A weather radar and wind profile radar networking wind field inversion method and system and storage medium

Abstract

The application provides a weather radar and wind profile radar networking wind field inversion method, system and readable storage medium, wherein the method comprises: in a networking system composed of a plurality of weather radars and wind profile radars, synchronously collecting radial velocity and reflectivity factor data of the weather radars in the networking system, and horizontal wind component profile data and vertical wind component profile data of the wind profile radars, to form an original data set; interpolating the radial velocity, reflectivity factor data and vertical profile data into the same inversion grid to obtain model input data; constructing a double-branch fusion deep learning model for extraction of the model input data, the double-branch fusion deep learning model comprising a DenseNet network for extracting spatial features of the radial velocity and reflectivity factor, and a Transformer network for extracting vertical time sequence features of the wind profile data, and outputting a three-dimensional wind field vector after feature fusion of the extracted spatial features and vertical time sequence features.

Description

Technical Field

[0001] This application relates to the field of meteorological radar detection technology, and in particular to a method, system and storage medium for wind field inversion by networking weather radar and wind profiler radar. Background Technology

[0002] Three-dimensional wind fields are core parameters describing atmospheric dynamic processes. High-precision, high-resolution three-dimensional wind field observation data play an irreplaceable role in the mechanism analysis, short-term forecasting and early warning, and disaster prevention and mitigation of severe convective weather. However, existing wind field detection and inversion technologies are limited by the physical limitations of single detection methods or the performance bottlenecks of traditional algorithms, making it difficult to obtain complete, real-time, and accurate wind field information in complex weather scenarios. Specifically: single radar detection systems have inherent defects. Although weather radars (mainly Doppler weather radars) have the scanning capability of large area and high spatiotemporal resolution, their detection principle determines that they can only obtain the projection component of the wind vector in the radar radial direction, i.e., radial velocity. Due to the lack of tangential velocity information, a single weather radar cannot independently invert the complete wind vector field; at the same time, due to the curvature of the earth and beam broadening, there is a significant detection blind zone in the near-surface layer (especially below 2 km).

[0003] In existing technologies, wind profiler radar can directly acquire detailed horizontal wind (u and v components) and vertical wind (w component) profiles above the station, but its detection mode is fixed-point vertical observation, and its horizontal coverage is extremely limited, making it difficult to characterize the three-dimensional spatial structure of small and medium-scale weather systems.

[0004] Secondly, traditional network-based inversion methods are limited by the bottlenecks of physical models and computational efficiency. To compensate for the shortcomings of single detection methods, the academic community has developed three-dimensional wind field inversion techniques based on multi-radar networks, with representative methods including three-dimensional variational (3D-Var) and optical flow methods. These methods essentially rely on artificially designed strong physical constraints (such as mass conservation and divergence-free assumptions) or pre-set smooth models to solve underdetermined equations. However, in practical operational applications, these methods face two major challenges: first, rigid physical models—in severe convective weather systems such as typhoons, strong squall lines, and supercells, atmospheric motion often does not meet idealized assumptions, leading to systematic biases in the inversion results and poor adaptability; second, low computational efficiency—variational methods typically require iterative solutions to the minimum cost function, with a single inversion taking tens of minutes, making it difficult to meet the operational needs of rapid updates for severe convective weather.

[0005] In recent years, some studies have attempted to introduce machine learning (such as support vector machines and shallow neural networks) into wind field inversion, but most of these efforts are limited to post-processing or parameterization improvements of single radar data, such as using them only for velocity deblurring or interpolation blinding. These methods have failed to fundamentally solve the problems of weather radar having spatial surface coverage but incomplete information and wind profiler radar having accurate information but only point observations. They lack the ability to mine the deep coupling characteristics of these two types of heterogeneous radar data, resulting in insufficient accuracy and robustness of the inversion results.

[0006] In summary, the field urgently needs a new method that can deeply integrate the advantages of spatial scanning by weather radar with the advantages of vertical detection by wind profiler radar. Summary of the Invention

[0007] This application discloses a method and system for wind field inversion by networking weather radar and wind profiler radar. It integrates the advantages of data from both types of radar with the feature mining capabilities of deep learning, and solves the problems of poor accuracy, poor real-time performance, and weak adaptability to complex weather in the existing technology.

[0008] This application also discloses a storage medium that can implement the steps in a weather radar and wind profiler radar network wind field inversion method.

[0009] Other objectives and advantages of this application can be further understood from the technical features disclosed herein.

[0010] To achieve one, some, or all of the above objectives or other objectives, in a first aspect, this application provides a method for wind field inversion using a network of weather radar and wind profiler radar, comprising: in a network system composed of several weather radars and wind profiler radars, acquiring the geographical coordinates, detection parameters, and temporal information of the weather radar stations and the wind profiler radar stations; determining the resolution and range of the three-dimensional inversion grid covered by the network system; synchronously collecting radial velocity and reflectivity factor data of the weather radars in the network system, as well as horizontal wind component profile data and vertical wind component profile data of the wind profiler radars, to form a raw dataset; and processing the raw dataset to separate different radar... The asynchronous data from the radar is aligned to a unified timestamp, and the polar coordinate data of the collected weather radar data points is converted into geographic coordinates. The radial velocity, reflectivity factor data, and vertical profile data are interpolated into the same inversion grid to obtain the model input data. A dual-branch fusion deep learning model is constructed for extracting the model input data. This model includes a DenseNet network for extracting the spatial features of the radial velocity and reflectivity factor, and a Transformer network for extracting the vertical temporal features of the wind profile data. The extracted spatial features and vertical temporal features are fused to output a three-dimensional wind field vector.

[0011] The network system includes at least two S / X-band weather radars and at least three wind profiler radars, with the wind profiler radars arranged in a triangular layout. The distance between radar stations in the network system is greater than 60 km, the coverage area is greater than 500 km × 500 km, and the vertical detection range is 100 m to 10 km. The weather radars and wind profiler radars are synchronized in time and space through a synchronization module. The time accuracy of the data recorded by the weather radars and the wind profiler radars is less than or equal to 1 second, and the resolution of the three-dimensional inversion grid is 100 m × 100 m × 50 m.

[0012] The weather radar scanning mode is PPI scanning with a time step of 1Hz, and the wind profiler radar detection mode is fixed-frequency continuous detection. The asynchronous data from different radars are aligned to a unified timestamp, including selecting each scanning time of the weather radar as the target time t, finding two adjacent times t1 and t2 in the time series of the wind profiler radar that are close to the target time t, where t1≤t≤t2, and applying a linear interpolation formula to calculate the interpolation result of the target time for each wind profiler variable.

[0013] The polar coordinate data of the weather radar data points includes azimuth, elevation, and distance, while the geographic coordinates include longitude, latitude, and altitude. After the polar coordinate data of the weather radar data points is converted into geographic coordinates, the radial velocity and reflectivity factor data of the weather radar data points are unified with the vertical profile data of the wind profile radar data points into the same inversion grid using bilinear interpolation.

[0014] The processing of the original dataset also includes quality control processing of the dataset interpolated to the same inversion grid, removing abnormal data based on a signal-to-noise ratio ≥3dB; deblurring the radial velocity collected by the weather radar using a U-Net convolutional neural network; and performing Min-Max normalization processing on the dataset after quality control processing to map the feature values ​​to the [0, 1] interval.

[0015] The DenseNet network comprises multiple dense blocks and transition layers stacked alternately. The transition layers include convolutional layers that compress the output of each dense block and pooling layers that perform pooling operations. The features of each grid point in the inversion grid are input into the DenseNet network, and the spatial correlation features between radial velocity and reflectivity factor are extracted through feature reuse. The Transformer network comprises multiple encoders. The Transformer network captures the temporal dependency features of wind profile data in the vertical direction for the features of each input grid point. The cross-source fusion module uses a self-attention mechanism to perform weighted fusion of the features output by the DenseNet network and the Transformer network. The fused features are mapped into a three-dimensional wind field vector through a regression layer. The output three-dimensional wind field vector includes zonal wind components, meridional wind components, and vertical wind components.

[0016] During training, the dual-branch fusion deep learning model is input with a labeled historical dataset. The labeled data is generated by fusing radiosonde observation data and lidar measurement data. A joint loss function of wind speed and wind direction is used as the loss function for model training. The formula for the loss function is as follows: L=α×MSE(u_pred,u_true)+α×MSE(v_pred,v_true)+β×MSE(w_pred,w_true)+γ×MAE(θ_pred,θ_true); Where α, β, and γ are weighting coefficients, u, v, and w are the zonal wind component, meridional wind component, and vertical wind component, respectively, θ is the wind direction, MSE is the mean square error, and MAE is the mean absolute error.

[0017] The inversion method further includes post-processing and verification output of the generated three-dimensional wind field vector. The verification output includes layer error correction and accuracy verification. Post-processing includes denormalizing the output three-dimensional wind field vector to restore the standardized result to the actual physical quantity. Error correction includes training a height adaptive correction module based on transfer learning to correct the wind field results at different height layers. Accuracy correction includes calculating the root mean square error of wind speed inversion and wind direction error of the three-dimensional wind field vector based on radiosonde data, and setting the verification threshold for the root mean square error of wind speed inversion and wind direction error. The actual physical quantity after error correction and accuracy verification is then gridded to generate three-dimensional wind field data in a standard format.

[0018] Another aspect of the present invention provides a networked wind field inversion system for weather radar and wind profiler radar, comprising: a network detection module, which acquires the geographic coordinates, detection parameters, and temporal information of the weather radar stations and the wind profiler radar stations in a networked system composed of several weather radars and wind profiler radars, and determines the resolution and range of the three-dimensional inversion grid covered by the networked system; a data preprocessing module, which aligns asynchronous data from different radars to a unified timestamp, converts the polar coordinate data of the collected weather radar data points into geographic coordinates, and interpolates them with the vertical profile data of the collected wind profiler radar data points into the same inversion grid to obtain model input data; and a deep learning inversion module, which constructs a dual-branch fusion deep learning model, including extracting the... The system employs a DenseNet network to describe the spatial features of radial velocity and reflectivity factor, and a Transformer network to extract the vertical temporal features of wind profile data. The extracted spatial and vertical temporal features are fused to output a three-dimensional wind field vector. A post-processing and output module performs error correction and accuracy verification on the output three-dimensional wind field vector before outputting it. A storage and training module are also included. The storage module stores labeled historical datasets and model parameters. The labeled data is generated by fusing radiosonde observation data and lidar measurement data. The training module uses a joint wind speed and direction loss function as the loss function for model training, training the dual-branch fusion deep learning model based on the stored historical dataset.

[0019] In another aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps in the above-described method for inverting wind fields in a network of weather radar and wind profiler radar.

[0020] Compared with existing technologies, the technical solution of this application has at least the following beneficial effects: 1. This invention constructs a dual-branch deep learning model to deeply explore the cross-source complementary features between the spatial coverage advantage of weather radar and the vertical detection accuracy of wind profiler radar, and introduces a transfer learning error correction module based on atmospheric stratification characteristics, breaking through the bottleneck of limited inversion accuracy of traditional physical models in complex atmospheric environments. 2. This invention abandons the complex iterative solution process of traditional variational assimilation methods, adopts an end-to-end deep learning inference architecture, and combines lightweight model design with inference acceleration. The single-batch three-dimensional wind field inversion takes no more than 3 seconds, achieving an efficiency improvement of three orders of magnitude compared with traditional methods, which can meet the operational real-time requirements of rapid update cycles for severe convective weather. 3. This invention does not rely on manually designed physical constraints (such as mass conservation, divergence-free assumptions, etc.). Through training with massive historical samples (covering typical severe convective weather such as typhoons, squall lines, and supercells), the model autonomously learns the multi-scale nonlinear mapping relationship of atmospheric motion. Compared with traditional machine learning methods, this invention shows stronger robustness and generalization ability when dealing with missing data, noise interference, and extreme weather scenarios. 4. The weather radar and wind profiler radar networking wind field inversion method and deep learning model architecture of the present invention have good elastic expansion capabilities. On the one hand, it supports the dynamic addition, reduction and layout optimization of radar sites, and can adapt to the mixed networking requirements of different bands (S / X / C bands) and different types of radars. On the other hand, it supports incremental training and model iterative updates based on new data, which facilitates the continuous upgrading of business systems and has broad business promotion and application prospects.

[0021] To make the above and other objects, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the specific embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the wind field inversion method for a network of weather radar and wind profiler radar according to the present invention.

[0024] Figure 2 This is a flowchart of the network setup and data acquisition process in the wind field inversion method of weather radar and wind profiler radar network of the present invention.

[0025] Figure 3 This is a flowchart of the data preprocessing process in the wind field inversion method of the weather radar and wind profiler radar network of the present invention.

[0026] Figure 4 This is a flowchart of the dual-branch fusion deep learning model inversion in the wind field inversion method of the weather radar and wind profiler radar network of the present invention.

[0027] Figure 5 This is a flowchart of the post-processing and verification output in the wind field inversion method of the weather radar and wind profiler radar network of the present invention.

[0028] Figure 6 This is an architecture diagram of the dual-branch fusion deep learning model of the present invention.

[0029] Figure 7 This is a comparison chart of the wind field inversion results of the weather radar and wind profiler radar network method of the present invention and the three-dimensional variational method. Detailed Implementation

[0030] The foregoing and other technical contents, features, and effects of this application will be clearly presented in the following detailed description of a preferred embodiment with reference to the accompanying drawings. The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate; this is merely a method of distinguishing objects with the same attributes in the embodiments of this application.

[0031] Example 1 Example 1 provides a wind field inversion method using a network of weather radar and wind profiler radar. The method includes acquiring the geographic coordinates, detection parameters, and time series information of weather radar stations and wind profiler radar stations in a network system composed of several weather radars and wind profiler radars, and determining the resolution and range of the three-dimensional inversion grid covered by the network system.

[0032] The synchronous acquisition network system collects radial velocity (the speed of wind moving towards or away from the weather radar), reflectivity factor data (the intensity and location of precipitation particles) of the weather radar (also known as meteorological radar, mostly Doppler radar, which works by emitting electromagnetic pulses into the atmosphere and then receiving the echoes reflected back by particles in the air, and then calculating the required parameters), as well as horizontal wind component profile data and vertical wind component profile data of the wind profiler radar. The profile data provided by the wind profiler radar is the speed and direction of wind at different altitude layers in the atmosphere, forming the raw dataset.

[0033] The original dataset is processed by aligning asynchronous data from different radars to a unified timestamp, converting the polar coordinate data of the collected weather radar data points into geographic coordinates, and interpolating them with the vertical profile data of the collected wind profile radar data points into the same inverted 3D grid to obtain the model input data. A dual-branch fusion deep learning model is constructed for extracting model input data. The dual-branch fusion deep learning model includes a DenseNet network for extracting the spatial features of the radial velocity and reflectivity factor, and a Transformer network for extracting the vertical temporal features of the wind profile data. The extracted spatial features and the vertical temporal features are fused to output a three-dimensional wind field vector.

[0034] Unlike traditional physical inversion methods (such as the 3D variational method), which require manually setting physical equations such as smoothing constraints or mass conservation, this application extracts spatial features of radial velocity and reflectivity factors from weather radar output and vertical temporal features from wind profiler radar output. These extracted spatial and vertical temporal features are then fused, correlated within the same 3D grid, and their weights are dynamically calculated to obtain the fused features. The fused features are then mapped end-to-end directly to a 3D wind vector. This method utilizes deep learning. Although the model does not directly input 3D inversion physical equations, during machine learning, learning from a massive input dataset (using radiosonde and lidar data as labels) allows the DenseNet and Transformer networks to learn the physical balance between "spatial structure" and "vertical variation." This allows the high precision of the wind profiler radar to "teach" the model to understand the high-resolution data from the weather radar, thereby generating a high-resolution, gridded 3D wind field vector that conforms to physical laws.

[0035] The following section provides a detailed explanation of the wind field inversion method for a network of weather radar and wind profiler radar provided in Embodiment 1, with reference to the accompanying drawings.

[0036] like Figure 1-6 As shown in Example 1, a wind field inversion method for a network of weather radar and wind profiler radar includes the following steps: Step 1: Network setup and data collection.

[0037] like Figure 2 As shown, network setup and data collection include the following steps.

[0038] Step 1-1: Construct a hybrid network system of weather radar and wind profiler radar.

[0039] The network system consists of several weather radars and wind profiler radars. The network system includes at least two S / X band weather radars and at least three wind profiler radars, with the wind profiler radars arranged in a triangular layout. The distance between radar stations in the network system is greater than 60 km, the coverage area is greater than 500 km × 500 km, and the vertical detection range is 100 m to 10 km.

[0040] Simultaneously, it acquires the geographical coordinates, detection parameters, and time-series information of weather radar stations and wind profiler radar stations.

[0041] Steps 1-2: Synchronously acquire raw data from multi-source radar.

[0042] The synchronous acquisition system collects radial velocity (wind speed towards or away from the weather radar), reflectivity factor data (intensity and location of precipitation particles) from weather radar (also known as meteorological radar, mostly Doppler radar, which works by emitting electromagnetic pulses into the atmosphere and then receiving the echoes reflected back by air particles to calculate the required parameters), and horizontal and vertical wind component profile data from the wind profiler radar. The profile data provided by the wind profiler radar shows the speed and direction of wind at different altitudes in the atmosphere. This results in the final raw dataset.

[0043] Among them, the wind profiler radar adopts a continuous detection mode with a vertical resolution greater than 50m, and the weather radar adopts a PPI scanning mode with a sampling frequency of 1Hz.

[0044] Steps 1-3: Synchronize timing processing, with a spatiotemporal synchronization accuracy of ≤1s, and determine the parameters of the three-dimensional inversion mesh.

[0045] The acquired weather radar and wind profiler radar time-series information is processed for time-series synchronization to ensure that the synchronized data meets the time-space synchronization accuracy of ≤1s, that is, the time accuracy of the data recorded by the weather radar and the wind profiler radar is less than or equal to 1s.

[0046] When performing synchronization timing processing, a GPS synchronization module can be used to achieve time and space synchronization.

[0047] Simultaneously, the parameters of the 3D inversion grid are determined. Taking one implementation method as an example, the determined 3D inversion grid resolution can be 100m × 100m × 50m. The obtained 3D inversion grid resolution is the smallest unit size in space of the final inverted 3D data product. During subsequent data processing, the data output by the weather radar and the data output by the wind profiler radar can be interpolated to each grid point in the 3D inversion grid.

[0048] Step 2: Data preprocessing.

[0049] Step 2-1: Spatiotemporal registration.

[0050] Step 2-1-1: Align the wind profiler radar data and weather radar data to a unified timestamp via linear interpolation in time. This includes selecting each scan time of the weather radar as the target time t, finding two adjacent times t1 and t2 in the time series of the wind profiler radar that are close to the target time t, where t1≤t≤t2, and applying the linear interpolation formula to calculate the interpolation result for the target time for each wind profiler variable.

[0051] The linear interpolation formula is as follows: V(t)=V1+(V2-V1)×((t-t1) / (t2-t1)); Where t represents the current time, t1 represents the previous time, t2 represents the next time, V1 represents the value at the previous time, V2 represents the value at the next time, and V(t) represents the interpolation result at the target time.

[0052] Step 2-1-2: Spatially, the polar coordinate data of the weather radar data points includes azimuth, elevation, and distance, while the geographic coordinates include longitude, latitude, and altitude. After converting the polar coordinate data of the weather radar data points to geographic coordinates, the vertical profile data of the weather radar data points (including radial velocity and reflectivity factor) and the wind profile radar data points are unified into the same inversion grid using bilinear interpolation.

[0053] Bilinear interpolation involves first performing linear interpolation in the x-direction (horizontal direction), then performing linear interpolation in the y-direction (height direction) (using the result of the linear interpolation in the x-direction), ultimately obtaining the bilinear interpolation result. The linear interpolation steps are as follows: based on the coordinates of each point in a regular grid, for any data point coordinates (weather radar and wind profiler radar data points), find the grid points surrounding the target data point, and use the bilinear interpolation method to insert the target data point's value into each grid point until every grid point has a value. Bilinear interpolation is existing technology and will not be described in detail here.

[0054] Step 2-2: Data quality control.

[0055] The datasets interpolated to the same inversion grid are subjected to quality control processing, and abnormal data are removed based on a signal-to-noise ratio ≥3dB.

[0056] The radial velocity data collected by weather radar is deblurred using a U-Net convolutional neural network.

[0057] Due to the inherent sampling limitations of radar, the maximum unambiguous velocity range it can measure is finite. When the actual wind speed exceeds this range, the measured value will "fold" back into this interval, forming an erroneous velocity value (ambiguous velocity). Deambiguity is the process of restoring this folded value to the true value. Mathematically, their relationship is: V_true=V_aliased+2n×V_Nyquist; Where V_true is the true value, V_aliased is the meridional velocity value directly measured by the radar, V_Nyquist is the maximum unambiguous range, and n is the number of folds.

[0058] In the above formula, the key to deblurring is calculating the number of folds. The U-Net convolutional neural network, with its encoder-decoder structure and skip connections, can effectively fuse global contextual information and local features. In deblurring problems, determining the number of folds for a point requires considering both the wind speed changes in its surrounding area and the features of each individual pixel. The characteristics of the U-Net network are well-suited for pixel-level problem diagnosis in radar velocity images.

[0059] The deblurring process mainly includes three core stages: designing the U-Net network model, constructing a training dataset and inputting it into the U-Net network model for training, and finally performing inference to output the number of folds. The deblurring value can then be obtained by using the deblurring formula mentioned above.

[0060] Steps 2-3: Data standardization processing.

[0061] The dataset after quality control is normalized using Min-Max to map the feature values ​​to the interval [0, 1] to eliminate the influence of dimensions.

[0062] Steps 2-4: Generate the input dataset for the deep learning model.

[0063] The preprocessed dataset has a unified data format, ensuring that the data structure matches the input layer of the dual-branch fusion deep learning model.

[0064] Step 3: Inversion of the dual-branch fusion deep learning model.

[0065] Step 3-1: Construct a dual-branch fusion deep learning model.

[0066] The constructed dual-branch fusion deep learning model includes a weather radar branch and a wind profiler radar branch. The weather radar branch uses the DenseNet network to extract spatial features of radial velocity and reflectivity factors obtained from weather radar scans. The extracted spatial features include low-level features (e.g., reflectivity gradient, radial velocity gradient, texture roughness), mid-level features (e.g., echo morphology, velocity pairs, vertical profile morphology), and high-level features (e.g., storm 3D structure, rotation features, convergence / divergence fields, storm relative airflow). These different features correspond to different input data. By learning from a large amount of labeled data, the DenseNet network can learn the correlation between different input data and features, thereby completing the extraction operation.

[0067] The DenseNet network consists of multiple dense blocks and transition layers stacked alternately. The dense blocks include convolutional layers, normalization layers, etc., and use the ReLU function as the activation function. The transition layers include convolutional layers that compress the output of each dense block and pooling layers that perform pooling operations. The features of each grid point in the inverted grid are input into the DenseNet network. Each layer of the DenseNet network can directly access the gradients of all previous layers, allowing shallow features (such as edges and textures) to be reused by deeper layers, promoting feature transfer. Spatial correlation features between radial velocity and reflectivity factor are extracted through feature reuse.

[0068] The wind profiler radar branch uses a Transformer network to extract the temporal correlation features of the profiler data output by the wind profiler radar.

[0069] The Transformer network comprises multiple encoders. For each grid point in the input, the Transformer network captures the temporal dependencies of the wind profile data in the vertical direction. These captured vertical temporal features include vertical gradient features (the rate of change of wind speed / direction with altitude; the Transformer network's self-attention mechanism automatically learns gradient information by comparing wind vector differences between adjacent layers), vertical correlation features (the coupling relationship between wind fields at different altitudes), vertical profile morphology features (the overall shape and structural pattern of the wind profile), fluctuation and periodic features (vertical fluctuation phenomena, such as gravity waves and inertial oscillations), and anomaly and mutation features (drastic changes in the wind field at a certain altitude). During model training, the Transformer network learns the relationships between different input data and features by studying a large amount of labeled data, thus completing the extraction operation.

[0070] Step 3-2: Model training.

[0071] A training dataset is constructed by collecting radar data and corresponding radiosonde data (the gold standard in meteorology, obtained by releasing balloons carrying radiosondes to directly measure temperature, humidity, air pressure, and wind field along their ascent) and lidar data (devices that use laser beams for remote sensing, emitting laser pulses and receiving scattered signals from aerosols in the atmosphere) from the network coverage area. A data training set with sample labels is built, consisting of high-precision wind field data fused from the radiosonde and lidar data. During training, the model continuously attempts to invert the wind field based on the input radar data, generating a prediction result. The system compares the prediction result with the sample labels (real wind field), calculates the error (loss function), and adjusts the internal parameters based on the calculated error to make the next prediction result closer to the label value.

[0072] Because radiosonde data is accurate but sparse, and lidar data is accurate but weather-dependent, combining the two yields highly accurate sample data. The fusion method complements the temporal dimension: sparse radiosonde data serves as the skeleton for calibrating system errors, while continuous lidar data fills in the temporal details. The constructed sample training dataset is divided into training, validation, and test sets in an 8:1:1 ratio.

[0073] When training a dual-branch fusion deep learning model, a joint loss function of wind speed and direction (physical model constraint) is used as the loss function for model training. The formula for the loss function is as follows: L=α×MSE(u_pred,u_true)+α×MSE(v_pred,v_true)+β×MSE(w_pred,w_true)+γ×MAE(θ_pred,θ_true); Where α, β, and γ are weighting coefficients, u, v, and w are the zonal wind component, meridional wind component, and vertical wind component, respectively, θ is the wind direction, MSE is the mean square error, and MAE is the mean absolute error.

[0074] Step 3-3: Feature extraction and cross-source attention fusion.

[0075] The spatial correlation features of radial velocity and reflectivity factor were extracted using the DenseNet network, and the temporal dependence features of wind profile in the vertical direction were captured using the Transformer network.

[0076] The cross-source fusion module employs a self-attention mechanism to weightedly fuse the features output by the DenseNet network and the Transformer network. Cross-source fusion includes three steps: feature reconstruction, dynamic weighting, and information integration. Feature reconstruction: The spatial features output by the weather radar branch are expanded into H×W×D (the location information of grid points in the 3D grid) spatial locations, transforming them into a long sequence. Each element in each sequence corresponds to a specific grid point in 3D space, yielding the "spatial structure description" of the weather radar. For the temporal correlation features output by the wind profiler radar branch, the extracted vertical profile features are copied to every 3D grid point in its surrounding influence area, obtaining the "vertical variation description" of the wind profiler radar.

[0077] Dynamic weighting: A self-attention mechanism is used to generate K, Q, and V vectors. The "spatial structure description" obtained from feature reconstruction is used as the query value (Q value), and the "vertical variation description" is used as the key value (K value) and the numerical value (V value). The K value represents the association of the vertical variation information carried, and the V value represents the vertical variation content carried. The module calculates the similarity between the Q value and the K value, and normalizes the result using the Softmax function, transforming it into weight coefficients between [0, 1]. Based on the calculated weight coefficients, the actual content of the vertical information (V value) is weighted.

[0078] Information integration: After the above calculations, each grid point in the three-dimensional space has a brand-new, integrated feature vector. The feature vector includes the spatial structure (the relative position of the grid point in the weather system) and the vertical pattern (whether the wind above the grid point increases or decreases with increasing height). The integrated information is output and sent to the regression layer for feature mapping.

[0079] Steps 3-4: End-to-end inversion, outputting preliminary results of the three-dimensional wind field vector.

[0080] The fused features are mapped to a 3D wind field vector through a regression layer (fully connected layer or 1×1 convolution). The output 3D wind field vector includes a zonal wind component u, a meridional wind component v, and a vertical wind component w. The calculation formula for the fully connected layer is as follows: (u, v, w) T =W×f+b; Where W is the weight matrix, with each row corresponding to an output component (u, v, w) and each column corresponding to a certain dimension of the input feature, and b is the bias vector.

[0081] Dual-branch fusion deep learning model system architecture as follows Figure 6 As shown, the working principle of this system architecture has been explained in detail in step 3 above, and will not be repeated here.

[0082] Step 4: Post-processing and verification of output.

[0083] Step 4-1: Data denormalization process to restore the standardized results to the actual physical quantities.

[0084] Step 4-2: Stratification error correction.

[0085] Error correction includes training a height-adaptive correction module based on transfer learning to correct the wind field results at different height levels.

[0086] The physical properties of the atmosphere vary greatly at different altitudes, so a uniform correction model cannot be used to handle all altitudes. Instead, the atmosphere needs to be divided into several layers vertically, and each layer needs to be treated separately. For example, the altitude from 100m to 10km can be divided into 20 layers, and the wind field results for each layer can be corrected separately.

[0087] In step 4, a general inversion model covering the entire altitude range has been trained. This model is then divided into multiple parts, each corresponding to a specific altitude level. For each altitude level, the shallow network parameters (responsible for extracting general meteorological features) are frozen. Only the last few layers (responsible for outputting u, v, and w) are used as correction data for further training, allowing the general model to learn the error patterns specific to that layer. During correction, the focus is on correcting errors in the low-altitude (≤2km) wind shear region.

[0088] Step 4-3: Accuracy calibration.

[0089] Based on radiosonde data, the root mean square error of wind speed inversion and the wind direction error of the three-dimensional wind field vector are calculated, and the verification thresholds for the root mean square error of wind speed inversion and the wind direction error are set.

[0090] For example, if the root mean square error (RMSE) of wind speed inversion is ≤1.5 m / s and the wind direction error is ≤4, the inversion result is considered to meet the requirements. Otherwise, new data is collected to iteratively optimize and train the model.

[0091] Step 4-4: Mesh processing to generate standard format three-dimensional wind field data.

[0092] The verified 3D wind field data are written into the 3D grid one by one to form 3D wind field data.

[0093] Steps 4-5: Generate derivative products such as vorticity, divergence, and lift helicity.

[0094] The three-dimensional wind field data generated in step 4-4 is encapsulated in NetCDF4 format, and derivative products such as vorticity, divergence, and lifting helicity are generated. It supports visualization (three-dimensional wind field streamline diagram, vertical profile diagram) and is adapted to the meteorological business system interface for data output and reading.

[0095] Example 2 Example 2 uses a specific example to illustrate a wind field inversion method for a weather radar and wind profiler radar network as described in Example 1.

[0096] Example 2 provides a method for wind field inversion using a network of weather radar and wind profiler radar, which includes the following steps: Step 1: Network setup and data collection.

[0097] A hybrid network system was established, consisting of two S-band weather radars (to supplement low-altitude coverage) and three ground-based wind profiler radars (triangular layout to improve horizontal resolution), with a station spacing of 60km, a coverage area of ​​500km×500km, and a vertical detection range of 100m-10km.

[0098] The GPS synchronization module achieves time and space synchronization with a synchronization accuracy of 0.5s; the weather radar adopts PPI scanning mode with a sampling frequency of 1Hz and a radial velocity detection range of -64m / s to 64m / s, outputting radial velocity and reflectivity factor data; the wind profiler radar adopts continuous detection mode with a vertical resolution of 50m, outputting horizontal wind (u, v) and vertical wind (w) components.

[0099] Step 2: Data preprocessing.

[0100] Step 2-1: Spatiotemporal registration.

[0101] In terms of time, the 1Hz data from the wind profiler radar and the scan data from the weather radar were aligned to the same timestamp through linear interpolation; in terms of space, the polar coordinate (azimuth, elevation, distance) data from the weather radar were converted into geographic coordinates (longitude, latitude, altitude), and the two types of data were unified into a 100m×100m×50m three-dimensional grid using bilinear interpolation.

[0102] Step 2-2: Data quality control.

[0103] A 3dB SNR threshold was set to remove noisy data; a U-Net-based velocity deblurring algorithm was used to correct the radial velocity blurring region, with a correction accuracy of ≥95%.

[0104] Steps 2-3: Data standardization processing.

[0105] The radial velocity, reflectivity factor, and wind profile components (horizontal and vertical wind components) were normalized to the Min-Max range and mapped to the [0,1] interval. At the same time, the original statistical information was retained for post-processing inverse normalization to eliminate the influence of dimensions.

[0106] Steps 2-4: Generate the input dataset for the deep learning model.

[0107] Step 3: Inversion of the dual-branch fusion deep learning model.

[0108] Step 3-1: Construct a dual-branch fusion deep learning model.

[0109] Model structure: The dual-branch fusion model has a total of approximately 8 million parameters, with an input dimension of (C×H×W×D) (C is the number of features = 4, H / W / D are the grid dimensions), and an output dimension of (3×H×W×D) (3 corresponds to the u, v, and w components). Weather radar branch: It adopts the DenseNet-121 network, which contains 4 dense blocks, and extracts the spatial correlation features between radial velocity and reflectivity factor through feature reuse; Wind profiler radar branch: Employs a Transformer network containing 6 encoder layers to capture the temporal dependency features of the wind profiler in the vertical direction; Cross-source fusion module: It adopts a self-attention mechanism to perform weighted fusion of the dual-branch output features (dimension 256), and the attention weights are dynamically adjusted based on feature relevance.

[0110] Step 3-2: Model training.

[0111] Radar data and corresponding radiosonde and lidar data from the network coverage area over the past three years were collected to construct a training set containing over 100,000 samples (divided into training / validation / test sets in an 8:1:1 ratio). The sample labels were high-precision wind field data fused from radiosonde and lidar data. Training parameters: The Adam optimizer was used with an initial learning rate of 1e-4 and a cosine annealing learning rate decay strategy; batch size = 32, training epochs = 50, and an early stopping strategy was adopted (training stopped if the validation set showed no decrease in loss after 5 epochs).

[0112] Step 3-3: Feature extraction and cross-source attention fusion.

[0113] The spatial correlation features of radial velocity and reflectivity factor are extracted using a DenseNet network, and the temporal dependency features of the wind profile in the vertical direction are captured using a Transformer network. A cross-source fusion module employs a self-attention mechanism to weightedly fuse the features output by the DenseNet and Transformer networks.

[0114] Steps 3-4: End-to-end inversion, outputting preliminary results of the three-dimensional wind field vector.

[0115] Step 4: Post-processing and verification of output.

[0116] Step 4-1: Data denormalization process, which restores the standardized results output by the model to the actual physical quantities.

[0117] Step 4-2: Stratification error correction.

[0118] Based on the transfer learning training height adaptive correction model, the height of 100m-10km is divided into 20 layers, and the wind field results of each layer are corrected separately, with a focus on correcting the error in the low-altitude (<2km) wind shear area.

[0119] Step 4-3: Accuracy calibration.

[0120] Based on radiosonde data, the test set validation results are as follows: wind speed RMSE = 1.2 m / s, wind direction error = 3.8°, and correlation coefficient R0. 2 =0.92.

[0121] Step 4-4: Mesh processing to generate standard format three-dimensional wind field data.

[0122] The verified 3D wind field data are written into the 3D grid one by one to form 3D wind field data.

[0123] Steps 4-5: Generate derivative products such as vorticity, divergence, and lift helicity.

[0124] The three-dimensional wind field data generated in step 4-4 is encapsulated in NetCDF4 format, and derivative products such as vorticity, divergence, and lifting helicity are generated. It supports visualization (three-dimensional wind field streamline diagram, vertical profile diagram) and is adapted to the meteorological business system interface for data output and reading.

[0125] The method described in Example 2 was used to perform wind direction inversion at different altitudes, and a traditional three-dimensional variational method was used for wind direction inversion at different altitudes as a comparative example. The inversion results of the two methods are as follows: Figure 7 As shown.

[0126] Combination Figure 7 It can be observed that, in terms of wind speed: the curve obtained by the method of this invention closely matches the radiosonde reference curve, with a root mean square error (RMSE) of 1.2 m / s; the traditional method's curve deviation (RMSE) is 1.8 m / s, 50% higher than that of this invention. In terms of wind direction: the wind direction error of the method of this invention is only 3.8°, and the curve almost coincides with the reference; while the wind direction error of the traditional method reaches 6.5°, with particularly significant deviations at mid-to-high altitudes (4km-8km). In terms of vertical variation: the three types of data show consistent trends in vertical height (wind speed exhibits sinusoidal fluctuations, and wind direction exhibits cosine fluctuations), verifying the physical consistency of the inversion results and highlighting the high accuracy of this invention across all altitude levels. Figure 7 This intuitively demonstrates the superior accuracy of the method of this invention compared to traditional methods in wind field inversion, especially in the low-altitude (<2km) wind shear region and the middle and upper atmosphere, where the inversion results are closer to the actual sounding data.

[0127] Example 3 Example 3 provides a networked wind field inversion system combining weather radar and wind profiler radar, comprising: The network detection module acquires the geographic coordinates, detection parameters, and temporal information of the weather radar stations and the wind profiler radar stations in a network system composed of several weather radars and wind profiler radars, and determines the resolution and range of the three-dimensional inversion grid covered by the network system. The data preprocessing module aligns asynchronous data from different radars to a unified timestamp, converts the polar coordinate data of the collected weather radar data points into geographic coordinates, and interpolates them with the vertical profile data of the collected wind profile radar data points into the same inversion grid to obtain the model input data. The deep learning inversion module constructs a dual-branch fusion deep learning model, including a DenseNet network for extracting the spatial features of the radial velocity and reflectivity factor, and a Transformer network for extracting the vertical temporal features of the wind profile data. The extracted spatial features and the vertical temporal features are fused to output a three-dimensional wind field vector. The post-processing and output module performs error correction and accuracy verification on the output three-dimensional wind field vector before outputting it. The system includes a storage module and a training module. The storage module stores labeled historical datasets and model parameters. The labeled data is generated by fusing radiosonde observation data and lidar measured data. The training module uses a joint loss function of wind speed and wind direction as the loss function for model training and trains the dual-branch fusion deep learning model based on the stored historical dataset.

[0128] The weather radar and wind profiler radar network wind field inversion system provided in Example 3 is used to implement the steps of the weather radar and wind profiler radar network wind field inversion method in Example 1.

[0129] Example 4 Example 4 provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the steps in the wind field inversion method of a weather radar and wind profiler radar network in Example 1.

[0130] It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from the principles thereof, and such improvements and modifications also fall within the scope of protection of the claims of this application. It should be understood that certain features of this disclosure described in the context of individual embodiments for clarity may also be provided in combination in a single embodiment. Conversely, various features of this disclosure described in the context of individual embodiments for clarity may also be provided individually or in any suitable combination or as part of any other described embodiment of this disclosure.

Claims

1. A method for wind field inversion using a network of weather radar and wind profiler radar, characterized in that, This includes acquiring the geographic coordinates, detection parameters, and temporal information of the weather radar stations and the wind profiler radar stations in a network system composed of several weather radars and wind profiler radars, and determining the resolution and range of the three-dimensional inversion grid covered by the network system. The radial velocity and reflectivity factor data of the weather radar in the network system, as well as the horizontal wind component profile data and vertical wind component profile data of the wind profiler radar, are collected synchronously to form the original dataset. The original dataset is processed by aligning asynchronous data from different radars to a unified timestamp, converting the polar coordinate data of the collected weather radar data points into geographic coordinates, and interpolating the radial velocity, reflectivity factor data and vertical profile data into the same inversion grid to obtain the model input data. A dual-branch fusion deep learning model is constructed for extracting model input data. The dual-branch fusion deep learning model includes a DenseNet network for extracting the spatial features of the radial velocity and reflectivity factor, and a Transformer network for extracting the vertical temporal features of the wind profile data. The extracted spatial features and the vertical temporal features are fused to output a three-dimensional wind field vector.

2. The wind field inversion method of a weather radar and wind profiler radar network according to claim 1, characterized in that, The network system includes at least two S / X band weather radars and at least three wind profiler radars, with the wind profiler radars arranged in a triangular layout. The radar stations in the network system are spaced more than 60 km apart, the coverage area is greater than 500 km × 500 km, and the vertical detection range is 100 m - 10 km. The weather radar and wind profiler radar are synchronized in time and space through a synchronization module. The time accuracy of the data recorded by the weather radar and the wind profiler radar is less than or equal to 1 second, and the resolution of the three-dimensional inversion grid is 100m×100m×50m.

3. The wind field inversion method for a network of weather radar and wind profiler radar according to claim 2, characterized in that, The weather radar scanning mode is PPI scanning with a time step of 1Hz, and the wind profiler radar detection mode is fixed frequency continuous detection. Aligning asynchronous data from different radars to a unified timestamp involves selecting each scan time of the weather radar as the target time t, finding two adjacent times t1 and t2 in the time series of the wind profiler radar that are close to the target time t, where t1≤t≤t2, and applying a linear interpolation formula to calculate the interpolation result of the target time for each wind profiler variable.

4. The wind field inversion method of a weather radar and wind profiler radar network according to claim 3, characterized in that, The polar coordinate data of the weather radar data points includes azimuth, elevation, and distance; the geographic coordinates include longitude, latitude, and altitude. After the polar coordinate data of the weather radar data points are converted into geographic coordinates, the radial velocity and reflectivity factor data of the weather radar data points are unified with the vertical profile data of the wind profile radar data points into the same inversion grid using bilinear interpolation.

5. The wind field inversion method of a weather radar and wind profiler radar network according to claim 1, characterized in that, The processing of the original dataset also includes quality control processing of the dataset interpolated to the same inversion grid, removing abnormal data based on a signal-to-noise ratio ≥3dB; The radial velocity collected by the weather radar is deblurred using a U-Net convolutional neural network; The dataset after quality control is subjected to Min-Max normalization to map the feature values ​​to the interval [0, 1].

6. The wind field inversion method of a weather radar and wind profiler radar network according to claim 1, characterized in that, The DenseNet network consists of multiple dense blocks and transition layers stacked alternately. The transition layer includes a convolutional layer that compresses the output of each dense block and a pooling layer that performs pooling operations. The features of each grid point in the inversion grid are input into the DenseNet network, and the spatial correlation features between radial velocity and reflectivity factor are extracted through feature reuse. The Transformer network includes multiple encoders, and the Transformer network captures the temporal dependency features of the wind profile data in the vertical direction based on the features of each grid point in the input. The cross-source fusion module employs a self-attention mechanism to weightedly fuse the features output by the DenseNet network and the Transformer network. The fused features are then mapped into a three-dimensional wind field vector through a regression layer. The output three-dimensional wind field vector includes zonal wind components, meridional wind components, and vertical wind components.

7. The wind field inversion method for a network of weather radar and wind profiler radar according to claim 6, characterized in that, During training, the dual-branch fusion deep learning model is input with a labeled historical dataset. The labeled data is generated by fusing radiosonde observation data and lidar measurement data. A joint loss function of wind speed and wind direction is used as the loss function for model training. The formula for the loss function is as follows: L=α×MSE(u_pred,u_true)+α×MSE(v_pred,v_true)+β×MSE(w_pred,w_true)+γ×MAE(θ_pred,θ_true); Where α, β, and γ are weighting coefficients, u, v, and w are the zonal wind component, meridional wind component, and vertical wind component, respectively, θ is the wind direction, MSE is the mean square error, and MAE is the mean absolute error.

8. The wind field inversion method of a weather radar and wind profiler radar network according to claim 1, characterized in that, The inversion method also includes post-processing and verification output of the generated three-dimensional wind field vector, and the verification output includes layer error correction and accuracy verification. Post-processing includes inverse normalization of the output three-dimensional wind field vector to restore the normalized result to the actual physical quantity; The error correction includes training a height adaptive correction module based on transfer learning to correct the wind field results at different height levels; The accuracy correction includes using radiosonde data as a reference to calculate the root mean square error of wind speed inversion and the wind direction error of the three-dimensional wind field vector, and setting the verification threshold for the root mean square error of wind speed inversion and the wind direction error. The actual physical quantities after error correction and accuracy verification are gridded to generate standard format three-dimensional wind field data.

9. A networked wind field inversion system combining weather radar and wind profiler radar, characterized in that, include: The network detection module acquires the geographic coordinates, detection parameters, and temporal information of the weather radar stations and the wind profiler radar stations in a network system composed of several weather radars and wind profiler radars, and determines the resolution and range of the three-dimensional inversion grid covered by the network system. The data preprocessing module aligns asynchronous data from different radars to a unified timestamp, converts the polar coordinate data of the collected weather radar data points into geographic coordinates, and interpolates them with the vertical profile data of the collected wind profile radar data points into the same inversion grid to obtain the model input data. The deep learning inversion module constructs a dual-branch fusion deep learning model, including a DenseNet network for extracting the spatial features of the radial velocity and reflectivity factor, and a Transformer network for extracting the vertical temporal features of the wind profile data. The extracted spatial features and the vertical temporal features are fused to output a three-dimensional wind field vector. The post-processing and output module performs error correction and accuracy verification on the output three-dimensional wind field vector before outputting it. The system includes a storage module and a training module. The storage module stores labeled historical datasets and model parameters. The labeled data is generated by fusing radiosonde observation data and lidar measured data. The training module uses a joint loss function of wind speed and wind direction as the loss function for model training and trains the dual-branch fusion deep learning model based on the stored historical dataset.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps in the wind field inversion method of a weather radar and wind profiler radar network as described in any one of claims 1-8.

Images