Wind field three-dimensional correction and uncertainty prediction method based on double-branch cross-attention and spatio-temporal joint modeling

Abstract

The invention relates to a wind field three-dimensional correction and uncertainty prediction method based on double-branch cross-attention and spatio-temporal joint modeling, and the method comprises the steps: firstly obtaining wind field grid data output by a numerical weather prediction mode and point observation data of a wind field, and obtaining the observation data of each observation point; a quality weight is given to each piece of observation data; then defining an observation operator of the observation data; then, a deep learning model based on double-branch coding-trans-attention fusion-spatio-temporal joint modeling is constructed, and training is carried out; and finally, inputting observation data of a past time step into the trained deep learning model, and outputting a corrected three-dimensional wind field of a future time step and prediction uncertainty of each observation point. According to the method, global interpolation can be avoided, multi-source heterogeneous data are dynamically fused, physical consistency is ensured, and reliable uncertainty information can be provided.

Description

Technical Field

[0001] This invention relates to a method for three-dimensional correction and uncertainty prediction of wind fields based on dual-branch cross-attention and spatiotemporal joint modeling, belonging to the field of intelligent fusion of meteorological data and numerical model correction technology. Background Technology

[0002] High-precision, high-resolution three-dimensional correction and forecasting of meteorological wind fields are crucial for applications such as wind energy resource assessment, aviation safety, and weather warnings. Current technologies mainly rely on the fusion of numerical weather prediction models and observational data, i.e., "observation + model interpolation correction," or direct global interpolation, but these generally have the following limitations: Accumulated interpolation error: Traditional methods heavily rely on interpolating sparse, unconventional observations (such as wind profiler radar and scatterometers) onto the model grid, or vice versa. This "global interpolation" introduces significant errors in sparse data areas and complex terrain areas, and the errors accumulate with spatial distance and the interpolation process, causing the interpolation error to increase sharply with spatial distance and observation sparsity.

[0003] Insufficient capture of dynamic relationships: Existing machine learning-based correction methods mostly use static bias correction models. Model bias correction generally relies on static spatial mapping, which cannot fully learn and utilize the complex dynamic nonlinear relationship between observations and model background fields that changes with the evolution of weather systems.

[0004] Lack of physical consistency: Purely data-driven correction results may violate basic constraints of atmospheric dynamics and thermodynamics (such as mass conservation and wind field divergence characteristics), resulting in physical unreliability. Physical consistency (such as wind field divergence and boundary layer structure) is difficult to guarantee, limiting its application in business.

[0005] The methods suffer from weak uncertainty quantification capabilities: most provide only a single correction result, lacking a quantitative assessment of the reliability of the correction, and have insufficient uncertainty output capabilities, making it difficult to reliably quantify the correction quality. Users cannot know in which regions or under what weather conditions the correction results are more reliable, which introduces risks into decision-making. Summary of the Invention

[0006] In order to solve the problems existing in the prior art, the present invention provides a three-dimensional wind field correction method that can avoid global interpolation, dynamically fuse multi-source heterogeneous data, ensure physical consistency, and provide reliable uncertainty information.

[0007] To achieve the above objectives, the technical solution proposed in this invention is: a method for three-dimensional correction and uncertainty prediction of wind fields based on dual-branch cross-attention and spatiotemporal joint modeling, characterized in that it includes: Step 1: Obtain the wind field grid data and point observation data of the wind field output from the numerical weather prediction model, and obtain the observation data for each observation point: ; in, x i , y i , z i Longitude, latitude, and altitude coordinates q i For observation time, o i For the observed values, t i For quality marks, For observation uncertainty; Step 2: Define the observation operator H for the observed data; ; in, G For grid field, Let be the coordinates of the observation point, ( l,m,n ) is the grid cell index where the point is located, d l d m d n These represent the interpolation ranges for the current observation point in the longitude, latitude, and altitude directions, respectively. Step 3: Construct a deep learning model based on dual-branch encoding, cross-attention fusion, and spatiotemporal joint modeling, and train it. Step 4, Past K The observed data at each time step is input into the trained deep learning model, which outputs the future... t Corrected three-dimensional wind field at each time step and prediction uncertainty at each observation point.

[0008] A further design of the above technical solution is as follows: In step 1, a quality weight is assigned to each observation data point: ; in, w i It is the first i The quality weight of each observation point q i It is a quality mark. Y It is the attenuation coefficient.

[0009] The deep learning model in step 2 is specifically: For observation points i The observation data is encoded using a dual-branch encoding method: Logarithmic Weather Prediction 3D Tensor Encode the data to obtain high-dimensional grid features: Where B is the batch size, T is the number of time steps, L is the latitude dimension, W is the longitude dimension, D is the number of altitude layers, and C is the number of meteorological variables. C hidden The hidden feature dimension is the output dimension of the encoding network; Encode the feature vector of each observation point i to obtain the enhanced observation point features: Among them, E point For the feature set of observation points; Cross-attention fusion of high-dimensional grid features and enhanced observation point features: Flatten the high-dimensional grid features into a sequence: ; Using a Cross-Attention mechanism where the observation point is the query and the high-dimensional grid features are the key / value: ; ; The fusion characteristics were obtained: ; Where Q is the Query feature of the observation point, K is the Key feature of the grid feature, V is the Value feature of the grid feature, and W... Q W K W V Let N be the projection weight matrix. obs D represents the number of observation points. q D k D v The feature vector dimension of Q / K / V; To integrate and update the high-dimensional grid features, To integrate and update the features of the observation points; The fused feature sequence from the past K time steps: Enter PredFormer and add the position code: ; Where P is the comprehensive positional code for each sequence element, including the spatial positional code PE. space and time location encoding PE time ;PE space PE is used to characterize the spatial distribution location information of the input fused features. time Used to characterize the chronological order of a sequence at historical time steps; Using a Transformer encoder layer, spatiotemporal dependencies are captured simultaneously through a self-attention mechanism: ; Output the future t The 3D wind field correction at each moment: ; in It is a three-dimensional wind field correction.

[0010] In step 2, the numerical weather prediction three-dimensional tensor is encoded using a 3D CNN or neural operator.

[0011] In step 2, the feature vector of observation point i is Among them, local background feature b i Extracted through observation operators: , The feature vectors of all observation points constitute the feature set of observation points. ; The feature set of observation points is encoded through a Point Transformer or a graph attention network.

[0012] The loss function of the deep learning model in step 2 includes: Main supervisor's loss: , Where v is the wind speed vector, θ is the wind direction, α is the weight, and N is the number of observation points; v i pred Let v be the predicted wind speed vector at the i-th observation point. i obs θ is the measured wind speed vector; i pred Let θ be the predicted wind direction at that point. i obs This is the measured wind direction; Divergence constraint: ; Where Ω is the spatial computational domain defined by the physical constraints of the wind field; v NWP Δv is the velocity vector of the original wind field of the model, and Δv is the correction of the model output wind field to the model wind field. Energy spectrum uniformity: ' Uncertainty quantification loss: ' Where E(k) is the energy spectrum distribution of the corrected wind field at different wavenumbers k, E ref (k) is the reference energy spectrum, σ i O represents the uncertainty of the model's prediction result for the i-th observation point or grid point. i This represents the observed true value, and its dimension is consistent with the model's output prediction value; The predicted value output by the model; Therefore, the total loss is: , where λ1, λ2, and λ3 are hyperparameters.

[0013] The corrected 3D wind field output from step 4 is adjusted with the following constraints: ; Among them, v model The corrected three-dimensional wind speed vector field output from step 4, For mass flux divergence.

[0014] The beneficial effects of this invention are as follows: This invention designs a dual-branch data encoding system with grid branch and point cloud branch to process the background field of regular gridded numerical weather prediction (NWP) and the observation data (such as wind profiles and ground stations) with irregular spatial distribution, respectively. This avoids interpolating observations onto the grid from the source, thus avoiding interpolation errors. Furthermore, through dynamic fusion and spatiotemporal joint modeling, it captures the details of wind field evolution more accurately.

[0015] This invention introduces a differentiable observation operator H for precise supervision. During training and inference, the grid field output by the network is directly sampled at the observation point location and compared with the real observation to calculate the loss. This unifies the training and inference process and ensures the consistency between the model optimization goal and the business application scenario.

[0016] This invention introduces a cross-attention dynamic feature fusion mechanism, utilizing a Cross-Attention mechanism with observation points as queries and grid features as keys / values ​​to achieve dynamic and adaptive fusion of observation information and pattern background field. This mechanism can intelligently extract relevant information from the background field to enhance the observation representation based on the reliability and spatial location of the observation, and vice versa.

[0017] This invention constructs a spatiotemporal integrated PredFormer prediction model, employing a Transformer-based PredFormer encoder to model the fused spatiotemporal sequence and directly output 3D wind field corrections for multiple future time points. This model can simultaneously capture temporal evolution patterns and spatial relationships, achieving end-to-end spatiotemporal prediction.

[0018] This invention explicitly introduces physical regularization terms such as wind field divergence constraints and energy spectrum consistency into the loss function to guide the model to generate physically more reasonable results. Simultaneously, through quantile regression or negative log-likelihood loss, the model can output the corrected uncertainty for each grid point. Detailed Implementation

[0019] The present invention will now be described in detail with reference to specific embodiments. Example

[0020] The wind field three-dimensional correction and uncertainty prediction method based on dual-branch cross-attention and spatiotemporal joint modeling in this embodiment includes: Step 1: Acquire data and preprocess it; 1.1 Data source; Grid data: Wind field grid data (3D grid field) output by numerical weather prediction models, with variables including zonal wind U, meridional wind V, vertical velocity W, temperature T, specific humidity Q, etc., and dimensions of [time step, latitude, longitude, altitude, variable].

[0021] Wind field point observation data: including sparse and irregular observations provided by wind profiler radar, ground meteorological stations, etc.

[0022] The data structure for each observation point is as follows: ; in, x i , y i , z i Longitude, latitude, and altitude coordinates q i For observation time, o i These are observed values ​​(such as wind speed, wind direction, etc.). t i For quality marks, This represents the observation uncertainty.

[0023] 1.2 Data quality control; The observed data undergoes threshold checks and continuity checks, and each observation is assigned a quality weight. ; Among them, w i q is the mass weight of the i-th observation point. i It is a quality mark (usually q) i =0 indicates the best quality), and Υ is the attenuation coefficient.

[0024] Step 2: Define the observation operator H; The observation operator is the core of this method to avoid global interpolation. It is a differentiable sampling function used to accurately obtain arbitrary spatial locations from a continuous grid field G. The value of .

[0025] This embodiment uses trilinear interpolation: ; Where G is the grid field, Let (l, m, n) be the coordinates of the observation point, and (l, m, n) be the coordinates of point p. i The index of the grid cell in which it is located.

[0026] The differentiability of this operator ensures that the gradient can be backpropagated from the observation point to the entire grid field during training.

[0027] The observation operator H is used to accurately sample physical quantities (such as wind speed and direction) at the spatial location of the observation point from a three-dimensional grid field. The implementation of this patent's observation operator is typically trilinear interpolation, the basic idea of ​​which is: for any observation point, its spatial coordinates (longitude, latitude, altitude) (x...)... i ,y i , z i In 3D grid data (such as NWP mode output), there are no completely overlapping grid points. Therefore, weighted summation needs to be performed within the grid cell where the observation point is located. The specific procedure is as follows: Let the data of the grid field be G;

[0028] For a given observation point, its actual spatial coordinates are usually located within a certain grid cell (i.e., within the range of [l~l+1], [m~m+1], [n~n+1], not necessarily falling on integer grid points). Trilinear interpolation requires a weighted average of the values ​​of the 8 nearest grid points around that cell.

[0029] d l d m d n These represent the interpolation ranges for the current observation point in the longitude, latitude, and altitude directions, respectively. Generally, d... l = {l, l+1},d m = {m, m+1},d n = {n, n+1}, where l, m, and n are the lower bound indices of the grid cell containing the observation point in the longitude, latitude, and altitude directions, respectively. l ∈ d l The value of l is the grid index near the observation point (usually the lower and upper bound indices of the cell), i.e., l∈{l,l+1}m ∈ d. m The value of m is the index of the observation point in the latitudinal direction (similarly), i.e., m∈{m,m+1}n ∈d n The value of n is the index of the observation point in the height direction (similarly), i.e., n∈{n,n+1}.

[0030] Step 3: Construct a deep learning model based on dual-branch encoding, cross-attention fusion, and spatiotemporal joint modeling, and train it; 3.1 Perform dual-branch coding on the observation data of the observation points; Grid branch: For NWP three-dimensional tensor Where B is the batch size, T is the time step, L is the latitude dimension, W is the longitude dimension, D is the altitude layer, and C is the number of meteorological variables, these are encoded using a 3D CNN or neural operator (such as FNO) to obtain high-dimensional grid features: G represents the three-dimensional tensor of the grid field. This three-dimensional tensor fully contains all the spatial grids and meteorological variables output by the numerical weather prediction model. It is the data input form of the gridded background field of the wind field in the model encoding and fusion process.

[0031] Point Cloud Branch: For each observation point i, its feature vector is composed of its observed value, mass weight, uncertainty, and local background features: Local background features Extracted through observation operators: Feature set of all observation points Then, after passing through a Point Transformer or graph attention network, enhanced observation point features are obtained: .

[0032] 3.2 Cross-attention fusion of high-dimensional grid features and enhanced observation point features; Flatten the high-dimensional grid features into a sequence:

[0033] Using a Cross-Attention mechanism where the observation point is the query and the grid features are the key / value:

[0034] Output grid-enhanced features that fuse observation information and observation-enhanced features that fuse background field information: This is the fusion feature;

[0035] Where: Q (Query) is the Query feature of the observation point: the query feature obtained by linear transformation of the observation point feature set E_point through the weight matrix WQ, which is used to retrieve relevant information in the grid features; K (Key) is the Key feature of the grid feature: the key feature obtained by WK linear transformation of the high-dimensional grid feature F_grid, which is used to match the query to determine the relevance; V(Value) is the value feature of the grid feature: the value feature obtained by WV linear transformation of the high-dimensional grid feature F_grid, representing the extractable context information; W Q W K W VThe projected weight matrices are the learnable linear projected weight matrices for Query, Key, and Value, respectively. N obs Number of observation points; D q D k D v The feature vector dimension of Query, Key, and Value (the hidden layer dimension inside the model) is the new feature length (hidden layer dimension) of the Query / Key / Value vector obtained after linear projection (such as through a fully connected layer / weight matrix transformation).

[0036] The high-dimensional grid features of the fusion output, after combining observation information, can be used as input for subsequent wind field correction or further inference. Features of observation points after fusing background information of the mode are used to correct the output or for subsequent processing.

[0037] 3.3 Spatiotemporal PredFormer encoder; The fused feature sequence from the past K time steps: ; Enter PredFormer and add the position code: ; Where: P is the comprehensive positional code for each sequence element, representing the total positional encoding matrix or vector used by the input fused feature sequence to capture temporal and spatial context information, containing the spatial positional encoding PE. space and time location encoding PE time ;PE space The PredFormer model uses spatial and temporal location encoding to represent the spatial distribution of the input fusion features, while PEtime represents the chronological order of the sequence over historical time steps. By effectively combining spatial and temporal location encoding, the PredFormer model achieves spatiotemporal joint modeling and accurate prediction of wind field evolution.

[0038] E space Spatial Positional Encoding (SPE) describes the unique location information of each grid point or observation point in three-dimensional space (longitude, latitude, and altitude). Each spatial location is encoded with an independent vector, such as using sine-cosine encoding or learned embedding. The goal is to enable the model to distinguish different spatial points and understand spatial distribution and relationships.

[0039] E timeTemporal Positional Encoding (TPE) is used to describe the temporal order or absolute moment of each historical time step in a sequence. It encodes a time vector (sine-cosine or embedding) for each time step, which, combined with spatial encoding, is used to identify temporal information. The purpose is to help models capture sequence order, temporal evolution, and dynamic features.

[0040] Using a Transformer encoder layer, spatiotemporal dependencies are captured simultaneously through a self-attention mechanism: ; Output the 3D wind field corrections for the next τ time steps: ,in It is a three-dimensional wind field correction.

[0041] C hidden The dimension of the high-dimensional feature vector obtained after processing the original meteorological variable data (C-dimensional) through a deep learning encoding network (such as 3D CNN, FNO, etc.) is the feature representation length of each grid point after encoding.

[0042] C hidden It is a hyperparameter used in model design. It is usually greater than or equal to C, but the two are not directly mathematically related and are determined by the coding network structure.

[0043] For example, if the original number of meteorological variables C=6 (U, V, W, T, Q, P), after passing through the coding network, the child can be set to 64, 128, 256, etc., making the feature representation ability stronger.

[0044] When flattened into a sequence for use in structures such as Transformer / Cross-Attention, it represents the dimension of the input features at each step.

[0045] In other words: C represents the number of original meteorological variables (input dimension), C hidden This refers to the encoded high-level feature dimension (used within the deep model, as the input dimension after serialization). C represents the number of original meteorological variables, such as wind speed components, temperature, humidity, etc. hidden C represents the hidden feature dimension of the encoded network output, i.e., the length of the feature representation of each grid point after deep feature extraction and flattening into a sequence. hidden A value greater than C is usually an important parameter for improving the expressive power of the model.

[0046] 3.4 Loss Function Design; The total loss function of a deep learning model consists of multiple parts: Main supervisor loss (calculated at the observation point): , Where v is the wind speed vector, θ is the wind direction, and α is the weight; N is the number of observation points (or the number of batch samples); v i pred Let v be the predicted wind speed vector at the i-th observation point. i obs θ is the measured wind speed vector; i pred Let θ be the predicted wind direction at that point. i obs This is the measured wind direction.

[0047] Physical consistency loss: Divergence constraint:

[0048] Energy spectrum uniformity:

[0049] Uncertainty quantification loss (negative log-likelihood loss):

[0050] Total loss: , where λ1, λ2, and λ3 are hyperparameters used to balance the contributions of various losses.

[0051] Where: Ω is the spatial computational domain defined by the physical constraints of the wind field; v NWP Let be the velocity vector of the original wind field in the model, and Δv be the correction of the model output wind field to the model wind field; E(k) is the energy spectrum distribution of the corrected wind field at different wavenumbers k. ref (k) is the reference energy spectrum (which can be obtained by observation or theory) used for physical consistency energy spectrum constraints.

[0052] Step 4: Reasoning and Post-processing; Input the past step observation data (NWP field and observation point cloud) into the trained deep learning model, and the model directly outputs the corrected 3D wind field for the future step and the prediction uncertainty for each grid point: .

[0053] Ranhouran'ho inputs the wind field output from the model into a mass-consistent model under mass conservation constraints for fine-tuning, further ensuring physical constraints: ; ; in, Let ρ be the mass flux divergence in fluid mechanics, where ρ is density and V is velocity vector.

[0054] v modelThe corrected 3D wind speed vector field output in step 4 (including the corrected wind components at each spatial grid point), v is the final physically consistent wind speed field obtained after constraint optimization adjustment. It is based on "v model Based on this, further fine-tuning and optimization are performed to ensure that it meets physical constraints such as mass conservation (e.g., ∇⋅(ρv) = 0). ∇⋅(ρV) is the mass flux divergence, ρ is the air density, and ∇⋅(ρV) expresses the net rate of change of air mass flux in three-dimensional space. It is used to constrain the physical consistency of the corrected wind field, ensuring that it meets the mass conservation conditions of atmospheric dynamics. In other words, v model The corrected 3D wind speed vector field directly output by the deep learning model. v This represents the final three-dimensional wind speed field obtained after adjusting for physical constraints. The optimization objective is to ensure... v Under the premise of satisfying physical constraints such as the law of conservation of mass, let it be... v model The closest approach is achieved even when the constrained optimized wind field retains the model correction effect to the greatest extent in both spatial distribution and numerical value.

[0055] Corrected 3D wind field is a numerical prediction by the model of wind speed (or other wind field parameters) at future t-step grids or observation points.

[0056] Prediction uncertainty is the quantification of the confidence interval or probability distribution of the model for each corrected wind field prediction, indicating the reliability or error range of the prediction.

[0057] Specifically, the model is typically designed with two outputs: First output head: Predicted values ​​of physical quantities such as wind speed and wind direction (i.e., the corrected three-dimensional wind field, whether for each grid point or each observation point); The second output head is the uncertainty estimate (such as standard deviation, variance, quantile interval, etc.) corresponding to each predicted value (each spatial point, each future time step).

[0058] The model output includes: The corrected three-dimensional wind field at the next t time steps (i.e., the predicted values ​​of wind speed, wind direction, etc. at each forecast time and each spatial location). ② The prediction uncertainty parameters (such as prediction standard deviation, confidence interval or probability distribution parameters) for each spatial point and each time step are used to quantify the credibility and risk of the model output. The prediction uncertainty is an error quantification index of the model output correction value, which may include standard deviation, covariance, quantile interval or other probability distribution parameters, and is used to quantitatively assess the reliability of the correction results.

[0059] Step 5, Visualization: The final results can be visualized through three-dimensional streamlines, isosurfaces, etc., and uncertainty information can be superimposed to provide intuitive decision support for meteorological analysts.

[0060] The technical solution in this embodiment abandons global spatial interpolation and adopts a real-time fusion of multi-source wind field information based on "dual-branch (grid branch and point cloud branch) + cross-attention mechanism". It uses the spatiotemporally encoded PredFormer module to realize joint spatiotemporal inference of historical multi-step wind fields and observations, realizes three-dimensional high-resolution correction of wind fields and prediction of future deviations, and introduces the observation operator H to directly sample observation points for accurate model output, unify the training and inference process, avoid interpolation dependence and errors, and capture the details of wind field evolution more accurately through dynamic fusion and spatiotemporal joint modeling. The embedded physical consistency loss term (divergence, energy spectrum, boundary layer structure) ensures that the output wind field is physically compliant and meets the basic laws of atmospheric dynamics. The provided uncertainty quantification results provide users with the confidence level of the correction results and support risk decision-making. Moreover, the design of the point cloud branch naturally supports the dynamic changes in the location and number of observation points, and the model has stronger adaptability to different observation network layouts.

[0061] The technical solutions of the present invention are not limited to the above embodiments. All technical solutions obtained by equivalent substitution fall within the scope of protection claimed by the present invention.

Claims

1. A method for three-dimensional correction and uncertainty prediction of wind fields based on dual-branch cross-attention and spatiotemporal joint modeling, characterized in that, include: Step 1: Obtain the wind field grid data and point observation data of the wind field output from the numerical weather prediction model, and obtain the observation data for each observation point: ; in, x i , y i , z i Longitude, latitude, and altitude coordinates q i For observation time, o i For the observed values, t i For quality marks, For observation uncertainty; Step 2: Define the observation operator H for the observed data; ; in, G For grid field, Let be the coordinates of the observation point, ( l,m,n ) is the grid cell index where the point is located, d l d m d n These represent the interpolation ranges for the current observation point in the longitude, latitude, and altitude directions, respectively. Step 3: Construct a deep learning model based on dual-branch encoding, cross-attention fusion, and spatiotemporal joint modeling, and train it. Step 4, Past K The observed data at each time step is input into the trained deep learning model, which outputs the future... τ Corrected three-dimensional wind field at each time step and prediction uncertainty at each observation point.

2. The method for three-dimensional correction and uncertainty prediction of wind fields based on dual-branch cross-attention and spatiotemporal joint modeling as described in claim 1, characterized in that: In step 1, a quality weight is assigned to each observation data point: ； in, w i It is the first i The quality weight of each observation point q i It is a quality mark. Υ It is the attenuation coefficient.

3. The wind field three-dimensional correction and uncertainty prediction method based on dual-branch cross-attention and spatiotemporal joint modeling as described in claim 2, characterized in that: The deep learning model in step 2 is specifically: For observation points i The observation data is encoded using a dual-branch encoding method: Logarithmic Weather Prediction 3D Tensor Encode the data to obtain high-dimensional grid features: Where B is the batch size, T is the number of time steps, L is the latitude dimension, W is the longitude dimension, D is the number of altitude layers, and C is the number of meteorological variables. C hidden The hidden feature dimension is the output dimension of the encoding network; Encode the feature vector of each observation point i to obtain the enhanced observation point features: Among them, E point For the feature set of observation points; Cross-attention fusion of high-dimensional grid features and enhanced observation point features: Flatten the high-dimensional grid features into a sequence: ; Using a Cross-Attention mechanism where the observation point is the query and the high-dimensional grid features are the key / value: ； ； The fusion characteristics were obtained: ； Where Q is the Query feature of the observation point, K is the Key feature of the grid feature, V is the Value feature of the grid feature, and W... Q W K W V Let N be the projection weight matrix. obs D represents the number of observation points. q D k D v The feature vector dimension of Q / K / V; To integrate and update the high-dimensional grid features, To integrate and update the features of the observation points; The fused feature sequence of the past K time steps Enter PredFormer and add the position code: ; Where P is the comprehensive positional code for each sequence element, including the spatial positional code PE. space and time location encoding PE time ;PE space PE is used to characterize the spatial distribution location information of the input fused features. time Used to characterize the chronological order of a sequence at historical time steps; Using a Transformer encoder layer, spatiotemporal dependencies are captured simultaneously through a self-attention mechanism: ； Output the future τ The 3D wind field correction at each moment: ; in It is a three-dimensional wind field correction.

4. The wind field three-dimensional correction and uncertainty prediction method based on dual-branch cross-attention and spatiotemporal joint modeling as described in claim 3, characterized in that: In step 2, the numerical weather prediction three-dimensional tensor is encoded using a 3D CNN or neural operator.

5. The method for three-dimensional correction and uncertainty prediction of wind fields based on dual-branch cross-attention and spatiotemporal joint modeling as described in claim 4, characterized in that: In step 2, the feature vector of observation point i is Among them, local background feature b i Extracted through observation operators: , The feature vectors of all observation points constitute the feature set of observation points. ; The feature set of observation points is encoded through a Point Transformer or a graph attention network.

6. The method for three-dimensional correction and uncertainty prediction of wind fields based on dual-branch cross-attention and spatiotemporal joint modeling as described in claim 5, characterized in that: The loss function of the deep learning model in step 2 includes: Main supervisor's loss: , Where v is the wind speed vector, θ is the wind direction, α is the weight, and N is the number of observation points; v i pred Let v be the predicted wind speed vector at the i-th observation point. i obs θ is the measured wind speed vector; i pred Let θ be the predicted wind direction at that point. i obs This is the measured wind direction; Divergence constraint: ; Where Ω is the spatial computational domain defined by the physical constraints of the wind field; v NWP Δv is the velocity vector of the original wind field of the model, and Δv is the correction of the model output wind field to the model wind field. Energy spectrum uniformity: ' Uncertainty quantification loss: ' Where E(k) is the energy spectrum distribution of the corrected wind field at different wavenumbers k, E ref (k) is the reference energy spectrum, σ i O represents the uncertainty of the model's prediction result for the i-th observation point or grid point. i This represents the observed true value, and its dimension is consistent with the model's output prediction value; The predicted value output by the model; Therefore, the total loss is: , where λ1, λ2, and λ3 are hyperparameters.

7. The method for three-dimensional correction and uncertainty prediction of wind fields based on dual-branch cross-attention and spatiotemporal joint modeling as described in claim 6, characterized in that: The corrected 3D wind field output from step 4 is adjusted with the following constraints: ; Among them, v model The corrected three-dimensional wind speed vector field output from step 4, For mass flux divergence.

Images