Three-dimensional wind field retrieval method based on deep learning of single wind lidar

Abstract

The present application relates to the technical field of laser radar meteorological detection, and discloses a single Doppler wind lidar three-dimensional wind field inversion method based on deep learning, comprising the following steps: S1, obtaining original power spectrum data obtained by scanning and detecting the atmospheric wind field by a single Doppler wind lidar; S2, preprocessing the original power spectrum data, and extracting observation characteristic parameters including at least radial wind speed, signal-to-noise ratio and spectrum width; S3, inputting the observation characteristic parameters into a pre-trained deep learning wind field inversion model, wherein the deep learning wind field inversion model takes the observation characteristic parameters as input, and outputs three-dimensional wind vectors of three-dimensional space grid points with the position of the laser radar as the origin. Through the deep learning model, the three-dimensional wind vectors are directly inverted from the radial observation data of the single laser radar, without the cooperation of multiple devices, so that the system cost and deployment complexity are significantly reduced.

Description

Technical Field

[0001] This invention relates to the field of lidar meteorological detection technology, specifically a method for three-dimensional wind field inversion based on a single wind-measuring lidar using deep learning. Background Technology

[0002] As an active remote sensing device, wind lidar detects the backscattering signals of laser light by aerosol particles in the atmosphere and retrieves wind field information. It has advantages such as high spatiotemporal resolution and mobile deployment. Traditional wind field retrieval methods mostly rely on the collaborative observation of multiple lidars or are based on assumptions of physical models, which have problems such as high equipment cost, complex deployment, and retrieval accuracy limited by model assumptions.

[0003] Currently, a single lidar can typically only acquire radial wind speed along the beam direction, making it difficult to directly reconstruct a three-dimensional wind field. Existing technologies often employ methods such as velocity orientation display (VAD) or volumetric velocity processing (VVP) to estimate the three-dimensional wind field by fitting the radial wind speed distribution. However, these methods are highly dependent on the scanning mode, have long observation cycles, and exhibit significant inversion errors in areas with complex wind field structures or sparse data. Furthermore, traditional methods are poorly adaptable to factors such as noise, signal quality variations, and platform motion, and lack intelligent evaluation and fusion mechanisms for data quality, resulting in insufficient reliability of the inversion results in areas with low signal-to-noise ratios or missing data. Summary of the Invention

[0004] The purpose of this invention is to provide a method for three-dimensional wind field inversion based on a single wind-measuring lidar using deep learning, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a three-dimensional wind field inversion method based on a single wind-measuring lidar using deep learning, comprising the following steps:

[0006] S1. Obtain the raw power spectrum data obtained by scanning and detecting the atmospheric wind field using a single Doppler wind lidar.

[0007] S2. Preprocess the raw power spectrum data to extract observation feature parameters including at least radial wind speed, signal-to-noise ratio, and spectral width;

[0008] S3. Input the observed feature parameters into a pre-trained deep learning wind field inversion model, wherein the deep learning wind field inversion model takes the observed feature parameters as input and outputs a three-dimensional wind vector of three-dimensional spatial grid points with the location of the lidar as the origin.

[0009] S4. Based on the three-dimensional wind vector output by the deep learning wind field inversion model, reconstruct the three-dimensional wind field of the target area.

[0010] As a preferred embodiment of the present invention, the deep learning wind field inversion model is a neural network with an encoder-decoder structure, wherein:

[0011] The encoder is used to extract features and spatially encode the input observation feature parameters to form a latent space feature vector containing global wind field information.

[0012] The decoder is used to decode and regress the three-dimensional wind vector of the three-dimensional spatial grid points based on the latent space feature vector.

[0013] As a preferred technical solution of the present invention, the deep learning wind field inversion model adopts a multi-scale feature fusion mechanism during training. Specifically, a multi-level convolutional module is introduced into the encoder to extract the local wind field structure and global wind field distribution features at different spatial scales. The feature maps of different levels in the encoder are passed to the corresponding levels in the decoder through skip connections. The decoder adopts a progressive upsampling structure to gradually restore the spatial resolution of the three-dimensional wind field.

[0014] As a preferred embodiment of the present invention, the training process of the deep learning wind field inversion model includes:

[0015] S31. Obtain a training dataset, which includes high-resolution three-dimensional wind field data generated by a numerical weather prediction model, or true three-dimensional wind field data obtained by fusion of observations from multiple lidars.

[0016] S32. Based on the true value data of the three-dimensional wind field, simulated observation characteristic parameters corresponding to the scanning mode of the single Doppler wind lidar are generated through forward model simulation.

[0017] S33. Using the simulated observation feature parameters as input and the corresponding three-dimensional wind field ground truth data as supervision labels, the deep learning wind field inversion model is trained under supervision.

[0018] As a preferred embodiment of the present invention, the preprocessing in step S2 specifically includes the following steps:

[0019] S21. In the signal range outside the effective detection range of the lidar, automatically identify and evaluate the noise levels of electromagnetic background and ambient light, and determine a dynamic noise floor threshold.

[0020] S22. Compare the raw power spectrum data at each detection range gate with the noise floor threshold. When the signal-to-noise ratio is lower than the preset threshold, mark it as invalid or low-confidence data point and isolate or assign a special identifier in subsequent processing. Apply a bandwidth-adjustable bandpass filter to the valid raw power spectrum signal and lock the center frequency of the filter with the emission frequency of the lidar.

[0021] S23. Assign a comprehensive signal quality index to each valid detection data point. The signal quality index comprehensively considers the strength of the signal-to-noise ratio, the peak sharpness of the power spectrum signal, and the stability of the signal in time. Through the signal quality index, all detection data are divided into three quality levels: high, medium, and low, and different quality identifiers are attached to the data of each level.

[0022] S24. When the lidar is deployed on a mobile platform, attitude data from the platform's inertial navigation system is integrated, including changes in the platform's roll, pitch, and heading. Through the attitude data, the beam pointing vector of the lidar is geometrically corrected in real time at each moment, and the measurement data is uniformly compensated to a stable reference coordinate system.

[0023] As a preferred embodiment of the present invention, the preprocessing in step S2 further includes the following steps:

[0024] S25. Directly analyze and generate a series of core observation parameters from the spectrum data that has been processed as described above;

[0025] S26. Associate and bind the core observation parameters of each data point with its corresponding spatial metadata;

[0026] S27. Encapsulate the core observation parameters, quality identifiers, and spatial metadata of all probe points into a structured, georeferenced dataset according to spatial logical relationships.

[0027] As a preferred technical solution of the present invention, the core observation parameters include radial wind speed, signal-to-noise ratio and spectral width. The radial wind speed is derived from the displacement of the spectral peak relative to the transmission frequency. The signal-to-noise ratio is used to characterize the reliability of the signal strength. The spectral width is used to reflect the intensity of atmospheric turbulence or wind shear. The spatial metadata includes the azimuth angle of the detection point, the elevation angle of the detection point and the slant range of the detection point relative to the radar.

[0028] As a preferred embodiment of the present invention, the step S2 and step S3 include:

[0029] S2a. Define the target three-dimensional spatial grid: Define a three-dimensional rectangular coordinate system with the location of a single wind-measuring lidar as the origin. Within the coordinate system, delineate a three-dimensional cubic space covering the effective detection range of the lidar as the target area. Divide the target area at fixed intervals in the horizontal, vertical and radial directions to form a three-dimensional spatial grid composed of numerous uniformly sized cubic grid units.

[0030] S2b. Establish a spatial impact model based on data quality: By using the signal quality index allocated in the preprocessing, a positive correlation is made for the initial weight of each original observation data point. For each three-dimensional spatial grid cell, all original observation data points in the spatial neighborhood are identified. The initial weight of each original observation data point is further attenuated by the spatial distance from the original observation data point to the center point of the three-dimensional spatial grid cell.

[0031] S2c, Perform quality-weighted feature value assignment: The radial wind speed values ​​of all original observation data points within the neighborhood of each 3D spatial grid cell are weighted and fused with the distance-attenuated weight values ​​to assign a radial wind speed feature value to the 3D spatial grid cell; the signal-to-noise ratio (SNR) values ​​of all original observation data points within the neighborhood of the same 3D spatial grid cell are weighted and fused with the distance-attenuated weights to assign a SNR feature value to that grid cell; the spectral width values ​​of all original observation data points within the neighborhood of the same 3D spatial grid cell are weighted and fused with the distance-attenuated weights to assign a spectral width feature value to that grid cell.

[0032] S2d, Generate Grid Confidence Label: Calculate the total weight of all original observation data points involved in the feature value calculation of each 3D spatial grid cell. Define the total weight as the data fullness of the 3D spatial grid cell. Based on the data fullness, assign a confidence level to each 3D spatial grid cell. When the data fullness of a 3D spatial grid cell is high, it is marked as high confidence. When the data fullness of a 3D spatial grid cell is low or zero, it is marked as low confidence or missing data.

[0033] As a preferred embodiment of the present invention, the step S2 and step S3 further include:

[0034] S2e. Constructing a regularized 3D feature tensor: The radial wind speed, signal-to-noise ratio, and spectral width feature values ​​of all obtained 3D spatial grid cells are filled into a 3D matrix that corresponds completely to the 3D spatial grid dimension. The confidence level information of each 3D spatial grid cell is stored in an independent 3D matrix. The 3D matrix containing the observed feature values ​​is combined with the 3D matrix containing the confidence level to form a complete and regularized 3D feature tensor.

[0035] As a preferred technical solution of the present invention, the method further includes post-processing optimization of the reconstructed three-dimensional wind field, specifically including: based on the confidence level of the three-dimensional spatial grid cell, performing spatial interpolation compensation on the wind vectors in low confidence or data missing areas, and smoothing the reconstructed three-dimensional wind field through wind field physical consistency constraints.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] 1. The three-dimensional wind vector can be directly inverted from the radial observation data of a single lidar through a deep learning model, without the need for multiple devices to work together, which significantly reduces system cost and deployment complexity.

[0038] 2. The neural network with an encoder-decoder structure, combined with multi-scale feature fusion and skip connection mechanism, can effectively capture local wind field structure and global distribution features, improving the accuracy of wind field reconstruction and the ability to restore spatial details.

[0039] 3. By introducing preprocessing mechanisms such as signal quality assessment, dynamic noise threshold, and mobile platform attitude correction, the robustness of the model under complex meteorological conditions and mobile platform environments is improved.

[0040] 4. Utilize numerical weather prediction or collaborative observation data as training ground values ​​to construct simulated observation feature parameters, thereby achieving end-to-end supervised learning and avoiding the excessive reliance on physical models in traditional methods.

[0041] 5. By constructing a three-dimensional spatial grid, using quality-weighted fusion, confidence level labeling, and post-processing optimization, intelligent compensation and wind field smoothing for sparse or low-quality data are achieved, thereby improving the overall physical consistency and reliability of the wind field.

[0042] 6. Integrating inertial navigation data enables real-time beam pointing correction. Combined with normalized 3D feature tensor input, it facilitates efficient and stable real-time inversion of 3D wind fields on mobile platforms. Attached Figure Description

[0043] Figure 1 This is an overall flowchart of the three-dimensional wind field inversion method based on a single wind-measuring lidar according to the present invention. Detailed Implementation

[0044] 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.

[0045] Example 1

[0046] Please see Figure 1 This invention provides a method for inverting three-dimensional wind fields using a single wind-measuring lidar based on deep learning, comprising the following steps:

[0047] S1. Obtain the raw power spectrum data obtained by scanning and detecting the atmospheric wind field using a single Doppler wind lidar.

[0048] S2. Preprocess the raw power spectrum data to extract observational characteristic parameters, including at least radial wind speed, signal-to-noise ratio, and spectral width;

[0049] S3. Input the observed feature parameters into a pre-trained deep learning wind field inversion model. The deep learning wind field inversion model takes the observed feature parameters as input and outputs a three-dimensional wind vector with the location of the lidar as the origin of the three-dimensional spatial grid points.

[0050] S4. Based on the three-dimensional wind vector output by the deep learning wind field inversion model, reconstruct the three-dimensional wind field of the target area.

[0051] Furthermore, the deep learning wind field inversion model is a neural network with an encoder-decoder structure, where:

[0052] The encoder is used to extract features and spatially encode the input observation feature parameters to form a latent space feature vector containing global wind field information;

[0053] The decoder is used to decode and regress the three-dimensional wind vector of the three-dimensional space grid points based on the latent space feature vector.

[0054] Furthermore, the deep learning wind field inversion model adopts a multi-scale feature fusion mechanism during training. Specifically, multi-level convolutional modules are introduced into the encoder to extract local wind field structure and global wind field distribution features at different spatial scales. Feature maps from different levels in the encoder are passed to the corresponding levels in the decoder through skip connections. A progressive upsampling structure is adopted in the decoder to gradually restore the spatial resolution of the three-dimensional wind field.

[0055] Furthermore, the training process of the deep learning wind field inversion model includes:

[0056] S31. Obtain the training dataset, which includes high-resolution three-dimensional wind field data generated by numerical weather prediction models, or true three-dimensional wind field data obtained by fusion of observations from multiple lidars.

[0057] S32. Based on the true value data of the three-dimensional wind field, simulated observation characteristic parameters corresponding to the scanning mode of a single Doppler wind lidar are generated through forward model simulation.

[0058] S33. Using simulated observation feature parameters as input and the corresponding three-dimensional wind field ground truth data as supervision labels, supervise the training of the deep learning wind field inversion model.

[0059] Furthermore, the preprocessing in step S2 specifically includes the following steps:

[0060] S21. In the signal range outside the effective detection range of the lidar, automatically identify and evaluate the noise levels of electromagnetic background and ambient light, and determine a dynamic noise floor threshold.

[0061] S22. Compare the raw power spectrum data of each detection range gate with the noise floor threshold. When the signal-to-noise ratio is lower than the preset threshold, mark it as invalid or low confidence data point and isolate or assign a special label in subsequent processing. Apply a bandwidth-adjustable bandpass filter to the valid raw power spectrum signal and lock the center frequency of the filter with the transmission frequency of the lidar.

[0062] S23. Assign a comprehensive signal quality index to each valid detection data point. The signal quality index takes into account the strength of the signal-to-noise ratio, the peak sharpness of the power spectrum signal, and the stability of the signal in time. Through the signal quality index, all detection data are divided into three quality levels: high, medium, and low, and different quality identifiers are attached to the data of each level.

[0063] S24. When the lidar is deployed on a mobile platform, it integrates attitude data from the platform's inertial navigation system, including changes in the platform's roll, pitch, and heading. Through the attitude data, it performs real-time geometric correction on the lidar's beam pointing vector at each moment, and uniformly compensates the measurement data to a stable reference coordinate system.

[0064] Furthermore, the preprocessing in step S2 also includes the following steps:

[0065] S25. Directly analyze and generate a series of core observation parameters from the spectrum data that has been processed as described above;

[0066] S26. Associate and bind the core observation parameters of each data point with its corresponding spatial metadata;

[0067] S27. Encapsulate the core observation parameters, quality identifiers, and spatial metadata of all probe points into a structured, georeferenced dataset according to spatial logical relationships.

[0068] Furthermore, the core observation parameters include radial wind speed, signal-to-noise ratio (SNR), and spectral width. Radial wind speed is derived from the displacement of the spectral peak relative to the transmission frequency. SNR is used to characterize the reliability of signal strength. Spectral width is used to reflect the intensity of atmospheric turbulence or wind shear. Spatial metadata includes the azimuth angle of the detection point, the elevation angle of the detection point, and the slant range of the detection point relative to the radar.

[0069] Furthermore, the interval between steps S2 and S3 includes:

[0070] S2a. Define the target three-dimensional spatial grid: Define a three-dimensional rectangular coordinate system with the location of a single wind-measuring lidar as the origin. Within the coordinate system, delineate a three-dimensional cubic space covering the effective detection range of the lidar as the target area. Divide the target area at fixed intervals in the horizontal, vertical and radial directions to form a three-dimensional spatial grid composed of numerous uniformly sized cubic grid units.

[0071] S2b. Establish a spatial impact model based on data quality: By using the signal quality index allocated in the preprocessing, a positive correlation is made for the initial weight of each original observation data point. For each three-dimensional spatial grid cell, all original observation data points in the spatial neighborhood are identified. The initial weight of each original observation data point is further attenuated by the spatial distance from the original observation data point to the center point of the three-dimensional spatial grid cell.

[0072] S2c, Perform quality-weighted feature value assignment: The radial wind speed values ​​of all original observation data points within the neighborhood of each 3D spatial grid cell are weighted and fused with the distance-attenuated weight values ​​to assign a radial wind speed feature value to the 3D spatial grid cell; the signal-to-noise ratio (SNR) values ​​of all original observation data points within the neighborhood of the same 3D spatial grid cell are weighted and fused with the distance-attenuated weights to assign a SNR feature value to that grid cell; the spectral width values ​​of all original observation data points within the neighborhood of the same 3D spatial grid cell are weighted and fused with the distance-attenuated weights to assign a spectral width feature value to that grid cell.

[0073] S2d, Generate Grid Confidence Label: Calculate the total weight of all original observation data points involved in the feature value calculation of each 3D spatial grid cell. Define the total weight as the data fullness of the 3D spatial grid cell. Based on the data fullness, assign a confidence level to each 3D spatial grid cell. When the data fullness of a 3D spatial grid cell is high, it is marked as high confidence. When the data fullness of a 3D spatial grid cell is low or zero, it is marked as low confidence or missing data.

[0074] Furthermore, the interval between steps S2 and S3 also includes:

[0075] S2e. Constructing a regularized 3D feature tensor: The radial wind speed, signal-to-noise ratio, and spectral width feature values ​​of all obtained 3D spatial grid cells are filled into a 3D matrix that corresponds completely to the 3D spatial grid dimension. The confidence level information of each 3D spatial grid cell is stored in an independent 3D matrix. The 3D matrix containing the observed feature values ​​is combined with the 3D matrix containing the confidence level to form a complete and regularized 3D feature tensor.

[0076] Furthermore, the method also includes post-processing optimization of the reconstructed 3D wind field, specifically including: spatial interpolation compensation of wind vectors in low-confidence or data-missing areas based on the confidence level of the 3D spatial grid cells, and smoothing of the reconstructed 3D wind field through wind field physical consistency constraints.

[0077] Example 2

[0078] Based on Example 1, this example further provides a specific implementation of a deep learning-based single-unit wind-measuring lidar three-dimensional wind field inversion system. This system is deployed on a mobile lidar platform equipped with an inertial navigation unit and is used for real-time inversion of the three-dimensional wind field of the urban boundary layer. The specific steps include:

[0079] 1. Hardware configuration:

[0080] It employs a 1550nm wavelength Doppler wind lidar with a detection range of 0.1-5km and a range resolution of 30m.

[0081] The scanning mode alternates between PPI (planar position display) and RHI (distance and height display), covering azimuth angle 0-360° and elevation angle 0-90°;

[0082] The platform is equipped with a high-precision IMU (Inertial Measurement Unit) that outputs roll, pitch, and heading data at a frequency of 100Hz.

[0083] 2. Data preprocessing and feature extraction:

[0084] According to steps S21-S24, the dynamic noise threshold is calculated in real time, and data points with a signal-to-noise ratio <-15dB are filtered out.

[0085] A bandpass filter with a bandwidth of 200MHz is applied to the effective signal to lock the transmission frequency;

[0086] Based on S25-S27, radial wind speed, signal-to-noise ratio, and spectral width are extracted and bound to azimuth, elevation, and slant range to form a structured dataset.

[0087] 3. 3D Mesh Construction and Feature Tensor Generation:

[0088] The target area is set to have a length × width × height of 1.5km × 1.5km × 1.0km and a grid resolution of 50m × 50m × 50m.

[0089] Based on S2a-S2e, a three-dimensional feature tensor is constructed with dimensions of 30×30×20. Each grid cell contains radial wind speed, signal-to-noise ratio, spectral width, and confidence level.

[0090] 4. Model Inference and Wind Field Reconstruction:

[0091] Using a pre-trained encoder-decoder model (approximately 1.2M parameters), the input is the aforementioned three-dimensional feature tensor;

[0092] The output is a 3D wind vector field (u, v, w) of 30×30×20×3.

[0093] The three-dimensional wind field of the target area is reconstructed based on the three-dimensional wind vector output by the deep learning wind field inversion model (corresponding to step S4). Based on the confidence level of the three-dimensional spatial grid cells, spatial interpolation compensation is performed on the wind vectors of low confidence or data missing areas. At the same time, the reconstructed three-dimensional wind field is smoothed by wind field physical consistency constraints (corresponding to the post-processing optimization step of claim 10).

[0094] 5. Real-time output and visualization:

[0095] The system achieves real-time inference on the NVIDIA Jetson AGX Orin platform with a frame rate ≥1Hz;

[0096] The wind field results are displayed in real time through a web interface, supporting various visualization methods such as wind speed vectors, streamline diagrams, and vertical profiles.

[0097] Comparison Example

[0098] To verify the effectiveness of the method of the present invention, the following comparative example is set up:

[0099] Comparison with Example 1 (traditional VAD method): The velocity azimuth display method is used. Based on the same set of radial wind speed data, the horizontal wind field is inverted through sine fitting, while ignoring the vertical wind speed;

[0100] Comparison with Example 2 (traditional VVP method): The volume velocity processing method is used, and the three-dimensional wind field is inverted based on least squares fitting, assuming that the wind field is uniformly distributed within the scanning volume;

[0101] Comparison Example 3 (DL method without confidence): The same deep learning model as the present invention is used, but confidence labeling and post-processing optimization are not introduced.

[0102] Experimental Design and Data Collection

[0103] The experiment was conducted in March 2024 at a meteorological observation center in northern China. One fixed lidar and three mobile lidars were used for simultaneous observation to acquire true data. The experiment included three weather conditions: sunny, cloudy, and rainy, with wind speeds ranging from 1 to 15 m / s. The comparison indicators included:

[0104] Horizontal wind speed error (MAE, m / s);

[0105] Wind direction error (MAE, °);

[0106] Vertical wind speed error (MAE, m / s);

[0107] Data effectiveness (%)

[0108] Calculation time (ms / frame).

[0109] Table 1: Comparison of Mean Absolute Error (MAE) of Different Methods in Horizontal Wind Speed ​​Inversion (Unit: m / s)

[0110]

[0111] Table 2: Comparison of Mean Absolute Error (MAE) of Different Methods in Wind Direction Inversion (Unit: °)

[0112]

[0113] Table 3: Comparison of Vertical Wind Speed ​​Inversion Error and Data Validity

[0114]

[0115] Table 4: Comparison of computational efficiency

[0116]

[0117] Experimental Data Analysis Explanation

[0118] 1. Accuracy Advantage: The present invention has significantly lower inversion errors for horizontal wind speed, wind direction and vertical wind speed than traditional methods and unoptimized DL methods, especially under medium and high wind speed conditions. This shows that the present invention effectively improves the accuracy and reliability of wind field inversion through multi-scale feature fusion, confidence labeling and post-processing optimization.

[0119] 2. Data efficiency: The present invention maintains a data efficiency of up to 98.5% even under complex weather conditions, which is better than other methods, indicating that it has stronger noise suppression and data compensation capabilities.

[0120] 3. Real-time analysis: Although the inference time of this invention is slightly longer than that of the VAD method, it still meets the real-time processing requirements and is much faster than the VVP method. In mobile platform deployment scenarios, this invention achieves a good balance between accuracy and efficiency.

[0121] 4. Summary of technical effects: This invention achieves high-precision, high-robustness, and real-time operation of three-dimensional wind field inversion through a technical approach of single lidar + deep learning + quality optimization.

[0122] 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 retrieving three-dimensional wind field based on deep learning of single wind lidar, characterized in that, The method comprises the following steps: S1, obtaining original power spectrum data obtained by scanning and detecting an atmospheric wind field by a single Doppler wind lidar; S2, preprocessing the original power spectrum data to extract observation characteristic parameters including at least radial wind speed, signal-to-noise ratio and spectral width; S3, inputting the observation characteristic parameters into a pre-trained deep learning wind field inversion model, wherein the deep learning wind field inversion model takes the observation characteristic parameters as input and outputs a three-dimensional wind vector of a three-dimensional space grid point with the position of the lidar as the origin; Between the steps S2 and S3, the following steps are included: S2a, defining a target three-dimensional space grid: defining a three-dimensional rectangular coordinate system with the position of the single wind lidar as the origin, and in the coordinate system, delimiting a three-dimensional cubic space covering the effective detection range of the lidar as a target area, and dividing the target area in the horizontal direction, vertical direction and radial distance by fixed intervals to form a three-dimensional space grid composed of many cubic grid units of uniform size; S2b, establishing a spatial influence model based on data quality: through the signal quality index assigned in the preprocessing, performing forward association of the initial weight for each original observation data point, identifying all original observation data points in the spatial neighborhood of each three-dimensional space grid unit, and further attenuating the initial weight of each original observation data point through the spatial distance from the original observation data point to the center point of the three-dimensional space grid unit; S2c, performing quality-weighted feature value assignment: weighting and fusing the radial wind speed values of all original observation data points in the neighborhood of each three-dimensional space grid unit with the weight values after distance attenuation to assign radial wind speed feature values to the three-dimensional space grid unit; weighting and fusing the signal-to-noise ratio values of all original observation data points in the neighborhood of the same three-dimensional space grid unit with the weight values of the data points after distance attenuation correction to assign signal-to-noise ratio feature values to the grid unit; weighting and fusing the spectral width values of all original observation data points in the neighborhood of the same three-dimensional space grid unit with the weight values of the data points after distance attenuation correction to assign spectral width feature values to the grid unit; S2d, generating a grid confidence identification: counting the total weight of all original observation data points participating in the feature value calculation of each three-dimensional space grid unit, defining the total weight as the data fullness of the three-dimensional space grid unit, and assigning a confidence level to each three-dimensional space grid unit according to the size of the data fullness, when the data fullness of the three-dimensional space grid unit is high, it is marked as high confidence, when the data fullness of the three-dimensional space grid unit is low or zero, it is marked as low confidence or data missing; S4, reconstructing a three-dimensional wind field of the target area according to the three-dimensional wind vector output by the deep learning wind field inversion model.

2. The deep learning-based single wind lidar three-dimensional wind field retrieval method according to claim 1, characterized in that, The deep learning wind field inversion model is a neural network with an encoder-decoder structure, wherein: the encoder is used for feature extraction and spatial coding of the input observation characteristic parameters to form a hidden space feature vector containing global wind field information; The decoder is used to decode and regress the three-dimensional wind vector of the three-dimensional space grid point according to the latent space feature vector.

3. The deep learning-based single wind lidar three-dimensional wind field retrieval method according to claim 2, characterized in that, In the training process of the deep learning wind field inversion model, a multi-scale feature fusion mechanism is adopted, specifically: a multi-level convolution module is introduced in the encoder to extract local wind field structure and global wind field distribution features at different spatial scales, respectively; features at different levels in the encoder are transmitted to the corresponding levels in the decoder through a jump connection; and a progressive upsampling structure is adopted in the decoder to gradually restore the spatial resolution of the three-dimensional wind field.

4. The deep learning-based single wind lidar three-dimensional wind field retrieval method according to claim 1, characterized in that, The training process of the deep learning wind field inversion model comprises: S31, obtaining a training data set, wherein the training data set comprises high-resolution three-dimensional wind field data generated by a numerical weather prediction model, or three-dimensional wind field true value data obtained by cooperative observation and fusion of multiple laser radars; S32, based on the three-dimensional wind field true value data, simulating and generating simulated observation feature parameters corresponding to the scanning mode of the single Doppler wind lidar through a forward model; S33, performing supervised training on the deep learning wind field inversion model by taking the simulated observation feature parameters as input and taking the corresponding three-dimensional wind field true value data as a supervision label.

5. The deep learning-based single wind lidar three-dimensional wind field retrieval method according to claim 1, characterized in that, The preprocessing in step S2 specifically comprises the following steps: S21, automatically identifying and evaluating the noise level of electromagnetic background and ambient light in the signal interval outside the effective detection distance of the laser radar, and determining a dynamic noise floor threshold; S22, comparing the original power spectrum data at each detection distance gate with the noise floor threshold, marking the data points with a signal-to-noise ratio lower than a preset threshold as invalid or low-confidence data points, isolating or assigning special identifiers to the invalid data points in subsequent processing, and applying a bandwidth-adjustable bandpass filter to the valid original power spectrum signal, with the center frequency of the filter being locked to the transmission frequency of the laser radar; S23, assigning a comprehensive signal quality index to each valid detection data point, wherein the signal quality index comprehensively considers the strength of the signal-to-noise ratio, the peak sharpness of the power spectrum signal, and the stability of the signal in time, and dividing all detection data into high, medium and low quality levels by the signal quality index, and attaching different quality identifiers to the data at each level; S24, when the laser radar is deployed on a mobile platform, integrating attitude data from the platform inertial navigation system, including roll, pitch and heading changes of the platform, and performing real-time geometric correction on the beam pointing vector of the laser radar at each time through the attitude data to compensate the measurement data to a stable reference coordinate system.

6. The deep learning-based single wind lidar three-dimensional wind field retrieval method according to claim 5, characterized in that, The preprocessing in step S2 further comprises the following steps: S25, directly analyzing and generating a series of core observation parameters from the spectrum data processed in the foregoing steps; S26, associating and binding the core observation parameters of each data point with its corresponding spatial metadata; S27, encapsulating the core observation parameters, quality identifiers and spatial metadata of all detection points into a structured, geographically referenced data set according to the spatial logical relationship.

7. The deep learning-based single wind lidar three-dimensional wind field retrieval method according to claim 6, characterized in that, The core observation parameters include radial wind speed, signal-to-noise ratio and spectral width, wherein the radial wind speed is obtained from the displacement amount of the spectral peak relative to the transmission frequency, the signal-to-noise ratio is used to characterize the reliability of the signal strength, and the spectral width is used to reflect the atmospheric turbulence intensity or wind shear condition, and the spatial metadata include the azimuth angle of the detection point, the elevation angle of the detection point and the slant range of the detection point relative to the radar.

8. The deep learning-based single wind lidar three-dimensional wind field retrieval method according to claim 1, characterized in that, The step S2 and the step S3 further include: S2e, constructing a regularized three-dimensional feature tensor: filling the obtained radial wind speed, signal-to-noise ratio and spectral width feature values of all three-dimensional space grid units into a three-dimensional matrix corresponding to the three-dimensional space grid dimension, respectively, storing the confidence level information of each three-dimensional space grid unit in an independent three-dimensional matrix, and combining the three-dimensional matrix containing the observation feature values with the three-dimensional matrix containing the confidence levels to form a complete and regularized three-dimensional feature tensor.

9. The deep learning-based single wind lidar three-dimensional wind field retrieval method according to claim 1, characterized in that, The method further includes post-processing optimization of the reconstructed three-dimensional wind field, specifically including: based on the confidence level of the three-dimensional space grid unit, performing spatial interpolation compensation on the wind vector in the low-confidence or data-missing area, and performing smoothing processing on the reconstructed three-dimensional wind field through wind field physical consistency constraint.

Images