Wind measurement radar data correction method based on high-precision topographic data

Abstract

The invention discloses a wind measurement radar data correction method based on high-precision topographic data. According to the technical scheme, the wind measurement radar data correction method is characterized by comprising the steps of 1, data preprocessing; 2, modeling a high-precision terrain generation model; step 3, performing CFD pre-simulation and terrain correction coefficient matrix; 4, modeling a data correction model; and establishing a data correction model by using a space-time convolutional neural network coupled LSTM mixed structure, learning terrain-wind field nonlinear mapping, and outputting a corrected three-dimensional wind field. 5, correcting the data in real time; and inputting real-time wind measurement radar data into the data correction model, optimizing the radar data through online matching of the terrain correction coefficient matrix under the current wind direction and wind speed, and outputting a correction result. The method is mainly used for solving the problems of high wind speed measurement error and low data analysis efficiency in a dynamic environment.

Description

Technical Field

[0001] This invention relates to the field of wind-measuring radar technology, and in particular to a method for correcting wind-measuring radar data based on high-precision terrain data. Background Technology

[0002] Wind-measuring radar is used to measure wind direction and speed at high altitudes. Existing wind-measuring radars are easily affected by topographical distortions such as mountains and valleys in complex terrain, leading to errors in wind speed measurement, such as abnormal wind speed acceleration / deceleration, enhanced turbulence and disordered flow field, and failing to reflect the true wind field gradient.

[0003] Existing technologies utilize GIS (Geographic Information System) or remote sensing to acquire 3D data of complex terrain, combining this with digital elevation models to analyze the geometric features of mountains and valleys, and establishing a terrain impact model. Static CFD is then calculated to simulate airflow patterns within the terrain, predicting wind speed acceleration / deceleration zones and turbulence intensity, and generating a terrain correction coefficient table to compensate for terrain distortion. This approach can significantly improve the accuracy and reliability of wind field measurements by wind-measuring radar in complex terrain. However, this static CFD correction method is mostly based on RANS models for simulation calculations, resulting in a large computational load. Furthermore, the wind speed measurement correction error tends to be high in dynamic wind field changing environments, presenting certain limitations. Summary of the Invention

[0004] To address the above shortcomings, this invention provides a solution that can resolve the problems of high wind speed measurement errors and low data analysis efficiency in dynamic environments.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a wind radar data correction method based on high-precision topographic data, comprising the following steps: Step 1: Data preprocessing; Spatiotemporally align the wind radar scan data, satellite DEM (Digital Elevation Model) data, and UAV observation data, align all data with timestamps and transform them to the same coordinate system to obtain multi-source aligned data; Point cloud data is obtained by performing point cloud processing on the UAV observation data in the multi-source aligned data; at the same time, the wind speed is calculated based on the wind measurement radar scanning data, and the wind speed anomaly area is identified by wind speed anomaly detection. The wind speed anomaly area is mapped to the geographic coordinate system in the satellite DEM data through data mapping, and marked as the area to be corrected by binary mask. Step 2: High-precision terrain generation modeling; A terrain generation model with a hybrid structure of PointNet and LSTM networks is used to generate a high-precision DEM. Specifically, the terrain generation model combines wind speed anomaly areas, low-resolution information from satellite DEM data, and high-precision point cloud features. The PointNet network is used to process the LiDAR point cloud data from the UAV to extract local terrain geometric features. At the same time, the LSTM network is used to model the temporal changes of the terrain. Combining the satellite DEM data and the output of the LSTM network, a high-precision DEM is predicted. Through regular grid conversion, the DEM grid vertices are directly mapped to 3D points, and adjacent points are connected to form triangular patches, realizing the conversion from high-precision DEM to 3D terrain mesh. By calculating the elevation gradient and curvature of the grid vertices, a ridge detection algorithm is used to mark key terrain features in the 3D terrain mesh. Step 3: CFD Pre-simulation: First, import the 3D terrain mesh and repair mesh topology defects; smooth the terrain surface; define the computational domain and set boundary conditions based on key terrain features; divide the 3D terrain mesh into flow field analysis meshes based on the mesh generation strategy; use the physical model to calculate the flow field and obtain relevant wind field parameters; define correction coefficients and calculate correction coefficients for discrete points within the computational domain; the correction coefficients include wind speed correction coefficients and wind direction deflection coefficients; map the mesh cells generated by the CFD pre-simulation to regular meshes, and store the correction coefficients in each mesh to obtain the terrain correction coefficient matrix; Step 4: Data Correction Model Building; A data correction model is established using a hybrid structure of spatiotemporal convolutional neural network coupled with LSTM to learn the nonlinear mapping between terrain and wind field, and output the corrected 3D wind field; The loss function of the data correction model includes MAE and physical constraint terms, and the physical constraint terms are generated based on the terrain correction coefficient matrix; Step 5: Real-time data correction; Input real-time wind radar data into the data correction model, optimize the radar data by matching the terrain correction coefficient matrix under the current wind direction and speed online, and output the correction results.

[0006] Furthermore, the point cloud processing of the UAV observation data in step one specifically includes noise reduction and downsampling processes.

[0007] Furthermore, the wind speed anomaly detection in step one includes turbulence intensity calculation, wind speed gradient analysis, and continuous anomaly detection; The turbulence intensity calculation is as follows: if the turbulence intensity in a certain area is greater than a threshold, it is marked as a strong turbulence area. The turbulence intensity calculation is based on the 10-minute average wind speed and standard deviation, and the calculation formula is as follows: (1); in, For turbulence intensity, For wind speed standard deviation, Average wind speed; The wind speed gradient analysis identifies abrupt change regions by calculating the rate of change of wind speed between adjacent measuring points, as shown in the following formula: (2); in, This represents the horizontal / vertical components of the radial wind speed V. This represents the gradient along the horizontal direction x. This represents the gradient along the vertical direction y. Indicates the rate of change of wind speed; The continuous anomaly detection specifically means that when the wind speed gradient value is greater than the threshold, it is determined that there may be an obstacle or a sudden change in terrain.

[0008] Furthermore, the high-precision terrain generation model in step two includes an input layer, a PointNet module, an LSTM module, a residual module, and an output layer; The input layer is used to access spatiotemporally aligned multi-source data; The PointNet module performs hierarchical feature extraction on UAV LiDAR point cloud data, and gradually aggregates local features through the SetAbstraction layer to form a 1024-dimensional global representation describing the entire point cloud, obtaining a global feature vector that represents the macroscopic shape of the terrain; it also extracts features from UAV observation data to obtain the geometric feature vector and key point coordinates of each point cloud. The LSTM module learns the patterns of terrain evolution, and then uses an attention mechanism to weight key time nodes. After multi-source fusion of the input data, it outputs the true value of the current terrain. The residual module predicts the difference between the true value of the current terrain and the satellite DEM; The output layer outputs the difference predicted by the residual module.

[0009] Furthermore, in step three, the CFD pre-simulation uses either LES large eddy simulation or RANS-LES hybrid model to obtain steady-state / transient wind fields under different wind directions.

[0010] Furthermore, for discrete points within the computational domain The formula for calculating the wind speed correction coefficient is as follows: (5); in, This is the wind speed correction factor. The wind speed is when there is terrain. Wind speed when the terrain is flat or the terrain is negligible. To compute discrete points within the domain, pair the discrete points within the domain. The formula for calculating the wind deflection coefficient is as follows: (6); in, This is the wind deflection coefficient. For pressure when there is terrain, Pressure when the terrain is flat or the terrain is ignored.

[0011] Furthermore, the data correction model modeling in step four includes an input layer, a spatiotemporal convolution module, an LSTM module, and an output layer; The input layer is a spatiotemporal cube, and the real-time wind speed and wind direction data of each radar scanning point are concatenated with the TCCM coefficient of the corresponding location to form a multidimensional feature vector; The spatiotemporal convolution module includes 3D convolutional layers and temporal convolutional layers. The 3D convolutional layers extract spatial features from the multidimensional feature vectors obtained from the input layer, while the temporal convolutional layers capture short-time sequence correlations and output spatiotemporal feature maps. The LSTM module includes a bidirectional LSTM layer and a residual correction unit; the bidirectional LSTM layer has 128 hidden units, which captures the wind field evolution pattern based on the spatiotemporal feature map; the residual correction unit calculates the correction amount of wind speed / direction according to the terrain correction coefficient and the current wind field evolution pattern. The output layer converts the correction values ​​output by the LSTM module into the final correction result, and outputs the corrected three-dimensional wind field.

[0012] Furthermore, the training process of the data correction model in step four includes the following steps: (1) Obtain real-time meteorological auxiliary data: wind direction, wind speed, temperature, humidity and air pressure; (2) Data set construction: The dataset is constructed based on the steady-state / transient wind field under different wind directions pre-simulated by CFD, meteorological auxiliary data and original wind measurement radar scan data.

[0013] (3) Model training: Divide the dataset constructed in step (2) into training set, test set and validation set, and iteratively train and fine-tune the initial data correction model according to the set hyperparameters and loss function to obtain the trained data correction model.

[0014] A computer-readable storage medium, characterized in that the computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to execute the wind radar data correction method based on high-precision terrain data as described above.

[0015] A processor, characterized in that the processor is configured to run a program, wherein the program executes the wind radar data correction method based on high-precision terrain data as described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. By integrating wind measurement radar scanning data, satellite DEM data, and UAV observation data, high-precision terrain data is generated through terrain generation models. This achieves multi-source data synergy and complementarity, including wide-area satellite coverage, detailed UAV data, and real-time radar anomaly detection. This provides more comprehensive coverage of terrain features and can dynamically adapt to terrain changes, thus providing an accurate data foundation for the physical authenticity of subsequent wind field calculations.

[0017] 2. A 3D terrain mesh is generated based on high-precision terrain data. Combining key terrain features within the mesh, CFD pre-simulation is performed using LES large eddy simulation or a RANS-LES hybrid model to generate steady-state / transient wind field and terrain correction coefficient matrices. The boundary conditions for the simulation can be dynamically updated based on real-time meteorological data, maintaining simulation timeliness and effectively reducing simulation errors under complex terrain. Simultaneously, the terrain correction coefficient matrix provides physical constraints for data correction model building and real-time data correction, improving the accuracy of data correction.

[0018] 3. A hybrid modeling approach using a spatiotemporal convolutional network (ST-CNN) and an LSTM network is employed to obtain a data correction model capable of real-time correction of radar data. The LSTM module of the model can learn temporal variations in the wind field, overcoming the fixed wind direction limitation of static CFD and improving the model's dynamic adaptability to complex terrain. Simultaneously, the model is deployed to edge devices, enabling real-time inference and correction of radar data, meeting the real-time terrain correction needs of scenarios such as wind power and drones. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0020] Figure 1 This is a flowchart illustrating a wind radar data correction method based on high-precision terrain data according to the present invention. Figure 2 This is a schematic diagram illustrating the process of modeling the terrain generation model in this invention; Figure 3 This is a schematic diagram illustrating the process of modeling the data correction model in this invention. Detailed Implementation

[0021] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0022] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0023] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. Furthermore, the technical features involved in the different embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0024] This invention provides a wind radar data correction method based on high-precision terrain data, which can solve the problems of high wind speed measurement error and low data analysis efficiency in dynamic environments.

[0025] like Figure 1 Specifically, it includes the following steps: Step 1: Data preprocessing; Spatiotemporally align the wind radar scan data, satellite DEM data, and UAV observation data, align all data with timestamps and transform them to the same coordinate system to obtain multi-source aligned data, which facilitates the subsequent generation of terrain data.

[0026] The UAV observation data in the multi-source aligned data is processed into point cloud data to obtain UAV LiDAR point cloud data. The point cloud processing specifically includes denoising and downsampling, which reduces data noise while reducing computational load.

[0027] Data preprocessing also includes wind speed anomaly detection. Wind speed is calculated based on wind radar scan data, and anomaly areas are identified through wind speed anomaly detection. These anomaly areas (polar coordinates) are mapped to a geographic coordinate system (UTM) using data mapping, and marked as areas to be corrected using binary masks (1 = anomaly area, 0 = normal area). Wind speed anomaly detection includes turbulence intensity calculation, wind speed gradient analysis, and continuous anomaly detection.

[0028] If the turbulence intensity of a certain area is greater than the threshold, it is marked as a strong turbulence area, where the turbulence intensity threshold for mountainous areas is 0.2 and the turbulence intensity threshold for plains is 0.1.

[0029] The turbulence intensity is calculated based on the 10-minute average wind speed and standard deviation, using the following formula: (1); in, For turbulence intensity, For wind speed standard deviation, This represents the average wind speed.

[0030] The wind speed gradient analysis calculates the rate of change of wind speed between adjacent measuring points. When the wind speed gradient value is greater than the threshold (5 m / s for wind speed at 100 meters), it is identified as a sudden change region. The sudden change region is an obstacle or a change in terrain.

[0031] The specific calculation for the wind speed gradient analysis is as follows: (2); in, This represents the horizontal / vertical components of the radial wind speed V. This represents the gradient along the horizontal direction x. This represents the gradient along the vertical direction y. This indicates the rate of change of wind speed.

[0032] Step 2: High-precision terrain generation model modeling; combining radar wind field anomalies, low-resolution satellite DEM information with high-precision point cloud features, and using the PointNet network to process the UAV's LiDAR point cloud data to extract local terrain geometric features; like Figure 2 The high-precision terrain generation model includes an input layer, a PointNet module, an LSTM module, a residual module, and an output layer. The input layer is used to access spatiotemporally aligned multi-source data; The PointNet module performs hierarchical feature extraction on UAV LiDAR point clouds from multiple source data sources. It aggregates local features step by step through the Set Abstraction layer to form a 1024-dimensional global representation describing the entire point cloud, obtaining a global feature vector that represents the macroscopic shape of the terrain. It also extracts features from UAV observation data to obtain the geometric feature vector and key point coordinates of each point cloud. The LSTM module learns the patterns of terrain evolution, and then uses an attention mechanism to weight key time nodes. After multi-source fusion of the input data, it outputs the true value of the current terrain. The residual module predicts the difference between the true value of the current terrain and the satellite DEM; The residual module shown predicts the difference (Δh) between the true value of the current terrain and the satellite DEM, instead of directly outputting the elevation. The purpose is to reduce the difficulty of model learning and improve training efficiency. The comparison results are shown in Table 1.

[0033] Table 1. Performance Comparison of Direct Elevation Prediction and Residual Prediction Models: The formula for calculating the difference between the true value of the current terrain and the satellite DEM is as follows: (3); in, Represents satellite DEM data, This represents the true value of the current terrain output by the LSTM module.

[0034] The output layer adds the difference predicted by the residual module to the satellite DEM to obtain a high-precision DEM and outputs it.

[0035] Since satellite DEMs already contain large-scale low-frequency information about the terrain (such as mountain outlines and valley orientations), directly predicting elevations may lead to non-physical results (such as abrupt changes in cliff heights). Therefore, the model only needs to learn the detail correction (Δh) rather than reconstructing the entire terrain from scratch, which significantly reduces training complexity.

[0036] The formula for calculating a high-precision DEM is shown below: (4); This represents the high-precision DEM predicted by the model. Represents satellite DEM data, This represents the difference between the true value of the current terrain and the satellite DEM.

[0037] Table 2 shows examples of satellite DEM correction in actual terrain scenarios: Table 2 Satellite DEM Correction Table Furthermore, after predicting the high-precision DEM, a regular mesh transformation is used to directly map the DEM raster vertices to 3D points (X = column index × resolution, Y = row index × resolution, Z = elevation value). Adjacent points are then connected to form triangular patches (Delaunay triangulation), achieving the conversion from the high-precision DEM to a 3D terrain mesh. Finally, by calculating the elevation gradient and curvature of the mesh vertices, a ridge detection algorithm is used to mark key terrain features in the 3D terrain mesh.

[0038] The loss function of the terrain generation model includes MSE loss, TV-Loss (total variational regularization), and Radar consistency loss.

[0039] The MSE loss refers to the error between the predicted elevation and the true LiDAR value. The TV-Loss (Total Variation Regularization) smooths out abrupt terrain noise. The Radar consistency loss is used to force the predicted terrain to match anomaly areas in the radar wind field.

[0040] The loss function L is shown in the following formula: in, Let MSE be the total loss function of the model, and let MSE represent the mean squared error. This represents the true value of the drone observation data. This indicates total variational regularization. Indicates radar consistency loss, Indicates mean square error. Show total variational regularization weights, This represents the radar consistency weight.

[0041] The terrain generation model of this invention adopts a core architecture of "PointNet++ spatial feature extraction + LSTM fusion with temporal dynamics + multi-source residual learning," enabling high-precision terrain prediction at the centimeter level. Its performance is significantly better than methods relying solely on single data sources such as satellite DEMs. The model also boasts excellent timeliness, supporting minute-level updates to quickly respond to environmental changes. Specifically, when radar detects anomalies in the wind field, it can trigger a UAV to periodically scan the anomalous area (e.g., once a week). Subsequently, the system can quickly update the model based on the UAV scan data and utilize the hidden state transfer mechanism of LSTM in a lightweight incremental learning manner (e.g., freezing PointNet++ weights and only fine-tuning LSTM layers), thereby achieving continuous and accurate perception of terrain dynamics.

[0042] Step 3: CFD pre-simulation and terrain correction coefficient matrix; The CFD pre-simulation process is as follows: First, import the 3D terrain mesh and repair mesh topology defects (such as holes and overlapping surfaces). Then, smooth the terrain surface (preserving macroscopic undulations and removing minor noise) to prevent mesh quality deterioration.

[0043] Secondly, based on key terrain features, the computational domain is defined and boundary conditions are set. The computational domain range is typically 5 to 10 times the length of the terrain feature (e.g., 8 times the length of the terrain extending along the wind direction), ensuring that the entrance, exit, and lateral boundaries are far from the terrain's influence zone. Boundary conditions include the entrance boundary, exit boundary, lateral boundary, and terrain surface.

[0044] Finally, based on the meshing strategy, the three-dimensional terrain mesh is divided into flow field analysis meshes. Specifically, a boundary layer mesh is generated on the terrain surface to capture boundary layer flow separation and reattachment. Using Octree or RBF mesh deformation technology, the mesh is automatically densified in areas corresponding to key terrain features, balancing accuracy and computational cost.

[0045] The key terrain feature types are shown in Table 3: Table 3 Key Terrain Feature Types This invention employs either LES (Large Eddy Simulation) or a RANS-LES hybrid model in its CFD physical model. Flow field calculations are performed using the physical model to obtain relevant wind field parameters. Flow field characteristics are visualized to plot wind speed vector fields, pressure contour maps, and turbulence intensity distributions, identifying terrain-induced airflow separation zones (such as leeward slope eddies) and acceleration zones (such as ridge crests). Simultaneously, parameters such as wind speed profiles, pressure coefficients, and shear stresses are extracted along the terrain surface. The terrain's blockage ratio is calculated to assess the overall impact of terrain on the wind field. By summarizing these parameters, steady-state and transient wind fields under different wind directions are obtained.

[0046] The terrain correction coefficient matrix is ​​obtained by mapping the grid cells generated by CFD pre-simulation to a regular grid, and each grid stores the wind speed correction coefficient and wind direction deflection coefficient.

[0047] For discrete points within the computational domain The formula for calculating the wind speed correction coefficient is as follows: (5); in, This is the wind speed correction factor. The wind speed is when there is terrain. Wind speed when the terrain is flat or the terrain is negligible. Discrete points within the domain For discrete points within the computational domain The formula for calculating the wind deflection coefficient is as follows: (6); in, This is the wind deflection coefficient. For pressure when there is terrain, Pressure when the terrain is flat or the terrain is ignored.

[0048] Step 4: Data correction model building; A data correction model is built using a hybrid structure of spatiotemporal convolutional neural network (ST-CNN) coupled with LSTM, which learns the nonlinear mapping between terrain and wind field and outputs the corrected three-dimensional wind field.

[0049] like Figure 3 The data correction model includes an input layer, a spatiotemporal convolution module, an LSTM module, and an output layer. The input layer is a spatiotemporal cube. The real-time wind speed and wind direction data of each radar scanning point are concatenated with the TCCM coefficients (such as wind speed correction coefficients and wind direction offset coefficients) at the corresponding locations to form a multidimensional feature vector. The concatenation method for multidimensional feature vectors is as follows: If the radar data contains a 3D wind field (u, v, w), and the TCCM contains 2 correction coefficients (wind speed correction value, terrain roughness coefficient), then the feature vector of each point is [u, v, w, wind speed correction coefficient, terrain roughness coefficient]. u represents the wind speed component along the x-axis of the horizontal coordinate system; v represents the wind speed component along the y-axis of the horizontal coordinate system; and w represents the wind speed component along the z-axis of the vertical coordinate system.

[0050] The spatiotemporal convolution module includes 3D convolutional layers and temporal convolutional layers. The 3D convolutional layers extract spatial features from the multidimensional feature vectors obtained from the input layer, while the temporal convolutional layers capture short-time sequence correlations and output spatiotemporal feature maps. The LSTM module includes a bidirectional LSTM layer and a residual correction unit; the bidirectional LSTM layer has 128 hidden units, which captures the wind field evolution pattern based on the spatiotemporal feature map; the residual correction unit calculates the correction amount of wind speed / direction according to the terrain correction coefficient and the current wind field evolution pattern. The output layer is a fully connected layer that converts the correction values ​​output by the LSTM module into the final correction results (such as 3D wind field and terrain residuals), and outputs the corrected 3D wind field.

[0051] The loss function of the data correction model includes MAE and physical constraint terms. The physical constraint terms are generated based on the terrain correction coefficient matrix and include penalized wind field divergence loss and local consistency loss.

[0052] The training process of the data correction model includes the following steps: (1) Obtain real-time meteorological auxiliary data: wind direction, wind speed, temperature, humidity, and air pressure; (2) Data set construction: The dataset is constructed based on the steady-state / transient wind field under different wind directions pre-simulated by CFD, meteorological auxiliary data and original wind measurement radar scan data.

[0053] (3) Model training: Divide the dataset constructed in step (2) into training set, test set and validation set, and iteratively train and fine-tune the initial data correction model according to the set hyperparameters and loss function to obtain the trained data correction model.

[0054] During training, a transfer learning strategy is adopted, using the constructed dataset to pre-train the model, and then adding the actual test data for joint training to achieve model fine-tuning and improve generalization ability.

[0055] The data correction model can also be modeled using a hybrid network structure that couples a spatiotemporal convolutional neural network (ST-CNN) with a Transformer.

[0056] Step 5: Real-time data correction; Input real-time wind radar data into the data correction model, optimize the radar data by matching the terrain correction coefficient matrix under the current wind direction and speed online, and output the correction results.

[0057] The specific steps are as follows: First, the real-time radar data undergoes preprocessing such as noise filtering. Then, the preprocessed radar data tensor and the TCCM coefficient tensor are concatenated along the channel dimension to form a multi-channel input tensor. Spatial convolution is performed on the multi-channel tensor to capture the spatial correlation between terrain features and wind field distribution (such as the wind speed attenuation pattern on the leeward slope of a mountain). LSTM is used to process the time-dimensional data to learn the wind field's variation over time, and the correction weights are dynamically adjusted based on the coefficients in the terrain correction matrix (such as the difference in the impact of terrain on the wind field at different times). Corrected wind speed and direction data are generated through a fully connected layer, with the output dimension consistent with the input radar data (e.g., [N, T, 2]).

[0058] The dimensions of each tensor are defined as follows: The radar data tensor has dimensions of [N, T, 2] (N is the number of grid points, T is the time step, and 2 corresponds to the wind speed and wind direction components). The TCCM coefficient tensor has dimensions of [N, T, K] (K is the terrain feature dimension, such as slope, altitude, roughness, etc.). The spliced ​​tensor has dimensions of [N, T, 2+K], forming a multi-channel input that includes measured data and terrain features.

[0059] This invention proposes a high-precision dynamic wind field modeling and real-time correction method based on multi-source data fusion. First, by fusing multi-source data from wind-measuring radar, satellite DEM, and UAVs, a terrain generation model is used to predict high-precision terrain, achieving synergy between wide-area coverage, detail enhancement, and real-time anomaly detection, providing a dynamically updated and reliable data foundation for wind field calculation. Next, a 3D mesh is generated based on this terrain data. Combining key terrain features, CFD pre-simulation is performed using LES large eddy simulation or a RANS-LES hybrid model to generate steady-state / transient wind field and terrain correction coefficient matrices. This matrix provides physical constraints for subsequent data correction models, and the pre-simulation boundary conditions can be dynamically updated based on real-time meteorological data, effectively improving the timeliness and accuracy of simulations under complex terrain. Finally, a hybrid network composed of spatiotemporal convolutional modules and LSTM is constructed as the data correction model. This model can learn the spatiotemporal variation patterns of the wind field and overcome the limitations of static simulation by utilizing the physical constraints provided by CFD, achieving accurate and real-time correction of radar data. The final model is deployed on edge devices, meeting the urgent needs of wind power, UAVs, and other fields for real-time terrain and wind field correction.

[0060] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for correcting wind-measuring radar data based on high-precision topographic data, characterized in that, Includes the following steps: Step 1: Data preprocessing; Spatiotemporally align the wind measurement radar scan data, satellite DEM data, and UAV observation data, align all data with timestamps, and transform them to the same coordinate system to obtain multi-source aligned data; Point cloud data is obtained by performing point cloud processing on the UAV observation data in the multi-source aligned data; at the same time, the wind speed is calculated based on the wind measurement radar scanning data, and the wind speed anomaly area is identified by wind speed anomaly detection. The wind speed anomaly area is mapped to the geographic coordinate system in the satellite DEM data through data mapping, and marked as the area to be corrected by binary mask. Step 2: High-precision terrain generation modeling; A terrain generation model with a hybrid structure of PointNet and LSTM networks is used to generate a high-precision DEM. Specifically, the terrain generation model combines wind speed anomaly areas, low-resolution information from satellite DEM data, and high-precision point cloud features. The PointNet network is used to process the LiDAR point cloud data from the UAV to extract local terrain geometric features. At the same time, the LSTM network is used to model the temporal changes of the terrain. Combining the satellite DEM data and the output of the LSTM network, a high-precision DEM is predicted. Through regular grid conversion, the DEM grid vertices are directly mapped to 3D points, and adjacent points are connected to form triangular patches, realizing the conversion from high-precision DEM to 3D terrain mesh. By calculating the elevation gradient and curvature of the grid vertices, a ridge detection algorithm is used to mark key terrain features in the 3D terrain mesh. Step 3: CFD Pre-simulation: First, import the 3D terrain mesh and repair mesh topology defects; smooth the terrain surface; define the computational domain and set boundary conditions based on key terrain features; Based on a mesh generation strategy, a 3D terrain mesh is divided into a flow field analysis mesh; a physical model is used to calculate the relevant wind field parameters. Define correction coefficients and calculate correction coefficients for discrete points within the computational domain; the correction coefficients include wind speed correction coefficients and wind direction deflection coefficients; map the grid cells generated by the CFD pre-simulation to a regular grid, and store the correction coefficients in each grid to obtain a terrain correction coefficient matrix; Step 4: Data Correction Model Building; A data correction model is established using a hybrid structure of spatiotemporal convolutional neural network coupled with LSTM to learn the nonlinear mapping between terrain and wind field, and output the corrected 3D wind field; The loss function of the data correction model includes MAE and physical constraint terms, and the physical constraint terms are generated based on the terrain correction coefficient matrix; Step 5: Real-time data correction; Input real-time wind radar data into the data correction model, optimize the radar data by matching the terrain correction coefficient matrix under the current wind direction and speed online, and output the correction results.

2. The wind radar data correction method based on high-precision topographic data according to claim 1, characterized in that, The point cloud processing of the UAV observation data in step one specifically includes noise reduction and downsampling processes.

3. The wind radar data correction method based on high-precision topographic data according to claim 1, characterized in that, The wind speed anomaly detection in step one includes turbulence intensity calculation, wind speed gradient analysis, and continuous anomaly detection. The turbulence intensity calculation is as follows: if the turbulence intensity in a certain area is greater than a threshold, it is marked as a strong turbulence area. The turbulence intensity calculation is based on the 10-minute average wind speed and standard deviation, and the calculation formula is as follows: （1）； in, For turbulence intensity, For wind speed standard deviation, Average wind speed; The wind speed gradient analysis identifies abrupt change regions by calculating the rate of change of wind speed between adjacent measuring points, as shown in the following formula: （2）； in, This represents the horizontal / vertical components of the radial wind speed V. This represents the gradient along the horizontal direction x. This represents the gradient along the vertical direction y. Indicates the rate of change of wind speed; The continuous anomaly detection specifically means that when the wind speed gradient value is greater than the threshold, it is determined that there may be an obstacle or a sudden change in terrain.

4. The wind radar data correction method based on high-precision topographic data according to claim 1, characterized in that, The high-precision terrain generation model in step two includes an input layer, a PointNet module, an LSTM module, a residual module, and an output layer. The input layer is used to access spatiotemporally aligned multi-source data; The PointNet module performs hierarchical feature extraction on UAV LiDAR point cloud data, and gradually aggregates local features through the SetAbstraction layer to form a 1024-dimensional global representation describing the entire point cloud, obtaining a global feature vector that represents the macroscopic shape of the terrain; it also extracts features from UAV observation data to obtain the geometric feature vector and key point coordinates of each point cloud. The LSTM module learns the patterns of terrain evolution, and then uses an attention mechanism to weight key time nodes. After multi-source fusion of the input data, it outputs the true value of the current terrain. The residual module predicts the difference between the true value of the current terrain and the satellite DEM; The output layer outputs the difference predicted by the residual module.

5. The wind radar data correction method based on high-precision topographic data according to claim 1, characterized in that, In step three, the CFD pre-simulation uses either LES large eddy simulation or RANS-LES hybrid model to obtain steady-state / transient wind fields under different wind directions.

6. The wind radar data correction method based on high-precision topographic data according to claim 1, characterized in that, For discrete points within the computational domain The formula for calculating the wind speed correction coefficient is as follows: （5）； in, This is the wind speed correction factor. The wind speed is when there is terrain. Wind speed when the terrain is flat or the terrain is negligible. To compute discrete points within the domain, pair the discrete points within the domain. The formula for calculating the wind deflection coefficient is as follows: （6）； in, This is the wind deflection coefficient. For pressure when there is terrain, Pressure when the terrain is flat or the terrain is ignored.

7. The wind radar data correction method based on high-precision topographic data according to claim 1, characterized in that, The data correction model modeling in step four includes an input layer, a spatiotemporal convolution module, an LSTM module, and an output layer; The input layer is a spatiotemporal cube, and the real-time wind speed and wind direction data of each radar scanning point are concatenated with the TCCM coefficient of the corresponding location to form a multidimensional feature vector; The spatiotemporal convolution module includes 3D convolutional layers and temporal convolutional layers. The 3D convolutional layers extract spatial features from the multidimensional feature vectors obtained from the input layer, while the temporal convolutional layers capture short-time sequence correlations and output spatiotemporal feature maps. The LSTM module includes a bidirectional LSTM layer and a residual correction unit; the bidirectional LSTM layer has 128 hidden units, which captures the wind field evolution pattern based on the spatiotemporal feature map. The residual correction unit calculates the correction amount for wind speed / direction based on the terrain correction coefficient and the current wind field evolution. The output layer converts the correction values ​​output by the LSTM module into the final correction result, and outputs the corrected three-dimensional wind field.

8. The wind radar data correction method based on high-precision topographic data according to claim 1, characterized in that, The training process of the data correction model in step four includes the following steps: (1) Obtain real-time meteorological auxiliary data: wind direction, wind speed, temperature, humidity and air pressure; (2) Dataset construction: The dataset is constructed based on the steady-state / transient wind field under different wind directions pre-simulated by CFD, meteorological auxiliary data and original wind measurement radar scan data. (3) Model training: Divide the dataset constructed in step (2) into training set, test set and validation set, and iteratively train and fine-tune the initial data correction model according to the set hyperparameters and loss function to obtain the trained data correction model.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device containing the computer-readable storage medium to perform the wind radar data correction method based on high-precision terrain data as described in any one of claims 1 to 8.

10. A processor, characterized in that, The processor is used to run a program, wherein the program executes the wind radar data correction method based on high-precision terrain data as described in any one of claims 1 to 8.

Images