Radar echo extrapolation processing method and system based on hypergraph enhanced diffusion model

Abstract

The present invention provides a radar echo extrapolation processing method and system based on a hypergraph-enhanced diffusion model. The method comprises obtaining a radar echo raw data sequence for consecutive time periods, constructing a hypergraph structure on the radar echo raw data sequence to obtain a radar echo hypergraph structure, integrating the radar echo hypergraph structure into a preset diffusion model framework to obtain a hypergraph-enhanced diffusion model, and performing extrapolation on the radar echo raw data sequence based on the hypergraph-enhanced diffusion model to obtain a radar echo extrapolated data sequence for a future time period. The radar echo extrapolated data sequence includes multiple extrapolated data frames that are chronologically continuous with the radar echo raw data sequence. The present invention can guide the spatiotemporal evolution direction of echo features by embedding multivariate correlation information in hypergraph features, avoiding the extrapolation trend deviation caused by traditional diffusion models relying solely on the statistical distribution of data, and ensuring that the generated radar echo extrapolated data sequence is more consistent with the actual evolution characteristics of radar echoes in terms of spatiotemporal distribution.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a radar echo extrapolation processing method and system based on a hypergraph enhanced diffusion model. Background Art

[0002] In meteorological business, radar echo data is an important basis for short-term and near-term weather forecasts. Radar echo extrapolation processing refers to a technology that uses specific technical means to predict the spatial distribution and evolution trend of radar echoes in the future based on historical and current radar echo data sequences. The radar echo extrapolation methods in the existing technology include traditional methods such as optical flow method and cross-correlation method, as well as methods based on deep learning. Among them, traditional methods mainly extrapolate by capturing the translation, expansion and other motion characteristics of echoes, but it is difficult to accurately characterize the nonlinear evolution process and complex generation and disappearance changes of echoes. Deep learning-based methods such as convolutional neural networks and recurrent neural networks can accurately describe the nonlinear evolution process and complex generation and disappearance changes of echoes. Although the extrapolation accuracy has been improved to a certain extent, it is still limited in processing the multi-region coordinated change relationship in radar echo data. It is difficult to effectively capture the multivariate correlation characteristics between three or more spatial regional units, resulting in the model being unable to fully pass through the complex spatial correlation information in the echo data. When generative models such as the grid structure input diffusion model are directly used, the model mainly relies on the statistical distribution characteristics of the data for extrapolation, and lacks explicit modeling and passing of the structured correlation relationship between echo regions, which makes the extrapolation results prone to deviations in the spatial morphological evolution and intensity change trend of the echo, affecting the accuracy and reliability of radar echo extrapolation. Summary of the Invention

[0003] The present invention provides a radar echo extrapolation processing method and system based on a hypergraph enhanced diffusion model.

[0004] In a first aspect, an embodiment of the present invention provides a radar echo extrapolation processing method based on a hypergraph enhanced diffusion model, comprising: obtaining a radar echo raw data sequence of a continuous time period, wherein the radar echo raw data sequence is composed of a plurality of radar echo data frames arranged in chronological order, and each radar echo data frame contains spatially distributed echo intensity information and a corresponding acquisition time identifier; constructing a hypergraph structure for the radar echo raw data sequence to obtain a radar echo hypergraph structure, wherein the radar echo hypergraph structure comprises a plurality of hypervertices and hyperedges connecting the hypervertices, and the hypervertices are used to characterize the spatial regions in the radar echo data. domain unit, the hyperedge is used to characterize the multivariate correlation relationship between different spatial area units; the radar echo hypergraph structure is integrated into a preset diffusion model framework to obtain a hypergraph enhanced diffusion model, the hypergraph enhanced diffusion model processes the radar echo hypergraph structure through a hypergraph feature extraction module to generate a hypergraph embedding feature containing multivariate correlation information; based on the hypergraph enhanced diffusion model, the radar echo original data sequence is extrapolated to obtain a radar echo extrapolated data sequence for a future time period, the radar echo extrapolated data sequence includes multiple extrapolated data frames that are time-sequentially continuous with the radar echo original data sequence.

[0005] In a second aspect, an embodiment of the present invention provides a computer system, comprising: a memory storing a computer program; and a processor configured to load the computer program to implement the radar echo extrapolation processing method based on the hypergraph enhanced diffusion model as described above.

[0006] The radar echo extrapolation processing method based on the hypergraph enhanced diffusion model provided by the present invention obtains the radar echo original data sequence of continuous time periods and constructs the radar echo hypergraph structure, integrates the hypergraph structure into the diffusion model framework to form a hypergraph enhanced diffusion model, and then performs extrapolation processing to obtain the radar echo extrapolation data sequence of the future time period. By constructing a radar echo hypergraph structure including hypervertices and hyperedges, the hyperedge can effectively capture the coordinated change relationship between multiple spatial area units in the radar echo data through the characterization ability of multi-dimensional correlation relationships, so that the hypergraph embedded features can carry richer correlation information, providing a data basis that is closer to the actual meteorological evolution law for the extrapolation processing, and integrating the radar echo hypergraph structure into the diffusion model framework, generating a hypergraph feature extraction module through the hypergraph feature extraction module. The hypergraph embedding features containing multivariate correlation information enable the diffusion model to enhance the model's ability to learn the complex evolution laws of radar echoes and improve the generation reliability of extrapolated data sequences by combining not only the spatiotemporal statistical characteristics of the original data but also the structural correlation information in the hypergraph embedding features during the iterative sampling process. At the same time, when performing extrapolation processing based on the hypergraph enhanced diffusion model, the hypergraph embedding features are input into the generation module to generate the radar echo extrapolation data sequence for the future time period through multi-step iterative sampling. The spatiotemporal evolution direction of the echo characteristics can be guided by the multivariate correlation information in the hypergraph embedding features, avoiding the extrapolation trend deviation caused by the traditional diffusion model relying solely on the statistical distribution of the data, so that the generated radar echo extrapolation data sequence is more consistent with the actual evolution characteristics of the radar echo in terms of spatiotemporal distribution. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 This is a flowchart of a radar echo extrapolation processing method based on a hypergraph enhanced diffusion model provided by an embodiment of the present invention.

[0008] Figure 2 It is a schematic diagram of the composition of a computer system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0009] See also Figure 1 , Figure 1 A flowchart of a radar echo extrapolation processing method based on a hypergraph enhanced diffusion model provided in an embodiment of the present invention is provided. The method can be executed by a computer system and includes the following steps:

[0010] Step S100: Acquire a radar echo raw data sequence of a continuous period. The radar echo raw data sequence consists of a plurality of radar echo data frames arranged in chronological order. Each radar echo data frame contains spatially distributed echo intensity information and a corresponding acquisition time identifier.

[0011] A radar echo raw data sequence is a collection of radar echo data frames arranged in chronological order. A radar echo data frame represents radar echo information collected at a specific moment. The spatially distributed echo intensity information describes the radar echo intensity at different spatial locations at that moment, reflecting the target object's ability to reflect radar waves. The acquisition time stamp specifies the acquisition time of each data frame, ensuring accurate temporal sequence and time intervals during processing.

[0012] Acquiring a continuous sequence of radar echo raw data can be done in real time using radar equipment. For example, in a weather monitoring scenario, a weather radar is used to monitor targets such as precipitation particles in the atmosphere. The radar continuously collects echo data at certain time intervals, forming continuous radar echo data frames.

[0013] Step S200: constructing a hypergraph structure for the radar echo raw data sequence to obtain a radar echo hypergraph structure. The radar echo hypergraph structure includes multiple hypervertices and hyperedges connecting the hypervertices. The hypervertices are used to represent the spatial area units in the radar echo data, and the hyperedges are used to represent the multivariate association relationships between different spatial area units.

[0014] A hypergraph is an extension of a graph structure. Unlike traditional graphs, hyperedges can connect multiple vertices, allowing for more flexible representation of multi-dimensional relationships. In an embodiment of the present invention, the radar echo hypergraph structure is an abstract representation of spatial area units in radar echo data and their multi-dimensional association relationships. Hypervertices correspond to spatial area units in radar echo data. These spatial area units are obtained by spatially dividing the radar echo data. Each spatial area unit has certain characteristics such as spatial range and echo intensity. Hyperedges connect multiple hypervertices, reflecting the multi-dimensional association relationships between these spatial area units. This association relationship may be formed based on factors such as spatial position and echo intensity changes.

[0015] The purpose of constructing a hypergraph structure for the raw radar echo data sequence is to explore the complex correlations between different spatial regions in the radar echo data, so that this correlation information can be better utilized in the subsequent diffusion model for radar echo extrapolation. Two specific implementation methods are given below.

[0016] As a first implementation, step S200 constructs a hypergraph structure for the radar echo raw data sequence to obtain a radar echo hypergraph structure, which can be specifically implemented as follows:

[0017] Step S201: performing spatial region division on each radar echo data frame in the radar echo raw data sequence to obtain a plurality of spatial region units, each of which corresponds to a continuous spatial sub-region in the radar echo data frame, and the boundary of the spatial region unit is determined by the echo intensity gradient change characteristic.

[0018] Spatial region segmentation involves spatially segmenting radar echo data frames into multiple contiguous subregions, which are referred to as spatial region units. The echo intensity gradient characteristic reflects the spatial rate of change of echo intensity. By analyzing this characteristic, the boundaries of spatial region units can be determined. A large echo intensity gradient indicates a more dramatic change in echo intensity at that location, potentially corresponding to a different meteorological phenomenon or target object. Therefore, this location can be used as the boundary of a spatial region unit.

[0019] Image segmentation methods can be used to partition radar echo data frames into spatial regions. For example, a gradient-based segmentation algorithm first calculates the echo intensity gradient value for each pixel in the radar echo data frame. The pixels are then classified according to a preset gradient threshold. Pixels with gradient values ​​greater than the threshold are considered to be on the boundary of different spatial regions; pixels with gradient values ​​less than the threshold are classified into the same spatial region. For example, suppose a radar echo data frame contains a region of heavy rainfall and a region of light rainfall. By analyzing the echo intensity gradient variation characteristics, these two regions can be divided into different spatial regions.

[0020] Step S202: extracting features from each spatial region unit to obtain spatial region features, where the spatial region features include echo intensity distribution features, shape features, and positional relationship features of the spatial region unit with adjacent spatial region units.

[0021] Spatial region features are a set of characteristics that describe the properties of spatial region units. Echo intensity distribution features reflect the distribution of echo intensity within a spatial region unit, such as average echo intensity and standard deviation. These features can help understand meteorological information such as precipitation intensity in the region. Shape features describe the geometric shape of a spatial region unit, such as area, perimeter, and circularity, helping to determine the morphological characteristics of the region. Positional relationship features with adjacent spatial region units reflect the relative positional relationship of a spatial region unit to other surrounding spatial region units, such as adjacency and distance, and are important for analyzing the mutual influence between different regions.

[0022] There are many methods that can be used to extract features from each spatial area unit. For echo intensity distribution features, the average and standard deviation of the echo intensity of all pixels within the spatial area unit can be calculated. For shape features, the geometric feature calculation method in image processing can be used. For example, by calculating the boundary pixels of the spatial area unit to determine its perimeter and area, and then calculate shape indicators such as circularity. For positional relationship features with adjacent spatial area units, the distance and adjacent relationship with adjacent spatial area units can be calculated by analyzing the center coordinates and boundary information of the spatial area unit. For example, in a radar echo data frame containing multiple precipitation areas, feature extraction is performed on the spatial area unit corresponding to each precipitation area to obtain features such as its echo intensity distribution, shape, and positional relationship with adjacent areas.

[0023] Step S203: each spatial region unit is used as a super vertex in the radar echo hypergraph structure, and the attribute information of the super vertex is represented by the corresponding spatial region feature.

[0024] When constructing the radar echo hypergraph, spatial region units are abstracted as hypervertices, allowing for convenient representation and processing of these spatial region units within the hypergraph. The attributes of these hypervertices are represented by their corresponding spatial region features, which reflect the specific characteristics of each spatial region unit within the hypergraph.

[0025] Converting spatial area units to supervertices can be achieved by establishing a mapping relationship. For example, each spatial area unit is assigned a unique identifier, which serves as the identifier of the supervertice. At the same time, the spatial area characteristics of the spatial area unit are stored in the hypergraph as the attribute information of the supervertice. Assuming that there are 100 spatial area units in a radar echo hypergraph structure, there will be 100 corresponding supervertices. The attribute information of each supervertice includes the echo intensity distribution characteristics and shape characteristics of the corresponding spatial area unit, as well as the positional relationship characteristics with adjacent spatial area units.

[0026] Step S204: Based on radar echo data frames with different time tags in the radar echo raw data sequence, determine the cross-time correlation relationship between super-vertices. The cross-time correlation relationship is determined by the echo intensity change trend and position offset characteristics of the same spatial area unit in continuous time frames.

[0027] Cross-temporal correlations reflect the connections between spatial units at different points in time. The echo intensity trends of the same spatial unit in consecutive time frames can reflect meteorological changes in that area. For example, an increase or decrease in echo intensity may indicate an increase or decrease in precipitation. Position shift characteristics reflect the positional changes of spatial units at different points in time, such as the movement of precipitation areas. By analyzing echo intensity trends and position shift characteristics, cross-temporal correlations between supervertices can be determined.

[0028] A tracking algorithm can be used to determine the cross-temporal correlation between supervertices. For example, a tracking algorithm based on feature matching is used to match and track the features of each spatial area unit in continuous radar echo data frames. For the same spatial area unit, its echo intensity change rate and position offset in different time frames are calculated. If the echo intensity change rate and position offset meet certain conditions, it is considered that there is a cross-temporal correlation between the supervertices in different time frames of the spatial area unit. Suppose that in a radar echo data sequence, a precipitation area gradually moves eastward in several consecutive time frames, and the echo intensity gradually increases. By analyzing the echo intensity change trend and position offset characteristics of the supervertices corresponding to the precipitation area in different time frames, the cross-temporal correlation between these supervertices can be determined.

[0029] Step S205: Based on the cross-time association relationship and the spatial position relationship characteristics between the spatial area units, construct a hyperedge connecting multiple supervertices, each hyperedge connects at least three supervertices to represent the coordinated change relationship between the multi-dimensional spatial area units.

[0030] Hyperedges are constructed to represent multivariate relationships between different spatial units. Each hyperedge connects at least three hypervertices, enabling a more comprehensive reflection of the coordinated changes between these multivariate spatial units. Cross-temporal relationships and spatial location relationships are crucial for constructing hyperedges. By comprehensively considering these two relationships, the interactions between different spatial units can be more accurately captured. Cluster analysis can be used to construct hyperedges connecting multiple hypervertices. First, the similarity between hypervertices is calculated based on cross-temporal relationships and spatial location relationships. Then, a clustering algorithm is used to cluster hypervertices with high similarity into a single cluster, and hypervertices within the same cluster are connected by hyperedges. For example, the DBSCAN (Density-Based Spatial Clustering Application) algorithm can be used to calculate a distance matrix based on the cross-temporal relationships and spatial location relationships between hypervertices. Hypervertices with distances below a certain threshold are grouped into the same cluster, and hypervertices within each cluster are connected by a hyperedge. Assuming that in a radar echo hypergraph structure, there are some spatial area units that have similar echo intensity change trends in continuous time frames and are spatially adjacent, then the hypervertices corresponding to these spatial area units can be connected by a hyperedge to represent the coordinated change relationship between them.

[0031] Step S206: assigning weights to the hyperedges to obtain weighted hyperedges. The weights of the weighted hyperedges are determined by the strength of the cross-temporal correlation and the closeness of the spatial position relationship.

[0032] The weight of a hyperedge reflects the strength of the relationship between the hypervertices connecting the two ends of the hyperedge. The strength of the cross-temporal relationship reflects the closeness of the connection between spatial units at different points in time, while the closeness of the spatial position relationship reflects the spatial proximity of spatial units. By comprehensively considering these two factors, a reasonable weight can be assigned to a hyperedge.

[0033] A weighted summation approach can be used to assign weights to hyperedges. First, quantitative indicators for the strength of the cross-temporal correlation and the closeness of the spatial relationship are calculated separately. For example, the strength of the cross-temporal correlation can be measured by the consistency of the echo intensity change rate of the same spatial unit in consecutive time frames, while the closeness of the spatial relationship can be measured by the center distance between the spatial units. Then, different weights are assigned to these two quantitative indicators and a weighted sum is taken to obtain the hyperedge weight. Suppose that in a radar echo hypergraph structure, there are two hyperedges. If the spatial units corresponding to the hypervertices of one hyperedge have very consistent echo intensity change trends in consecutive time frames and are spatially close, then the weight of this hyperedge will be relatively large. If the spatial units corresponding to the hypervertices of the other hyperedge have less consistent echo intensity change trends and are spatially far apart, then the weight of this hyperedge will be relatively small.

[0034] Step S207: Integrate the hypervertices, weighted hyperedges, and attribute information of the hypervertices to obtain a radar echo hypergraph structure.

[0035] Integrating hypervertices, weighted hyperedges, and hypervertex attribute information combines the components obtained in the previous steps into a complete radar echo hypergraph structure. This integration allows spatial regions and their multivariate relationships within the radar echo data to be uniformly represented as a hypergraph, facilitating subsequent processing and analysis.

[0036] The integration of hypervertices, weighted hyperedges, and hypervertex attribute information can be achieved through data structures. For example, a graph database can be used to store the radar echo hypergraph structure, with hypervertices as nodes, weighted hyperedges as edges, and hypervertex attribute information stored as node attributes. During the integration process, it is necessary to ensure that the correspondence between hypervertices, weighted hyperedges, and attribute information is correct. Assuming that a radar echo hypergraph structure has 100 hypervertices and 50 weighted hyperedges, organizing and storing these hypervertices, weighted hyperedges, and corresponding attribute information according to a specific data structure will result in a complete radar echo hypergraph structure.

[0037] As a second implementation, step S200 constructs a hypergraph structure for the radar echo raw data sequence to obtain a radar echo hypergraph structure, which can be specifically implemented as follows:

[0038] Step S210: performing spatiotemporal correlation analysis on the radar echo raw data sequence to determine a plurality of spatial region groups with coordinated variation characteristics, each of which includes at least three spatial region units that vary synchronously in the time dimension.

[0039] Spatiotemporal correlation analysis comprehensively considers the temporal and spatial correlations of radar echo data to identify groups of spatial regions with co-variation characteristics. Co-variation characteristics indicate that characteristics such as echo intensity within these spatial regions exhibit similar trends in the temporal dimension, such as simultaneous increases or decreases. Each spatial region group contains at least three spatial regions, which more accurately reflects the co-variation relationships between multiple spatial regions.

[0040] Temporal and spatial correlation analysis of raw radar echo data sequences can be performed using a combination of time series analysis and cluster analysis. First, the echo intensity data for each spatial unit is treated as a time series. Time series analysis methods (such as autocorrelation and cross-correlation) are used to calculate the temporal correlation between different spatial units. Then, a clustering algorithm is used to group spatial units with high temporal correlations into a single category, forming a spatial group. For example, in a meteorological monitoring scenario, analysis of radar echo data over a period of time reveals that the echo intensity of several precipitation areas exhibits similar temporal trends. The spatial units corresponding to these precipitation areas are then clustered into a single spatial group.

[0041] Step S220: taking the spatial area units in each spatial area group as an associated supervertex set, constructing an initial hyperedge based on the associated supervertex set, and the connection range of the initial hyperedge covers all spatial area units in the associated supervertex set.

[0042] The associated supervertex set is a collection of supervertices corresponding to spatial area units with co-variation characteristics. The initial hyperedge is an edge connecting all supervertices in the associated supervertex set. Its connection range covers all spatial area units in the associated supervertex set and is used to represent the multivariate association relationship between these spatial area units.

[0043] Constructing an initial hyperedge based on an associated supervertex set can be achieved through graph construction. Each hypervertex in the associated supervertex set is treated as a node in the graph, and connections are then established between these nodes to form an initial hyperedge. For example, in a radar echo hypergraph structure, a spatial region group contains five spatial region units. The supervertices corresponding to these five spatial region units are used as the associated supervertex set. A hyperedge is established between these five hypervertices, and the connection range of this hyperedge covers these five hypervertices, forming an initial hyperedge.

[0044] Step S230: performing correlation strength calculation on the initial hyperedge to obtain a hyperedge correlation strength value. The hyperedge correlation strength value is calculated by jointly calculating the similarity and the change phase difference of the echo intensity change curves of each spatial area unit in the correlation hypervertex set.

[0045] The hyperedge correlation strength reflects the strength of the correlation between the hypervertices connecting the two ends of the initial hyperedge. The similarity of the echo intensity change curves of each spatial unit in the associated hypervertex cluster reflects the degree of similarity in the echo intensity change trends of these spatial units, while the phase difference reflects the temporal order of their changes. By comprehensively considering these two factors, the hyperedge correlation strength value can be calculated more accurately.

[0046] The following method can be used to calculate the association strength of the initial hyperedge. First, the echo intensity data of each spatial area unit in the associated hypervertex set is plotted as an echo intensity change curve. Then, the similarity between these curves is calculated using a curve similarity calculation method (such as the dynamic time warping algorithm, the Pearson correlation coefficient, etc.). At the same time, the phase difference of the echo intensity change curve of each spatial area unit is calculated. Finally, the curve similarity and phase difference are comprehensively calculated to obtain the hyperedge association strength value. For example, in a radar echo hypergraph structure, there is an initial hyperedge connecting three hypervertices. If the echo intensity change curves of the corresponding spatial area units have a high similarity and a small phase difference, then the association strength value of the initial hyperedge will be relatively large.

[0047] Step S240: screening the initial hyperedges according to the hyperedge association strength values, and retaining the initial hyperedges whose hyperedge association strength values ​​are greater than a preset threshold as valid hyperedges.

[0048] The preset threshold is a pre-set critical value used to determine whether the correlation strength of the initial hyperedge is strong enough. By screening, those initial hyperedges with weak correlation strength can be removed, and only the valid hyperedges with strong correlation strength are retained, making the radar echo hypergraph structure more concise and effective.

[0049] Screening initial hyperedges based on their association strength can be accomplished through a comparison operation. The association strength value of each initial hyperedge is compared with a preset threshold. If the association strength value is greater than the preset threshold, the initial hyperedge is retained as a valid hyperedge; if the association strength value is less than or equal to the preset threshold, the initial hyperedge is discarded. For example, if the preset threshold is 0.5 and there are 10 initial hyperedges in a radar echo hypergraph structure, and 3 of them have association strength values ​​greater than 0.5, these 3 initial hyperedges are retained as valid hyperedges, and the remaining 7 initial hyperedges are discarded.

[0050] Step S250: performing spatial topology verification on valid hyperedges, and eliminating valid hyperedges with spatial topology conflicts. Spatial topology conflicts include situations where the spatial area units connected by the hyperedges are not adjacent in physical space and have no indirect association paths.

[0051] Spatial topology verification checks whether the topological relationships between spatial units connected by valid hyperedges are reasonable in physical space. Spatial topology conflicts indicate that the spatial units connected by hyperedges are not physically adjacent and lack indirect paths. In this case, the hyperedges may not conform to the actual physical relationships and need to be removed.

[0052] Spatial analysis methods can be used to verify the spatial topology of valid hyperedges. First, determine the location information of each spatial area unit connected by a valid hyperedge in physical space. Then, check whether these spatial area units are adjacent or whether there are indirect connection paths. For example, using geographic information system (GIS) technology, the location information of spatial area units can be visualized on a map to check the spatial relationship between them. If it is found that the spatial area units connected by a valid hyperedge are not adjacent in physical space and there is no indirect connection path, such as if there are other unrelated areas in between, then the valid hyperedge is eliminated.

[0053] Step S260: Integrate the screened and verified valid hyperedges with the corresponding hypervertices to obtain a radar echo hypergraph structure containing multi-element association relationships.

[0054] The effective hyperedges that have been screened and verified are those with strong correlation strength and reasonable spatial topological relationship. By integrating these effective hyperedges with the corresponding hypervertices, a complete radar echo hypergraph structure containing multiple correlation relationships can be obtained.

[0055] Integrating screened and verified valid hyperedges with their corresponding hypervertices can be achieved through data structures. For example, a graph database can be used to store the radar echo hypergraph structure, with hypervertices as nodes and valid hyperedges as edges, organized and stored according to a specific data structure. During the integration process, it is necessary to ensure that the connection between valid hyperedges and their corresponding hypervertices is correct.

[0056] Step S300: Integrate the radar echo hypergraph structure into a preset diffusion model framework to obtain a hypergraph enhanced diffusion model. The hypergraph enhanced diffusion model processes the radar echo hypergraph structure through a hypergraph feature extraction module to generate a hypergraph embedding feature containing multivariate correlation information.

[0057] The preset diffusion model framework is a predefined model structure used for data processing and analysis. The radar echo hypergraph structure is integrated into the preset diffusion model framework to introduce the multivariate correlation information in the radar echo data into the diffusion model, thereby enhancing the model's processing capabilities for radar echo data. The hypergraph feature extraction module is specifically designed to process the radar echo hypergraph structure. It extracts hypergraph embedding features containing multivariate correlation information from the hypergraph structure. These features can better reflect the correlation relationships between different spatial regions in the radar echo data.

[0058] Integrating the radar echo hypergraph structure into the pre-set diffusion model framework can be achieved through model fusion. A hypergraph feature extraction module is added to the pre-set diffusion model framework. The radar echo hypergraph structure is passed as input to the hypergraph feature extraction module. After processing, hypergraph embedding features are generated. These features are then fused with other features in the diffusion model framework to obtain a hypergraph-enhanced diffusion model. For example, in a deep learning-based diffusion model framework, the radar echo hypergraph structure is input into the hypergraph feature extraction module to generate hypergraph embedding features. These features are then concatenated with the convolutional layer features in the diffusion model. After subsequent processing, the hypergraph-enhanced diffusion model is obtained.

[0059] As an implementation manner, step S300 can be specifically implemented as the following steps S310 to S360:

[0060] Step S310: Obtain a preset basic diffusion model, which includes an input layer, a feature transformation layer, and an output layer. The feature transformation layer is used to generate target data through an iterative denoising and de-noising process.

[0061] The default basic diffusion model is a pre-trained model with a specific structure and functionality. The input layer is the model's entry point, receiving input data. The feature transformation layer is the model's core, processing the input data through iterative denoising and de-noising to generate target data. The output layer is the model's exit, outputting the processed results.

[0062] Obtaining a preset basic diffusion model can be achieved through model downloading or pre-training. For example, a pre-trained diffusion model can be downloaded from an open-source deep learning model library. This model consists of an input layer, a feature transformation layer, and an output layer. In the weather radar echo extrapolation scenario, radar echo data is input to the input layer of the basic diffusion model. After iterative denoising and de-noising in the feature transformation layer, the predicted radar echo data is obtained from the output layer.

[0063] Step S320: inserting a hypergraph feature extraction module into the feature transformation layer of the basic diffusion model, the hypergraph feature extraction module includes a super-vertex embedding submodule, a hyper-edge information aggregation submodule and a cross-layer feature fusion submodule.

[0064] The hypergraph feature extraction module is designed to extract useful features from the radar echo hypergraph structure. It consists of a hypervertex embedding submodule, a hyperedge information aggregation submodule, and a cross-layer feature fusion submodule. The hypervertex embedding submodule converts hypervertex attribute information into a low-dimensional, dense vector representation; the hyperedge information aggregation submodule aggregates information about hypervertices connected by hyperedges; and the cross-layer feature fusion submodule fuses hypergraph information with existing hierarchical features from the feature transformation layer of the underlying diffusion model.

[0065] Inserting the hypergraph feature extraction module into the feature transformation layer of the basic diffusion model can be achieved by modifying the model. In the code of the basic diffusion model, the code of the hypergraph feature extraction module is added at the appropriate position of the feature transformation layer.

[0066] Step S330: Input the radar echo hypergraph structure into the super-vertex embedding submodule, and convert the attribute information of the super-vertex into a low-dimensional dense super-vertex embedding vector through the graph embedding algorithm. The super-vertex embedding vector retains the key information of the spatial region characteristics.

[0067] A graph embedding algorithm converts the attribute information of nodes (supervertices) in a graph structure into a low-dimensional vector representation. This conversion compresses the supervertices' attributes into a low-dimensional space while preserving key information about the spatial region's characteristics. The supervertices' embedding vectors are low-dimensional, dense representations of the supervertices' attributes, facilitating subsequent processing and analysis.

[0068] The radar echo hypergraph structure is input into the super-vertex embedding submodule, and the attribute information of the super-vertex is converted into a low-dimensional dense super-vertex embedding vector through the graph embedding algorithm. The following specific steps can be adopted.

[0069] As an implementation manner, step S330 may be specifically implemented as the following steps S331 to S335:

[0070] Step S331: Standardize the attribute information of the super-vertex to eliminate the dimensional differences between different types of features and obtain standardized attribute features.

[0071] Hypervertex attribute information includes various types of features, such as echo intensity distribution, shape, and positional relationships. These features may have different dimensions, which can affect subsequent processing and analysis. Normalization converts these different types of features into features with the same dimensions and range, eliminating dimensional differences.

[0072] Standardizing super-vertex attribute information can be done using common standardization methods, such as Z-score standardization. First, calculate the mean and standard deviation of each feature. Then, subtract the mean from each feature value and divide it by the standard deviation to obtain the standardized feature value. For example, for the echo intensity distribution and shape features of a super-vertex, calculate their mean and standard deviation, respectively. Then, apply the Z-score standardization formula to obtain the standardized attribute features.

[0073] Step S332: input the standardized attribute features into a preset embedding neural network, where the embedding neural network includes at least two fully connected layers and one nonlinear activation layer, and the number of neurons in the fully connected layer gradually decreases from the input dimension to the target embedding dimension.

[0074] The pre-defined embedding neural network is a neural network used to convert standardized attribute features into low-dimensional embedding vectors. It consists of at least two fully connected layers and a nonlinear activation layer. The number of neurons in the fully connected layers gradually decreases from the input dimension to the target embedding dimension. This compresses the high-dimensional standardized attribute features into a low-dimensional space.

[0075] The standardized attribute features are input into the preset embedding neural network, which performs a nonlinear transformation on the input features. For example, assuming the dimension of the standardized attribute features is 100 and the target embedding dimension is 10, the first fully connected layer of the embedding neural network has 80 neurons, the second fully connected layer has 30 neurons, and finally a nonlinear activation layer (such as the ReLU activation function) performs a nonlinear transformation to convert the input standardized attribute features into a low-dimensional embedding vector.

[0076] Step S333: Perform nonlinear transformation on the standardized attribute features through embedding neural network to generate preliminary embedding vectors.

[0077] Embedding neural networks perform nonlinear transformations on standardized attribute features to learn complex relationships between features, thereby generating more expressive preliminary embedding vectors. Nonlinear activation layers play a key role in this process, introducing nonlinear factors that enable neural networks to handle more complex problems.

[0078] When performing nonlinear transformations on standardized attribute features using an embedding neural network, the standardized attribute features are first fed into the first fully connected layer. After linear transformation and weighted summation, an intermediate result is obtained. This intermediate result is then fed into a nonlinear activation layer for nonlinear transformation. Finally, the nonlinearly transformed result is fed into the second fully connected layer for another linear transformation and weighted summation to obtain the initial embedding vector.

[0079] Step S334: Based on the hyperedge connection relationship in the radar echo hypergraph structure, a super-vertex similarity matrix is ​​constructed. The element values ​​in the super-vertex similarity matrix are determined by the hyper-edge weight values ​​and spatial distances between corresponding super-vertices.

[0080] The supervertex similarity matrix reflects the similarity between supervertices. The matrix elements represent the degree of similarity between corresponding supervertices. Hyperedge weights reflect the strength of the association between supervertices, while spatial distances reflect the physical proximity of the corresponding spatial regions. By combining these two factors, we can more accurately determine the similarity between supervertices.

[0081] When constructing a super-vertex similarity matrix based on the hyper-edge connection relationship in the radar echo hypergraph structure, you can first traverse all the hyper-edges in the radar echo hypergraph structure, and for each hyper-edge connected to two hyper-vertices, record the hyper-edge weight value between them. Then, calculate the center coordinates of the spatial area unit corresponding to each hyper-vertex, and calculate the spatial distance between the hyper-vertices based on the center coordinates. Finally, the hyper-edge weight value and the spatial distance are comprehensively calculated to obtain the element value in the super-vertex similarity matrix. For example, for super-vertex i and super-vertex j, the hyper-edge weight value between them is w, and the spatial distance is d, then the element value of the i-th row and j-th column in the super-vertex similarity matrix can be calculated by the formula s=w / (1+d).

[0082] Step S335: Smoothing the preliminary embedding vector using the super-vertex similarity matrix so that the embedding vectors corresponding to closely connected super-vertices are closer in the vector space, thereby obtaining the super-vertex embedding vector.

[0083] Smoothing the preliminary embedding vectors can bring the embedding vectors corresponding to closely connected supervertices closer together in vector space, thereby better reflecting the associations between supervertices. Larger values ​​in the supervertice similarity matrix indicate greater similarity between the corresponding supervertices. Smoothing the preliminary embedding vectors can bring the embedding vectors of these supervertices closer together in vector space.

[0084] Specifically, the supervertex similarity matrix can be normalized so that its element values ​​are in the range [0, 1]. Then, a weighted average of the preliminary embedding vectors is performed based on the normalized supervertex similarity matrix. For each supervertex's preliminary embedding vector, its similarity to other supervertices is used as a weight, and the preliminary embedding vectors of other supervertices are weighted and summed to obtain the smoothed supervertex embedding vector.

[0085] Step S340: Input the super-vertex embedding vector and the weighted hyper-edge into the hyper-edge information aggregation sub-module, aggregate the neighborhood super-vertex information of each super-vertex through the hypergraph convolution operation, and obtain the hyper-edge aggregation feature. The hyper-edge aggregation feature contains the influence information of the multi-dimensional association relationship on the super-vertex feature.

[0086] The hyperedge information aggregation submodule aggregates information about hypervertices connected by hyperedges. Through hypergraph convolution, information about neighboring hypervertices is aggregated onto the current hypervertice, yielding hyperedge aggregate features. These features incorporate the impact of multi-dimensional relationships on hypervertice features, providing a more comprehensive reflection of the characteristics of a hypervertice within these relationships.

[0087] The hypervertex embedding vector and weighted hyperedge are input into the hyperedge information aggregation submodule. The process of aggregating the neighborhood hypervertex information of each hypervertex through the hypergraph convolution operation can be further refined into the following steps S341 to S345:

[0088] Step S341: For each supervertex, determine all hyperedges directly connected to it to form an associated hyperedge set.

[0089] The associated hyperedge set is a set of all hyperedges directly connected to a certain hypervertex. By determining the associated hyperedge set, the neighborhood information of the hypervertex can be clarified, providing a basis for subsequent information aggregation.

[0090] For each hypervertex, we can determine all hyperedges directly connected to it by traversing the radar echo hypergraph structure. For each hypervertex, we examine all hyperedges in the hypergraph structure, identify the hyperedges connected to it, and add these hyperedges to the associated hyperedge set. For example, in a radar echo hypergraph structure with 100 hypervertices and 50 hyperedges, for one of the hypervertex, we traverse the hyperedge list and find that it has three hyperedges connected to it. These three hyperedges are then added to the associated hyperedge set.

[0091] Step S342: For each hyperedge in the associated hyperedge set, obtain all hypervertices connected to the hyperedge to form a hyperedge vertex set.

[0092] A hyperedge vertex set is a collection of all hypervertices connected to a hyperedge, containing all hypervertices associated with the current hyperedge. For each hyperedge in the associated hyperedge set, all hypervertices connected to it can be retrieved using the hyperedge's connection information. In the radar echo hypergraph structure, each hyperedge records the hypervertices it connects to, and this information can be accessed to obtain the hyperedge vertex set. For example, if a hyperedge in the associated hyperedge set connects five hypervertices, these five hypervertices form the hyperedge vertex set.

[0093] Step S343: Calculate the contribution weight of each super vertex in the super edge vertex set to the current super vertex. The contribution weight is determined by the super edge weight value and the feature similarity between the super vertices.

[0094] The contribution weight reflects the contribution of each hypervertex in the hyperedge vertex set to the current hypervertex feature aggregation. It is determined by both the hyperedge weight and the feature similarity between hypervertices. The hyperedge weight reflects the strength of the association between hypervertices, while the feature similarity reflects the similarity between the hypervertex features. By combining these two factors, the contribution weight can be calculated more accurately.

[0095] As an implementation manner, step S343 may be specifically implemented as the following steps S3431 to S3437:

[0096] Step S3431: Obtain the super-vertex embedding vector of each super-vertex in the hyper-edge vertex set and the super-vertex embedding vector of the current super-vertex.

[0097] The supervertex embedding vector is a low-dimensional, dense representation of the supervertex's attribute information, encompassing the supervertex's key features. Obtaining the supervertex embedding vector for each supervertex in the hyperedge vertex set and the current supervertex is necessary for subsequent calculations of feature similarity and contribution weights.

[0098] Obtaining a supervertex embedding vector can be achieved by querying the storage structure of supervertex embedding vectors. In the previous step, the supervertex attribute information has been converted into a supervertex embedding vector and stored. The corresponding supervertex embedding vector can be queried using the supervertex identification information. For example, in a dictionary storing supervertex embedding vectors, the corresponding supervertex embedding vector can be obtained by searching the supervertex number.

[0099] Step S3432: Calculate feature similarity between the super-vertex embedding vector of each super-vertex in the hyper-edge vertex set and the super-vertex embedding vector of the current super-vertex to obtain a feature similarity value.

[0100] Feature similarity calculations can measure the similarity of features between hypervertices. Methods for calculating feature similarity include cosine similarity and Euclidean distance. By calculating feature similarity values, we can understand the degree of feature similarity between each hypervertex in the hyperedge vertex set and the current hypervertex. Cosine similarity can be used to calculate feature similarity between the hypervertex embedding vectors of each hypervertex in the hyperedge vertex set and the hypervertex embedding vector of the current hypervertex.

[0101] Step S3433: Obtain the weighted hyperedge weight value of the hyperedge corresponding to the hyperedge vertex set, where the weighted hyperedge weight value is a preset weight value of the hyperedge in the radar echo hypergraph structure.

[0102] The weighted hyperedge weight value reflects the strength of the association between the hypervertices connected by the hyperedge, and is pre-set when constructing the radar echo hypergraph structure. Obtaining the weighted hyperedge weight value of the hyperedge corresponding to the hyperedge vertex set can be achieved by querying the hyperedge information in the radar echo hypergraph structure. Specifically, the identification information of the hyperedge corresponding to the hyperedge vertex set can be determined first. Then, the information of the hyperedge is searched in the radar echo hypergraph structure to obtain the weighted hyperedge weight value. For example, in a radar echo hypergraph structure, the information of the hyperedge is stored in a list, and the information of each hyperedge includes the identifier of the hyperedge, the connected hypervertices, and the weighted hyperedge weight value, etc. The corresponding weighted hyperedge weight value can be queried through the identifier of the hyperedge.

[0103] Step S3434: perform weighted product of the feature similarity value and the weighted hyperedge weight value to obtain the preliminary contribution weight. The weighted product is processed as a direct multiplication operation of the feature similarity value and the weighted hyperedge weight value.

[0104] The initial contribution weight is calculated by multiplying the feature similarity value by the weighted hyperedge weight. This factor takes into account the influence of the feature similarity between hypervertices and the strength of the hyperedge association on the current hypervertice feature aggregation. The weighted product of the feature similarity value and the weighted hyperedge weight can be simply multiplied by the feature similarity value and the weighted hyperedge weight.

[0105] Step S3435: Perform centrality analysis on all hypervertices in the hyperedge vertex set to obtain the centrality index of each hypervertex. The centrality index is calculated comprehensively based on the number of hyperedges connected to the hypervertex in the radar echo hypergraph structure and the average weight value of the corresponding hyperedges. The larger the centrality index value, the higher the core degree of the hypervertex in the multi-dimensional association relationship.

[0106] Centrality analysis can evaluate the core degree of a hypervertex in the radar echo hypergraph structure. The larger the centrality index value, the more important the hypervertex is in the multivariate association relationship and the greater its influence on other hypervertices.

[0107] Specifically, we can first count the number of hyperedges connected to each hypervertex in the radar echo hypergraph structure. Then, calculate the average weight of the hyperedges connected to each hypervertex. Finally, the number of hyperedges and the average weight are calculated together to obtain the centrality index. For example, for hypervertex i in the hyperedge vertex set, it is connected to 3 hyperedges, and the weights of these 3 hyperedges are 0.5, 0.6, and 0.7 respectively. Then the average weight is (0.5+0.6+0.7) / 3=0.6. Assuming that the calculation formula of the centrality index is c=n×w, where n is the number of hyperedges and w is the average weight, the centrality index of hypervertex i is 3×0.6=1.8.

[0108] Step S3436: Adjust the preliminary contribution weight according to the centrality index to obtain an adjustment coefficient. The adjustment coefficient is the result of normalization of the centrality index value. The normalization is achieved by dividing the centrality index value by the sum of the centrality index values ​​of all hypervertices in the hyperedge vertex set.

[0109] The adjustment coefficient is used to adjust the initial contribution weights so that supervertices with larger centrality index values ​​contribute more to the current supervertice feature aggregation. Normalization can convert centrality index values ​​to the range of [0, 1] to facilitate subsequent calculations.

[0110] The process of adjusting the initial contribution weights based on the centrality index is as follows: First, calculate the sum of the centrality index values ​​of all hypervertices in the hyperedge vertex set. Then, divide the centrality index value of each hypervertice by this sum to obtain the adjustment coefficient.

[0111] Step S3437: Multiply the preliminary contribution weight by the adjustment coefficient to obtain the contribution weight of each super vertex in the hyperedge vertex set to the current super vertex. The contribution weight comprehensively reflects the influence of the hyperedge association strength, feature similarity and super vertex coreness on the current super vertex feature aggregation.

[0112] The contribution weight comprehensively considers the impact of hyperedge association strength, feature similarity, and hypervertex coreness on the current hypervertex feature aggregation. By multiplying the initial contribution weight with the adjustment coefficient, a more accurate contribution weight can be obtained. The process of multiplying the initial contribution weight with the adjustment coefficient can be directly multiplying the initial contribution weight and the adjustment coefficient.

[0113] Step S344: performing weighted summation on the hypervertex embedding vectors of all hypervertices in the hyperedge vertex set based on the contribution weights to obtain a single hyperedge aggregation vector.

[0114] The single hyperedge aggregation vector is obtained by weighted summing of the hypervertex embedding vectors of all hypervertices in the hyperedge vertex set. It integrates the information of the hypervertices connected by the hyperedge and reflects the influence of the hyperedge on the current hypervertex.

[0115] Specifically, we first traverse all hypervertices in the hyperedge vertex set. For each hypervertex, we multiply its hypervertex embedding vector by its corresponding contribution weight. We then sum all weighted hypervertex embedding vectors to obtain the single hyperedge aggregation vector. For example, if there are three hypervertices in the hyperedge vertex set, with hypervertex embedding vectors v1, v2, and v3, and contribution weights w1, w2, and w3, respectively, then the single hyperedge aggregation vector v = w1 × v1 + w2 × v2 + w3 × v3.

[0116] Step S345: average pooling is performed on the single hyperedge aggregation vectors corresponding to all hyperedges in the associated hyperedge set to obtain a hyperedge aggregation feature. The hyperedge aggregation feature is used to comprehensively reflect the feature aggregation result of the current hypervertex in the multi-dimensional association relationship.

[0117] Average pooling can average multiple vectors to obtain a comprehensive feature representation. Average pooling of the single hyperedge aggregate vectors corresponding to all hyperedges in the associated hyperedge set can comprehensively consider the impact of multiple hyperedges on the current hypervertex, obtaining a more comprehensive hyperedge aggregate feature.

[0118] Specifically, we first add the single-hyperedge aggregation vectors corresponding to all hyperedges in the associated hyperedge set. Then, we divide the sum by the number of hyperedges to obtain the hyperedge aggregation feature. For example, if there are three hyperedges in the associated hyperedge set, and their corresponding single-hyperedge aggregation vectors are v1, v2, and v3, respectively, then the hyperedge aggregation feature v = (v1+v2+v3) / 3.

[0119] Step S350: The hyperedge aggregation features are input into the cross-layer feature fusion submodule and fused with the original hierarchical features in the feature transformation layer of the basic diffusion model to obtain enhanced features of the fused hypergraph information. The fusion process is achieved through feature splicing and attention weighted summation.

[0120] The cross-layer feature fusion submodule is used to fuse hypergraph information with existing hierarchical features in the feature transformation layer of the basic diffusion model to enhance the model's processing capabilities for radar echo data. Feature concatenation combines hyperedge aggregation features with existing hierarchical features along the feature dimension. Attention weighted summation uses the attention mechanism to perform a weighted summation of the concatenated features, enabling the model to focus more on important features.

[0121] The process of inputting hyperedge aggregation features into the cross-layer feature fusion submodule for fusion is as follows: First, the hyperedge aggregation features and the original hierarchical features from the basic diffusion model feature transformation layer are concatenated along the feature dimension to produce a concatenated feature vector. Then, an attention mechanism is used to calculate the attention weight for each feature dimension, which reflects the importance of that feature dimension. Finally, the concatenated feature vector and the attention weights are weighted summed to obtain enhanced features that incorporate hypergraph information. The attention mechanism is then used to calculate the attention weight for each dimension, and the feature vector and attention weights are weighted summed to produce the enhanced features.

[0122] Step S360: Input the enhanced features into the subsequent network layers of the basic diffusion model, and adjust the denoising sampling process of the diffusion model so that the enhanced features can continuously guide the spatiotemporal evolution direction of the echo features during iterative denoising, thereby obtaining a hypergraph enhanced diffusion model.

[0123] Subsequent network layers of the basic diffusion model, such as the denoising network, feed enhanced features into these layers to fine-tune the diffusion model's denoising sampling process. These enhanced features incorporate hypergraph information, which continuously guides the spatiotemporal evolution of echo features during iterative denoising, enabling the model to better predict future changes in radar echoes.

[0124] Specifically, the enhanced features can be first input into the denoising network, which then removes noise from the features based on the enhanced features. During the iterative denoising process, the enhanced features continuously influence the direction of denoising, guiding the echo features towards a more reasonable spatiotemporal evolution. For example, in the scenario of weather radar echo extrapolation, the enhanced features can guide the denoising process towards predicting the movement and intensity changes of precipitation areas, thereby obtaining more accurate prediction results. After multiple iterative denoising, the hypergraph enhanced diffusion model is obtained.

[0125] Step S400: extrapolating the radar echo raw data sequence based on the hypergraph enhanced diffusion model to obtain a radar echo extrapolated data sequence for a future period, wherein the radar echo extrapolated data sequence includes a plurality of extrapolated data frames that are chronologically continuous with the radar echo raw data sequence.

[0126] After the previous steps of construction and training, the hypergraph-enhanced diffusion model has a stronger ability to process radar echo data. Using this model, we can extrapolate the raw radar echo data sequence to predict radar echo conditions in the future. This produces an extrapolated radar echo data sequence, which consists of multiple extrapolated data frames that are chronologically continuous with the raw radar echo data sequence. Each extrapolated data frame represents the radar echo condition at a certain moment in the future.

[0127] The following two implementations can be used to perform extrapolation on the radar echo raw data sequence based on the hypergraph enhanced diffusion model.

[0128] As a first implementation method of the step, step S400, extrapolating the radar echo original data sequence based on the hypergraph enhanced diffusion model to obtain the radar echo extrapolated data sequence of the future time period can be specifically implemented as the following steps S410 to S470:

[0129] Step S410: convert the last N radar echo data frames in the radar echo raw data sequence into a model input tensor. The dimensions of the model input tensor include a time dimension, a space dimension, and an intensity dimension, and N is an integer greater than or equal to 3.

[0130] The model input tensor converts radar echo data into a format suitable for the hypergraph-enhanced diffusion model. It contains time, space, and intensity dimensions. The time dimension represents the temporal order of the radar echo data, the space dimension represents the spatial distribution of the radar echo data, and the intensity dimension represents the intensity information of the radar echo. The last N radar echo data frames are selected as input to utilize the most recent radar echo information for extrapolation. N is an integer greater than or equal to 3, which provides sufficient temporal information for model learning and prediction.

[0131] Specifically, we can first determine the last N radar echo data frames. Then, we organize each radar echo data frame according to the spatial and intensity dimensions to form a two-dimensional matrix. Finally, we stack these N two-dimensional matrices along the time dimension to obtain the model input tensor. For example, assuming the spatial dimensions of the radar echo data frame are 100×100 and the intensity dimension is 1, we select the last three radar echo data frames, convert each data frame into a 100×100×1 two-dimensional matrix, and then stack these three two-dimensional matrices along the time dimension to obtain a 3×100×100×1 model input tensor.

[0132] Step S420: Input the model input tensor into the input layer of the hypergraph enhanced diffusion model, and convert it into an initial feature tensor that can be processed by the diffusion model through feature mapping processing of the input layer.

[0133] The feature mapping processing of the input layer is the process of converting the model input tensor into the initial feature tensor that can be processed by the diffusion model. It can preprocess the input data to make it more suitable for model learning and processing.

[0134] Specifically, the model input tensor is fed into the neurons in the input layer, which perform a weighted summation and linear transformation on the input data. Then, an activation function performs a nonlinear transformation to produce an initial feature tensor. For example, if the input layer has 50 neurons and the model input tensor has a dimension of 3 × 100 × 100 × 1, then after the weighted summation, linear transformation, and nonlinear transformation in the input layer, an initial feature tensor with a dimension of 50 is obtained.

[0135] Step S430: Input the initial feature tensor into the feature transformation layer of the hypergraph enhanced diffusion model. In each diffusion step, the hypergraph feature extraction module extracts features from the radar echo hypergraph structure, generates hypergraph embedding features, and fuses the hypergraph embedding features with the intermediate features of the current diffusion step.

[0136] The feature transformation layer of the hypergraph-enhanced diffusion model processes the initial feature tensor through an iterative denoising and denoising process. During each diffusion step, the hypergraph feature extraction module extracts hypergraph embedding features from the radar echo hypergraph structure. These features contain multivariate correlation information within the radar echo data. Fusing the hypergraph embedding features with the intermediate features from the current diffusion step allows the model to fully utilize the hypergraph information during processing, improving its predictive capabilities. Specifically, the initial feature tensor is first input into the denoising network of the feature transformation layer, which applies noise to the initial feature tensor. During each diffusion step, the hypergraph feature extraction module extracts hypergraph embedding features from the radar echo hypergraph structure. These hypergraph embedding features are then fused with the intermediate features from the current diffusion step through feature concatenation and attention-weighted summation. Finally, the fused features are input into the denoising network for denoising, updating the intermediate feature tensor.

[0137] As an implementation manner, step S430 may be specifically implemented as the following steps S431 to S435:

[0138] Step S431: At the beginning of each diffusion step, the intermediate feature tensor of the current step and the corresponding diffusion time step parameter are obtained.

[0139] The intermediate feature tensor is the feature tensor processed in the diffusion step, which records the processing results of the current step. The diffusion time step parameter represents the time interval of the current diffusion step and is used to control the speed and direction of the diffusion process.

[0140] Retrieving the intermediate feature tensor and the corresponding diffusion time step parameters at the beginning of each diffusion step can be achieved through the model state record. During model execution, the intermediate feature tensor and diffusion time step parameters for each diffusion step are recorded. At the beginning of each diffusion step, the intermediate feature tensor and diffusion time step parameters for the current step are retrieved from the record. For example, in a model with 10 diffusion steps, at the beginning of the fifth diffusion step, the intermediate feature tensor and the corresponding diffusion time step parameters for the fifth step are retrieved from the model state record.

[0141] Step S432: Input the radar echo hypergraph structure into the hypergraph feature extraction module, generate a hypervertex embedding vector through the hypervertex embedding submodule, and generate a hyperedge aggregation feature through the hyperedge information aggregation submodule.

[0142] The hypergraph feature extraction module extracts useful features from the radar echo hypergraph structure through a hypervertex embedding submodule and a hyperedge information aggregation submodule. The hypervertex embedding submodule converts the attribute information of hypervertices into low-dimensional, dense hypervertex embedding vectors. The hyperedge information aggregation submodule aggregates the information of hypervertices connected by hyperedges to generate hyperedge aggregated features.

[0143] The process of inputting the radar echo hypergraph structure into the hypergraph feature extraction module and generating the hypervertex embedding vector and hyperedge aggregation feature is similar to the process described in the previous steps S330-S345. First, the supervertex embedding submodule processes the attribute information of the supervertex to generate a supervertex embedding vector. Then, the hyperedge information aggregation submodule generates the hyperedge aggregation feature based on the supervertex embedding vector and the hyperedge connection relationship. For example, the hypergraph feature extraction module receives the radar echo hypergraph structure as input, the supervertex embedding submodule converts the attribute information of the supervertex into a supervertex embedding vector, and the hyperedge information aggregation submodule generates the hyperedge aggregation feature through a series of calculations based on the supervertex embedding vector and the hyperedge connection relationship.

[0144] Step S433: Input the hyperedge aggregation feature into the cross-layer feature fusion submodule, perform dimension matching with the intermediate feature tensor, so that the spatial dimension and channel dimension of the hyperedge aggregation feature are consistent with the intermediate feature tensor, and obtain the matched hyperedge aggregation feature.

[0145] The cross-layer feature fusion submodule matches the dimensions of the hyper-edge aggregate features and the intermediate feature tensors to ensure that they have the same dimensions when fused, so that operations such as feature splicing and attention weighted summation can be performed correctly.

[0146] Specifically, we can first analyze the spatial and channel dimensions of the intermediate feature tensor. Then, we adjust the hyperedge aggregation features based on the dimensions of the intermediate feature tensor to align their spatial and channel dimensions with those of the intermediate feature tensor. For example, if the intermediate feature tensor has a spatial dimension of 100×100 and a channel dimension of 50, and the hyperedge aggregation features have a spatial dimension of 50×50 and a channel dimension of 20, we can adjust the spatial dimension of the hyperedge aggregation features to 100×100 and the channel dimension to 50 through upsampling or other methods, thus obtaining the matched hyperedge aggregation features.

[0147] Step S434: performing element-wise multiplication on the matched hyperedge aggregation features and the intermediate feature tensor to obtain weighted fusion features. The element-wise multiplication enhances the feature components in the intermediate feature tensor that are related to the multi-dimensional association relationship of the hypergraph.

[0148] The element-level multiplication operation is to multiply the matched hyperedge aggregation features with the corresponding elements of the intermediate feature tensor. Through this operation, the feature components related to the multi-dimensional correlation relationship of the hypergraph in the intermediate feature tensor can be enhanced, highlighting the influence of hypergraph information on the intermediate features.

[0149] The process of performing element-wise multiplication operation on the matched hyper-edge aggregation feature and the intermediate feature tensor may be to directly multiply the matched hyper-edge aggregation feature and the corresponding elements of the intermediate feature tensor.

[0150] Step S435: Input the weighted fusion features into the denoising network of the feature transformation layer, perform denoising in combination with the diffusion time step parameter, update the intermediate feature tensor, and complete the feature transformation of the current diffusion step.

[0151] The denoising network of the feature transformation layer denoises the features according to the weighted fusion features and the diffusion time step parameters, and updates the intermediate feature tensor so that the features gradually approach the real radar echo features.

[0152] Specifically, the weighted fusion features are first input into the denoising network, which removes the noise from the weighted fusion features according to the diffusion time step parameter. The denoising network then outputs the denoised features as the new intermediate feature tensor, completing the feature transformation for the current diffusion step. For example, in a diffusion step, the weighted fusion features are input into the denoising network, which removes the noise from the weighted fusion features according to the diffusion time step parameter and outputs the denoised features as the new intermediate feature tensor.

[0153] Step S440: De-noising the fused intermediate features through the denoising network of the feature transformation layer, gradually reducing the noise level in the features while retaining the multivariate correlation information in the hypergraph embedding features.

[0154] The denoising network in the feature transformation layer gradually reduces the noise level in the features during multiple iterative denoising processes, while preserving the multivariate correlation information in the hypergraph embedded features. This enables the model to better capture the complex correlations in radar echo data, thereby improving the accuracy of radar echo extrapolation. During each denoising operation, the denoising network estimates and removes noise based on the current state of the intermediate features and the diffusion time step parameter. Because the hypergraph embedded features are integrated into the intermediate features during the fusion process, the denoising network removes noise while preserving the multivariate correlation information contained therein.

[0155] The denoising network can adopt the architecture of a convolutional neural network (CNN), utilizing convolutional layers, batch normalization layers, and activation function layers (such as ReLU) to process intermediate features. In each denoising iteration, the convolutional layers extract and transform the intermediate features. The batch normalization layers help accelerate model convergence and improve model stability, and the activation function layers introduce nonlinear factors to enhance the model's expressiveness. For example, the denoising network can be designed to include multiple convolutional and batch normalization layers. The kernel size, stride, and padding of each layer can be adjusted according to actual needs. When processing intermediate features, the network transforms them based on learned weight parameters, gradually reducing the impact of noise while preserving the multivariate correlation information carried by the hypergraph embedding features.

[0156] Step S450: After completing the preset number of diffusion steps, the final denoised features are input into the output layer of the hypergraph enhanced diffusion model. Through the deconvolution operation and intensity mapping of the output layer, the radar echo extrapolation data frame at the first moment in the future is generated.

[0157] The preset number of diffusion steps is predetermined during the model training and setup phase and determines the number of iterations the model will use to denoise and process the features. After the preset number of diffusion steps, the noise in the intermediate features has been reduced to a low level, while retaining sufficient hypergraph multivariate correlation information. The deconvolution operation in the output layer restores the denoised features from the low-dimensional feature space to the same spatial dimensions as the original radar echo data frame, while intensity mapping converts the feature values ​​into actual radar echo intensity values.

[0158] Deconvolution is the opposite of convolution, increasing the spatial size of the feature map. In the output layer, the kernel size, stride, and padding of the deconvolution layer are designed based on the input feature dimensions and the target output spatial dimensions. For example, if the input features have a spatial dimension of 50×50 and the target output radar echo data frame has a spatial dimension of 100×100, the deconvolution layer can increase the spatial size of the feature map to 100×100 by setting appropriate parameters. Intensity mapping maps the deconvolution feature values ​​to the range of radar echo intensity. This can be achieved through a linear transformation or a nonlinear function, such as a fully connected layer. The deconvolution feature vector is input to the fully connected layer. After weighted summation and an activation function (such as a sigmoid function to map the output value to the range [0, 1]), it is scaled according to the actual radar echo intensity range to obtain the extrapolated radar echo data frame for the first moment in the future.

[0159] Step S460: The generated radar echo extrapolation data frame at the first future moment is used as a new input, and is combined with the other data frames in the radar echo original data sequence except the first data frame to form a new input sequence. The above extrapolation steps are repeated to generate the radar echo extrapolation data frame at the second future moment.

[0160] By incorporating the generated radar echo extrapolation data frame for the first moment in the future into the new input sequence, the latest prediction information can be used to perform extrapolation for the next moment. This allows the model to take into account the latest radar echo conditions in each extrapolation, thereby improving the accuracy and continuity of the extrapolation.

[0161] Specifically, the first data frame is first removed from the radar echo original data sequence because it is the oldest information. Then, the generated radar echo extrapolated data frame for the first moment in the future is added to the remaining data frames to form a new input sequence. Next, the new input sequence is processed according to the process of steps S410-S450, that is, the last N data frames in the new input sequence are converted into model input tensors, input into the hypergraph enhanced diffusion model, and after iterative denoising and feature fusion in the feature transformation layer, the radar echo extrapolated data frame for the second moment in the future is finally generated through the output layer. For example, the original radar echo data sequence contains 10 data frames. After generating the extrapolated data frame for the first moment in the future, the first data frame in the original sequence is removed and the extrapolated data frame is added to the remaining 9 data frames to form a new input sequence containing 10 data frames, and then the next round of extrapolation is performed.

[0162] Step S470: The extrapolation steps are iteratively performed in sequence until a preset number of radar echo extrapolation data frames for future time periods are generated, and all the extrapolation data frames are arranged in chronological order to obtain a radar echo extrapolation data sequence.

[0163] Sequentially and iteratively executing the extrapolation steps can continuously generate radar echo extrapolation data frames for multiple future moments. The preset number of future time period radar echo extrapolation data frames is pre-set based on actual needs. For example, the number of future moments to be extrapolated can be determined based on the time range of the weather forecast.

[0164] In each iteration, the extrapolated data frame generated last time is used as the new input, and together with the remaining data frames in the original data sequence, a new input sequence is formed. The extrapolation step is then repeated. When the number of extrapolated data frames generated reaches the preset number, these extrapolated data frames are arranged in chronological order to obtain the radar echo extrapolated data sequence. For example, if the preset radar echo data frames for five future moments need to be extrapolated, after five iterations of extrapolation, extrapolated data frames for the 1st, 2nd, 3rd, 4th, and 5th future moments are generated, respectively. These five data frames are arranged in chronological order to obtain a radar echo extrapolated data sequence containing five data frames.

[0165] As a second implementation, step S400 performs extrapolation on the radar echo original data sequence based on the hypergraph enhanced diffusion model to obtain the radar echo extrapolated data sequence for the future period, which can be specifically implemented as follows:

[0166] Step S401: updating the radar echo hypergraph structure, and adjusting the attribute information of the hypervertices and the weight values ​​of the hyperedges according to the latest radar echo data frame in the radar echo raw data sequence.

[0167] Radar echo data changes over time. To ensure that the radar echo hypergraph accurately reflects the latest radar echo conditions, it must be updated. Adjusting the hypervertices' attributes and hyperedge weights based on the latest radar echo data frames allows the hypergraph to better capture the dynamic changes and multivariate relationships in radar echo data.

[0168] A specific updating implementation method is given below.

[0169] As an implementation manner, step S401 may be specifically implemented as the following steps S4011-S4018:

[0170] Step S4011: obtaining the radar echo data frame with the latest time stamp in the radar echo raw data sequence as an updated reference data frame.

[0171] The update reference data frame is the key data used to update the radar echo hypergraph structure. It contains the latest radar echo information. By obtaining the latest radar echo data frame with the latest time stamp, we can ensure that the update operation is based on the latest radar echo situation.

[0172] To obtain the radar echo data frame with the latest timestamp in a radar echo raw data sequence, one can query the timestamp information in the data sequence. For example, in a list of radar echo data frames, each data frame records its acquisition time. By comparing these timestamps, the latest data frame can be found and used as the updated reference data frame.

[0173] Step S4012: performing spatial region division on the updated reference data frame to obtain updated spatial region units. The division method of the updated spatial region units is consistent with the spatial region units corresponding to the initial super-vertices.

[0174] Spatial region partitioning involves spatially segmenting the updated reference data frame to create updated spatial region units. To ensure consistency between the updated hypergraph structure and the original hypergraph structure, the partitioning of the updated spatial region units must be consistent with the spatial region units corresponding to the initial hypervertices.

[0175] The updated reference data frame can be spatially segmented using the same method as in step S201, i.e., determining the boundaries of the spatial region units based on echo intensity gradient variation characteristics. For example, a gradient-based segmentation algorithm can be used to calculate the echo intensity gradient value for each pixel in the updated reference data frame. Pixels are then classified based on a preset gradient threshold. Pixels with gradient values ​​greater than the threshold are defined as the boundaries of the spatial region units, while pixels with gradient values ​​less than the threshold are grouped into the same spatial region unit.

[0176] Step S4013: extracting the spatial region features of the updated spatial region units as updated attribute features.

[0177] The update attribute feature is a set of features that describe the characteristics of the updated spatial area unit, which includes the echo intensity distribution feature, shape feature and position relationship feature of the updated spatial area unit with adjacent spatial area units.

[0178] The spatial region features of the updated spatial region unit can be extracted using the same method as in step S202. For example, for echo intensity distribution features, the mean and standard deviation of the echo intensities of all pixels within the updated spatial region unit are calculated. For shape features, the area, perimeter, circularity, and other properties of the updated spatial region unit are calculated using geometric feature calculation methods used in image processing. For positional relationship features with adjacent spatial region units, the distance and adjacency relationship with adjacent spatial region units are calculated by analyzing the center coordinates and boundary information of the updated spatial region unit.

[0179] Step S4014: compare the updated attribute features with the original attribute information of the corresponding super vertex, and calculate the feature difference. The feature difference is calculated comprehensively by cosine similarity and Euclidean distance.

[0180] Feature difference reflects the degree of difference between the updated attribute features and the original attribute information of the corresponding super vertex. The feature difference can be calculated by combining cosine similarity and Euclidean distance to measure the difference between features from different perspectives.

[0181] Specifically, the updated attribute features and the original attribute information are first represented as vectors. Next, their cosine similarity and Euclidean distance are calculated. Cosine similarity measures the directional similarity between two vectors, while Euclidean distance measures the spatial distance between two vectors. Finally, the cosine similarity and Euclidean distance are combined to calculate the feature difference.

[0182] Step S4015: When the feature difference is greater than the preset difference threshold, the original attribute information of the corresponding super vertex is replaced with the updated attribute feature to complete the update of the super vertex attribute.

[0183] The preset difference threshold is a pre-set critical value used to determine whether the difference between the updated attribute features and the original attribute information is sufficiently large. When the feature difference exceeds the preset difference threshold, it indicates that there is a significant difference between the updated attribute features and the original attribute information, and the updated attribute features need to replace the original attribute information to ensure the accuracy of the supervertex attribute information. When the feature difference exceeds the preset difference threshold, the updated attribute features are directly assigned to the corresponding supervertex attribute information, completing the supervertex attribute update.

[0184] Step S4016: Based on the updated super-vertex attribute information, recalculate the association strength value of the hyperedge. The recalculation process is consistent with the method of initial hyper-edge weight assignment.

[0185] The association strength value of a hyperedge reflects the strength of the association relationship between the hypervertices connected at both ends of the hyperedge. Recalculating the association strength value of the hyperedge based on the updated hypervertices attribute information can make the weight value of the hyperedge more accurately reflect the multivariate association relationship in the current radar echo data. The recalculation of the association strength value of the hyperedge can adopt the same method as in step S206, that is, it is determined by the strength of the cross-time association relationship and the closeness of the spatial position relationship. For example, based on the updated supervertices attribute information, the echo intensity change trend and position offset characteristics of the same spatial area unit in the continuous time frame are calculated to determine the cross-time association relationship between the supervertices; the center coordinates of the spatial area unit corresponding to the supervertices are calculated to determine the closeness of the spatial position relationship. Then, the strength of the cross-time association relationship and the closeness of the spatial position relationship are comprehensively calculated to obtain the association strength value of the hyperedge.

[0186] Step S4017: Adjust the weight value of the hyperedge according to the recalculated association strength value. When the adjusted association strength value is less than the preset threshold, remove the corresponding hyperedge, and construct a new hyperedge based on the updated super-vertex attribute information to supplement the newly emerged multi-dimensional association relationship.

[0187] Adjusting the weight of hyperedges allows the hypergraph structure to better reflect dynamic changes in radar echo data. When the adjusted correlation strength value is less than a preset threshold, it indicates that the correlation between the hypervertices connected by the hyperedge is weak, and the hyperedge can be removed to simplify the hypergraph structure. Furthermore, constructing new hyperedges based on the updated hypervertice attribute information can supplement newly emerging multivariate correlations.

[0188] A specific implementation method for constructing a new hyperedge is given below.

[0189] As an implementation manner, step S4017 may be specifically implemented as the following steps S40171-S40179:

[0190] Step S40171: Perform a spatial neighborhood search on all updated supervertices to determine the spatial neighborhood supervertex set of each supervertex. The spatial neighborhood supervertex set includes other supervertices whose distance from the current supervertex in physical space is less than a preset distance threshold. The distance is calculated by the Euclidean distance of the center coordinates of the spatial area unit corresponding to the supervertex.

[0191] Spatial neighborhood search searches for other hypervertices in physical space that are close to the current hypervertice. The set of spatial neighborhood hypervertices contains these close hypervertices. The preset distance threshold is a pre-set critical value used to determine whether the distance between hypervertices is close enough.

[0192] The spatial neighborhood search of all updated supervertices can be performed by traversing the supervertex list, calculating the center coordinates of the spatial area unit corresponding to each supervertex, and then calculating the Euclidean distance between the center coordinates to find other supervertices whose distance is less than the preset distance threshold, and forming these supervertices into a spatial neighborhood supervertex set.

[0193] Step S40172: Perform feature change trend analysis on each spatial neighborhood super-vertex set, extract the spatial area feature change sequence of each super-vertex in the continuous time frame, and calculate the synchronization index of the change sequence. The synchronization index is determined comprehensively through the dynamic time warping distance and Pearson correlation coefficient of the change sequence. The larger the synchronization index value, the more consistent the feature change trend of the corresponding super-vertex.

[0194] Feature change trend analysis aims to identify supervertices with consistent feature change trends across consecutive time frames. The synchronicity index measures the consistency of supervertice feature change trends. A comprehensive analysis using the dynamic time warping distance and the Pearson correlation coefficient can more accurately reflect the synchronicity of feature change trends.

[0195] Specifically, the method first extracts the spatial feature change sequences of each super-vertex in continuous time frames, such as the echo intensity change sequence or the shape feature change sequence. Then, the dynamic time warping distance and Pearson correlation coefficient of the change sequence are calculated. The dynamic time warping distance measures the similarity between two time series, while the Pearson correlation coefficient measures the linear correlation between two variables. Finally, the dynamic time warping distance and the Pearson correlation coefficient are combined to obtain a synchronization index. For example, for the echo intensity change sequences of two super-vertices, their dynamic time warping distance and Pearson correlation coefficient s are calculated. The synchronization index is then calculated using the formula s = a × (1-dtw) + b × r (where dtw is the dynamic time warping distance, r is the Pearson correlation coefficient, and a and b are weighting coefficients).

[0196] Step S40173: Filtering the spatial neighborhood supervertex sets whose synchronization index values ​​are greater than a preset synchronization threshold to form candidate supervertex groups, each candidate supervertex group containing at least three supervertices.

[0197] The preset synchronization threshold is a pre-set critical value used to select sets of spatially neighboring supervertices with consistent feature change trends. Candidate supervertex groups are composed of sets of spatially neighboring supervertices whose synchronization index values ​​exceed the preset synchronization threshold. Each candidate supervertex group contains at least three supervertices, allowing for the construction of hyperedges to represent multivariate relationships.

[0198] Screening the spatial neighborhood supervertex sets whose synchronization index values ​​are greater than a preset synchronization threshold can be done by traversing all spatial neighborhood supervertex sets, comparing the synchronization index value of each spatial neighborhood supervertex set with the preset synchronization threshold, and forming the spatial neighborhood supervertex sets whose synchronization index values ​​are greater than the preset synchronization threshold into candidate supervertex groups.

[0199] Step S40174: Perform multivariate association pre-verification on each candidate super-vertex group to verify whether there is a potential collaborative change relationship between the super-vertex groups in the historical time frame. The potential collaborative change relationship is determined by the consistency of the common occurrence frequency and change amplitude of the historical feature change sequence.

[0200] Multivariate association pre-verification is to ensure that there are real multivariate associations between the supervertices in the candidate supervertex group. Potential co-variation relationships can be determined by the co-occurrence frequency and consistency of the change magnitude of the historical feature change sequence.

[0201] The following method can be used to pre-verify the multivariate association of each candidate super-vertex group. First, extract the historical feature change sequence of each super-vertex in the candidate super-vertex group. Then, count the common occurrence frequency of the historical feature change sequence, that is, the number of times the feature change sequences of these super-vertices appear simultaneously in the historical time frame. At the same time, calculate the consistency of the change amplitude of the feature change sequence, for example, by comparing statistics such as the standard deviation of the feature change sequence. Finally, based on the consistency of the common occurrence frequency and the change amplitude, jointly judge whether there is a potential collaborative change relationship. For example, if the historical feature change sequence of each super-vertex in a candidate super-vertex group appears simultaneously in most historical time frames, and the consistency of the change amplitude is high, then it can be considered that this group of super-vertices has a potential collaborative change relationship.

[0202] Step S40175: retain the candidate supervertex group pre-verified by the multi-element association, use all supervertices in the group as connection objects of the new hyperedge, and construct an initial new hyperedge.

[0203] There is a potential collaborative change relationship between the hypervertices in the candidate hypervertex group pre-verified by multi-association. Using these hypervertices as the connection objects of new hyperedges can construct hyperedges that can reflect multi-association relationships.

[0204] Constructing an initial new hyperedge by using all supervertices in a candidate supervertex group pre-verified by multi-element associations as the connection targets of a new hyperedge can be achieved through graph construction. In the radar echo hypergraph structure, each supervertex in a candidate supervertex group is treated as a node in the graph, and connections are established between these nodes to form an initial new hyperedge. For example, if a candidate supervertex group contains four supervertices, these four supervertices are used as the connection targets of a new hyperedge, and a hyperedge is established between these four supervertices to form the initial new hyperedge.

[0205] Step S40176: Calculate the association strength of the initial new hyperedge to obtain a new hyperedge association strength value. The new hyperedge association strength value is calculated by the synchronization index value of the candidate hypervertex group, the inverse of the spatial neighborhood distance, and the strength of the historical potential collaborative change relationship.

[0206] The new hyperedge association strength value reflects the strength of the association between the hypervertices connected by the new hyperedge. Calculating the new hyperedge association strength value by combining the synchrony index value of the candidate hypervertex group, the inverse of the spatial neighborhood distance, and the strength of the historical potential co-variation relationship can more comprehensively consider the association factors between hypervertices. Specifically, the synchrony index value, the inverse of the spatial neighborhood distance, and the strength of the historical potential co-variation relationship of the candidate hypervertex group are first obtained. These indicators are then combined to calculate the new hyperedge association strength value, for example, using the formula s=a×syn+b×(1 / d)+c×h (where syn is the synchrony index value, d is the spatial neighborhood distance, h is the strength of the historical potential co-variation relationship, and a, b, and c are weight coefficients).

[0207] Step S40177: Perform spatial topology conflict detection on the initial new hyperedge to detect whether there is a spatial topology conflict among the supervertices connected by the hyperedge. Spatial topology conflict includes the situation where the spatial area units corresponding to the supervertices are discontinuously distributed in the physical space or isolated by other supervertices.

[0208] Spatial topology conflict detection is used to ensure that the topological relationships between the hypervertices connected by the new hyperedge are reasonable in physical space. Spatial topology conflicts can cause the associations between hypervertices connected by the hyperedge to be inconsistent with the actual physical situation and require detection and resolution.

[0209] Spatial analysis methods can be used to detect spatial topological conflicts in initial new hyperedges. For example, using Geographic Information System (GIS) technology, the location information of the spatial area units corresponding to the hypervertices can be visualized on a map to examine the spatial relationships between them. If the spatial area units corresponding to the hypervertices connected by an initial new hyperedge are found to be discontinuous in physical space or isolated by other hypervertices, the initial new hyperedge is considered to have a spatial topological conflict.

[0210] Step S40178: Eliminate initial new hyperedges with spatial topology conflicts, normalize the new hyperedge association strength values ​​of the remaining initial new hyperedges, and obtain new hyperedge weight values.

[0211] Eliminating initial new hyperedges with spatial topological conflicts can ensure the rationality of the hypergraph structure. Normalizing the new hyperedge association strength values ​​of the remaining initial new hyperedges can convert the association strength values ​​to the range of [0, 1] to facilitate subsequent processing.

[0212] The following method can be used to normalize the new hyperedge association strength values ​​of the remaining initial new hyperedges. First, find the maximum and minimum new hyperedge association strength values ​​among the remaining initial new hyperedges. Then, for each new hyperedge association strength value, normalize it using the formula w = (s - min) / (max - min) (where s is the new hyperedge association strength value, min is the minimum value, and max is the maximum value) to obtain the new hyperedge weight value.

[0213] Step S40179: assigning the new hyperedge weight value to the corresponding initial new hyperedge to obtain a new hyperedge that can be added to the radar echo hypergraph structure, thereby completing the supplement of the newly emerged multi-element association relationship.

[0214] Assigning new hyperedge weights to the corresponding initial hyperedges can give the new hyperedges reasonable weights and accurately reflect the associations between hypervertices. Adding these new hyperedges to the radar echo hypergraph structure can supplement the newly emerged multivariate associations.

[0215] The process of assigning the new hyperedge weight value to the corresponding initial new hyperedge can be to directly record the new hyperedge weight value into the information of the corresponding initial new hyperedge. For example, in the radar echo hypergraph structure, each hyperedge has a weight attribute. The new hyperedge weight value is assigned to the weight attribute of the corresponding initial new hyperedge to obtain a new hyperedge that can be added to the radar echo hypergraph structure.

[0216] Step S4018: Integrate the updated hypervertices, adjusted hyperedges, and newly constructed hyperedges to obtain an updated radar echo hypergraph structure.

[0217] By integrating the updated hypervertices, adjusted hyperedges and newly constructed hyperedges, a complete updated radar echo hypergraph structure can be obtained, which can accurately reflect the latest radar echo conditions and multivariate correlation relationships.

[0218] Integrating updated hypervertices, adjusted hyperedges, and newly constructed hyperedges can be achieved through data structures. For example, a graph database can be used to store the radar echo hypergraph structure, with updated hypervertices as nodes, and adjusted and newly constructed hyperedges as edges. These are organized and stored according to a specific data structure. During the integration process, it is necessary to ensure that the connections between hypervertices and hyperedges are correct.

[0219] Step S402: inputting the updated radar echo hypergraph structure into the hypergraph enhanced diffusion model, and generating dynamic hypergraph embedding features through the hypergraph feature extraction module. The dynamic hypergraph embedding features contain the latest multivariate association relationship information.

[0220] The hypergraph feature extraction module extracts useful features from the updated radar echo hypergraph structure to generate dynamic hypergraph embedding features. The dynamic hypergraph embedding features contain the latest multivariate relationship information and can reflect the dynamic changes in radar echo data.

[0221] The process of inputting the updated radar echo hypergraph structure into the hypergraph enhanced diffusion model and generating dynamic hypergraph embedding features is similar to the process described in steps S330-S345. First, the super-vertex embedding submodule converts the attribute information of the supervertex into a low-dimensional dense super-vertex embedding vector. Then, the hyper-edge information aggregation submodule generates hyper-edge aggregation features based on the super-vertex embedding vector and the hyper-edge connection relationship. Finally, the cross-layer feature fusion submodule fuses the hyper-edge aggregation features with the original hierarchical features in the feature transformation layer of the basic diffusion model to obtain the dynamic hypergraph embedding features.

[0222] Step S403: The dynamic hypergraph embedding features are fused with the spatiotemporal features of the radar echo original data sequence to obtain enhanced input features. The spatiotemporal features are extracted by performing temporal convolution and spatial convolution on the original data sequence.

[0223] Spatiotemporal features are extracted by performing temporal and spatial convolution on the raw radar echo data sequence. They contain characteristic information about the radar echo data in both time and space. Fusion of dynamic hypergraph embedding features with spatiotemporal features enables the model to utilize both hypergraph and spatiotemporal information, improving its predictive capabilities.

[0224] The dynamic hypergraph embedding features can be fused with the spatiotemporal features of the radar echo raw data sequence using a method called feature concatenation and attention-weighted summation. First, the dynamic hypergraph embedding features and spatiotemporal features are concatenated along the feature dimension to produce a concatenated feature vector. Then, an attention mechanism is used to calculate the attention weight for each feature dimension. The concatenated feature vector and the attention weights are weighted summed to obtain the enhanced input features.

[0225] Step S404: inputting the enhanced input features into the generation module of the hypergraph enhanced diffusion model, setting the extrapolation time step parameter, and dynamically adjusting the extrapolation time step parameter according to the historical evolution speed of the radar echo.

[0226] The generation module in the hypergraph-enhanced diffusion model is used to generate future radar echo data. The extrapolation time step parameter determines the time interval between the generation of future radar echo data. Dynamically adjusting the extrapolation time step parameter based on the historical evolution rate of radar echoes allows the model to more flexibly adapt to different radar echo variations.

[0227] Specifically, the enhanced input features are first fed into the neurons of the generation module, which then perform a weighted summation and linear transformation on the input data. The extrapolation time step parameter is then dynamically adjusted based on the historical evolution speed of the radar echo. For example, if the historical evolution speed of the radar echo is fast, the extrapolation time step parameter is set to a smaller value; if the historical evolution speed of the radar echo is slow, the extrapolation time step parameter is set to a larger value.

[0228] Step S405: Through the multi-step iterative sampling process of the generation module, in each sampling step, the spatial distribution change and time evolution direction of the echo intensity are guided by the dynamic hypergraph embedding feature to generate the sampling feature of the corresponding time step.

[0229] The generation module's multi-step iterative sampling process gradually generates future radar echo signatures. In each sampling step, the dynamic hypergraph embedding features guide the spatial distribution and temporal evolution of echo intensity, ensuring that the generated sampling signatures are more consistent with actual radar echo variations.

[0230] Specifically, in the first sampling step, the sampling features of the first time step are generated based on the enhanced input features and the extrapolated time step parameters. In subsequent sampling steps, the sampling features of the next time step are generated based on the sampling features generated in the previous sampling step and the dynamic hypergraph embedding features. In each sampling step, the dynamic hypergraph embedding features affect the generation process of the sampling features, guiding the spatial distribution changes and temporal evolution direction of the echo intensity. For example, in a generation process with five sampling steps, in the third sampling step, the sampling features of the third time step are generated based on the sampling features generated in the second sampling step and the dynamic hypergraph embedding features.

[0231] Step S406: Integrate the sampling features of all sampling steps, convert the sampling features into radar echo intensity values ​​through the intensity recovery algorithm, and obtain the radar echo extrapolation data sequence of the future period.

[0232] Integrating the sampling features of all sampling steps can generate a feature sequence containing multiple time steps. The intensity recovery algorithm can convert the sampling features into actual radar echo intensity values, thereby obtaining a radar echo extrapolation data sequence for the future period.

[0233] The sampling features from all sampling steps can be integrated using a feature concatenation method, where the sampling features from all sampling steps are concatenated in the time dimension to produce a feature sequence. The intensity recovery algorithm can convert the sampling features into radar echo intensity values ​​based on their distribution and statistical properties. For example, a linear or nonlinear function can be used to map the sampling features to the range of radar echo intensity values, resulting in a radar echo extrapolation data sequence for the future time period. For example, assuming that five sampling features are generated through five sampling steps, these five sampling features are concatenated in the time dimension to produce a feature sequence. This feature sequence can then be converted into a radar echo extrapolation data sequence for the next five time steps using the intensity recovery algorithm.

[0234] It is understandable that the various algorithms involved in the above-mentioned introductions of the embodiments of the present invention, such as the Euclidean distance algorithm, the cosine distance algorithm, the tracking algorithm, the segmentation algorithm, the clustering algorithm, etc., can all be learned from the relevant content in the prior art. In order to save space, they will not be expanded too much in the embodiments of the present invention. In addition, when implementing the scheme of the present invention, those skilled in the art can supplement the details according to the common knowledge in this field. For example, according to the common knowledge in this field, normalization can be used to eliminate dimensional conflicts before feature fusion, interpolation can be used to eliminate dimensional differences, and thresholds can be reasonably set based on historical data, experience or business scenario requirements. The model can be trained based on a general model training method, and the number of layers in the model structure can be set based on actual needs, the activation function can be selected, etc. The present invention will no longer provide redundant introductions to the overly detailed implementation process.

[0235] See also Figure 2 , Figure 2This is a schematic diagram of the structure of a computer system provided in an embodiment of the present invention. The computer system includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 may be connected via a bus or other means. The processor 101 (also known as the Central Processing Unit (CPU)) is the computing and control core of the computer system, capable of parsing various instructions within the computer system and processing various data within the computer system. The communication interface 102 may optionally include a standard wired interface or a wireless interface (such as Wi-Fi, a mobile communication interface, etc.), which can be used to send and receive data under the control of the processor 101. The communication interface 102 may also be used for data transmission and interaction within the computer system. The memory 103 is a storage device in the computer system for storing programs and data. It is understood that the memory 103 herein may include both the built-in memory of the computer system and, of course, the extended memory supported by the computer system. The memory 103 provides storage space, which stores the computer system's operating system, but this is not limited to this in the present invention.

[0236] In one embodiment, the processor 101 executes the radar echo extrapolation processing method based on the hypergraph enhanced diffusion model provided in the above embodiment of the present invention by running the computer program in the memory 103 .

Claims

1. A radar echo extrapolation processing method based on a hypergraph enhanced diffusion model, characterized in that: include: Acquire a radar echo raw data sequence for a continuous period, wherein the radar echo raw data sequence is composed of a plurality of radar echo data frames arranged in chronological order, each radar echo data frame including spatially distributed echo intensity information and a corresponding acquisition time identifier; Constructing a hypergraph structure on the radar echo raw data sequence to obtain a radar echo hypergraph structure, wherein the radar echo hypergraph structure includes a plurality of hypervertices and hyperedges connecting the hypervertices, wherein the hypervertices are used to represent spatial area units in the radar echo data, and the hyperedges are used to represent multivariate association relationships between different spatial area units; The radar echo hypergraph structure is integrated into a preset diffusion model framework to obtain a hypergraph enhanced diffusion model, and the hypergraph enhanced diffusion model processes the radar echo hypergraph structure through a hypergraph feature extraction module to generate a hypergraph embedding feature containing multivariate correlation information; The radar echo raw data sequence is extrapolated based on the hypergraph enhanced diffusion model to obtain a radar echo extrapolated data sequence for a future period, wherein the radar echo extrapolated data sequence includes a plurality of extrapolated data frames that are time sequentially continuous with the radar echo raw data sequence.

2. The method according to claim 1, wherein The step of constructing a hypergraph structure on the radar echo raw data sequence to obtain a radar echo hypergraph structure includes: Performing spatial region division on each radar echo data frame in the radar echo raw data sequence to obtain a plurality of spatial region units, each spatial region unit corresponding to a continuous spatial sub-region in the radar echo data frame, wherein a boundary of the spatial region unit is determined by an echo intensity gradient variation characteristic; Performing feature extraction on each of the spatial region units to obtain spatial region features; Each of the spatial region units is used as a super vertex in the radar echo hypergraph structure, and the attribute information of the super vertex is represented by the corresponding spatial region feature; Determining a cross-time correlation relationship between the supervertices based on radar echo data frames with different time tags in the radar echo raw data sequence, wherein the cross-time correlation relationship is determined by an echo intensity change trend and position offset characteristics of the same spatial area unit in consecutive time frames; Based on the cross-temporal association relationship and the spatial position relationship characteristics between the spatial area units, constructing a hyperedge connecting the multiple hypervertices, each hyperedge connecting at least three hypervertices, so as to characterize the coordinated change relationship between the multi-dimensional spatial area units; Assigning weights to the hyperedges to obtain weighted hyperedges, wherein the weights of the weighted hyperedges are determined by the strength of the cross-temporal association relationship and the closeness of the spatial position relationship; The hypervertices, the weighted hyperedges, and the attribute information of the hypervertices are integrated to obtain the radar echo hypergraph structure.

3. The method according to claim 1, wherein The step of constructing a hypergraph structure on the radar echo raw data sequence to obtain a radar echo hypergraph structure includes: Performing spatiotemporal correlation analysis on the radar echo raw data sequence to determine a plurality of spatial region groups having coordinated variation characteristics, each of the spatial region groups comprising at least three spatial region units that vary synchronously in the time dimension; Taking the spatial area units in each of the spatial area groups as an associated supervertex set, constructing an initial hyperedge based on the associated supervertex set, wherein the connection range of the initial hyperedge covers all the spatial area units in the associated supervertex set; Performing association strength calculation on the initial hyperedge to obtain a hyperedge association strength value, wherein the hyperedge association strength value is calculated by jointly calculating the similarity and change phase difference of the echo intensity change curves of each spatial area unit in the associated hypervertex set; Screening the initial hyperedges according to the hyperedge association strength value, and retaining the initial hyperedges whose hyperedge association strength value is greater than a preset threshold as valid hyperedges; Performing spatial topology verification on the valid hyperedges to eliminate valid hyperedges with spatial topology conflicts, wherein the spatial topology conflicts include situations where the spatial area units connected by the hyperedges are not adjacent in physical space and have no indirect association paths; The effective hyperedges that have been screened and verified are integrated with the corresponding hypervertices to obtain the radar echo hypergraph structure containing multi-element association relationships.

4. The method according to claim 2, wherein The radar echo hypergraph structure is integrated into a preset diffusion model framework to obtain a hypergraph enhanced diffusion model, including: Obtaining a preset basic diffusion model, wherein the basic diffusion model includes an input layer, a feature transformation layer, and an output layer, wherein the feature transformation layer is used to generate target data through an iterative denoising and de-noising process; Inserting a hypergraph feature extraction module into the feature transformation layer of the basic diffusion model, wherein the hypergraph feature extraction module includes a super-vertex embedding submodule, a super-edge information aggregation submodule, and a cross-layer feature fusion submodule; Inputting the radar echo hypergraph structure into the super-vertex embedding submodule, converting the attribute information of the super-vertex into a low-dimensional dense super-vertex embedding vector through a graph embedding algorithm, wherein the super-vertex embedding vector retains the key information of the spatial region feature; Inputting the super-vertex embedding vector and the weighted hyperedge into the super-edge information aggregation submodule, aggregating the neighborhood super-vertex information of each super-vertex through a hypergraph convolution operation to obtain a super-edge aggregation feature, wherein the super-edge aggregation feature includes information on the impact of multi-element associations on the super-vertex feature; Inputting the hyperedge aggregation features into the cross-layer feature fusion submodule, and fusing them with the original hierarchical features in the basic diffusion model feature transformation layer to obtain enhanced features of the fused hypergraph information, wherein the fusion process is achieved through feature splicing and attention weighted summation; The enhanced features are input into the subsequent network layers of the basic diffusion model, and the denoising sampling process of the diffusion model is adjusted so that the enhanced features continuously guide the spatiotemporal evolution direction of the echo features in iterative denoising, thereby obtaining the hypergraph enhanced diffusion model.

5. The method according to claim 4, wherein Inputting the radar echo hypergraph structure into the super-vertex embedding submodule and converting the attribute information of the super-vertex into a low-dimensional dense super-vertex embedding vector through a graph embedding algorithm includes: Normalizing the attribute information of the super vertex to obtain a standardized attribute feature; Inputting the standardized attribute features into a preset embedding neural network, wherein the embedding neural network comprises at least two fully connected layers and one nonlinear activation layer, wherein the number of neurons in the fully connected layer gradually decreases from the input dimension to the target embedding dimension; Performing a nonlinear transformation on the standardized attribute features through the embedding neural network to generate a preliminary embedding vector; Based on the hyperedge connection relationship in the radar echo hypergraph structure, a hypervertex similarity matrix is ​​constructed, wherein the element values ​​in the hypervertex similarity matrix are determined by the hyperedge weight values ​​and spatial distances between corresponding hypervertices; The preliminary embedding vector is smoothed by the super-vertex similarity matrix to obtain the super-vertex embedding vector.

6. The method according to claim 4, wherein The step of inputting the super-vertex embedding vector and the weighted hyper-edge into the hyper-edge information aggregation submodule and aggregating the neighborhood hyper-vertex information of each super-vertex through a hypergraph convolution operation to obtain a hyper-edge aggregation feature includes: For each super vertex, determine all hyperedges directly connected to it to form an associated hyperedge set; For each hyperedge in the associated hyperedge set, obtain all hypervertices connected by the hyperedge to form a hyperedge vertex set; Calculating the contribution weight of each super vertex in the super edge vertex set to the current super vertex, wherein the contribution weight is determined by the super edge weight value and the feature similarity between the super vertices; Performing weighted summation on the hypervertex embedding vectors of all hypervertices in the hyperedge vertex set based on the contribution weights to obtain a single hyperedge aggregation vector; Average pooling is performed on the single hyperedge aggregation vectors corresponding to all hyperedges in the associated hyperedge set to obtain the hyperedge aggregation feature, which is used to comprehensively reflect the feature aggregation result of the current hypervertex in the multi-element association relationship.

7. The method according to claim 1, wherein The step of extrapolating the radar echo original data sequence based on the hypergraph enhanced diffusion model to obtain a radar echo extrapolated data sequence for a future period includes: Convert the last N radar echo data frames in the radar echo raw data sequence into a model input tensor, where the dimensions of the model input tensor include a time dimension, a space dimension, and an intensity dimension, and N is an integer greater than or equal to 3; Inputting the model input tensor into the input layer of the hypergraph enhanced diffusion model, and converting it into an initial feature tensor that can be processed by the diffusion model through feature mapping processing of the input layer; Inputting the initial feature tensor into the feature transformation layer of the hypergraph enhanced diffusion model, in each diffusion step, the hypergraph feature extraction module extracts features from the radar echo hypergraph structure, generates hypergraph embedding features, and fuses the hypergraph embedding features with the intermediate features of the current diffusion step; De-noising the fused intermediate features through the denoising network of the feature transformation layer; After completing a preset number of diffusion steps, the final denoised features are input into the output layer of the hypergraph enhanced diffusion model, and the radar echo extrapolation data frame at the first moment in the future is generated through the deconvolution operation and intensity mapping of the output layer; The generated radar echo extrapolated data frame at the first moment in the future is used as a new input, and the generated data frame is combined with the other data frames in the radar echo original data sequence except the first data frame to form a new input sequence, and the above extrapolation steps are repeated to generate the radar echo extrapolated data frame at the second moment in the future; The extrapolation steps are iteratively performed in sequence until a preset number of radar echo extrapolation data frames for future time periods are generated, and all the extrapolation data frames are arranged in chronological order to obtain the radar echo extrapolation data sequence.

8. The method according to claim 7, wherein The initial feature tensor is input into the feature transformation layer of the hypergraph enhanced diffusion model. In each diffusion step, the hypergraph feature extraction module extracts features from the radar echo hypergraph structure to generate hypergraph embedding features, and fuses the hypergraph embedding features with the intermediate features of the current diffusion step, including: At the beginning of each diffusion step, the intermediate feature tensor of the current step and the corresponding diffusion time step parameter are obtained; Inputting the radar echo hypergraph structure into the hypergraph feature extraction module, generating a hypervertex embedding vector through the hypervertex embedding submodule, and generating a hyperedge aggregation feature through the hyperedge information aggregation submodule; Inputting the hyperedge aggregation feature into the cross-layer feature fusion submodule, and performing dimension matching with the intermediate feature tensor, so that the spatial dimension and channel dimension of the hyperedge aggregation feature are consistent with the intermediate feature tensor; Performing element-wise multiplication on the matched hyperedge aggregation features and the intermediate feature tensor to obtain weighted fusion features, wherein the element-wise multiplication enhances the feature components in the intermediate feature tensor that are related to the multivariate association relationship of the hypergraph; The weighted fusion feature is input into the denoising network of the feature transformation layer, denoising is performed in combination with the diffusion time step parameter, the intermediate feature tensor is updated, and the feature transformation of the current diffusion step is completed.

9. The method according to claim 1, wherein The step of extrapolating the radar echo original data sequence based on the hypergraph enhanced diffusion model to obtain a radar echo extrapolated data sequence for a future period includes: The radar echo hypergraph structure is updated, and attribute information of hypervertices and weight values ​​of hyperedges are adjusted according to the latest radar echo data frame in the radar echo raw data sequence; Inputting the updated radar echo hypergraph structure into the hypergraph enhanced diffusion model, and generating dynamic hypergraph embedding features through the hypergraph feature extraction module, wherein the dynamic hypergraph embedding features contain the latest multivariate association relationship information; The dynamic hypergraph embedding feature is fused with the spatiotemporal features of the radar echo raw data sequence to obtain enhanced input features, wherein the spatiotemporal features are extracted by performing temporal convolution and spatial convolution on the raw data sequence; Inputting the enhanced input features into a generation module of the hypergraph enhanced diffusion model, setting an extrapolation time step parameter, and dynamically adjusting the extrapolation time step parameter according to the historical evolution speed of the radar echo; Through the multi-step iterative sampling process of the generation module, in each sampling step, the spatial distribution change and temporal evolution direction of the echo intensity are guided by the dynamic hypergraph embedding feature, and the sampling features corresponding to the time step are generated; The sampling features of all sampling steps are integrated and converted into radar echo intensity values ​​to obtain the radar echo extrapolation data sequence for the future period.

10. A computer system, characterized in that: include: a memory storing a computer program; A processor, configured to load the computer program to implement the radar echo extrapolation processing method based on the hypergraph enhanced diffusion model according to any one of claims 1 to 9.

Images