen, the size of the input feature map in the three-dimensional spatial dimension is halved layer by layer, while the number of feature channels doubles layer by layer. The decoder is pre-defined as consisting of three stacked 3D transposed convolutional upsampling layers, with a pre-defined size of the 3D transposed convolutional kernel. The step size is set to 2 to recover the spatial dimension size of the feature map layer by layer, and at each corresponding spatial resolution level, a cross-layer skip connection is established with the encoder. Before performing the actual wind field inversion, the three-dimensional spatial generation model is first pre-trained. Specifically, this process involves acquiring high-precision hydrodynamic simulation wind field data or historical high-quality multi-radar joint observation wind field data to construct a real wind field dataset with complete spatial dimensions. This real wind field dataset is then processed according to the same preset voxel size and fixed grid dimensions (size fixed at [value missing]). Spatial partitioning is performed to obtain multiple real wind field segments with complete physical features, which are used as real wind field label data during the model training phase.

[0038] Three-dimensional continuous simulated missing intervals are randomly generated within the internal space of the real wind field block. The wind speed data located within the simulated missing intervals are masked and zeroed to simulate the data missing situation in real observation. At the same time, the positional features of each grid node are calculated according to the aforementioned relative position encoding rules. The zeroed wind speed data, radar line-of-sight vector and three-dimensional relative position encoding tensor are concatenated to generate the block feature tensor and the corresponding mask position tensor for training. The segmented feature tensor used for training and its corresponding mask position tensor are input into the untrained 3D spatial generation model. Through multi-scale hidden layer feature extraction by the encoder and cross-layer jump reconstruction mapping by the decoder, the final layer output of the decoder provides a reconstructed 3D wind field containing three wind speed vector components.

[0039] The error of the reconstructed predicted 3D wind field is evaluated using a preset loss function, and network parameters are optimized. In this embodiment, the preset loss function during the training phase includes a data reconstruction loss term and a physical constraint loss term. Three wind speed vector components are extracted from the reconstructed predicted 3D wind field, and their mean square error relative to the actual wind field label data at the corresponding grid nodes is calculated. The sum of the mean square errors of all grid nodes within the simulated missing range is used as the data reconstruction loss term to constrain the absolute accuracy of the network's output wind speed values.

[0040] The physical constraint loss term is constructed. This term is based on the fluid dynamics continuity equation, presupposing the physical assumption of an incompressible atmosphere at the microscale, where the fluid mass conservation law requires that the airflow divergence at any location in three-dimensional space approach zero. Each grid node in the reconstructed predicted three-dimensional wind field is traversed, and the three wind speed components corresponding to that node in the horizontal, vertical, and axial directions are denoted as follows: , and In this embodiment, the three-dimensional central difference method is used to calculate the three-dimensional spatial divergence of each grid node. The specific divergence calculation formula is as follows: ; in, The spatial coordinate grid index is ( The three-dimensional spatial divergence of the nodes of ) , and These represent the preset physical spacing between adjacent grids along the horizontal, vertical, and longitudinal axes, respectively.

[0041] The squared average of the 3D spatial divergence of all grid nodes within the reconstructed predicted 3D wind field is calculated and output as a penalty term, constituting the physical constraint loss term. The calculated data reconstruction loss term and the physical constraint loss term are weighted and summed according to preset weight coefficients to obtain the final total loss function value. This total loss function value is then used to execute a backpropagation algorithm, iteratively updating the network weight parameters of the encoder and decoder using a preset adaptive gradient optimizer. This forward and backward propagation process is repeated until the total loss function value converges to a preset threshold. The network parameters are then fixed, and a pre-trained 3D spatial generation model is output. The preset weight coefficients adopt an adaptive dynamic weight strategy based on backpropagation gradient, so that the weight coefficients of the physical constraint loss term dynamically increase from the initial 0.01 to the upper limit of 0.5 as the training iterations increase. This ensures that the network learns the coarse-grained wind field numerical distribution first in the early stage of training, and then refines the fine-grained continuity features through strong physical constraints in the later stage of training. The method for determining and setting the convergence of the total loss function value to a preset threshold is as follows: An early stopping mechanism from deep learning is introduced, and an independent real wind field validation set is fed into the network to calculate the validation loss. Instead of setting the preset threshold as an absolute static value, it is defined as the relative gradient of the total loss function value on the validation set. This embodiment presets the threshold as follows: when the decrease in the total loss function value on the validation set is less than 10 for 15 consecutive training iterations... -5When the model converges to the preset threshold, backpropagation and iterative training are immediately terminated, and the set of network weight parameters that minimizes the validation set loss is fixed. In the actual inference application stage of the model, the segmented feature tensor generated by real-time observation and processing of multiple radars is extracted and input into the pre-trained three-dimensional space generation model for hidden layer feature extraction; the corresponding mask region is obtained and spatial pooling downsampling is performed to generate a mask position tensor that matches the spatial dimension of the hidden layer features; the extracted hidden layer features and the mask position tensor are merged in the channel dimension and input into the decoder of the three-dimensional space generation model. The decoder fuses the hidden layer features with the high-frequency detail features of the shallow skip connections of the encoder, performs reconstruction mapping constrained by the aforementioned fluid physics rules, and outputs the reconstructed three-dimensional wind field corresponding to the mask region.

[0042] By constructing a three-dimensional convolutional encoder-decoder network with cross-layer skip connections, it is possible to effectively fuse multi-scale wind field spatial features. While deeply reconstructing the macroscopic topology of the wind field, the high-frequency edge details extracted by the shallow network are directly transmitted to the decoding end, effectively avoiding the feature smoothing and flow field detail loss problems that occur in traditional deep learning models after multiple upsampling and downsampling.

[0043] This invention generates a dynamic mask by extracting the resolution differences in the overlapping areas of multi-source radars, effectively avoiding the amplification of boundary errors caused by traditional global forced interpolation. By combining three-dimensional relative position encoding and mask zeroing, the model fully explores the three-dimensional spatial interaction laws of the wind field while preserving the original spatial topology. Finally, a three-dimensional generation model constrained by fluid dynamics equations is introduced, forcing the reconstruction process to follow the law of conservation of fluid mass, fundamentally eliminating abrupt changes in wind field vectors and data cliffs, and realizing wind field inversion.

[0044] Example 2: This embodiment applies the microscale three-dimensional wind field inversion method based on three-dimensional laser wind radar to the three-dimensional wind field evolution analysis of the wake of a large wind turbine. Three three-dimensional laser wind-measuring radars deployed at different ridge locations within the wind farm were used to coordinate the scanning centers to the wake core area behind the rotor of a large wind turbine in the target area, acquiring discrete wind field data including radial wind speed and radar line-of-sight vector. A unified three-dimensional Cartesian coordinate system was established with the center of the bottom of the wind turbine tower as the origin, and the radial wind speed point sets of the three radars in the original spherical coordinate system were accurately mapped to this coordinate system. The electromechanical scanning data of each radar were strictly synchronized and aligned to the same reference time in the time dimension using a second-order Lagrange polynomial interpolation algorithm. The scanning point density and spatial sampling interval of the three radars were calculated in the wake overlap detection area, and resolution difference characteristics reflecting the observation quality of the wake area were generated by point-by-point comparison. Based on the extracted resolution difference features, spatial gradient calculation is performed using the three-dimensional Sobel operator within the wake three-dimensional spatial region to capture the area where the observation resolution is sharply reduced due to the high-speed rotation of the wind turbine blades and strong turbulence at the wake edge. Defective boundary points with gradient values ​​greater than the judgment threshold are extracted, and the geometric surface of these boundary points is reconstructed using the three-dimensional concave hull algorithm to construct a closed three-dimensional polyhedron closely attached to the wind turbine wake blind zone, enclosing the defective boundary interval. Then, using the ray cross-judgment algorithm, the three-dimensional spatial grid points located inside the defective boundary interval are assigned a mask label 0 (to be reconstructed), and the outside is assigned a mask label 1 (visible), generating a dynamic mask region covering the wind turbine wake observation blind zone. In the established three-dimensional Cartesian coordinate system, the wake region of the wind turbine is divided into three-dimensional spatial grids according to a preset voxel size of 10 meters and a sliding step size of 160 meters corresponding to the actual physical space. Local data blocks with strictly fixed dimensions of 32×32×32 grid nodes are extracted from the discrete wind field data. For each extracted wake block, the tensor is aligned bitwise according to the generated dynamic mask region. The radial wind speed values ​​at the grid nodes with the first identifier value (mask of 0) are forcibly erased and set to zero, while the actual observed wind speeds at the nodes with the second identifier value (mask of 1) are completely preserved. A complete input tensor is constructed without destroying the overall spatial topology of the wake, thus isolating the shading and distortion data. For the aforementioned 32×32×32 local data block, with its physical center position as the origin of the coordinate system, the relative coordinates of each grid node within the block along the x-axis, y-axis, and z-axis are calculated and normalized by dividing by the maximum physical distance on one side. The high-frequency sinusoidal position coding function containing multi-band mapping is used to perform high-frequency feature mapping on the normalized continuous relative coordinates. The coding results of each axis are spliced ​​and fused along the channel dimension to generate a three-dimensional position coding tensor. This tensor is then sequentially spliced ​​with the radial wind speed and radar line-of-sight direction vector corresponding to the grid node along the channel dimension to generate a block feature tensor containing wake region geometric projection and high-frequency spatial position information. The sliced ​​feature tensor containing wake features is input into a pre-trained encoder-decoder model based on a 3D convolutional neural network. The encoder extracts multi-scale hidden layer features of the wake region through multi-layer 3D convolution and merges them with the downsampled mask position tensor before sending them to the decoder. The decoder performs reconstruction mapping by combining the high-frequency turbulent edge details passed by the cross-layer skip connection. Under the forced guidance of the physical constraint loss term composed of the fluid dynamics continuity equation, it strictly follows the fluid mass conservation theorem to deduce the true vector of the mask zeroing region and outputs a 3D continuous reconstructed wind speed vector field in the horizontal, vertical and longitudinal directions within the wake blind zone of the wind turbine.

[0045] 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 method for microscale three-dimensional wind field inversion based on three-dimensional laser wind radar, characterized in that, include: Discrete wind field data is acquired by multiple three-dimensional laser wind radars, including radial wind speed and corresponding radar line-of-sight direction vector; the discrete wind field data is spatiotemporally registered in a unified coordinate system to obtain spatial coordinate and resolution difference features; based on the resolution difference features, the defect boundary interval is located and the defect boundary interval is marked as a mask region. Based on the spatial coordinates, local data blocks of fixed dimensions are extracted from the discrete wind field data, and the local data blocks are divided into visible non-masked data and masked data to be reconstructed according to the mask region. Based on the spatial coordinates and the radar line-of-sight direction vector, a three-dimensional relative position encoding calculation is performed on the local data blocks to obtain a block feature tensor containing geometric projection information, and a mask position tensor is generated by combining the mask region. The extracted feature tensor is input into a pre-trained 3D spatial generation model for hidden layer feature extraction. The extracted hidden layer features are then merged with the mask position tensor and input into the decoder of the 3D spatial generation model to perform a reconstruction mapping constrained by physical rules, thereby generating the reconstructed 3D wind field corresponding to the mask region.

2. The microscale three-dimensional wind field inversion method based on three-dimensional laser wind radar according to claim 1, characterized in that, The discrete wind field data consists of radial wind speed point sets and their corresponding radar line-of-sight vectors obtained by multiple three-dimensional laser wind measuring radars in spherical coordinates. The radar line-of-sight vectors include spatial azimuth and elevation information. The multiple three-dimensional laser wind measuring radars are deployed at different spatial locations in the target detection area, with their scanning centers pointing towards the target detection area, so that the scanning fields of multiple radars form a three-dimensional spatial overlap area within the target detection area. Each radar performs coordinated scanning of azimuth and elevation angles through a dual-axis optical scanning mirror to obtain radial wind speed point sets in different line-of-sight directions in the target detection area.

3. The microscale three-dimensional wind field inversion method based on three-dimensional laser wind radar according to claim 1, characterized in that, The unified coordinate system is a three-dimensional Cartesian coordinate system; the spatiotemporal registration includes spatial registration and temporal registration. The spatial registration process involves mapping the radial wind speed point sets of each radar in the original spherical coordinate system to the three-dimensional Cartesian coordinate system using the coordinate transformation formula from spherical coordinate system to three-dimensional Cartesian coordinate system to generate spatial coordinates; the temporal registration process involves performing time synchronization interpolation on the scanning data of each radar according to their respective timestamp information to align the data points of each radar to the same time reference; the resolution difference feature acquisition process involves calculating the local spatial resolution index of the discrete wind field data of each radar in the three-dimensional spatial overlap area formed by the scanning fields of multiple three-dimensional laser wind measuring radars. The local spatial resolution index includes at least one of scanning point density, adjacent point spacing, and angle sampling interval. The resolution difference feature is generated by comparing the local spatial resolution indexes of different radars point by point.

4. The microscale three-dimensional wind field inversion method based on three-dimensional laser wind radar according to claim 1, characterized in that, The mask region acquisition process involves calculating the spatial gradient within the three-dimensional spatial region corresponding to the discrete wind field data based on the resolution difference feature, and obtaining the resolution difference gradient value at each spatial location. Based on the resolution difference gradient value, the defect boundary points are marked, and the defect boundary interval is formed by the defect boundary points; all spatial positions within the defect boundary interval are assigned mask labels in a three-dimensional Cartesian coordinate system to generate the mask region; The mask identifier is a binary symbol. The mask identifier corresponding to the spatial position within the defect boundary interval is set to a first identifier value, and the mask identifier corresponding to the spatial position outside the defect boundary interval is set to a second identifier value. The first identifier value and the second identifier value are 0 and 1, respectively.

5. The microscale three-dimensional wind field inversion method based on three-dimensional laser wind radar according to claim 1, characterized in that, The local data block is a three-dimensional data sub-block containing a fixed number of grid nodes extracted from the discrete wind field data after three-dimensional spatial meshing in a three-dimensional Cartesian coordinate system based on the spatial coordinates and according to a preset voxel size and block step size. The three-dimensional data sub-block has a preset number of grid dimensions along the x-axis, y-axis and z-axis, and the number of grid dimensions in each dimension is consistent in the local data block of the fixed dimension. The local data block is divided into visible non-masked data and masked data to be reconstructed according to the mask region.

6. The microscale three-dimensional wind field inversion method based on three-dimensional laser wind radar according to claim 5, characterized in that, The process of dividing the visible non-masked data and the masked data to be reconstructed is to match and determine each spatial position in the local data block based on the mask identifier. The determination process is to divide the data at the spatial position with the first identifier value into the masked data to be reconstructed and set its value to zero; divide the data at the spatial position with the second identifier value into the visible non-masked data, retain its original value, and construct an input tensor with complete spatial dimensions.

7. The microscale three-dimensional wind field inversion method based on three-dimensional laser wind radar according to claim 1, characterized in that, The three-dimensional relative position encoding is based on the spatial coordinates of each grid node within the local data block. The relative coordinates of each grid node are calculated and normalized using the center position of each block as the origin. Then, the normalized relative coordinates in the x, y, and z axes are encoded using a sinusoidal position encoding function. The encoding results of each axis are then concatenated and fused along the channel dimension to generate a three-dimensional position encoding tensor that matches the spatial dimension of the local data block. The block feature tensor is formed by concatenating the radial wind speed and radar line-of-sight vector corresponding to the local data block with the three-dimensional position encoding tensor along the channel dimension.

8. The microscale three-dimensional wind field inversion method based on three-dimensional laser wind radar according to claim 1, characterized in that, The three-dimensional space generation model is an encoder-decoder network built on a three-dimensional convolutional neural network. The encoder is composed of stacked three-dimensional convolutional downsampling layers, which are used to extract hidden layer features at multiple scales from the input segmented feature tensor. The decoder is composed of stacked 3D transposed convolutional upsampling layers. It fuses multi-scale features with the encoder through cross-layer skip connections. This is used to merge the hidden layer features with the mask position tensor and then perform conditional decoding mapping to generate the reconstructed 3D wind field corresponding to the mask region. The loss function of the 3D space generation model during the training phase includes a data reconstruction loss term and a physical constraint loss term. The physical constraint loss term is constructed based on the fluid dynamics continuity equation. It uses the 3D spatial divergence of the output reconstructed 3D wind field as a penalty term to constrain the output reconstructed 3D wind field of the 3D space generation model to satisfy the fluid mass conservation theorem.

Images