A grid point correction method, device and medium for spatial meteorological data

Abstract

The application discloses a kind of grid point revision methods, equipment and medium of space weather data, it is related to meteorological artificial intelligence technical field, including, original space weather data is collected and is preprocessed;Grid point geographic coordinates, observation station and satellite pixel are extracted from the original space weather data after pre-processing and defined as node, construct multimodal weather spatiotemporal heterogeneity diagram;By graph neural network, the deep feature extraction and fusion processing of multimodal weather spatiotemporal heterogeneity diagram are carried out, obtain high-dimensional feature fusion vector, and according to high-dimensional feature fusion vector, calculate preliminary residual error estimate value, obtain preliminary residual error field data.The application is constructed multimodal weather spatiotemporal heterogeneity diagram, and utilizes graph neural network to carry out deep feature extraction and fusion processing, realizes the spatiotemporal correlation modeling and nonlinear feature learning of multi-source heterogeneous data, enhances data representation capability, improves preliminary residual error estimate accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of meteorological artificial intelligence, and in particular to a grid point correction method for spatial meteorological data, equipment and a medium. BACKGROUND

[0002] In the field of spatial meteorological data analysis, the grid point correction technology is a key link to realize the conversion of sparse observation data to continuous spatial field. The traditional method mainly relies on various types of spatial interpolation algorithms and variants, which interpolate and extrapolate the station observation data through mathematical optimal estimation theory to generate regular latitude and longitude grid data. In recent years, with the wide application of numerical weather prediction models and reanalysis data, the data assimilation and fusion method based on background field and observation residual has become one of the mainstream technologies in the field, which aims to improve the accuracy and spatial consistency of grid meteorological products by fusing multi-source observation information and prior background field.

[0003] However, the existing method still has some deficiencies. In the process of multi-source heterogeneous data fusion, the traditional method is difficult to effectively depict the complex spatio-temporal correlation and physical consistency between different types of data, resulting in insufficient information utilization; relying on mathematical interpolation or statistical optimization, the physical conservation law of meteorology is not used as a constraint, so that in the extreme weather or sparse observation area, the interpolation result may be physically unreliable (such as temperature field violating the law of thermodynamics). SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a grid point correction method for spatial meteorological data to solve the problems of insufficient spatio-temporal correlation description and lack of physical constraints in multi-source meteorological data fusion.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides a grid point correction method for spatial meteorological data, which comprises collecting and preprocessing original spatial meteorological data; extracting grid geographical coordinates, observation stations and satellite pixels from the preprocessed original spatial meteorological data and defining them as nodes to construct a multi-modal meteorological spatio-temporal heterogeneous graph; performing deep feature extraction and fusion processing on the multi-modal meteorological spatio-temporal heterogeneous graph through a graph neural network to obtain a high-dimensional feature fusion vector, and calculating a preliminary residual estimate value according to the high-dimensional feature fusion vector to obtain preliminary residual field data; setting a meteorological physics equation constraint term, combining the preliminary residual field data and the grid geographical coordinates, constructing a physical information neural network framework and initializing network parameters to form a physical information neural network; inputting the preliminary residual field data and the grid geographical coordinates into the physical information neural network for forward propagation calculation to obtain high-precision spatial meteorological correction analysis field data.

[0008] As a preferred scheme of the grid point correction method of spatial meteorological data, wherein: the original spatial meteorological data includes observation data of ground meteorological stations, upper air sounding stations, meteorological satellites, weather radars, and background fields and elevation data;

[0009] The preprocessing includes spatio-temporal alignment interpolation, normalization standardization and missing value filling.

[0010] As a preferred scheme of the grid point correction method of spatial meteorological data, wherein: the construction of the multi-modal meteorological spatio-temporal heterogeneous graph is as follows,

[0011] Extracting grid geographic coordinate data, observation station data and satellite remote sensing data from the preprocessed original spatial meteorological data and creating nodes to generate a grid node set, a station node set and a satellite node set;

[0012] Assigning background fields and elevation data to the grid node set, assigning ground meteorological station and upper air sounding station observation data to the station node set, and assigning meteorological satellite observation data to the satellite node set to obtain attribute grid node set, attribute station node set and attribute satellite node set;

[0013] Calculating the spatial distance between the attribute grid node set, the attribute station node set and the attribute satellite node set, and establishing a spatial proximity connection relationship to generate a multi-connection relationship edge set;

[0014] Integrating the attribute grid node set, the attribute station node set, the attribute satellite node set and the multi-connection relationship edge set to construct a multi-modal meteorological spatio-temporal heterogeneous graph.

[0015] As a preferred scheme of the grid point correction method of spatial meteorological data, wherein: the step of obtaining the high-dimensional feature fusion vector is as follows,

[0016] Initializing the features of each type of node in the multi-modal meteorological spatio-temporal heterogeneous graph through a graph neural network to generate a node initial feature matrix;

[0017] Based on the node initial feature matrix, the connection weights between different types of nodes are calculated to generate a weighted adjacency matrix;

[0018] Using the weighted adjacency matrix to perform message passing and feature aggregation across nodes of different types to generate an updated node feature matrix;

[0019] According to the updated node feature matrix, feature propagation and fusion are performed through a multi-layer graph neural network to generate a high-dimensional feature fusion vector.

[0020] As a preferred embodiment of the gridded correction method for space meteorological data described in this invention, the steps for obtaining preliminary residual field data are as follows:

[0021] A linear transformation is performed on the high-dimensional feature fusion vector to calculate the initial residual scalar value for each grid point;

[0022] The geographic coordinates and elevation data of each grid point are extracted from the preprocessed raw space meteorological data, and feature combination is performed with the preliminary residual scalar values ​​of each grid point to generate a composite feature vector set.

[0023] The composite feature vector set is regularly reorganized according to the spatial grid position and spatially aligned with the background field data to obtain preliminary residual field data.

[0024] In a preferred embodiment of the grid correction method for space meteorological data described in this invention, the steps for forming a physical information neural network are as follows:

[0025] Extract the initial residual scalar values ​​of each grid point from the initial residual field data, and combine them with the corresponding grid point geographic coordinates to generate an input feature vector set;

[0026] The grid geographic coordinates of observation stations and the corresponding ground meteorological and upper-air sounding station observation data are extracted from the preprocessed raw space meteorological data to generate the training target dataset.

[0027] Thermodynamic equations and fluid dynamics equations are defined based on the conservation laws of meteorological physics, and then discretized to generate a discrete calculation framework for the physical equations.

[0028] Based on the dimensional features of the input feature vector set, the structure of the physical information neural network is defined, and a network architecture computation graph is generated.

[0029] The network architecture computation graph is used to perform forward propagation computation on the input feature vector set to obtain the network prediction value, and the mean square error is calculated in combination with the training target dataset to generate the data fitting loss term;

[0030] Based on the set of input feature vectors, the residuals are calculated using the physical equation discrete computation framework to obtain the physical regularization loss term, which is then fused with the data fitting loss term to generate a joint optimization loss function.

[0031] All weight parameters are initialized based on the network architecture computation graph, and the weight parameters are reassigned to generate an initial network parameter set;

[0032] By integrating the joint optimization loss function, the set of input feature vectors, and the initialization set of network parameters, a physical information neural network is constructed.

[0033] As a preferred embodiment of the grid-based correction method for space meteorological data described in this invention, the steps of inputting the preliminary residual field data and grid-based geographic coordinates into a physical information neural network for forward propagation calculation are as follows:

[0034] The preliminary residual field data and grid point geographic coordinates are spliced ​​together according to spatial location to generate a composite input feature vector;

[0035] The composite input feature vector is input into the physical information neural network, the output of the meteorological element correction values ​​of each grid point is generated, and the meteorological element correction tensor is generated by integrating them.

[0036] As a preferred embodiment of the grid correction method for space meteorological data described in this invention, the high-precision space meteorological correction analysis field data is obtained by fusing the meteorological element correction tensor with the background field data.

[0037] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the grid correction method for space meteorological data as described in the first aspect of the present invention.

[0038] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the grid correction method for space meteorological data as described in the first aspect of the present invention.

[0039] The beneficial effects of this invention are as follows: By constructing a multimodal meteorological spatiotemporal heterogeneous map and using graph neural networks for deep feature extraction and fusion processing, the spatiotemporal correlation modeling and nonlinear feature learning of multi-source heterogeneous data are realized, enhancing data representation capabilities and improving the accuracy of preliminary residual estimation; by constructing a physical information neural network framework, the deep integration of physical conservation laws and data-driven methods is realized, which is used to constrain the network output to conform to thermodynamic and fluid dynamic laws, generating physically consistent and highly reliable space meteorological correction analysis field data. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Fig. 1 A flowchart for a grid correction method for space meteorological data.

[0042] Fig. 2A flowchart for constructing a multimodal meteorological spatiotemporal heterogeneous map.

[0043] Fig. 3 This is a flowchart for feature extraction and initial residual field data generation.

[0044] Fig. 4 A flowchart for constructing a physical information neural network and acquiring space meteorological correction analysis field data. Detailed Implementation

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0046] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0047] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0048] Reference Figs. 1-4 As an embodiment of the present invention, this embodiment provides a grid correction method for space meteorological data, comprising the following steps:

[0049] S1. Collect raw space meteorological data and perform preprocessing;

[0050] Raw space meteorological data includes observational data from ground meteorological stations, upper-air sounding stations, meteorological satellites, and weather radars, as well as background field and elevation data;

[0051] It should be noted that ground-based meteorological station observation data refers to real-time data of near-surface meteorological elements directly measured by temperature, humidity, air pressure, wind speed, and wind direction sensors from meteorological observation stations distributed on the Earth's surface; upper-air sounding station observation data refers to vertical profile data obtained by directly detecting the height, temperature, humidity, wind speed, and wind direction of various atmospheric pressure layers through the release of radiosondes and drop-in radiosondes; meteorological satellite observation data refers to cloud and precipitation data obtained through inversion algorithms from Earth observations by remote sensing instruments (such as imagers and spectrometers) on polar-orbiting and geostationary meteorological satellites. Water, temperature, and humidity parameters; weather radar observation data refers to the reflectivity, radial velocity, and spectral width obtained after signal processing and data inversion from electromagnetic waves emitted by weather radar and echo reception, reflecting the basic data of precipitation cloud structure and dynamic characteristics; background field data refers to global and regional gridded meteorological element analysis field data generated based on the fusion of multi-source observation data and physical laws, providing the initial state of spatial distribution of meteorological elements; elevation data refers to the surface elevation data obtained through satellite remote sensing mapping and topographic mapping, recording topographic relief characteristics in the form of a regular grid;

[0052] Preprocessing includes spatiotemporal alignment interpolation, normalization, and missing value imputation;

[0053] It should be noted that spatiotemporal alignment interpolation refers to uniformly interpolating the original spatial meteorological data to the same target time point and regular geographic latitude and longitude grid points to ensure the consistency of the data spatiotemporal reference; normalization refers to converting the observation data of ground meteorological stations, upper-air sounding stations, meteorological satellites and weather radars to a uniform numerical range (such as [0,1] or a mean of 0 and a variance of 1) through the Z-score method to eliminate dimensional differences; missing value imputation refers to using statistical methods (such as climate averages) to fill the observation data of ground meteorological stations, upper-air sounding stations, meteorological satellites and weather radars to ensure data integrity.

[0054] S2. Extract grid point geographic coordinates, observation stations and satellite pixels from the preprocessed raw space meteorological data and define them as nodes to construct a multimodal meteorological spatiotemporal heterogeneous map;

[0055] Extract gridded geographic coordinate data, observation station data, and satellite remote sensing data from the preprocessed raw space meteorological data and create nodes to generate gridded node sets, station node sets, and satellite node sets.

[0056] Furthermore, the latitude and longitude coordinates and background field data of each grid point are extracted from the preprocessed raw space meteorological data to generate a grid point geographic coordinate data list. Based on the grid point geographic coordinate data list, a node is created for each grid point and assigned corresponding latitude and longitude coordinates and background field data, generating a grid point node set. The latitude and longitude coordinates and field-observed meteorological element values ​​(such as temperature, air pressure, humidity, wind speed, and wind direction) of each station are extracted from the preprocessed raw space meteorological data to generate an observation station data list. Based on the observation station data list, a node is created for each observation station and assigned corresponding latitude and longitude coordinates and field-observed meteorological element values, generating a station node set. The latitude and longitude coordinates and remote sensing inversion physical quantity values ​​(such as precipitation, air temperature, reflectivity, and radial velocity) of each pixel are extracted from the preprocessed raw space meteorological data to generate a satellite remote sensing data list. Based on the satellite remote sensing data list, a node is created for each satellite pixel and assigned latitude and longitude coordinates and remote sensing inversion physical quantity values, generating a satellite node set.

[0057] Assign background field and elevation data to the grid node set, assign observation data from ground meteorological stations and upper-air sounding stations to the station node set, assign meteorological satellite observation data to the satellite node set, and obtain attribute grid node set, attribute station node set, and attribute satellite node set;

[0058] Furthermore, background field data and elevation data are extracted from the preprocessed raw space meteorological data and assigned to grid node sets to generate attribute grid node sets; ground meteorological station observation data and upper-air sounding station observation data are extracted from the preprocessed raw space meteorological data and assigned to station node sets to generate attribute station node sets; meteorological satellite observation data are extracted from the preprocessed raw space meteorological data and assigned to satellite node sets to generate attribute satellite node sets.

[0059] Calculate the spatial distance between the set of attribute grid nodes, the set of attribute site nodes, and the set of attribute satellite nodes, establish spatial proximity connections, and generate a set of multi-connection edges;

[0060] Furthermore, the latitude and longitude coordinates of each node are extracted from the attribute grid node set, attribute station node set, and attribute satellite node set. These coordinates are then transformed into Cartesian coordinates using projection transformation. Simultaneously, the Euclidean distance formula is used to calculate the planar distance between each attribute station node and all attribute grid nodes, generating a station-grid distance matrix. Similarly, the planar distance between each attribute satellite node and all attribute grid nodes is calculated, generating a satellite-grid distance matrix. Finally, the station-grid distance matrix is ​​compared with a preset distance threshold: if the planar distance between a station node and a grid node is less than or equal to a certain threshold, the distance is considered lower. When the distance threshold is met, spatial connection edges are established between station nodes and grid nodes. When the planar distance between a station node and a grid node is greater than the distance threshold, no connection edge is established, and a set of station-grid connection edges is generated. The satellite-grid distance matrix is ​​compared with the distance threshold: when the planar distance between a satellite node and a grid node is less than or equal to the distance threshold, spatial connection edges are established between the satellite node and the grid node; when the planar distance between a satellite node and a grid node is greater than the distance threshold, no connection edge is established, and a set of satellite-grid connection edges is generated. The set of station-grid connection edges and the set of satellite-grid connection edges are merged to generate a set of multi-connection edges.

[0061] The expression for calculating the Euclidean distance between nodes is:

[0062] ;

[0063] in, It is the Euclidean distance between nodes, representing the straight-line distance between two nodes in a Cartesian coordinate system; , It is the Cartesian coordinate of the first node; , These are the Cartesian coordinates of the second node;

[0064] It should be noted that the preset distance threshold is an empirical value set based on the attenuation characteristics of the spatial correlation of meteorological elements and the actual observation network density. The value range is 10-200 kilometers. 10 kilometers ensures the capture of small and medium-scale weather processes (such as convective activity) and avoids excessively sparse connections. 200 kilometers covers the effective influence range of large-scale weather processes (such as fronts and cyclones) and avoids the introduction of irrelevant noise due to excessive distance.

[0065] By integrating the set of attribute grid nodes, the set of attribute station nodes, the set of attribute satellite nodes, and the set of multi-connection edges, a multimodal meteorological spatiotemporal heterogeneous map is constructed.

[0066] Furthermore, based on the attribute grid node set, attribute site node set, and attribute satellite node set, a type identifier (such as "grid", "site", "satellite") is assigned to each node to generate a typed node set; according to the multi-connection edge set, a connection type identifier (such as "site-grid connection", "satellite-grid connection") is assigned to each edge to generate a typed edge set; a unique node identifier is assigned to each node in the typed node set and the type attribute is recorded to generate an identifiable node list, and each edge in the typed edge set is converted into a tuple composed of a source node identifier and a target node identifier, while recording the connection type attribute to generate a typed edge list; using a graph data structure, all nodes in the identifiable node list are inserted as vertices into the graph data structure, and all edges in the typed edge list are inserted as connections into the graph data structure to construct a basic graph structure containing all vertices and edges, generating the original heterogeneous graph framework, and loading the corresponding attribute data (such as background field data, field observed meteorological element values, and remote sensing inverted physical quantity values) for each node in the original heterogeneous graph framework to generate a multimodal meteorological spatiotemporal heterogeneous map.

[0067] S3. Deep feature extraction and fusion processing of multimodal meteorological spatiotemporal heterogeneous maps are performed using graph neural networks to obtain high-dimensional feature fusion vectors. Preliminary residual estimates are then calculated based on the high-dimensional feature fusion vectors to obtain preliminary residual field data.

[0068] The graph neural network is used to initialize the features of nodes of various types in the multimodal meteorological spatiotemporal heterogeneous graph, generating the node initial feature matrix;

[0069] Furthermore, attribute data of all grid nodes, station nodes, and satellite nodes are extracted from the multimodal meteorological spatiotemporal heterogeneous map to generate a node attribute dataset. Based on the node attribute dataset, the input and output dimensions of the fully connected layer are set for each node type (grid, station, and satellite) according to the corresponding attribute dimensions, generating type-specific fully connected layer configuration parameters. Based on the type-specific fully connected layer configuration parameters, the Xavier initialization method is used to initialize the weight matrix and bias vector for each fully connected layer, generating a type-specific set of initialized weights and biases. For each node in the node attribute dataset, the corresponding fully connected layer is selected according to the node type. At the same time, the attribute data corresponding to each node is used as input to perform a linear transformation and obtain the linear transformation results. A nonlinear activation function (such as the ReLU function) is applied to the linear transformation results to generate activated feature values. The activated feature values ​​of each node are combined into a feature vector and arranged in order of node identifier to generate the initial feature matrix of the node.

[0070] Based on the initial feature matrix of the nodes, the connection weights between different types of nodes are calculated, and a weighted adjacency matrix is ​​generated.

[0071] Furthermore, feature vectors of all nodes are extracted from the initial feature matrix to generate a set of node feature vectors. Source and target node index pairs of all connections are extracted from the set of multi-connection edges to generate a list of node pairs for weight calculation. Based on this list, source and target node feature vectors are extracted from the node feature vector set, and each pair is concatenated and input into the attention mechanism computation layer. Attention scores are calculated using a fully connected layer and the LeakyReLU activation function, and then normalized to generate a standardized attention weight set. Based on the standardized attention weight set and the node index information in the multi-connection edge set, each weight value is filled into the elements corresponding to the source and target node positions in the adjacency matrix to generate a weighted adjacency matrix.

[0072] The expression for calculating feature similarity scores is:

[0073] ;

[0074] in, It is a node With nodes The unnormalized attention score between them; It is the source node eigenvectors; The target node eigenvectors; This refers to the vector concatenation operation, which means concatenating the source node... Feature vectors and target nodes The feature vectors are concatenated along the feature dimension; It is a trainable weight matrix used to map the concatenated high-dimensional features to a new space; It is a trainable bias vector that, in conjunction with a trainable weight matrix, provides an offset in linear transformations; It is a non-linear activation function; It is a trainable attention vector used to map activated features to a scalar score; This represents the transpose operation, used to transpose the attention vector. Convert a column vector to a row vector;

[0075] A weighted adjacency matrix is ​​used to perform message passing and feature aggregation across node types, generating an updated node feature matrix.

[0076] Furthermore, the source node indices and corresponding attention weights of each target node are extracted from the weighted adjacency matrix to generate a node connection weight list. Based on the source node indices in the node connection weight list, feature vectors of all source nodes are extracted from the initial feature matrix of the node to generate a neighbor feature set. Using the attention weights in the node connection weight list, the feature vectors in the neighbor feature set are weighted and summed to generate an aggregated neighbor feature vector. The aggregated neighbor feature vector is then concatenated with the feature vector of the target node itself to generate a fused feature vector. The fused feature vector is input into a fully connected layer for linear transformation and nonlinear activation processing to generate a node feature update vector. All node feature update vectors are then recombined according to the node order to generate an updated node feature matrix.

[0077] Based on the updated node feature matrix, feature propagation and fusion are performed iteratively through a multi-layer graph neural network to generate a high-dimensional feature fusion vector.

[0078] Furthermore, the updated node feature matrix is ​​used as the input to the initial layer node feature matrix. Based on the weighted adjacency matrix and the updated node feature matrix, attention weights are applied to the features of all neighboring nodes of each target node to generate the aggregated feature matrix of the current layer. The aggregated feature matrix of the current layer is then linearly transformed through a trainable fully connected layer and nonlinearly processed using the ReLU activation function to generate the node feature matrix of the next layer. The node feature matrix of the next layer is used as the new node feature matrix of the current layer, and the weighted aggregation, linear transformation, and nonlinear activation processing are repeated to complete the iterative propagation of the preset number of layers. The node feature matrix generated by the last iteration is output as a high-dimensional feature fusion vector.

[0079] It should be noted that the preset number of layers is an empirical value determined through historical data verification and network performance optimization experiments. It is usually set to 2 to 6 layers based on the spatial correlation attenuation distance of meteorological elements and the message transmission efficiency of graph neural networks.

[0080] A linear transformation is performed on the high-dimensional feature fusion vector to calculate the initial residual scalar value for each grid point;

[0081] Furthermore, feature vectors corresponding to each grid point are extracted from the high-dimensional feature fusion vector to generate a grid point feature vector set. Based on the dimensionality of the grid point feature vector set, the input dimension of the fully connected layer is set, and the output dimension is 1, to obtain the linear transformation layer structure parameters. According to the linear transformation layer structure parameters, the Xavier initialization method is used to generate a weight matrix and a zero-initialized bias vector. Using the weight matrix and the zero-initialized bias vector, a linear transformation calculation is performed on each feature vector in the grid point feature vector set. The feature vector of each grid point is multiplied by the weight vector and combined with the bias vector to obtain the linear output value of each grid point. The linear output value of each grid point is then subjected to range constraints (e.g., limited to the range [-1, 1]) to generate the preliminary residual scalar value for each grid point.

[0082] The geographic coordinates and elevation data of each grid point are extracted from the preprocessed raw space meteorological data, and feature combination is performed with the preliminary residual scalar values ​​of each grid point to generate a composite feature vector set.

[0083] Furthermore, geographic coordinates and elevation data of each grid point are extracted from the preprocessed raw spatial meteorological data to generate a list of grid point geographic coordinates and elevation data. The latitude and longitude coordinates and elevation data in the list of grid point geographic coordinates and elevation data are matched and aligned with the corresponding preliminary residual scalar values ​​in the preliminary residual scalar values ​​of each grid point according to their spatial location. The matched data (longitude, latitude, elevation data, and preliminary residual scalar values) of each grid point are standardized to eliminate dimensional differences. The standardized longitude, latitude, elevation, and preliminary residual scalar values ​​of each grid point are concatenated in sequence to form a multidimensional feature vector. The multidimensional feature vectors of all grid points are arranged according to the spatial grid order to generate a composite feature vector set.

[0084] The composite feature vector set is regularly reorganized according to the spatial grid position and spatially aligned with the background field data to obtain preliminary residual field data;

[0085] Furthermore, the latitude and longitude coordinates and composite feature values ​​of each grid point are extracted from the composite feature vector set to generate a grid point coordinate feature list. Based on the grid point coordinate feature list, the irregularly distributed composite feature values ​​are mapped to regular latitude and longitude grid points using the Kriging spatial interpolation algorithm to generate a gridded feature field. Grid data with the same spatial range and resolution are extracted from the background field data to generate a background field grid. The latitude and longitude coordinates and resolutions of the gridded feature field and the background field grid are compared to verify spatial alignment: if latitude and longitude coordinate deviations are found, the gridded feature field is resampled using bilinear interpolation; if resolution inconsistencies are found, the gridded feature field is downsampled using the nearest neighbor method to ensure complete alignment between the gridded feature field and the background field grid. The aligned gridded feature field is then output as the initial residual field data.

[0086] S4. Set the constraints of the meteorological physical equations, combine the preliminary residual field data and grid geographic coordinates, construct the physical information neural network framework and initialize the network parameters to form the physical information neural network.

[0087] Extract the initial residual scalar values ​​of each grid point from the initial residual field data, and combine them with the corresponding grid point geographic coordinates to generate an input feature vector set;

[0088] Furthermore, the preliminary residual scalar values ​​of each grid point are extracted from the preliminary residual field data to generate a list of preliminary residual scalar values, and the latitude and longitude coordinates of each grid point are extracted from the grid point geographic coordinate data to generate a grid point coordinate list. The list of preliminary residual scalar values ​​and the list of grid point coordinates are matched and aligned according to the grid point index to generate a grid point data matching list. Based on the grid point data matching list, the latitude and longitude coordinates of each grid point and the corresponding preliminary residual scalar value are concatenated into a multi-dimensional feature vector, and integrated to generate an input feature vector set.

[0089] The grid geographic coordinates of observation stations and the corresponding ground meteorological and upper-air sounding station observation data are extracted from the preprocessed raw space meteorological data to generate the training target dataset.

[0090] Furthermore, from the preprocessed raw space meteorological data, all grid locations with observation stations are selected, generating a grid index list of observation stations. Based on this list, the latitude and longitude coordinates of the corresponding grid points are extracted from the grid geographic coordinate data to generate a list of observation station coordinates. Temperature, air pressure, and humidity of the corresponding stations are extracted from ground meteorological station observation data as ground observation element values ​​to generate a list of ground observation data. Temperature, humidity, and wind speed of the corresponding stations are extracted from upper-air sounding station observation data as upper-air profile observation element values ​​to generate an upper-air observation data list. The observation station coordinate list, ground observation data list, and upper-air observation data list are matched and aligned according to the station index to generate a matched observation data list. Based on this matched observation data list, the latitude and longitude coordinates, ground observation element values, and upper-air observation element values ​​of each station are combined into a training sample vector to generate a training sample vector list. All training sample vectors are arranged in station order, and sample index identifiers are added to generate the training target dataset.

[0091] Thermodynamic equations and fluid dynamics equations are defined based on the conservation laws of meteorological physics, and then discretized to generate a discrete calculation framework for the physical equations.

[0092] Furthermore, based on the meteorological laws of energy and momentum conservation, thermodynamic and fluid dynamic control equations are established, and these equations are combined to construct a set of meteorological physical control equations. The partial differential terms in this set are discretized using the finite difference method: a forward difference scheme is used for the time derivative term, and a central difference scheme is used for the spatial derivative term, converting continuous differential operators into discrete difference operators and generating a set of discrete difference operators containing all difference calculation rules. The parameters required for discrete difference are calculated based on grid point geographic coordinate data: according to the longitude of adjacent grid points... The latitudinal coordinates are transformed using the Lambert projection transformation formula to convert the longitude difference between adjacent grid points into the actual east-west grid spacing, and the latitude difference between adjacent grid points into the north-south grid spacing. Based on the grid point latitude, the Coriolis force parameters are calculated using the Coriolis force calculation formula. The Sobel operator is used to calculate the terrain gradient using elevation data, and the slope parameters are calculated based on the terrain gradient, generating a difference parameter table containing grid spacing, Coriolis force parameters, and terrain parameters. The discrete difference operator set is integrated with the difference parameter table, and boundary condition processing schemes are added (such as using the zero normal gradient condition for solid wall boundaries) to generate a discrete calculation framework for the physical equations.

[0093] The expression for calculating the Coriolis force parameters is:

[0094] ;

[0095] in, It is the Coriolis force parameter, which characterizes the intensity of the deflection effect of the Earth's rotation on fluid motion (such as the atmosphere and ocean). It is the angular velocity of Earth's rotation, a physical constant representing the angular rate of Earth's rotation; It is grid point latitude;

[0096] The expression for calculating the slope parameter is:

[0097] ;

[0098] in, It's the slope; It is the horizontal (east-west) terrain gradient; It is the vertical (north-south) terrain gradient;

[0099] It should be noted that the thermodynamic and fluid dynamic control equations are established as follows: Based on the meteorological law of conservation of energy, a mathematical expression is defined as the rate of change of internal energy of a gas parcel equal to the sum of the non-adiabatic heating rate and the adiabatic compression and expansion power. Based on this mathematical expression, a temperature advection term is introduced to describe the horizontal heat transport effect, a radiation heating term quantifies the contribution of net radiation flux, and a non-adiabatic process term includes specific parameters of latent heat release and sensible heat transport. Partial differential operators are used to integrate the temperature advection term, radiation heating term, and non-adiabatic process term into the expression for the rate of change of internal energy, deriving the thermodynamic control equations containing spatiotemporal derivatives. Simultaneously, based on the law of conservation of momentum, a mathematical expression for the force balance relationship is defined. Based on this expression, a pressure gradient force term is introduced to represent the force driven by pressure differences, a Coriolis force term reflects the deflection effect caused by the Earth's rotation, and a turbulent viscous force term characterizes the momentum dissipation process. Through vector operations, the pressure gradient force term, Coriolis force term, and turbulent viscous force term are combined into the force balance relationship mathematical expression, establishing a vector-form fluid dynamic control equation.

[0100] It should be noted that the meteorological law of conservation of energy states that the rate of change of internal energy of an air parcel is equal to the sum of the non-adiabatic heating rate and the adiabatic compression / expansion power. It characterizes the quantitative relationship between temperature change and the energy exchange process of radiation, latent heat, and vertical motion, and is used to constrain the evolution of the temperature field in numerical weather prediction and data assimilation processes to ensure strict adherence to the principle of energy conservation. The law of conservation of momentum states that the rate of change of momentum of an air parcel is equal to the resultant force of the pressure gradient force, the Coriolis force, and the turbulent viscosity force. It characterizes the quantitative balance relationship between wind field changes and pressure distribution, the deflection effect of Earth's rotation, and the turbulent dissipation process. It is used to constrain the evolution of the wind field in numerical weather prediction and data assimilation processes to ensure strict adherence to the principle of momentum conservation.

[0101] Based on the dimensional features of the input feature vector set, the structure of the physical information neural network is defined, and a network architecture computation graph is generated.

[0102] Furthermore, the length and number of samples of the feature vectors in the input feature vector set are statistically analyzed to generate input dimension parameters. Based on these parameters, the input layer structure of the physical information neural network is set, ensuring that the number of neurons in the input layer matches the length of the feature vectors, thus generating input layer structure parameters. Based on the dimensionality of the input feature vector set and the requirements of the meteorological grid correction task, the number of hidden layers (e.g., 3-5 layers) and the number of neurons per layer (e.g., 128, 256, 512) are set according to historical experimental verification, generating hidden layer structure parameters. The activation function type is selected based on the smoothness and nonlinearity characteristics of the meteorological physical control equations (e.g., tanh for smooth problems, ReLU for piecewise linear problems); the activation function parameters are then configured. (For example, the negative slope of LeakyReLU is set to 0.01), and the output target (such as ground weather station observation data and upper-air sounding station observation data) is extracted from the training target dataset. The number of neurons in the output layer is determined according to the dimension of the output target (e.g., 1 for single-factor output and n for multi-factor output). At the same time, the activation function of the output layer is selected (linear activation is used for regression tasks), and the output layer structure parameters are generated. The input layer structure parameters, hidden layer structure parameters, activation function parameters and output layer structure parameters are integrated to construct a network layer structure configuration table. According to the network layer structure configuration table, the inter-layer connection relationship is defined using the API of the physical information neural network framework (such as TensorFlow or PyTorch), and the network architecture computation graph is generated.

[0103] The network architecture computation graph is used to perform forward propagation computation on the input feature vector set to obtain the network prediction value, and the mean square error is calculated in combination with the training target dataset to generate the data fitting loss term;

[0104] Furthermore, the set of input feature vectors is input into the input layer defined by the network architecture computation graph, where linear transformations and activation functions are performed to generate the first layer of hidden features. These first-layer hidden features are then input into subsequent hidden layers defined by the network architecture computation graph, and linear transformations and activation functions are performed sequentially for each hidden layer to generate the final hidden features. The final hidden features are then input into the output layer defined by the network architecture computation graph, where linear transformations are performed to obtain network predictions, which are then integrated to generate a set of network predictions. Real observations (such as ground weather station observations and upper-air sounding station observations) corresponding to the set of input feature vectors are extracted from the training target dataset to generate a set of real observations. Based on the set of network predictions and the set of real observations, the squared error between each network prediction and its corresponding real observation is calculated to generate a set of squared errors. The average of all error values ​​in the set of squared errors is then calculated to obtain the data fitting loss term.

[0105] Based on the set of input feature vectors, the residuals are calculated using the physical equation discrete computation framework to obtain the physical regularization loss term, which is then fused with the data fitting loss term to generate a joint optimization loss function.

[0106] Furthermore, the latitude and longitude coordinates, elevation data, and preliminary residual field data of each grid point are extracted from the input feature vector set to generate a grid point physical feature vector set. Then, using the discrete difference operator set in the physical equation discrete calculation framework, partial differential operations are performed on the grid point physical feature vector set to calculate the difference between the theoretical and actual values ​​of the meteorological physical control equations, generating a physical residual set. The squared mean of all residual values ​​in the physical residual set is calculated to generate a physical regularization loss term. Based on historical experimental verification results (such as the importance of physical constraints under different weather types), physical loss weight coefficients and data loss weight coefficients are set. Finally, the physical regularization loss term and the data fitting loss term are weighted and summed according to the physical loss weight coefficients and data loss weight coefficients to generate a joint optimization loss function.

[0107] All weight parameters are initialized based on the network architecture computation graph, and the weight parameters are reassigned to generate an initial network parameter set;

[0108] Furthermore, the dimensionality information of all weight matrices and bias vectors requiring initialization is extracted from the network architecture computation graph to generate a parameter dimension list. Based on this list, the Xavier initialization method is selected as the default initialization strategy, and an initialization method configuration is generated. According to the initialization method configuration and the input and output dimensions in the parameter dimension list, the reciprocal of the square root of the sum of the input and output dimensions is calculated as the standard deviation scaling factor for random number generation. This scaling factor is then used to scale the random number matrix generated by the standard Gaussian distribution, generating initial values ​​for the weight matrix that meet the variance requirements. For the bias vectors... The initialization strategy is zero-initialization, directly generating a vector structure with all zero values ​​to obtain the initial value of the bias vector. The initial values ​​of each weight matrix and bias vector are assigned to the corresponding weights and bias parameters in the network architecture computation graph to complete the parameter initialization and generate the initialized network parameters. The values ​​of the initialized network parameters are verified to be within a reasonable range (e.g., checking for gradient vanishing or exploding risks). If the verification fails, the initialization method is adjusted (e.g., using He initialization) and the initial values ​​and assignment process of each weight matrix and bias vector are repeated until a qualified set of initialized network parameters is generated. If the verification passes, the initialized network parameter set is output.

[0109] It should be noted that the reasonable range is set based on gradient stability theory and Xavier initialization method, and the value range is [-0.1, 0.1]. -0.1 is to prevent the gradient from vanishing due to the weight value being too small, and 0.1 is to prevent the gradient from exploding due to the weight value being too large.

[0110] By integrating the joint optimization loss function, the set of input feature vectors, and the initialization set of network parameters, a physical information neural network is constructed.

[0111] Furthermore, each weight matrix and bias vector in the initialized network parameter set is loaded into the network architecture computation graph to generate the parameter-initialized network architecture. The input feature vector set is integrated into the parameter-initialized network architecture as the network input data source to generate the data-ready network architecture. The joint optimization loss function is bound to the data-ready network architecture as the optimization objective function for network training to generate the loss ensemble network architecture. Based on the loss ensemble network architecture, the Adam optimizer is selected as the default optimizer type according to historical experimental verification results, and the optimal initial learning rate is determined by learning rate grid search (e.g., from 0.1 to 1e-5). The default optimizer type and the optimal initial learning rate are bound to the loss ensemble network architecture to generate the parameter configuration network architecture. The optimal initial learning rate of the parameter configuration network architecture is verified to be within the effective range (e.g., whether the optimal initial learning rate avoids gradient explosion or vanishing) to obtain the physical information neural network.

[0112] It should be noted that the effective range is set based on the deep learning gradient stabilization theory and historical experimental data, and the value range is 1e-5 to 0.1. 1e-5 is to prevent gradient vanishing and ensure that the parameters of the physical information neural network can be updated effectively at the lowest acceptable speed. 0.1 is to avoid gradient explosion and prevent the loss function from diverging during the optimization process.

[0113] It should be noted that training the physical information neural network is specifically as follows: the input feature vector set is divided into a training set and a validation set (e.g., randomly divided at an 80% / 20% ratio). The training set is injected into the physical information neural network for forward propagation. The mean square error between the network output value and the true observation value in the training target dataset is calculated to obtain the data fitting loss term. The physical regularization loss term is obtained by calculating the residual between the network output value and the meteorological physical control equation set using the physical equation discrete calculation framework. The data fitting loss term and the physical regularization loss term are weighted and fused according to preset weights to obtain the weighted total loss value. The Adam optimizer is used to minimize the joint loss with an initial learning rate (e.g., 0.001). The network weights are iteratively updated through backpropagation. After each iteration, the validation loss is calculated using the validation set. If the validation loss does not decrease continuously, early stopping is triggered. The network parameters with the minimum validation loss are saved, and the trained physical information neural network is output.

[0114] It should be noted that the preset weights refer to the balance coefficients of the physical regularization loss term and the data fitting loss term in the joint loss function during the training of the physical information neural network. Through historical experiments, these are hyperparameters that are preset according to the sensitivity of different meteorological elements and the strength of physical constraints, and are used to quantify the contribution ratio of physical laws and observational data to the optimization objective.

[0115] S5. Input the preliminary residual field data and grid point geographic coordinates into the physical information neural network, perform forward propagation calculations, and obtain high-precision space meteorological correction analysis field data.

[0116] The preliminary residual field data and grid point geographic coordinates are spliced ​​together according to spatial location to generate a composite input feature vector;

[0117] Furthermore, the preliminary residual scalar value of each grid point is extracted from the preliminary residual field data to generate a preliminary residual value list; the latitude and longitude coordinates of each grid point are extracted from the grid point geographic coordinate data to generate a grid point coordinate list; based on the spatial location index of the grid points, the preliminary residual value list and the grid point coordinate list are matched and aligned to generate a matched grid point data list, and for each grid point in the matched grid point data list, the preliminary residual scalar value and latitude and longitude coordinates are concatenated into a multi-dimensional feature vector to generate a composite feature vector for a single grid point; the composite feature vectors of all grid points are arranged according to the spatial grid order to generate a composite input feature vector;

[0118] The composite input feature vector is input into the physical information neural network, the output of the meteorological element correction values ​​of each grid point is generated, and the meteorological element correction tensor is generated by integrating them.

[0119] Furthermore, the composite input feature vector is input into the trained physical information neural network. The input layer receives the composite input feature vector and performs a linear transformation to obtain the linear transformation result. The hidden layer then sequentially performs linear calculations and ReLU activation function processing on the linear transformation result to obtain the hidden layer activation values. The output layer performs a linear transformation on the hidden layer activation values ​​to generate the original output values. The original output values ​​are then de-standardized to restore them to the actual meteorological element dimensions, generating standardized correction values. The standardized correction values ​​of all grid points are arranged in spatial grid index order to generate a two-dimensional correction matrix. This two-dimensional correction matrix is ​​then converted into a tensor format, and batch and feature dimensions are added to generate a three-dimensional meteorological element correction tensor. The consistency between the dimensions of the three-dimensional meteorological element correction tensor and the grid points of the target area is verified to ensure spatial coverage integrity. The meteorological element correction tensor is then output.

[0120] It should be noted that the target area grid is a regularized latitude and longitude coordinate matrix generated during the preprocessing stage based on a preset latitude and longitude range (such as 100° E, -120° N, 20° N, -40° N) and spatial resolution (such as 0.1° × 0.1°). It is obtained by dividing the geographic space at equal intervals and assigning unique grid identifiers, and serves as a unified benchmark for spatial alignment of all meteorological data.

[0121] The meteorological element correction tensor is fused with the background field data to obtain high-precision spatial meteorological correction analysis field data;

[0122] Furthermore, based on the preset target meteorological element type, the corresponding standardized correction values ​​are extracted from the meteorological element correction tensor to generate a standardized correction value matrix, and grid data of the same meteorological element type are extracted from the background field data to generate a background field matrix; the spatial dimensions of the standardized correction value matrix and the background field matrix are compared: if the spatial dimensions do not match, the standardized correction value matrix is ​​resampled to the background field grid resolution using bilinear interpolation; the standardized correction value matrix and the background field matrix with matched spatial dimensions are fused point-by-point by addition to generate a preliminary analysis field matrix, and the preliminary analysis field matrix is ​​then subjected to Gaussian filtering for spatial smoothing to generate a smoothed analysis field matrix; predefined material... A threshold table (including temperature, humidity, air pressure, and wind speed thresholds) is used to verify the meteorological element values ​​at each grid point in the smoothing analysis field matrix. If the meteorological element values ​​(temperature, humidity, air pressure, and wind speed) of a grid point all fall within the corresponding threshold, the grid point is considered to have passed verification. If any meteorological element value of a grid point is not within the threshold, the grid point is considered to have failed verification, and outliers in the grid point's meteorological element values ​​are corrected using neighboring grid point interpolation. All verified and corrected meteorological element values ​​in the smoothing analysis field matrix are converted to GRIB2 format and metadata (such as element name, unit, and spatial dimension) is added to generate high-precision spatial meteorological correction analysis field data.

[0123] It should be noted that the preset target meteorological element types are specific elements pre-selected based on meteorological operational needs (such as temperature forecasting, humidity analysis, or wind field correction), and are set based on elements with significant persistent deviations in historical data or operational priorities, including core meteorological variables such as temperature, humidity, air pressure, wind speed, and wind direction; the predefined physical threshold table is a reasonable range comparison table for each element based on the statistical extreme values ​​of historical meteorological observation data and basic meteorological principles (such as humidity cannot be negative and wind speed cannot be negative).

[0124] It should be noted that the temperature threshold is set based on extreme temperature data from global historical meteorological observation records, with a range of -100 to 60°C. -100°C ensures data integrity under extreme polar low-temperature environments, preventing data loss due to sensor measurement errors. 60°C avoids data anomalies under ultra-high-temperature environments in desert regions, preventing non-physical high-temperature values ​​caused by thermal radiation errors. The humidity threshold is set based on the thermodynamic principle of water vapor saturation (air water content cannot be lower than absolute dryness or exceed saturation), with a range of 0 to 100%. 0% ensures the physical authenticity of absolute dryness, preventing negative humidity values ​​from violating the law of conservation of atmospheric humidity. 100% prevents the overflow of supersaturated water vapor values, avoiding the generation of false data that violates thermodynamic equilibrium conditions. The air pressure threshold is based on global... The values ​​are set based on historical extreme values ​​of sea level pressure (such as the lowest pressure of typhoons at 870 hPa and the highest at 1083.8 hPa), with a range of 800 to 1100 hPa. The 800 hPa value covers the extreme low-pressure scenarios of super typhoon systems, preventing misjudgments of instrument malfunctions due to excessively low pressure. The 1100 hPa value accommodates the reasonable performance of extreme high-pressure systems, avoiding incorrect filtering of high-pressure peak data. The wind speed threshold is set based on the maximum wind speed record of the Earth's atmosphere (such as 135 m / s for tornadoes) with an extended safety buffer, with a range of 0 to 150 m / s. The 0 m / s value adheres to the physical definition of wind speed as a non-negative scalar, preventing wind direction data from being incorrectly converted into negative wind speed values. The 150 m / s value ensures the integrity of wind speed records for extreme weather such as tornadoes, avoiding the truncation of true data from strong convective systems.

[0125] This embodiment also provides a computer device applicable to the grid correction method for space meteorological data, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the grid correction method for space meteorological data as proposed in the above embodiment.

[0126] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0127] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the grid correction method for space meteorological data as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0128] In summary, this invention achieves spatiotemporal correlation modeling and nonlinear feature learning of multi-source heterogeneous data by constructing a multimodal meteorological spatiotemporal heterogeneous map and using graph neural networks for deep feature extraction and fusion processing, thereby enhancing data representation capabilities and improving the accuracy of preliminary residual estimation. Furthermore, by constructing a physical information neural network framework, it achieves a deep integration of physical conservation laws and data-driven methods, constraining the network output to conform to thermodynamic and fluid dynamic laws, and generating physically consistent and highly reliable space meteorological correction analysis field data.

[0129] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A grid correction method for space meteorological data, characterized in that: include, Collect raw space meteorological data and perform preprocessing; Grid geographic coordinates, observation stations, and satellite pixels are extracted from the preprocessed raw space meteorological data and defined as nodes to construct a multimodal meteorological spatiotemporal heterogeneous map; The graph neural network is used to perform deep feature extraction and fusion processing on multimodal meteorological spatiotemporal heterogeneous maps to obtain high-dimensional feature fusion vectors. Preliminary residual estimates are then calculated based on the high-dimensional feature fusion vectors to obtain preliminary residual field data. By setting constraints for the meteorological physics equations, combining preliminary residual field data and gridded geographic coordinates, constructing a physical information neural network framework, and initializing the network parameters, the physical information neural network is formed. The steps are as follows. Extract the initial residual scalar values ​​of each grid point from the initial residual field data, and combine them with the corresponding grid point geographic coordinates to generate an input feature vector set; The grid geographic coordinates of observation stations and the corresponding ground meteorological and upper-air sounding station observation data are extracted from the preprocessed raw space meteorological data to generate the training target dataset. Thermodynamic equations and fluid dynamics equations are defined based on the conservation laws of meteorological physics, and then discretized to generate a discrete calculation framework for the physical equations. Based on the dimensional features of the input feature vector set, the structure of the physical information neural network is defined, and a network architecture computation graph is generated. The network architecture computation graph is used to perform forward propagation computation on the input feature vector set to obtain the network prediction value, and the mean square error is calculated in combination with the training target dataset to generate the data fitting loss term; Based on the set of input feature vectors, the residuals are calculated using the physical equation discrete computation framework to obtain the physical regularization loss term, which is then fused with the data fitting loss term to generate a joint optimization loss function. All weight parameters are initialized based on the network architecture computation graph, and the weight parameters are reassigned to generate an initial network parameter set; By integrating the joint optimization loss function, the set of input feature vectors, and the initialization set of network parameters, a physical information neural network is constructed. The preliminary residual field data and grid point geographic coordinates are input into the physical information neural network for forward propagation calculation to obtain high-precision space meteorological correction analysis field data.

2. The grid correction method for space meteorological data as described in claim 1, characterized in that: The raw space meteorological data includes observation data from ground meteorological stations, upper-air sounding stations, meteorological satellites, and weather radars, as well as background field and elevation data; The preprocessing includes spatiotemporal alignment interpolation, normalization, and missing value imputation.

3. The grid correction method for space meteorological data as described in claim 1, characterized in that: The steps for constructing the multimodal meteorological spatiotemporal heterogeneous map are as follows: Extract gridded geographic coordinate data, observation station data, and satellite remote sensing data from the preprocessed raw space meteorological data and create nodes to generate gridded node sets, station node sets, and satellite node sets. Assign background field and elevation data to the grid node set, assign observation data from ground meteorological stations and upper-air sounding stations to the station node set, assign meteorological satellite observation data to the satellite node set, and obtain attribute grid node set, attribute station node set, and attribute satellite node set; Calculate the spatial distance between the set of attribute grid nodes, the set of attribute site nodes, and the set of attribute satellite nodes, establish spatial proximity connections, and generate a set of multi-connection edges; By integrating the set of attribute grid nodes, the set of attribute station nodes, the set of attribute satellite nodes, and the set of multi-connection edges, a multimodal meteorological spatiotemporal heterogeneous map is constructed.

4. The grid correction method for space meteorological data as described in claim 1, characterized in that: The steps for obtaining the high-dimensional feature fusion vector are as follows: The graph neural network is used to initialize the features of nodes of various types in the multimodal meteorological spatiotemporal heterogeneous graph, generating the node initial feature matrix; Based on the initial feature matrix of the nodes, the connection weights between different types of nodes are calculated, and a weighted adjacency matrix is ​​generated. A weighted adjacency matrix is ​​used to perform message passing and feature aggregation across node types, generating an updated node feature matrix. Based on the updated node feature matrix, feature propagation and fusion are performed iteratively through a multi-layer graph neural network to generate a high-dimensional feature fusion vector.

5. The grid correction method for space meteorological data as described in claim 1, characterized in that: The steps for obtaining preliminary residual field data are as follows: A linear transformation is performed on the high-dimensional feature fusion vector to calculate the initial residual scalar value for each grid point; The geographic coordinates and elevation data of each grid point are extracted from the preprocessed raw space meteorological data, and feature combination is performed with the preliminary residual scalar values ​​of each grid point to generate a composite feature vector set. The composite feature vector set is regularly reorganized according to the spatial grid position and spatially aligned with the background field data to obtain preliminary residual field data.

6. The grid correction method for space meteorological data as described in claim 1, characterized in that: The steps for inputting the preliminary residual field data and grid point geographic coordinates into the physical information neural network for forward propagation calculation are as follows: The preliminary residual field data and grid point geographic coordinates are spliced ​​together according to spatial location to generate a composite input feature vector; The composite input feature vector is input into the physical information neural network, the output of the meteorological element correction values ​​of each grid point is generated, and the meteorological element correction tensor is generated by integrating them.

7. The grid correction method for space meteorological data as described in claim 1, characterized in that: The high-precision spatial meteorological correction analysis field data is obtained by fusing the meteorological element correction tensor with the background field data.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the grid correction method for space meteorological data according to any one of claims 1 to 7.

9. 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 of the grid correction method for space meteorological data according to any one of claims 1 to 7.

Images