Residual error correction wind profile radar inversion method based on dynamic space-time field network

Abstract

The invention discloses a residual error correction wind profile radar inversion method based on a dynamic space-time field network, and relates to the technical field of meteorological radar data processing, and the method comprises the steps: constructing a four-dimensional observation tensor, fusing a Beidou sounding true value, and carrying out the residual error learning mechanism guided by physical prior. And synchronously extracting wind field spatial-temporal characteristics by using mask time self-attention and a graph neural network, and realizing high-precision and physical consistent inversion of the whole wind profile in combination with a composite loss function. According to the method, the problems of insufficient precision, vertical fracture and discontinuous time caused by the hypothesis failure of wind field horizontal uniformity in the traditional method can be effectively solved, and the wind field inversion quality under the complex meteorological condition is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of meteorological observation technology, and in particular to a residual correction wind profile radar inversion method based on a dynamic spatiotemporal field network. Background Technology

[0002] Wind profiler radar, a key component of modern meteorological observation systems, transmits multiple directional beams (typically 3 or 5 beams, including vertical beams and azimuth-orthogonal tilted beams) to measure Doppler radial velocities along each beam direction. Based on vector synthesis principles, it then infers wind field information such as atmospheric wind speed and direction. This technology provides high spatiotemporal resolution vertical wind profiler data, which has significant application value in numerical weather prediction, severe weather warnings, and atmospheric dynamics research. Specifically, accurate wind field profiles can significantly improve the ability of numerical models to simulate atmospheric motion processes, thereby enhancing the accuracy of short- to medium-term weather forecasts. Simultaneously, the real-time capture capability of wind field structures during extreme weather events such as strong convection, high winds, and dust storms provides crucial support for the early identification and warning of meteorological disasters. Furthermore, its continuous observation characteristics provide a high-quality data foundation for fundamental scientific research on atmospheric boundary layer evolution and convection triggering mechanisms.

[0003] Currently, the inversion of wind profiler radar wind measurement data mainly relies on traditional vector calculation methods, with typical schemes including the 3-beam method, the 4-beam method, and the 5-beam joint verification method. The 3-beam method utilizes two adjacent oblique beams to form a triangular relationship with the vertical beam, calculating the horizontal wind component through trigonometric functions. This is suitable for scenarios with rapidly changing airflow to improve temporal resolution. The 4-beam method independently calculates the U and V wind components using oblique beams in the east-west and north-south directions, avoiding the influence of ground clutter on the vertical beam, and typically has higher wind measurement accuracy. The 5-beam method combines the advantages of the former two, ensuring data integrity while maintaining inversion accuracy, and is the most widely adopted scheme in current operational applications. However, all of the above methods are based on the idealized assumption that the wind field is uniform in the horizontal direction. In reality, the atmosphere, especially under the influence of strong convection, wind shear, or complex terrain, often exhibits significant non-uniformity and three-dimensional structural characteristics, leading to a significant increase in inversion error for traditional algorithms under complex meteorological conditions.

[0004] More importantly, wind profiler radar errors originate from diverse and highly coupled sources, including differences in radar deployment locations, spatial separation effects between beams, beam deflection caused by atmospheric refraction, ground clutter pollution, and interference from extreme weather on signal quality. Existing quality control methods mostly employ strategies such as consistency checks, removal of suspicious data, or simple averaging. While these can suppress outliers to some extent, they inevitably lead to missing data, disrupting the vertical continuity and temporal integrity of the wind profile, making it difficult to effectively correct the inherent nature of the errors. Furthermore, traditional inversion algorithms heavily rely on the radar's own beam data and lack systematic integration of high-precision external observation information. Although some studies have attempted to introduce radiosonde data for post-evaluation or coarse calibration, a physically constrained error correction model guided by high-precision ground truth has not yet been established, limiting further improvements in inversion accuracy.

[0005] In summary, existing wind profiler radar inversion techniques have significant limitations in terms of theoretical assumptions, error processing mechanisms, and multi-source data fusion, making it difficult to meet the urgent need for high-precision, high-completeness wind field products in refined meteorological services. Therefore, there is an urgent need to develop a novel inversion method that can overcome the constraints of traditional vector algorithm assumptions, effectively integrate external high-precision observations, and fully explore the spatiotemporal evolution of wind fields, in order to achieve accurate, continuous, and physically consistent dynamic reconstruction of wind profiler data under complex atmospheric conditions. Summary of the Invention

[0006] The purpose of this invention is to provide a residual-corrected wind profile radar inversion method based on a dynamic spatiotemporal field network. It aims to overcome the limitations of traditional inversion methods in fully utilizing the temporal continuity and spatial correlation of wind fields and capturing nonlinear error structures. By constructing a dynamic observation map of time-altitude-beam and integrating physical constraints with deep learning, it achieves high-precision, spatiotemporally consistent dynamic inversion of wind fields under complex environments.

[0007] To achieve the above objectives, the present invention is implemented according to the following technical solution:

[0008] This invention includes the following steps:

[0009] S1: Organize the observation data of wind profiler radar at multiple times, altitudes, and beams into a four-dimensional tensor, and process the missing values ​​to generate a mask matrix, providing structured input for spatiotemporal modeling;

[0010] S2: The basic wind field is constructed using the traditional vector algorithm as a physical prior; at the same time, the high-precision ground truth obtained by the Beidou sounding system is used to calculate the full profile residual, which is then used as the training target label for the deep learning model.

[0011] S3: Extract wind field evolution features in the time dimension through Transformer, and extract inter-layer coupling features in the vertical space through graph neural network; use attention mechanism to deeply fuse spatiotemporal features with physical priors to generate a fused feature field with physical constraints.

[0012] S4: Input the fused features into the residual prediction head and output the full-height layer residual field at once; superimpose the predicted residual with the basic wind field to obtain the final high-precision wind field, and realize the synchronous correction of the error across the entire profile;

[0013] S5: Residual regression loss ensures accuracy, time smoothing loss constrains evolution stability, and spatial smoothing loss enhances vertical continuity; a time series partitioning strategy is used for training to ensure the model's generalization ability in unknown weather processes.

[0014] The beneficial effects of this invention are:

[0015] The above technical solution was validated using measured data from the Xilinhot Observatory in Inner Mongolia. Compared to the traditional five-beam vector inversion method, this invention reduces the root mean square error (RMSE) of average wind speed across all altitude layers from 2.53 m / s to 2.15 m / s, a relative reduction of 15.0%; and the mean absolute error (MAE) of wind direction from 26.0° to 24.1°, a relative reduction of 7.3%. Simultaneously, the output wind field shows no significant jumps in the time series and no interlayer breaks in the vertical profile, exhibiting good physical consistency. This method provides an engineering-feasible technical path for high-precision wind field inversion by wind profiler radar under complex meteorological conditions such as strong convection and low signal-to-noise ratio, and can provide more reliable data support for numerical weather prediction model assimilation, strong convection early warning systems, and boundary layer dynamics research. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0018] like Figure 1 As shown, the solution of this invention can be summarized into the following five interrelated steps, which cover the complete process from data preparation to model training and then to real-time application.

[0019] Step S1: Data Preparation and Feature Extraction

[0020] (1) Definition of observation node: K types of observations (radial velocity Vr, signal-to-noise ratio SNR, spectral width W) from T consecutive time points, H altitudes, and B beams of the wind profiler radar are combined into a four-dimensional tensor D(t,z,b,k) ∈ .in, .

[0021] (2) Missing measurement value processing: Time / height direction interpolation is used to process missing measurement values, and a mask matrix is ​​generated at the same time. ∈ Record data availability. Marking method 1 indicates that the original data is available, and 0 indicates that the point has been interpolated.

[0022] Step S2: Construct the basic wind field and dataset labels:

[0023] (1) Basic wind field generation: The basic wind field matrix of the entire height layer is generated using the traditional three-beam / five-beam inversion method. OE is short for Original Estimate.

[0024]

[0025] As a physical prior, it serves as the benchmark for final correction.

[0026] (2) Acquisition of real wind field (labeling): The wind profiler radar and the BeiDou sounding ground truth at the same site are spatiotemporally matched and aligned with the vertical height axis to generate a real wind field label matrix at all height levels. .

[0027]

[0028] (3) Dataset construction and target residual field calculation: The sample pair construction includes the radar observation tensor D, mask matrix M, and physical prior field at all times and altitudes. True value tags for BeiDou sounding spacecraft aligned with time and space. By performing correlation, a training sample pair of "observation field - true value field" is constructed.

[0029] Target label calculation: Calculate the deviation between the true values ​​of each layer of the full profile and the base field to obtain the target residual field for model training. ,

[0030] =[ ]

[0031] Step S3: Construct a spatiotemporal wind field representation model (DSTFN encoder):

[0032] (1) Feature dimension normalization: The preset feature embedding dimension is 1. By using linear transformation, the original observations and physical priors are mapped to... This space ensures that data from different sources are integrated within a unified semantic space.

[0033] (2) Temporal Encoder:

[0034] Method: Temporal Multi-Head Attention (TMA) mechanism of Transformer is used to extract time series features.

[0035] Input: A sequence of full-altitude observations over the past m time steps, D(t−m:t, (, B, K) and its corresponding mask matrix M(t−m:t, , B, K).

[0036] Operation Description: A masked temporal self-attention mechanism is employed to extract historical evolution features in parallel for each height layer. The mask is used to filter out invalid observation times when calculating the attention score, ensuring the purity of the temporal features.

[0037] Output: Time feature matrix:

[0038] (3) Spatial Encoder:

[0039] Method: Vertical coupling features are extracted using graph neural networks (GNN) and graph convolution.

[0040] Input: Current full-altitude observation data D(t, (t, B, K) and its corresponding mask matrix M(t, , B, K)

[0041] Operation Description: Graph convolution operations are performed using a graph neural network (GNN). H height layers are abstracted as graph nodes, and a neighborhood aggregation mechanism is used to capture the vertical wind shear and physical coupling features. Masking dynamically adjusts the convolution kernel weights to suppress error propagation in the vertical space.

[0042] Output: Spatial feature matrix:

[0043] (4) Multimodal physical fusion layer:

[0044] Method: An attention-weighted fusion mechanism is employed. This mechanism utilizes... As a query, it guides the model to learn and generate targets and dynamic weight vector and Realizing physical constraint information Nonlinear spatiotemporal characteristics , Deep integration.

[0045] enter: , ,

[0046] Operation Description:

[0047] The physical query field is constructed using a fully connected layer (MLP). Mapped to a query vector ,

[0048] Dynamic weight field generation will query the matrix. Respectively with the time feature matrix and spatial characteristic matrix Similarity calculation (matrix dot product) is performed, and the result is normalized using the Sigmoid activation function to generate a dynamic weight matrix for the entire height layer. and

[0049] Synchronous fusion of the entire profile: Element-wise weighting of spatiotemporal features is performed using generated dynamic weights, and then concatenated with the basic wind field features. Guided by the airport wind field, the model adaptively determines the contribution of temporal and spatial information at each altitude level to residual correction.

[0050] Output: Full-height layer fused feature matrix ,formula

[0051]

[0052] Step S4: Dynamic residual field prediction and wind field output

[0053] The core of this step is to directly map the fused "feature field" to the "residual field" to achieve a one-time prediction of the entire profile.

[0054] Residual prediction: Utilizes a residual prediction head (composed of a multilayer perceptron or convolutional layers) to fuse feature matrices across the entire height layer. Perform a nonlinear mapping to directly output the wind field residual matrix for all height layers at that moment. .

[0055] =[ ]

[0056] Wind field output: The initial wind field is superimposed with the predicted residual to obtain the final wind field. .

[0057] =

[0058] Step S5: Model Training and Loss Function

[0059] This step details the preparation of training data and the construction of the loss function for the DSTFN model in high-precision, spatiotemporally consistent inversion tasks.

[0060] (1) Construction of full-height sample pairs and field supervision mechanism

[0061] To support the model in learning the physical relationships in the vertical space, this invention adopts a "field-to-field" training mode:

[0062] Training input field: For each sample, input the full-height layer observation tensor D(tm:t, 1:H, B, K), the mass mask matrix M(tm:t, 1:H, B, K), and the basic wind field matrix at the current time and its historical periods in one go. .

[0063] Target residual field label: The target residual matrix covering the entire altitude layer is generated by subtracting the true values ​​of BeiDou sounding from the basic wind field. .

[0064] Significance of field supervision: Unlike traditional point-by-point training, this invention enables the model to perceive the co-evolution characteristics of errors between different height layers during the optimization process through matrix-level supervision.

[0065] (2) Construction of the composite loss function:

[0066] Model training employs a composite loss function that includes spatiotemporal smoothing constraints. The aim is to balance the model's prediction accuracy with the physical plausibility of the output wind field.

[0067] +

[0068] in, , , These are the weighting coefficients for each loss term. Each sub-loss term is defined as follows: Residual Regression Loss : Use mean squared error (MSE) to minimize the model prediction residuals residual with target The differences between them ensure the accuracy of wind field inversion.

[0069]

[0070] Where: H: Total number of height layers detected by wind profiler radar; z: Vertical height layer index, z∈ ;

[0071] The 2D wind field residual vector predicted by the model at time t and height z. Time smoothing loss To constrain the stability of the wind field evolution over time, calculate the full profile vector difference between adjacent time points within m consecutive time points to ensure that the model captures stable atmospheric signals;

[0072] The signal is a steady evolution of the atmosphere, rather than instantaneous random noise.

[0073]

[0074] Where: m is the time step of the model input (window length is m+1 time steps); i: time series index, which is traversed starting from the second time step in the window to calculate the adjacent differences; : The full-height layer output wind field matrix at time i;

[0075] Spatial smoothing loss This is the core physical constraint of the invention. Utilizing the characteristic of synchronous output across the entire profile, it directly affects the predicted wind field. Penalize continuity in high-dimensional dimensions.

[0076]

[0077] in: The final output wind field after correction at layer z is calculated as follows:

[0078] (3) Dataset partitioning and validation strategy: To validate the model's generalization ability under unknown weather processes, this invention implements a strict time series partitioning strategy:

[0079] Split ratio: The entire sample set is split into 80% training set and 20% test set.

[0080] Time series segmentation principle: Segmentation is based on chronological order. Continuous observation data from the first 80% of the time period is selected as the training set, and independent observation data from the last 20% of the time period is selected as the test set.

[0081] Technical advantages: This approach ensures that the test set includes complete weather processes that the model has not encountered, effectively avoiding data leakage and enabling a realistic assessment of the model's reliability in real-world business environments.

[0082] (4) Model training and testing:

[0083] Based on the constructed 80% training set, the model is input and end-to-end training of the spatiotemporal coding module and fusion layer is performed using a composite loss function. Throughout the training process, the integrity of the full-height layer matrix is ​​maintained, enabling the model to learn the residual distribution patterns across heights and time intervals. After training, the remaining 20% ​​test set is used for validation to ensure the model's ability to synchronously output high-precision, physically continuous full-profile wind fields.

[0084] Example:

[0085] This embodiment uses the wind profiler radar system located at the Xilinhot Observatory in Inner Mongolia Autonomous Region and the Beidou sounding system deployed at the same site as the observation platform to specifically illustrate the residual correction wind profiler radar inversion method based on dynamic spatiotemporal field network (DSTFN) proposed in this invention. However, this embodiment does not constitute a limitation on the scope of protection of this invention.

[0086] (I) Observation equipment and data sources:

[0087] Wind profiler radar configuration: The wind profiler radar used in this embodiment is a five-beam wind profiler radar, including one vertical beam and four oblique beams set at different azimuth angles, used to acquire atmospheric radial velocity information at different altitudes. The radar starts detection from an altitude of 150 m and uses both high and low operating modes.

[0088] Low-level mode: Height warehouse length (interlayer spacing) is 120 m;

[0089] High mode: The height of the storage room (interlayer spacing) is 240 m.

[0090] Model fusion strategy: For the highly overlapping parts of low-level and high-level models, this embodiment uniformly selects high-level model data as valid observation data to ensure the consistency and stability of data at different height levels.

[0091] Time and Feature Resolution: The wind profiler radar has a time resolution of once every 6 minutes, and the observations include radial velocity, signal-to-noise ratio (SNR), and spectral width (W).

[0092] True reference: The BeiDou sounding system at the same location conducts two balloon releases per day to obtain high-precision vertical wind field information, which serves as the real wind field reference data in this embodiment.

[0093] (II) Observation time range and sample construction:

[0094] Time span: The observation period selected in this embodiment is from October 2024 to October 2025, covering multiple seasons and different weather patterns, and has good representativeness.

[0095] Spatiotemporal alignment and matrix extraction:

[0096] Traversing the altitude layers, taking the time of the BeiDou weather balloon closest to the altitude layer as the center, the most recent radar observation data at the best matching time within a 5-minute window before and after.

[0097] Full-field feature construction: Based on the matching wind profiler radar data time, backtrack by 5 steps to construct a four-dimensional spatiotemporal observation tensor D(t-5:t, 1:H, B, K) containing 6 consecutive time points (all H height layers, 5 beams, and 3 types of physical features).

[0098] Masking Matrix: When reading raw wind profiler radar data, if a data point at a certain altitude is missing ( / / / / / / ), it is marked as 0. The data at that point is then obtained by interpolation using data from the adjacent altitude levels above and below. If data from adjacent altitude levels is also missing, the data at that point is obtained by interpolation using data from the preceding and following times. If all adjacent altitude levels are missing, the data from the most recent time point is used. A valid value is marked as 1. A mask is then generated.

[0099] Full-field label generation: Extract U and V wind components from BeiDou radiosonde data to generate a synchronized, full-altitude layer real wind field label matrix. .

[0100] (III) Basic wind field generation and residual construction:

[0101] After the sample construction is completed, the initial wind field matrix at all altitude levels is calculated using the traditional five-beam vector inversion algorithm. As the basic field.

[0102] Subsequently, the difference between the base field and the true field of the BeiDou sounding at the same time and altitude is calculated to obtain the full profile residual matrix, which serves as the learning target of the dynamic spatiotemporal field network in this invention.

[0103] (iv) DSTFN model training and wind field correction:

[0104] The constructed full-field observation tensor, basic wind field, mask, and residual labels are input into the DSTFN model. The model is then optimized end-to-end using a composite loss function, enabling the network to grasp the nonlinear mapping law from multibeam observation features to the wind field residuals at all heights.

[0105] In practical applications, after a new observation sequence is input, the model outputs the wind field residual matrix at all heights simultaneously.

[0106] The output residual is added to the traditional basic wind field to obtain the final corrected full profile high-precision wind field.

[0107] (V) Explanation of Implementation Results:

[0108] From the experimental process, compared with traditional vector computation physical models, the DSTFN model proposed in this invention can better capture the dynamic evolution of the wind field profile in time and space and its vertical coupling characteristics. This scheme adopts a physics-guided residual correction architecture, using traditional inversion results as physical priors (basic field). This guides deep learning models to accurately correct their errors, thereby achieving error correction without removing data and improving the accuracy and completeness of wind profile inversion.

[0109] Experimental results show that, compared with high-precision BeiDou sounding data, the DSTFN scheme of this invention achieves a significant improvement in wind field accuracy. The table below compares the average error of the DSTFN-corrected wind field (U and V winds converted to wind speed and direction) with the traditional vector-calculated base field across the entire profile:

[0110] Error index Traditional vector calculation (Baseline) DSTFN model predictions (corrected) Relative error reduced Root mean square error of wind speed (RMSE) 2.53 m / s 2.15 m / s 15.0% Mean Absolute Error (MAE) of Wind Direction 26.0° 24.1° 7.3%

[0111] The key innovation of this invention lies in:

[0112] (1) The residual modeling mechanism is adopted, and the traditional vector inversion results are used as the physical prior field. Only the deviation between the results and the high-precision Beidou sounding true value is learned, so that the deep network does not need to reconstruct the complete wind field from the original observation, but focuses on the nonlinear correction of the error structure, thereby improving the model training stability and cross-site transfer capability.

[0113] (2) Construct a four-dimensional dynamic spatiotemporal map field (time × height × beam × variable), capture the independent temporal evolution law of each height layer through the mask multi-head self-attention mechanism in the time encoder, and model the vertical wind shear and inter-layer coupling through the graph neural network in the spatial encoder to realize the synchronous inversion of the entire wind profile and overcome the vertical discontinuity problem caused by traditional layer-by-layer calculation.

[0114] (3) Deeply integrate the high-precision true value of Beidou sounding, construct residual labels for the entire altitude layer, and explicitly introduce time and space smoothing constraints in the loss function to force the model output to satisfy the continuous wind field of atmospheric motion physical laws from the optimization objective level;

[0115] (4) By identifying the original valid observation and interpolation data through the mask matrix, the information weights are dynamically adjusted in the time attention and graph convolution calculation, so that the model can complete the error correction while retaining all data structures, avoiding the missing profile measurements caused by the elimination of suspicious data in traditional quality control, and significantly improving the data availability.

[0116] The technical solutions of the present invention are not limited to the specific embodiments described above. Any technical modifications made in accordance with the technical solutions of the present invention fall within the protection scope of the present invention.

Claims

1. A residual-corrected wind profile radar inversion method based on a dynamic spatiotemporal field network, characterized in that, Includes the following steps: S1: Organize the observation data of wind profiler radar at multiple times, altitudes, and beams into a four-dimensional tensor, and process the missing values ​​to generate a mask matrix, providing structured input for spatiotemporal modeling; S2: The basic wind field is constructed using the traditional vector algorithm as a physical prior; at the same time, the high-precision ground truth obtained by the Beidou sounding system is used to calculate the full profile residual, which is then used as the training target label for the deep learning model. S3: Extract wind field evolution features in the time dimension using Transformer, and extract inter-layer coupling features in the vertical space using graph neural network; An attention mechanism is used to deeply fuse spatiotemporal features with physical priors to generate a fused feature field with physical constraints. S4: Input the fused features into the residual prediction head and output the full-height layer residual field at once; superimpose the predicted residual with the basic wind field to obtain the final high-precision wind field, and realize the synchronous correction of the error across the entire profile; S5: Residual regression loss ensures accuracy, time smoothing loss constrains evolution stability, and spatial smoothing loss enhances vertical continuity; a time series partitioning strategy is used for training to ensure the model's generalization ability in unknown weather processes.

2. The residual-corrected wind profile radar inversion method based on dynamic spatiotemporal field network according to claim 1, characterized in that: Step S1 specifically involves: constructing a four-dimensional tensor D(t,z,b,k) from K observations of the wind profiler radar over T consecutive time points, H altitudes, and B beams. The observed quantities include radial velocity Vr, signal-to-noise ratio SNR, and spectral width W; wherein, Missing measurements are processed using time / height interpolation, and a mask matrix is ​​generated simultaneously. ∈ Record data availability. Marking method 1 indicates that the original data is available, and 0 indicates that the point has been interpolated.

3. The residual-corrected wind profile radar inversion method based on dynamic spatiotemporal field network according to claim 2, characterized in that: Step S2 specifically involves: The basic wind field matrix at all heights is generated using the traditional three-beam / five-beam inversion method. , As a physical prior and the benchmark for final correction, the wind profiler radar is spatiotemporally matched with the BeiDou sounding ground truth at the same location, and aligned with the vertical height axis to generate a full-height layer real wind field label matrix. ; The sample pairs will be used to construct the radar observation tensor D, mask matrix M, and physical prior fields at all times and altitudes. True value tags for BeiDou sounding spacecraft, aligned with time and space. By performing correlation, a training sample pair of "observation field - true value field" is constructed. Target label calculation: Calculate the deviation between the true values ​​of each layer of the full profile and the base field to obtain the target residual field for model training. , =[ ]。 4. The residual-corrected wind profile radar inversion method based on dynamic spatiotemporal field network according to claim 3, characterized in that: Step S3 specifically involves: (1) Feature dimension normalization: The preset feature embedding dimension is By using linear transformation, the original observations and physical priors are mapped to... The space ensures that data from different sources are integrated within a unified semantic space; (2) Time encoder: Method: Time series features are extracted using the Transformer temporal attention mechanism; Input: A sequence of full-altitude observations over the past m time steps, D(t−m:t, (, B, K) and its corresponding mask matrix M(t−m:t, , B, K). Operation Description: A masked temporal self-attention mechanism is used to extract historical evolution features in parallel for each height layer; Output: Time feature matrix: (3) Spatial encoder: Method: Graph Neural Networks (GNNs) and Graph Convolution are used to extract vertically coupled features; Input: Current full-altitude observation data D(t, (t, B, K) and its corresponding mask matrix M(t, , B,K) Operation Description: Graph neural networks (GNNs) are used to perform graph convolution operations. H height layers are abstracted as graph nodes. The vertical wind shear and physical coupling features are captured through a neighborhood aggregation mechanism. The mask dynamically adjusts the weights of the convolution kernel to suppress the spread of errors in the vertical space. Output: Spatial feature matrix: (4) Multimodal physical fusion layer: Method: An attention-weighted fusion mechanism is employed, which utilizes... As a query, it guides the model to learn and generate targets and dynamic weight vector and Realize physical constraint information Nonlinear spatiotemporal characteristics , Deep integration; enter: , , Operation Description: Physical query field construction: using a fully connected layer (MLP) Mapped to a query vector , Dynamic weight field generation: Query matrix Respectively with the time feature matrix and spatial characteristic matrix Similarity is calculated, and the results are normalized using the Sigmoid activation function to generate a dynamic weight matrix for the entire height layer. and Synchronous fusion of the entire profile: The generated dynamic weights are used to perform element-level weighting of the spatiotemporal features and then spliced ​​with the basic wind field features; Output: Full-height layer fused feature matrix ,formula 。 5. The residual-corrected wind profile radar inversion method based on dynamic spatiotemporal field network according to claim 4, characterized in that: Step S4 specifically involves: fusing the feature matrix of the full-height layer using a residual prediction head composed of multilayer perceptrons or convolutional layers. Perform a nonlinear mapping to directly output the wind field residual matrix for all height layers at that moment. ; =[ ] The initial wind field is superimposed with the predicted residual to obtain the final wind field. ; = 。 6. The residual-corrected wind profile radar inversion method based on dynamic spatiotemporal field network according to claim 5, characterized in that: Step S5 specifically involves: For each sample, input the full-height observation tensor D(tm:t, 1:H, B, K), the mass mask matrix M(tm:t, 1:H, B, K), and the basic wind field matrix at the current time and its historical periods. ; The difference between the true values ​​of BeiDou sounding and the basic wind field is used to generate a target residual matrix covering the entire altitude layer. ; Model training employs a composite loss function that includes spatiotemporal smoothing constraints. + in, , , Here are the weighting coefficients for each loss term; each sub-loss term is defined as follows: Residual regression loss : Use mean squared error (MSE) to minimize the model prediction residuals residual with target The differences between them ensure the accuracy of wind field inversion; Where: H: Total number of height layers detected by wind profiler radar; z: Vertical height layer index, z∈ ; The 2D wind field residual vector predicted by the model at time t and height z. ]; Time smoothing loss To constrain the stability of the wind field evolution along the time axis, calculate the vector difference of the full profile between adjacent time points within a consecutive m time intervals. Where: m is the time step of the model input; i: time series index, which is traversed starting from the second time step in the window to calculate the adjacent differences; : The full-height layer output wind field matrix at time i; Spatial smoothing loss : in: The final output wind field after correction at layer z is calculated as follows: .

Images