Wind field data reconstruction method for wind lidar

Abstract

The present application relates to the technical field of laser radar atmospheric detection, and specifically discloses a wind field data reconstruction method of wind measurement laser radar. The wind field data reconstruction method of wind measurement laser radar comprises the following steps: collecting atmospheric echo signals of a target wind field by a wind measurement laser radar, calculating a u component and a v component of a horizontal wind speed, and constructing a wind field data set based on the u component and the v component of the horizontal wind speed; preprocessing wind field data of the wind field data set, and dividing the wind field data set into a training set, a validation set and a test set; constructing a physical information generative adversarial network model, which comprises a generator for generating a prediction result and a discriminator for evaluating the authenticity of the prediction result; training the physical information generative adversarial network model by using the training set and the validation set; inputting the test set into the trained physical information generative adversarial network model, and outputting reconstructed wind field data. The method improves the completeness and availability of wind field data.

Description

Technical Field

[0001] This invention relates to the field of lidar atmospheric detection technology, and in particular to a method for reconstructing wind field data using a wind-measuring lidar. Background Technology

[0002] Wind lidar is an important tool for acquiring high-resolution vertical wind field profiles. However, in actual operation, due to limitations in atmospheric aerosol concentration, weather conditions, and the signal-to-noise ratio of the optical system, observational data often exhibits random gaps at specific altitudes. Existing data reconstruction methods mostly employ linear interpolation or statistical regression. These methods often ignore the physical continuity of atmospheric motion in vertical space and its evolution over time, resulting in numerical jumps and poor physical consistency in the reconstructed wind field data compared to the original radar data. This makes it difficult to meet the needs of high-precision meteorological analysis and wind power regulation.

[0003] Wind lidar, as an important remote sensing wind sensing device, has been widely used in wind energy resource assessment, atmospheric boundary layer research, and wind farm operation monitoring. However, in practical applications, wind speed data acquired by lidar often suffers from missing data or unstable data quality due to factors such as insufficient aerosol concentration, precipitation, instrument signal attenuation, and environmental interference. This not only reduces the completeness and usability of the data but also significantly impacts subsequent applications such as wind resource assessment, numerical simulation, and intelligent control. Therefore, how to efficiently and accurately reconstruct missing lidar wind field data has become one of the most pressing problems to be solved in this field.

[0004] Existing data completion methods mainly include statistical interpolation-based methods (such as linear interpolation and Kriging interpolation), time series analysis-based methods (such as the ARIMA model), and deep learning-based methods that have emerged in recent years (such as recurrent neural networks and convolutional neural networks). Although these methods can achieve data completion to some extent, they still have significant limitations: on the one hand, traditional statistical methods are difficult to characterize the complex nonlinear dynamics of wind fields; on the other hand, while purely data-driven deep learning methods have strong fitting capabilities, they often ignore the physical laws governing wind field evolution, easily leading to insufficient physical consistency and generalization ability in the results, making it difficult to meet the needs of high-precision meteorological analysis and wind power regulation.

[0005] To address the aforementioned issues, in recent years, Physically Informed Neural Networks (PINNs) have demonstrated significant advantages in improving the physical consistency of models by incorporating physical constraints such as governing equations into the neural network training process. Meanwhile, Generative Adversarial Networks (GANs) exhibit excellent performance in data distribution learning and high-quality sample generation. However, using PINN or GAN alone still struggles to simultaneously achieve both data distribution approximation capability and physical constraint consistency. Summary of the Invention

[0006] Therefore, the purpose of this invention is to provide a wind field data reconstruction method using a wind-measuring lidar to improve the integrity and usability of wind field data.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] This invention provides a method for reconstructing wind field data using a wind-measuring lidar, comprising the following steps:

[0009] S1. Collect atmospheric echo signals of the target wind field using wind-measuring lidar, calculate the u and v components of the horizontal wind speed, and construct a wind field dataset based on the u and v component data of the horizontal wind speed.

[0010] S2. Preprocess the wind field data in the wind field dataset and divide the wind field dataset into a training set, a validation set, and a test set;

[0011] S3. Construct a physical information generative adversarial network model, wherein the physical information generative adversarial network model includes an attention-based generator and a gradient-aware discriminator, wherein the generator is used to generate prediction results and the discriminator is used to evaluate the authenticity of the generator's prediction results;

[0012] S4. Train the physical information generative adversarial network model using the training set and validation set; input the test set into the trained physical information generative adversarial network model and output the reconstructed wind field data.

[0013] Preferably, in step S1, the step of acquiring atmospheric echo signals of the target wind field using a wind-measuring lidar and calculating the u and v components of the horizontal wind speed includes:

[0014] The wind-measuring lidar emits laser pulses sequentially in four azimuth directions: east, south, west, and north, and receives echo signals formed by backscattering from aerosol particles, thereby obtaining the Doppler frequency shift f at different azimuth angles. d ;

[0015] according to The radial wind speeds at the four azimuth angles of east, south, west, and north were calculated respectively. , , , Then, the u and v components of the horizontal wind speed are calculated using the DBS synthesis formula:

[0016]

[0017]

[0018] Where λ represents the laser wavelength and r represents the radial direction.

[0019] Preferably, in step S2, the data of the wind field dataset is preprocessed, including:

[0020] Calculate the carrier-to-noise ratio (CNR) of wind field data:

[0021]

[0022] In the formula: η represents quantum efficiency, P r Where is the signal echo power, h is Planck's constant, f is the laser carrier frequency, and B represents the radar's receiving bandwidth.

[0023] The wind field data is filtered based on the carrier-to-noise ratio. When the carrier-to-noise ratio of the wind field data is less than a preset threshold, the wind field data is determined to be invalid and removed.

[0024] Calculate the mean of wind field data and standard deviation Anomaly detection is performed using the three-standard-deviation rule: when a certain observation value... satisfy If the wind field data is found to be outlier, it will be removed.

[0025] Preferably, in step S2, the preprocessing of the wind field dataset further includes:

[0026] Using a one-dimensional linear interpolation method, the missing positions are estimated using known observation points (z0, y0) and (z1, y1). Then, any missing position within the interval (z0, z1) can be calculated. Wind field data fill value at the location Represented as:

[0027]

[0028] In the formula: The missing wind field data filler value represents the missing location, z represents the corresponding height layer of the current missing data; z0 and z1 represent the two adjacent height layers before and after the location of the missing data, and y0 and y1 are the wind field data of height layers z0 and z1, respectively.

[0029] Preferably, in step S2, the preprocessing of the wind field data in the wind field dataset further includes:

[0030] The wind field data is normalized using the following formula:

[0031]

[0032] in, Represents wind field data, The mean of the wind field data. The standard deviation of wind field data This represents the normalized wind field data.

[0033] Preferably, in step S2, the preprocessing of the wind field dataset further includes: randomly setting a predetermined number of normalized wind field data as missing values, and using masks 0 and 1 to mark the missing data and the non-missing data respectively.

[0034] Preferably, in step S3, the generator adopts an improved 1D U-Net residual network architecture, introducing a multi-head self-attention mechanism and a residual learning strategy. A multi-head self-attention module is added to the 1D U-Net model to capture the long-range spatial dependence features of the wind field data across the entire altitude range. The generator's output uses residual learning logic, adding the generated residual terms to the original radar observations at the current moment to obtain the final predicted wind field data.

[0035] Preferably, in step S3, the discriminator employs a multi-layer one-dimensional convolutional neural network structure. The discriminator extracts the first-order and second-order vertical gradient features of the wind field profile in parallel at the input layer, and concatenates these gradient features with the original wind speed numerical features in the channel dimension. Through comprehensive discrimination of numerical distribution features and spatial rate of change features, the generator is forced to restore high-frequency detail features that conform to the statistical characteristics of atmospheric turbulence during the reconstruction process.

[0036] In terms of model structure design, this invention adopts a lightweight one-dimensional neural network architecture, built entirely on one-dimensional convolution. Compared to traditional two-dimensional or three-dimensional models, this significantly reduces the parameter scale and computational complexity, improving computational efficiency and deployment flexibility while maintaining modeling capabilities. Simultaneously, the one-dimensional processing method can directly adapt to LiDAR vertical profile data, avoiding redundant computations caused by high-dimensional modeling, and exhibiting good versatility and scalability. In the generator design, a lightweight one-dimensional UNet structure is used in conjunction with an attention mechanism. Multi-scale feature extraction is achieved through an encoder-decoder framework, enabling efficient fusion of global and local information with fewer parameters. An attention mechanism is introduced into the bottleneck layer to enhance the model's ability to perceive key wind field structures. In the discriminator, one-dimensional convolution and gradient enhancement strategies are combined, and first- and second-order derivative information is introduced through simple difference calculations. This effectively improves the model's ability to discriminate detailed changes in wind field and texture features without adding complex computations.

[0037] Preferably, in step S4, during the training process, the parameters of the physical information generative adversarial network model are iteratively optimized using a composite loss function, which includes the Navier-Stokes equation residual loss function, the baseline loss function, the temporal smoothing loss function, and the boundary smoothing loss function.

[0038] Preferably, in step S4, the test set is input into the trained physical information to generate an adversarial network model, and the reconstructed wind field data is output, including:

[0039] The test set data is input into the model to obtain the predicted wind field data, namely the u and v components of the predicted horizontal wind speed.

[0040] The discriminator performs inverse normalization on the model output:

[0041]

[0042] In the formula, This represents the wind field data after denormalization. For predicted wind field data; The mean of the predicted wind field data. The standard deviation of the predicted wind field data;

[0043] Introducing Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Pearson Correlation Coefficient A quantitative analysis was conducted on the difference between the predicted wind field data and the actual wind field data after inverse normalization.

[0044] Based on the predicted wind field data, the horizontal wind speed magnitude V and wind direction are synthesized according to the following formula. :

[0045]

[0046]

[0047] The results will also be visualized.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] This invention innovatively combines the Physical Information Neural Network (PINN) with the Generative Adversarial Network (GAN) to construct a hybrid model for wind lidar data incompleteness. On one hand, through adversarial training between the generator and discriminator in the GAN structure, the complex spatiotemporal distribution characteristics of wind field data are effectively learned, improving the reconstruction accuracy and detail representation of missing wind field data. On the other hand, the PINN framework is introduced to embed physical constraints related to wind field evolution into the loss function, thereby ensuring that the generated results satisfy basic physical laws. This fusion method, while balancing the advantages of data-driven approaches and physical constraints, significantly improves the model's generalization ability and stability under complex meteorological conditions.

[0050] The generative adversarial network model (PINN-GAN) with integrated physical constraints constructed in this invention significantly improves the data reconstruction accuracy and physical reliability of wind lidar in complex atmospheric environments. By introducing the Navier-Stokes equations as core constraints, the principles of mass and momentum conservation from fluid dynamics are embedded into the model training process, overcoming the limitations of traditional methods that rely solely on numerical fitting. This ensures the physical consistency of the reconstructed wind field in both the u and v components and effectively suppresses common non-physical oscillation problems in areas with missing data.

[0051] Compared to traditional methods and single deep learning models, the method of this invention can not only effectively reduce the error of missing data imputation, but also improve the physical rationality and spatial continuity of the reconstructed data, which is of great significance for improving the quality of lidar wind field data. At the same time, this method has good scalability and can be extended to other types of remote sensing data repair and physical field reconstruction problems, demonstrating high engineering application value and promising prospects for wider application. Attached Figure Description

[0052] Figure 1 A flowchart of a wind field data reconstruction method provided by an embodiment of the present invention using a wind-measuring lidar;

[0053] Figure 2 This is a structural diagram of the physical information generation adversarial network model in an embodiment of the present invention;

[0054] Figure 3 The image shows the reconstruction effect of horizontal wind speed and direction using a wind field data reconstruction method provided by an embodiment of the present invention, which utilizes wind measurement lidar. Detailed Implementation

[0055] To facilitate understanding of the present invention, a more comprehensive description will be given below with reference to specific embodiments. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of the present invention.

[0056] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0057] See Figure 1 This embodiment provides a method for reconstructing wind field data using a wind-measuring lidar, comprising three stages: data acquisition and preprocessing, construction and training of a Physical Information Generative Adversarial Network (PINN-GAN) model, and data reconstruction and evaluation.

[0058] I. Data Acquisition and Preprocessing Stage

[0059] First, perform step S1: acquire lidar wind field dataset: transmit pulse, receive echo, and synthesize wind field using DBS.

[0060] A wind-measuring lidar system emits laser pulses into the atmosphere, receives echo signals formed by backscattering from aerosol particles, and obtains radial wind speeds in different directions from the echo signals based on the Doppler frequency shift principle. The u and v components of the horizontal wind speed are then synthesized using Doppler beamsweeping (DBS) technology.

[0061] The wind-measuring lidar system sequentially emits laser pulses in four azimuth directions: east, south, west, and north, with a fixed elevation angle θ for each azimuth. By receiving the backscattered echo signals from aerosol particles in the atmosphere, it measures the Doppler frequency shift f in the east, south, west, and north directions. d .

[0062] According to the formula The radial wind speeds in the east, south, west, and north directions were calculated respectively. , , , , where λ is the laser wavelength and r represents the radial direction.

[0063] Subsequently, the u and v components of the horizontal wind speed varying with height were calculated using the DBS synthesis formula:

[0064]

[0065]

[0066] The U-component and V-component data profiles of the synthesized horizontal wind speed data are used to construct a wind field dataset.

[0067] Next, step S2 is performed: Data preprocessing: handling noisy data, outlier removal, missing value imputation, normalization, and random missing data labeling. Specifically:

[0068] S21. Filter wind field data based on carrier-to-noise ratio (CNR).

[0069] The formula for calculating the carrier-to-noise ratio is:

[0070]

[0071] In the formula: η represents quantum efficiency, P r denoted as signal echo power, h as Planck's constant, f as laser carrier frequency, and B as radar receiving bandwidth.

[0072] When the carrier-to-noise ratio of wind field data is less than a preset threshold (10dB in this embodiment), the wind field data is deemed invalid and discarded. Furthermore, outliers in the wind field data are identified and removed through statistical analysis. The specific method is as follows:

[0073] Calculate the mean of wind field data and standard deviation Anomaly detection is performed using the three-standard-deviation principle, i.e., when a given observation x satisfies... If an outlier is detected, the wind field data is identified as an outlier and removed. Outliers may be caused by equipment malfunctions, environmental interference, or other external factors. Removing these outliers can improve the quality and reliability of the wind field data.

[0074] S22. Fill in the missing data.

[0075] Since neural networks cannot directly process inputs with null values, it is necessary to initially fill in the missing parts of the wind field data. This invention uses a one-dimensional linear interpolation method to estimate the intermediate missing positions using known observation points (z0, y0) and (z1, y1). Therefore, any missing position within the interval (z0, z1) can be filled in. Wind field data fill value at the location Represented as:

[0076]

[0077] In the formula: The missing wind field data filler value represents the missing location, z represents the corresponding height layer of the current missing data; z0 and z1 represent the two adjacent height layers before and after the location of the missing data, and y0 and y1 are the wind field data of height layers z0 and z1, respectively.

[0078] The interpolated data generated in this step is only used as an auxiliary input for the model to capture the spatiotemporal context. The missing data and the non-missing data are marked by masks 0 and 1 respectively. The final reconstruction result will still be corrected by the physical residual generated by the neural network through the masking mechanism.

[0079] S23. Normalize the wind field data.

[0080] To eliminate the impact of differences in the dimensions of different data on the gradient update of the neural network, the wind field data is normalized to transform data with different dimensions and scales into the same dimension and scale range. This invention uses the Z-Score normalization method, and the specific formula is as follows:

[0081]

[0082] in, Represents wind field data, The mean of the wind field data. The standard deviation of wind field data This represents the normalized wind field data.

[0083] S24. Perform missing data processing on the normalized wind field data.

[0084] To enable the model to learn how to reconstruct wind fields, 20% of the normalized wind field data were randomly set as missing values. This artificially created missing data simulated the missing radar data phenomenon in real-world scenarios. Missing and non-missing data were then marked with masks of 0 and 1 respectively, allowing the model to distinguish between them.

[0085] S25. Randomly divide the wind field data into training, validation, and test sets in a ratio of 7:2:1 for subsequent model training and testing.

[0086] II. Model Building and Training Phase

[0087] Step S3: Construct and train a physical information generative adversarial network model (hereinafter referred to as PINN-GAN) to achieve high-fidelity reconstruction of wind field features missing from lidar.

[0088] Combination Figure 2 PINN-GAN consists of an attention-based generator and a gradient-aware discriminator.

[0089] The generator is based on the classic U-Net and has three improvements: First, the original U-Net model is modified from a two-dimensional network to a one-dimensional (1D) symmetric convolutional architecture, making it more consistent with the structural features of the vertical data profile of the wind field. This allows for efficient extraction of deep spatial features through multi-layer downsampling and upsampling, while reducing the model's hardware performance requirements. Second, an attention module is added, incorporating a multi-head attention mechanism to compensate for the limitations of local convolutions and accurately capture the long-range spatial dependence features of the wind field data across the entire altitude range. Finally, a residual learning strategy is introduced, whereby the model does not directly predict the wind field data, but rather, based on the initial linear interpolation results, only learns and predicts the deviation correction amount (i.e., residual) between the missing parts filled by these interpolations and the real wind field data.

[0090] The generator's input consists of 8-channel feature maps containing wind field data from the previous time t-1, the current time t, the next time t+1, the current observation mask, and normalized altitude coordinates. The generator's output employs residual learning logic, obtaining the final predicted wind field data by interpolating the generated residual terms with the initial padding values ​​obtained through linear interpolation. This approach allows the model to focus more on error correction in missing regions, reducing training complexity and further improving wind field reconstruction accuracy.

[0091] The discriminator employs a multi-layer one-dimensional convolutional neural network structure. By judging the authenticity of wind speed numerical distribution and gradient texture, it prompts the generator to produce data that better conforms to atmospheric dynamics. The discriminator extracts the first-order and second-order vertical gradient features of the wind field profile in parallel at the input layer, and concatenates these gradient features with the original wind speed numerical features along the channel dimension. Through comprehensive discrimination of numerical distribution features and spatial rate of change features, the generator is forced to restore high-frequency detail features that conform to the statistical characteristics of atmospheric turbulence during the reconstruction process.

[0092] To ensure that the reconstructed wind field conforms to atmospheric physical laws, this invention designs a composite loss function to optimize the model parameters. The composite loss function consists of a baseline loss, a physical loss function based on the Navier-Stokes (NS) equations, a boundary smoothing loss term, and a time-dependent loss term.

[0093] (1) Baseline Loss: The masked L1 loss is used to calculate the difference between the generated values ​​and the true values ​​within the effective observation region. The specific formula for the L1 loss function is as follows:

[0094]

[0095] Where i is the index value of the height layer. V represents the model's predicted wind field data (u or v component) at the i-th altitude level. i M represents the true value of the wind field data at the i-th altitude level. i This represents the effective observation mask for the i-th height layer.

[0096] (2) Boundary smoothing loss: This includes global second-order smoothing and first-order smoothing. By calculating the second and first derivatives of wind speed with respect to height, the gradient and numerical changes of the wind field profile at the boundary of the original radar data are constrained in the reconstruction result, preventing abnormal high-frequency oscillations and ensuring that the reconstruction result can be smoothly connected with the original radar data. The boundary smoothing loss can be expressed by the following formula:

[0097]

[0098] in, , Let represent the first-order smoothing function and the second-order smoothing function, respectively. The specific formulas are as follows:

[0099]

[0100]

[0101] in, This represents the predicted value of the wind field data at the i-th altitude level; , These are the predicted wind field data for the (i+1)th and (i-1)th altitude layers, respectively; b i The boundary weights are used to indicate whether the effective observation mask has changed; 0 indicates no change, and 1 indicates a change.

[0102] The edges of the observation mask are extracted using the max pooling operator to construct the boundary mask region. The first and second spatial derivatives of the reconstructed wind field within the boundary mask region are calculated. By minimizing the sum of squares of the first and second derivatives of the boundary region, the numerical jumps and curvature discontinuities between the original effective radar observations and the model-predicted values ​​at the hard fusion stitching interface are eliminated.

[0103] (3) Physical loss function based on Navier-Stokes (NS) equations:

[0104] The calculation process is as follows: First, the predicted values ​​output by the model are denormalized to restore the true physical dimensions of the predicted values, resulting in physical quantities including horizontal wind speed, vertical wind speed, and kinematic pressure. The partial derivatives of the physical quantities in vertical space are calculated using the central difference method, and the time partial derivatives are approximated using observations at adjacent times. The calculated derivatives are substituted into the Navier-Stokes equations consisting of the continuity equation, the horizontal momentum equation, and the vertical momentum equation to calculate the physical residuals of the equations, and the mean square error of the physical residuals is used as a regularization constraint term for model training.

[0105] This invention achieves strong constraints on the model output by calculating the residuals of the following four physical equations and converting them into MSE loss terms:

[0106] Continuity equation loss (mass conservation): In a one-dimensional vertical cylinder model, this can be simplified to a constraint on the vertical wind speed gradient. .

[0107] The loss in the horizontal momentum equation (X, Y components): Considering local acceleration, vertical convection term, and viscous diffusion term, we can obtain:

[0108]

[0109]

[0110] Vertical momentum equation residuals (Z direction): Introducing a pressure gradient term

[0111]

[0112] Where Re is the Reynolds number, representing the ratio of inertial force to viscous force. The optimal value is determined experimentally, and the Reynolds number used in this invention is 2.5 × 10⁻⁶. 4 The choice of this value is to ensure that the inertial term dominates in the Navier-Stokes equations while implicitly smoothing and regularizing the output of the neural network through the viscous diffusion term, thereby suppressing non-physical noise while preserving the turbulent characteristics of the wind field.

[0113] The model minimizes This forces the reconstructed wind field to approximate the actual fluid motion state.

[0114] (4) Temporal correlation loss: Using the wind field data at times t-1 and t+1 in the input channel as a reference, the prediction result at the current time t is constrained to ensure that the evolution of the wind field on the time axis conforms to the principle of continuity. The formula for temporal correlation loss is as follows:

[0115]

[0116] in, This is the predicted wind field value at the current time t. This is the wind field data at the altitude layer corresponding to the previous time t-1. This provides the wind field data at the corresponding altitude level for the next time step t+1. , These represent the effective data masks for the previous time t-1 and the next time t+1, respectively.

[0117] III. Data Reconstruction and Evaluation Phase

[0118] Step S4: Reconstruct the test data using the trained PINN-GAN model. This specifically includes:

[0119] S41: Input the test set data into the trained PINN-GAN model to obtain predicted wind field data to evaluate the model's generalization performance. Specifically, input the test set data preprocessed in step S2 into the generator and output the predicted wind field data, including the u and v components of the predicted horizontal wind speed.

[0120] S42: Perform inverse normalization on the output result. Inverse normalization is the inverse transform of normalization. The specific formula is as follows:

[0121]

[0122] In the formula, This represents the wind field data after denormalization. For predicted wind field data, The mean of the predicted wind field data. The standard deviation of the predicted wind field data is used to recover its physical dimensions.

[0123] S43: Multiple error evaluation indicators are introduced to quantitatively analyze the difference between the inversely normalized wind field data and the actual wind field data (actual radar observations, i.e., the u and v components calculated in step S1). These include root mean square error (RMSE), mean absolute error (MAE), and Pearson correlation coefficient. The root mean square error (RMSE) is used to quantitatively assess the absolute deviation between the predicted and actual wind field data. The specific formula is as follows:

[0124]

[0125] In the above formula, V represents the wind field data after inversion, where V is the actual wind field data and M is the effective observation mask.

[0126] Mean Absolute Error (MAE) is used to evaluate the mean absolute error between the denormalized wind field data and the actual wind field data, and can be expressed by the following formula:

[0127]

[0128] The Pearson correlation coefficient is used to assess the similarity in data trends between the denormalized wind field data and the actual wind field data. The formula is as follows:

[0129]

[0130] in, This represents the average value of the wind field data after inverse normalization. This represents the average value of the actual wind field data.

[0131] S44: Combine the u and v components of the wind field output by the model to obtain horizontal wind speed and horizontal wind direction for visualization. (Horizontal wind speed magnitude V and horizontal wind direction are also shown.) It can be calculated using the following formula:

[0132]

[0133]

[0134] The results are also visualized.

[0135] When outputting the final reconstructed data, the system uses element-by-element logic to determine whether the observation mask of the current height layer is 1, indicating that the data is valid, and the original normalized observation value of the lidar is directly retained. If the observation mask of the current height layer is 0, it indicates that the data is invalid or missing, and the predicted value output by the generator is filled in. This strategy ensures that the original observation accuracy of the radar in the high carrier-to-noise ratio detection range is not changed while repairing missing data.

[0136] This embodiment optimizes the data preprocessing and result output stages. It improves the reliability of input data through noise filtering and quality control, and adopts fusion and smoothing strategies to achieve a smooth transition between observed and predicted values ​​without changing the high signal-to-noise ratio observation data. This ensures the continuity and stability of the reconstruction results, thereby effectively improving data fidelity and data availability under low signal-to-noise ratio conditions.

[0137] This invention introduces a residual learning mechanism to reduce the difficulty of model learning and improve training stability and convergence speed. The overall method achieves a balance between structural simplicity, computational efficiency, and application flexibility while ensuring reconstruction accuracy. It is particularly suitable for resource-constrained environments and online wind field data processing scenarios, and has promising engineering application prospects.

[0138] Figure 3 To illustrate the reconstruction effect of horizontal wind speed and direction using the wind field data reconstruction method provided in this embodiment using a wind-measuring lidar, Figure 3 The left side shows the original radar observation results. Figure 3The right side shows the reconstruction results of the PINN-GAN model. The comparison results show that the reconstructed wind field data is highly consistent with the original radar observation data, maintaining the vertical variations and temporal evolution of wind speed and direction at all altitude levels. For missing regions in the original data, the model effectively recovers the data trends and layered structure, and the reconstruction results do not show significant distortion or outliers. Overall, the PINN-GAN model not only accurately preserves the main features and trends of the original wind field data but also reasonably reconstructs the wind speed and direction information in missing regions, indicating that this method has high reconstruction accuracy and good spatiotemporal consistency, meeting the needs of wind field data reconstruction and application.

[0139] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0140] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for reconstructing wind field data using a wind-measuring lidar, characterized in that, It includes the following steps: S1. Collect atmospheric echo signals of the target wind field using wind-measuring lidar, calculate the u and v components of the horizontal wind speed, and construct a wind field dataset based on the u and v component data of the horizontal wind speed. S2. Preprocess the wind field data in the wind field dataset and divide the wind field dataset into a training set, a validation set, and a test set; S3. Construct a physical information generative adversarial network model, wherein the physical information generative adversarial network model includes a generator and a discriminator, wherein the generator is used to generate prediction results and the discriminator is used to evaluate the authenticity of the generator's prediction results; S4. Train the physical information generative adversarial network model using the training set and validation set; input the test set into the trained physical information generative adversarial network model and output the reconstructed wind field data.

2. The wind field data reconstruction method of wind-measuring lidar according to claim 1, characterized in that, In step S1, the step of acquiring atmospheric echo signals of the target wind field using a wind-measuring lidar and calculating the u and v components of the horizontal wind speed includes: The wind-measuring lidar emits laser pulses sequentially in four azimuth directions: east, south, west, and north, and receives echo signals formed by backscattering from aerosol particles, thereby obtaining the Doppler frequency shift f at different azimuth angles. d ; according to The radial wind speeds at the four azimuth angles of east, south, west, and north were calculated respectively. , , , Then, the u and v components of the horizontal wind speed are calculated using the DBS synthesis formula: Where λ represents the laser wavelength and r represents the radial direction.

3. The wind field data reconstruction method of wind-measuring lidar according to claim 1, characterized in that, In step S2, the data of the wind field dataset is preprocessed, including: Calculate the carrier-to-noise ratio (CNR) of wind field data: In the formula: η represents quantum efficiency, P r Where is the signal echo power, h is Planck's constant, f is the laser carrier frequency, and B represents the radar's receiving bandwidth. The wind field data is filtered based on the carrier-to-noise ratio. When the carrier-to-noise ratio of the wind field data is less than a preset threshold, the wind field data is determined to be invalid and removed. Calculate the mean of wind field data and standard deviation Anomaly detection is performed using the three-standard-deviation rule: when a certain observation value... satisfy If the wind field data is found to be outlier, it will be removed.

4. The wind field data reconstruction method of wind-measuring lidar according to claim 3, characterized in that, Step S2, the preprocessing of the wind field dataset, further includes: Using a one-dimensional linear interpolation method, the missing positions are estimated using known observation points (z0, y0) and (z1, y1). Then, any missing position within the interval (z0, z1) can be calculated. Wind field data fill value at the location Represented as: In the formula: The missing wind field data filler value represents the missing location, z represents the corresponding height layer of the current missing data; z0 and z1 represent the two adjacent height layers before and after the location of the missing data, and y0 and y1 are the wind field data of height layers z0 and z1, respectively.

5. The wind field data reconstruction method of wind-measuring lidar according to claim 3, characterized in that, Step S2, the preprocessing of the wind field data in the wind field dataset, further includes: The wind field data is normalized using the following formula: in, Represents wind field data, The mean of the wind field data. The standard deviation of wind field data This represents the normalized wind field data.

6. The wind field data reconstruction method of wind-measuring lidar according to claim 5, characterized in that, In step S2, the data of the wind field dataset is preprocessed, which further includes: randomly setting a predetermined number of normalized wind field data as missing values, and using masks 0 and 1 to mark missing data and non-missing data respectively.

7. The wind field data reconstruction method of wind-measuring lidar according to claim 1, characterized in that, In step S3, the generator adopts an improved 1D U-Net residual network architecture and introduces a multi-head self-attention mechanism and a residual learning strategy.

8. The wind field data reconstruction method of wind-measuring lidar according to claim 1, characterized in that, In step S3, the discriminator adopts a multi-layer one-dimensional convolutional neural network structure.

9. The wind field data reconstruction method of wind-measuring lidar according to claim 1, characterized in that, In step S4, during the training process, the parameters of the physical information generative adversarial network model are iteratively optimized using a composite loss function, which includes the Navier-Stokes equation residual loss function, the baseline loss function, the temporal smoothing loss function, and the boundary smoothing loss function.

10. The wind field data reconstruction method of wind-measuring lidar according to claim 1, characterized in that, In step S4, the test set is input into the trained physical information to generate an adversarial network model, which outputs the reconstructed wind field data, including: The test set data is input into the model to obtain the predicted wind field data, namely the u and v components of the predicted horizontal wind speed. The discriminator performs inverse normalization on the model output: In the formula, This represents the wind field data after denormalization. For predicted wind field data; The mean of the predicted wind field data. The standard deviation of the predicted wind field data; Introducing Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Pearson Correlation Coefficient A quantitative analysis was conducted on the difference between the predicted wind field data and the actual wind field data after inverse normalization. Based on the predicted wind field data, the horizontal wind speed magnitude V and wind direction are synthesized according to the following formula. : The results will also be visualized.

Images