A hierarchical hypergraph neural network method and system for rainfall prediction

Abstract

This invention discloses a hierarchical hypergraph neural network method and system for rainfall prediction in the field of meteorological station rainfall prediction technology. The method includes: extracting features from acquired meteorological index data and meteorological text data of each meteorological station to obtain meteorological features for each station; performing relationship interaction on the meteorological features of each meteorological station to obtain relationship features for each station; and concatenating the meteorological features and relationship features of each station and inputting them into a feedforward neural network for rainfall prediction to obtain the predicted rainfall for each meteorological station. This invention, by introducing an attention mechanism with time decay and suppression effect capture capabilities and a dynamically adaptive hypergraph message passing mechanism, makes the modeling of the temporal dynamics of rainfall processes more accurate, adapts to dynamic weather environments, and achieves precise prediction of rainfall at meteorological stations.

Description

Technical Field

[0001] This invention relates to the field of meteorological station rainfall prediction technology, and in particular to a hierarchical hypergraph neural network rainfall prediction method and system. Background Technology

[0002] Traditional rainfall forecasting mainly relies on numerical weather prediction models, which extrapolate future weather by solving physical equations. However, in short-term forecasting, these models are limited by initial errors, parameterization schemes, and resolution, resulting in systematic biases that often require post-processing techniques for correction.

[0003] In recent years, machine learning has provided new avenues for rainfall forecasting. Deep learning methods, such as recurrent neural networks, convolutional neural networks, and Transformer models, can learn complex mappings from historical data, effectively capturing temporal dependencies and spatial features. However, most of these methods treat stations or grids as independent units, neglecting the widespread spatial correlations and high-order synergies within the atmospheric system.

[0004] Graph neural networks have partially improved spatial modeling by constructing binary relationship graphs between stations, but they still struggle to effectively represent the collective impact relationships in meteorology, such as the synergistic effect of a weather system on multiple stations in a region, or the combined effect of terrain on a group of stations. Modeling these high-order relationships using traditional graph structures requires building dense connections, which easily introduces noise and results in computational redundancy.

[0005] Hypergraph theory provides a new paradigm for this. A hypergraph allows a single hyperedge to connect any number of nodes, naturally representing group relationships such as "a cluster of stations affected by the same weather system." Hypergraph-based neural networks, through message passing between nodes and hyperedges, can fuse group information and embed prior meteorological knowledge.

[0006] However, the atmospheric system is highly dynamic, and meteorological correlations constantly evolve with the formation and dissipation of weather systems, leading to a dual instability in the hypergraph structure: on the one hand, the strength of relationships between nodes within the hyperedge changes in real time, exhibiting dynamic vulnerability; on the other hand, the hyperedge itself dynamically increases or decreases with the formation or dissipation of weather systems. Simultaneously, meteorological processes exhibit time decay characteristics—recent events are more indicative of short-term precipitation, but existing methods are insufficient in modeling this, limiting forecast timeliness. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a hierarchical hypergraph neural network rainfall prediction method and system. By introducing an attention mechanism with the ability to capture time decay and inhibition effects and a dynamically adaptive hypergraph message passing mechanism, the modeling of the temporal dynamics of rainfall processes becomes more accurate, which can adapt to dynamic weather environments and achieve accurate prediction of rainfall at meteorological stations.

[0008] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution:

[0009] In a first aspect, the present invention provides a hierarchical hypergraph neural network method for predicting rainfall, comprising:

[0010] Feature extraction was performed on the meteorological index data and meteorological text data of each meteorological station to obtain the meteorological characteristics of each meteorological station.

[0011] The meteorological characteristics of each meteorological station are analyzed to obtain the relationship characteristics of each meteorological station.

[0012] The meteorological characteristics and relational characteristics of each meteorological station are concatenated and input into a feedforward neural network to predict rainfall, thus obtaining the predicted rainfall for each meteorological station.

[0013] Optionally, the step of extracting features from the acquired meteorological index data and meteorological text data of each meteorological station to obtain the meteorological characteristics of each meteorological station includes:

[0014] The temperature, humidity, wind speed and wind direction data of each meteorological station were normalized by maximum and minimum values, and the rainfall was normalized by logarithmic normalization to obtain the normalized meteorological index data of each meteorological station.

[0015] The acquired meteorological text data from each meteorological station were normalized to obtain normalized meteorological text data for each meteorological station.

[0016] The normalized meteorological index data and normalized meteorological text data of each meteorological station are input into the multilayer perceptron mixer for feature splicing and feature mixing operations to obtain the fused features of each meteorological station.

[0017] The meteorological characteristics of each meteorological station are obtained by linear projection and multi-head attention calculation of the fusion features of each meteorological station using the time decay self-attention mechanism.

[0018] Optionally, the expression for the maximum and minimum value normalization process is as follows:

[0019] ,

[0020] in, This represents the data after normalization of the maximum and minimum values. This represents the data before normalization of the maximum and minimum values. express The minimum value in, express The maximum value in;

[0021] The logarithmic normalization process is achieved through the following formula:

[0022] ,

[0023] in, This represents the data before logarithmic normalization. This represents the data after logarithmic normalization.

[0024] Optionally, the fusion characteristics of each weather station are obtained using the following formula:

[0025] ,

[0026]

[0027] in, For weather station indexing, Indicates weather station The fusion characteristics Indicates transpose. Indicates splicing characteristics, This represents the trainable weights of the first fully connected layer. This represents the trainable weights of the second fully connected layer. express Activation function Presentation layer normalization operation, Indicates weather station Normalized meteorological index data, Indicates weather station Normalized meteorological text data, This represents a vector concatenation operation;

[0028] The meteorological characteristics of each meteorological station are obtained using the following formula:

[0029] ,

[0030] ,

[0031] ,

[0032] in, Represents a set of meteorological characteristics. The total number of weather stations. Indicates weather station meteorological characteristics, Indicates weather station The fusion characteristics Represents the set of fused features. Represents the query vector. Represents the key vector. Represents the position mask matrix, Represents a value vector. Indicates the query weight. Indicates key weight. Indicates value weight, Indicates the attention head index, This indicates the total number of heads of attention. Represents a symbolic function. This represents the activation function. Represents absolute value. Scaling factor This represents the deviation matrix.

[0033] Optionally, the step of performing relational interaction on the meteorological characteristics of each meteorological station to obtain relational characteristics of each meteorological station includes:

[0034] A multi-relationship hypergraph is constructed based on the relationship types between the meteorological stations; wherein, the multi-relationship hypergraph includes a set of meteorological station nodes, a set of hyperedges, and an association matrix;

[0035] Input the meteorological characteristics of each meteorological station into A multilayer perceptron was used to obtain initial feature representations of weather station nodes under different relation types; among them... Indicates the total number of relation types;

[0036] The initial weather station node feature representation is aggregated based on the correlation matrix to obtain the initial hyperedge feature representation.

[0037] Based on the initial weather station node feature representation and the initial hyperedge feature representation, the attention score between the weather station node and the hyperedge is calculated using the maskless hypergraph attention mechanism, and the correlation matrix is ​​updated using the attention score to obtain the updated correlation matrix;

[0038] Based on the updated association matrix and the initial weather station node feature representation, the weather station node feature representation under different relation types is obtained by using hypergraph convolution based on hyperedge aggregator.

[0039] Based on the node feature representation of the meteorological stations, the relationship features of each meteorological station are obtained using a hypergraph-level attention mechanism.

[0040] Optionally, the step of obtaining meteorological station node feature representations under different relation types using hypergraph convolution based on hyperedge aggregators, according to the updated association matrix and the initial meteorological station node feature representation, includes:

[0041] Based on the updated correlation matrix and the initial weather station node feature representation, the embedding representation of the hyperedge is calculated;

[0042] Calculate the comprehensive environmental characterization under different relationship types based on the initial meteorological station node feature representation;

[0043] Based on the comprehensive environmental characterization and the embedding representation of hyperedges, a gating network is used to obtain the hyperedge weight matrix;

[0044] The feature representations of weather station nodes under different relation types are calculated based on the hyperedge weight matrix, the updated correlation matrix, and the embedding representation of the hyperedge.

[0045] Optionally, the expression for the multi-relationship hypergraph is:

[0046] ,

[0047] ,

[0048] ,

[0049] ,

[0050] ,

[0051] in, This represents a multi-relationship hypergraph. For a set of relation types, For relation type indexing, Representing relation type The supergraph For the set of weather station nodes, Indicates the first Each weather station node, Indicates the total number of weather stations; Let be a hyperedge set, representing the connection relationships between weather station nodes; Representing relation type The set of superedges below, Representing relation type The correlation matrix below, , This represents the meteorological station node index. Representing relation type Next weather station node With super-edge The relationship between them .

[0052] Optionally, the initial weather station node feature representation is obtained by the following formula:

[0053] ,

[0054] in, Indicates the type used to handle relations. Multilayer perceptron, Indicates meteorological characteristics, Indicates in relation type Initial weather station node feature representation, , Indicates in relation type Next Initial weather station node characteristics representation for the day. , Indicates in relation type Next Weather station nodes of the day The initial feature representation;

[0055] The initial hyperedge feature representation is obtained by the following formula:

[0056] ,

[0057] in, Indicates in relation type Next The first day A super edge, Indicates the superedge The initial feature representation, Indicates belonging to a superedge The set of all nodes , This represents the meteorological station node index. Indicates in relation type Next Weather station nodes of the day The initial feature representation, Indicates average operation;

[0058] The attention score is obtained using the following formula:

[0059] ,

[0060] in, Indicates the first Weather Station Node With super-edge Attention score Indicates the first Weather Station Node With super-edge The correlation, , express Activation function It is a feedforward neural network. For learnable weight matrix, Indicates the first Weather Station Node meteorological characteristics, Indicates the first Weather Station Node With super-edge The correlation, As a filter factor, This represents the average correlation. ;

[0061] The updated correlation matrix is ​​obtained using the following formula:

[0062] ,

[0063] in, Representing relation type Next The correlation matrix of days, Representing relation type Next Weather Station Node With super-edge The relationship between them;

[0064] The embedding representation of the hyperedge is obtained by the following formula:

[0065] ,

[0066] in, Indicates in relation type Next Embedded representation of sky super edges Indicates in relation type The hyper-edge degree diagonal matrix below;

[0067] The relation type hyper-edge degree diagonal matrix The expression is as follows:

[0068] ,

[0069] in, Indicates in relation type The number of bottom edges, Denotes the first element on the diagonal of the hyper-dimension diagonal matrix. One element;

[0070] The comprehensive environmental characterization is obtained through the following formula:

[0071] ,

[0072] in, Indicates in relation type Next A comprehensive environmental characterization of the sky;

[0073] The hyperedge weight matrix is ​​obtained by the following formula:

[0074] ,

[0075] in, Indicates in relation type The hyperedge weight matrix below, , Represents a diagonal matrix. Represents the hyperedge weight matrix The first diagonal One element, express and splicing, Represents the hyperbolic tangent function. This represents the activation function. Indicates the bias term. The weight matrix is ​​a learnable weight matrix;

[0076] The feature representation of the meteorological station node is obtained by the following formula:

[0077] ,

[0078] in, Indicates in relation type Next The characteristics of weather station nodes on a given day are represented. Represents the degree matrix of weather station nodes;

[0079] The relationship characteristics of the various meteorological stations are obtained through the following formula:

[0080] ,

[0081] in, Indicates the first Weather Station Node Relationship characteristics, Indicates the first Weather Station Node Relationship types The importance of , For learnable weight vectors, Indicates in relation type Next Weather station nodes of the day The characteristics are represented.

[0082] Optionally, the predicted rainfall is obtained using the following formula:

[0083] ,

[0084] in, Indicates weather station The predicted rainfall, This represents a dense layer in a feedforward neural network. Indicates weather station Corresponding meteorological station nodes Relationship characteristics, Indicates weather station Meteorological characteristics.

[0085] In a second aspect, the present invention provides a hierarchical hypergraph neural network rainfall prediction system, comprising:

[0086] The feature extraction module is used to extract features from the acquired meteorological index data and meteorological text data of each meteorological station to obtain the meteorological features of each meteorological station.

[0087] The relationship interaction module is used to: perform relationship interaction on the meteorological characteristics of each meteorological station to obtain the relationship characteristics of each meteorological station;

[0088] The rainfall prediction module is used to: concatenate the meteorological characteristics and relational characteristics of each meteorological station and input them into a feedforward neural network to predict rainfall, thereby obtaining the predicted rainfall for each meteorological station.

[0089] Compared with existing technologies, the beneficial effects achieved by this invention are as follows:

[0090] (1) An attention mechanism with time decay characteristics was designed. By introducing a decay function, the timeliness of meteorological information value is explicitly modeled. Specifically, the decay coefficient is calculated using the current hidden layer state and location information, and the attention weights are readjusted. Combined with the relationship between meteorological stations, it can better capture the strength of meteorological correlation between stations and the changing pattern of rainfall trend over time. At the same time, negative attention weights are supported to capture the inhibition effect between meteorological elements, thereby improving the model's accuracy in modeling meteorological time dynamics.

[0091] (2) A three-stage hyperedge-sensitive message passing mechanism is proposed. This mechanism contains two core components: On the one hand, the maskless hypergraph attention mechanism realizes the autonomous reconstruction of topological connection relationship by evaluating the correlation strength between nodes and hyperedges in real time, so that the model can accurately strengthen effective meteorological correlations and adaptively weaken the noise interference caused by pseudo-correlations; On the other hand, the novel gating mechanism realizes the dynamic evaluation and adaptive selection of the importance of hyperedges by encoding real-time atmospheric environmental information into control signals, so as to activate or suppress specific hyperedges according to the weather system state. This comprehensive design enables the message passing mechanism to effectively cope with hyperedge instability. Attached Figure Description

[0092] Figure 1 This is a flowchart of a hierarchical hypergraph neural network rainfall prediction method provided according to an embodiment of the present invention;

[0093] Figure 2 This is a flowchart of feature extraction provided according to an embodiment of the present invention;

[0094] Figure 3 This is a flowchart illustrating the relationship interaction process provided by an embodiment of the present invention;

[0095] Figure 4 A flowchart for rainfall prediction provided according to an embodiment of the present invention. Detailed Implementation

[0096] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations thereof. In the absence of conflict, the embodiments and technical features in the embodiments can be combined with each other.

[0097] It should be noted that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0098] Example 1

[0099] This invention discloses a hierarchical hypergraph neural network method for predicting rainfall, with reference to... Figure 1 As shown, the specific steps include the following:

[0100] S1, extract features from the acquired meteorological index data and meteorological text data of each meteorological station to obtain the meteorological characteristics of each meteorological station;

[0101] S2, perform relationship interaction on the meteorological characteristics of each meteorological station to obtain the relationship characteristics of each meteorological station;

[0102] S3, the meteorological characteristics and relational characteristics of each meteorological station are concatenated and input into the feedforward neural network to predict rainfall, and the predicted rainfall of each meteorological station is obtained.

[0103] Specifically, this embodiment first obtains the meteorological observation data sequence of the target location meteorological station and its surrounding meteorological stations within the same historical period, including time series data of multi-dimensional meteorological elements such as rainfall, temperature, humidity, and air pressure.

[0104] Then, the Pearson correlation coefficient was used to perform correlation analysis between rainfall and historical data to confirm meteorological indicators. The formula for calculating the Pearson correlation coefficient is as follows:

[0105] ,

[0106] in, This indicates the correlation between meteorological elements and rainfall. Indicates the length of the time series. Indicates the meteorological elements in the first Weather observations for the day, Indicates the rainfall in the th Observations over the days Represents the mean of meteorological elements. This represents the average rainfall.

[0107] Based on correlation analysis, this embodiment collected five key meteorological factors for the target weather station and surrounding weather stations. ,in, Indicates weather station In the A collection of meteorological factors of the day Indicates weather station In the Rainfall of the day, Indicates weather station exist The temperature of the day, Indicates weather station In the The humidity of the sky Indicates weather station In the The wind speed of the day, Indicates weather station In the The wind direction in the sky.

[0108] Considering that meteorological forecast texts and weather report texts (including official weather forecasts from meteorological bureaus, outputs from international meteorological models (such as ECMWF and GFS), interpretation reports of meteorological radar and satellite images, and analyses from professional meteorological websites) also have a significant impact on rainfall prediction, this embodiment also includes meteorological text data (meteorological forecast texts and weather report texts) in the feature construction scope.

[0109] However, meteorological texts typically contain a large number of technical terms and uncertain descriptions, and their processing faces multiple challenges: the diverse sources of information (such as official forecasts from meteorological bureaus, outputs from international meteorological models, and reports from local meteorological stations) make it difficult to guarantee reliability and consistency; the prevalence of technical terms and probabilistic descriptions can easily lead to misunderstandings; and the differences in the spatiotemporal scales of language descriptions further increase the difficulty of extracting accurate forecast information from the text.

[0110] To overcome the above problems, this embodiment extracts five key features from meteorological text to characterize the relationship between the text and rainfall in a specific area. These features include: the correlation between meteorological events and rainfall at meteorological stations, the degree of certainty in the description of meteorological events, the importance of meteorological events, the predictive impact of meteorological events on rainfall levels, and the expected duration of meteorological events.

[0111] ,

[0112] in, Indicates weather station In the Meteorological text features of the day Indicates weather events and weather stations The correlation of rainfall is a value between 0 and 1; This indicates the degree of certainty in the description of a meteorological event, with a value between 0 and 1. 0 represents an uncertain description such as "possible" and 1 represents a certain description such as "certain" or "continuous". Indicate the importance level of the meteorological event; This indicates the predictive impact of meteorological events on rainfall levels, estimates the specific rainfall levels that may be caused by the meteorological text description, and normalizes to the range of 0-1. This indicates the expected duration of the impact of a weather event, expressed in hours and normalized to the range of 0-1.

[0113] In step S1, feature extraction is performed on the acquired meteorological index data and meteorological text data of each meteorological station to obtain the meteorological characteristics of each meteorological station, with reference to... Figure 2 As shown, it includes:

[0114] S1.1, the temperature, humidity, wind speed, and wind direction data from the acquired meteorological indicator data of each meteorological station are normalized by maximum and minimum values, and the rainfall is normalized by logarithmic normalization to obtain the normalized meteorological indicator data of each meteorological station:

[0115] The expression for the maximum and minimum value normalization process is as follows:

[0116] ,

[0117] in, This represents the data after normalization of the maximum and minimum values. This represents the data before normalization of the maximum and minimum values. express The minimum value in, express The maximum value in;

[0118] The logarithmic normalization process is achieved through the following formula:

[0119] ,

[0120] in, This represents the data before logarithmic normalization. This represents data after logarithmic normalization.

[0121] S1.2, Normalize the acquired meteorological text data of each meteorological station to obtain normalized meteorological text data of each meteorological station.

[0122] S1.3, the normalized meteorological index data and normalized meteorological text data of each meteorological station are input into the multilayer perceptron mixer for feature splicing and feature mixing operations to obtain the fused features of each meteorological station:

[0123] ,

[0124]

[0125] in, For weather station indexing, Indicates weather station The fusion characteristics Indicates transpose. Indicates splicing characteristics, This represents the trainable weights of the first fully connected layer. This represents the trainable weights of the second fully connected layer. express Activation function Presentation layer normalization operation, Indicates weather station Normalized meteorological index data, Indicates weather station Normalized meteorological text data, This represents a vector concatenation operation. In this embodiment, a multilayer perceptron mixer (MLP-Mixer) is used to learn the correlation between feature vectors, thereby enabling the extraction of high-level features. The MLP-Mixer can exchange information between feature dimensions without interfering with the time dimension, allowing homogeneous and heterogeneous indicators in the meteorological system to communicate closely, thereby enhancing the expressive power of the model.

[0126] S1.4, The fusion features of each weather station are linearly projected and multi-head attention is calculated using a time decay self-attention mechanism to obtain the meteorological features of each weather station.

[0127] Self-attention mechanisms calculate the degree of correlation between elements within a sequence and aggregate information based on this correlation weight, thereby generating a new representation for each element. This mechanism has successfully replaced recursive structures in language modeling and is widely used in time series modeling. However, self-attention mechanisms can only capture the similarity between hidden states and cannot effectively model the dynamic characteristics of the decay of the impact of meteorological events over time.

[0128] For example, the impact of short-term rainfall forecasts issued by meteorological bureaus on current rainfall predictions weakens over time, and the latest international meteorological model outputs are more valuable than historical data. However, when processing historical sequences, self-attention mechanisms cannot automatically assign higher weights to recent meteorological information; instead, they treat the similarity of all days equally, which is inconsistent with the actual timeliness of meteorological information. The state of the current day in a meteorological time series is influenced by the previous day, exhibiting both excitation effects (such as the strengthening of a persistent low-pressure system leading to an increased probability of rainfall) and inhibition effects (such as the establishment of a high-pressure system inhibiting convective activity and reducing the likelihood of rainfall; or the intrusion of a dry air mass inhibiting cloud development). Because the attention weights generated by self-attention mechanisms are always non-negative, they struggle to characterize the inhibition effects (i.e., negative correlations) existing in meteorological systems, thus limiting the model's ability to describe complex meteorological feedback mechanisms.

[0129] This embodiment uses the Reaction Point Process (RPPS) to enhance the self-attention mechanism by introducing adjustment and decay parameters to calculate the bias term, and then readjusting the attention value.

[0130] The intensity function of the reflected point process is:

[0131] ,

[0132] in, Indicates the first The basic strength of the sky, Indicates the first On the day that historical event occurred, and These represent the excitation and inhibition effects on a specific historical day, respectively. These inhibition effects weaken over time, prompting us to incorporate this concept into self-attention mechanisms, as the point process of self-attention mechanisms inherently detects excitation and inhibition. Furthermore, the positional information of the time series plays a crucial role in the runtime variations between different measurement modes. Self-attention mechanisms with sinusoidal relative position embeddings in the input perform poorly due to the loss of precise distance information, while the decay effect reflecting the point process provides accurate positional information for self-attention mechanisms.

[0133] The specific implementation process of the time-decay self-attention mechanism in this embodiment is as follows:

[0134] First, linear projection is used to obtain... :

[0135] ,

[0136] ,

[0137] in, The total number of weather stations. Represents the set of fused features. Represents the query vector. Represents the key vector. Represents a value vector. Indicates the query weight. Indicates key weight. Indicates the value weight;

[0138] Then learn the decay parameter based on V. and adjusting parameters And further calculate the deviation term ;

[0139] , ,

[0140] ,

[0141] in, It is a learnable matrix. It is a value vector In the The value of the day, Used to control the daily decay rate; A value greater than 0 indicates an excitatory effect. A value less than 0 indicates an inhibitory effect. Indicates the deviation term The Line number Column elements;

[0142] The computational process for multi-head attention can be represented as follows:

[0143] ,

[0144] in, Represents a set of meteorological characteristics. The total number of weather stations. Indicates weather station meteorological characteristics, Indicates weather station The fusion characteristics Representation of the position mask matrix:

[0145] ,

[0146] ,

[0147] in, Represents the position mask matrix The Line number Column elements.

[0148] In step S2, refer to Figure 3 As shown, based on meteorological features and predefined relationship knowledge, this embodiment constructs a multi-relationship hypergraph to model higher-order associations between meteorological phenomena. Subsequently, a maskless hypergraph attention mechanism is used to dynamically adjust the weight distribution of nodes in hyperedges to enhance the model's ability to perceive key nodes. Next, in the interaction within the hypergraph, a hypergraph convolution operation based on a hyperedge aggregator is adopted. This process is divided into two stages: first, information aggregation from nodes to hyperedges is performed; then, information aggregation from hyperedges to nodes is performed. In the information aggregation from hyperedges to nodes, an important hyperedge is first dynamically filtered through a novel gating mechanism. Then, to further address the homogenization problem in "hyperedge-node" message passing (i.e., the model may treat all nodes within the same hyperedge as a homogeneous whole and assign consistent influence weights), this embodiment normalizes the node features. This processing avoids erroneous correlation learning due to differences in node contribution, thereby obtaining a more accurate node feature representation, which helps to enhance the accuracy of hyperedge-node message passing. Finally, in the interaction between hypergraphs, an attention mechanism at the hypergraph level is used to coordinate the information contributions from different relational hypergraphs. This mechanism can adaptively weight and aggregate multi-source information to calculate the importance of each hypergraph.

[0149] 1. Construction of multi-relationship hypergraphs:

[0150] In this embodiment, the purpose of constructing the hypergraph is to establish meteorological domain knowledge hyperedges and node indexes for the hypergraph using predefined relational knowledge; the interaction within the hypergraph is used to learn the relationship of rainfall between stations within a specific hypergraph, while the interaction between hypergraphs is used to integrate information from different hypergraphs; this embodiment injects domain knowledge into the hyperedges between meteorological stations through the following three relationships:

[0151] Neighborhood Hypergraph: Geographically proximate meteorological stations are influenced by local atmospheric processes (such as mesoscale convection), and their precipitation processes exhibit high spatial correlation in terms of occurrence time, intensity, and duration. To capture this distance-determined synergistic effect, this embodiment constructs a neighborhood hypergraph, where each hyperedge... The hypergraph connects the groups of stations with the closest geodesic distance; this hypergraph ensures that when the model predicts rainfall at a station, it can give priority to the latest observations of its surrounding stations, providing a basic spatial correlation constraint for the prediction.

[0152] Topographic hypermap: Meteorological stations in the same topographic region often exhibit similar precipitation patterns due to the overall influence of the topography. To utilize this characteristic, this embodiment constructs a set of topographic hyperedges, where each hyperedge... It connects all meteorological stations belonging to the same terrain type (including elevation, slope, and aspect). Terrain directly regulates the dynamic lifting and thermal distribution of airflow through factors such as elevation, aspect, and slope, thereby systematically affecting the spatial distribution and intensity of precipitation. This results in stations within the same super-border exhibiting a high degree of consistency in precipitation statistical characteristics and physical mechanisms.

[0153] Pressure System Hypermap: Pressure systems are typically closely related to weather changes and precipitation processes. When a pressure system (such as a trough, ridge, or frontal system) develops, relevant meteorological stations often experience similar precipitation changes, thereby increasing the probability and intensity of precipitation. Therefore, this embodiment constructs a set of pressure system hyperedges, each hyperedge... Meteorological stations affected by the same type of pressure system are connected. Pressure systems dominate regional precipitation patterns by driving large-scale vertical motion, convergence and divergence, and changes in atmospheric stability. This hypermap helps to capture the co-evolution of precipitation fields in synoptic-scale systems.

[0154] Hidden Hypergraph: Predefined explicit relationships may not fully capture the complex high-order interaction patterns between sites. To uncover the underlying relational structures, this embodiment further utilizes historical site data to learn an implicit hypergraph, denoted as […]. This allows for the automatic identification and enhancement of unstated high-order dependencies in the data.

[0155] This results in the following multi-relation hypergraph:

[0156] ,

[0157] ,

[0158] ,

[0159] ,

[0160] ,

[0161] in, This represents a multi-relationship hypergraph. For a set of relation types, For relation type indexing, Indicates the relation type is The supergraph For the set of weather station nodes, Indicates the first Each weather station node, Indicates the total number of weather stations; Let be a hyperedge set, representing the connection relationships between weather station nodes; Representing relation type The set of superedges below, Representing relation type The correlation matrix below, , This represents the meteorological station node index. Representing relation type Next weather station node With super-edge The relationship between them .

[0162] 2. HyperMap Internal Interaction

[0163] While introducing a multi-relationship hypergraph helps discover the interactions between intelligent nodes grouped under different relations, this approach has two limitations: it cannot dynamically adjust the nodes on the hyperedges, nor can it adjust the number of hyperedges in a timely manner. To address these issues, this embodiment introduces maskless hypergraph attention and a hyperedge aggregator in the multi-relationship hypergraph interaction stage.

[0164] First, input the meteorological characteristics of each meteorological station into... A multilayer perceptron was used to obtain initial feature representations of weather station nodes under different relation types;

[0165] ,

[0166] in, Indicates the total number of relation types. Indicates the type used to handle relations. Multilayer perceptron, Indicates meteorological characteristics, Indicates in relation type Initial weather station node feature representation, , Indicates in relation type Next Initial weather station node characteristics representation for the day. , Indicates in relation type Next Weather station nodes of the day The initial feature representation;

[0167] Subsequently, the initial weather station node feature representations are aggregated based on the correlation matrix to obtain the initial hyperedge feature representations:

[0168] ,

[0169] in, Indicates in relation type Next The first day A super edge, Indicates the superedge The initial feature representation, Indicates belonging to a superedge The set of all nodes , This represents the meteorological station node index. Indicates in relation type Next Weather station nodes of the day The initial feature representation, This indicates the average operation.

[0170] (1) Maskless hypergraph attention mechanism:

[0171] In rainfall forecasting tasks, to capture the differences in the influence of each meteorological station on its related stations at different times, a hypergraph attention mechanism is often used to learn a pseudo-dynamic correlation matrix. However, this dynamism is limited to defining a predefined topology and cannot adapt to the complex changes in the actual atmospheric environment. In practical applications, even stations belonging to the same topographic region or pressure system may not have effective correlations due to differences in local meteorological conditions, and may even introduce noise interference.

[0172] To address this issue, this embodiment proposes a maskless hypergraph attention mechanism. Unlike traditional hypergraph attention mechanisms, this embodiment allows each meteorological station node to perform attention calculations with all hyperedges and dynamically readjusts the correlation matrix. Specifically, for each station node... First, based on the current site Attention coefficients are calculated based on meteorological and hyperedge features to quantify all station nodes. By assigning different levels of importance to hyperedges, the system adaptively strengthens physically significant connections and weakens invalid or redundant connections during message passing.

[0173] ,

[0174] in, Indicates the first Weather Station Node With super-edge The correlation, express Activation function It is a feedforward neural network. The weight matrix is ​​a learnable weight matrix;

[0175] Then, a threshold function is designed to filter out invalid hyperedge relationships:

[0176] ,

[0177] in, Indicates the first Weather Station Node With super-edge The attention score represents the relative importance weight that should be assigned to different parts during the information integration process. Indicates the first Weather Station Node meteorological characteristics, Indicates the first Weather Station Node With super-edge The correlation, As a filter factor, This represents the average correlation. .

[0178] Finally, the attention score is used to update the association matrix, resulting in the updated association matrix:

[0179] ,

[0180] in, Representing relation type Next The correlation matrix of days, Representing relation type Next Weather Station Node With super-edge The relationship between them.

[0181] Through the maskless hypergraph attention mechanism, this embodiment can dynamically adjust the node distribution on each hyperedge: both filtering out invalid nodes on the original hyperedge and adding new nodes that are strongly associated with the hyperedge.

[0182] (2) Hypergraph convolution based on hyperedge aggregator:

[0183] However, hypergraph convolution does not fully consider the interference of hyperedge validity on message passing in rainfall prediction applications. It assumes that all predefined hyperedges are equally important, which inevitably introduces noise into meteorological analysis. In real meteorological systems, the performance of different types of physical correlations is significantly affected by the overall atmospheric environment.

[0184] For example, during the development of severe convective weather or under conditions of abundant water vapor, the orographic lifting effect and local convective activity are often significantly enhanced, and related superedges play a crucial role in precipitation forecasting. However, under stable atmospheric stratification or arid conditions, the influence of these physical processes is significantly weakened, and they may even produce misleading signals. This coordinated movement pattern of meteorological elements is closely related to the macro-atmospheric environment, specifically manifested as follows: when the atmosphere is unstable, stations within the orographic superedge often exhibit coordinated convective development characteristics; when monsoon circulation is established, the correlation of water vapor transport superedges is significantly enhanced; and when weather systems are active, the importance of pressure response superedges becomes prominent. Conversely, under stable weather conditions, the effectiveness of these superedges is greatly reduced. Including them in the message transmission process not only fails to improve forecast accuracy but also introduces unnecessary noise interference. Based on these observations, this embodiment designs a gating mechanism that integrates weather environment to achieve adaptive selection of superedges. The specific implementation process is as follows:

[0185] In a given study area For each meteorological station, dynamic meteorological observation data is converted into time data using weights associated with station characteristics. The comprehensive environmental characterization below; this comprehensive environmental characterization captures the overall atmospheric state by aggregating the characteristics of all stations, expressed by the formula:

[0186] ,

[0187] in, Indicates in relation type Next The comprehensive environmental characterization of the sky, In relation types Next Weather station nodes of the day The characteristics are represented.

[0188] Then, message passing from nodes to hyperedges is performed using node representations and the association matrix to generate the embedded representation of the hyperedge:

[0189] ,

[0190] in, Indicates in relation type Next Embedded representation of sky super edges Indicates in relation type The following hyper-edge degree diagonal matrix:

[0191] ,

[0192] in, Indicates in relation type The number of bottom edges, Denotes the first element on the diagonal of the hyper-dimension diagonal matrix. One element;

[0193] Furthermore, by combining the comprehensive environmental representation with the hyperedge representation, the hyperedge weight matrix is ​​calculated through a gated network:

[0194] ,

[0195] in, Indicates in relation type The hyperedge weight matrix below, , Represents a diagonal matrix. Represents the hyperedge weight matrix The first diagonal One element, express and splicing, This represents the hyperbolic tangent function, compressing the output to the range [-1, 1] to represent the positive or negative effects of the hyperedge (such as enhancing or suppressing precipitation). This represents the activation function. Indicates the bias term. The weight matrix is ​​learnable; this mechanism allows the model to dynamically adjust the contribution of the hyperedge based on real-time meteorological conditions (such as the passage of fronts or the active period of monsoons), and prioritizes the associations with significant physical meaning, thereby reducing noise interference and improving the accuracy of rainfall forecasting.

[0196] Finally, based on the hyperedge weight matrix, the updated association matrix, and the embedding representation of the hyperedge, the feature representation of the weather station nodes under different relation types is calculated:

[0197] ,

[0198] in, Indicates in relation type Next The characteristics of weather station nodes on a given day are represented. This represents the degree matrix of weather station nodes.

[0199] 3. Interaction between hypergraphs

[0200] To coordinate the importance of information from different hypergraphs, this embodiment selectively aggregates information from different hypergraphs using an attention mechanism. This embodiment uses a hypergraph-level attention mechanism to fuse them together, specifically including:

[0201] First, calculate the importance of each hypergraph:

[0202] ,

[0203] in, Indicates the first Weather Station Node Relationship types The importance of For learnable weight vectors, Indicates in relation type Next Weather station nodes of the day The characteristics are represented.

[0204] Based on the importance of each hypergraph, all types of hypergraph information are weighted and aggregated to obtain the relationship characteristics of each weather station:

[0205] ,

[0206] in, Indicates the first Weather Station Node The relationship characteristics.

[0207] In step S3, refer to Figure 4 As shown, this embodiment generates sites using a feedforward neural network (FFN). Predicted rainfall:

[0208] ,

[0209] in, Indicates weather station The predicted rainfall, This represents a dense layer in a feedforward neural network. Indicates weather station Corresponding meteorological station nodes Relationship characteristics, Indicates weather station Meteorological characteristics.

[0210] In summary, this embodiment incorporates the dynamic impact of meteorological events by introducing a time-decaying attention function, and utilizes a negative weighting mechanism to capture the inhibitory effect of meteorological events on predicted rainfall, thus making the modeling of the temporal dynamics of rainfall processes more accurate. At the relational modeling level, the three-stage hypergraph message passing mechanism, through the synergy of a maskless attention mechanism and a gating mechanism, enables the existing hypergraph structure to adapt to dynamic weather environments. This mechanism can both reconstruct node-hyperedge relationships in real time to strengthen effective connections and weaken noise interference, and dynamically adjust the importance of hyperedges according to weather conditions, effectively avoiding the feature contamination problem caused by hyperedge failure in traditional methods. Combined with a node-level normalization strategy, a prediction method that can adapt to weather evolution and accurately quantify relational strength is finally constructed.

[0211] Example 2:

[0212] Based on the same inventive concept as Embodiment 1, this embodiment of the invention discloses a hierarchical hypergraph neural network rainfall prediction system, comprising:

[0213] The feature extraction module is used to extract features from the acquired meteorological index data and meteorological text data of each meteorological station to obtain the meteorological features of each meteorological station.

[0214] The relationship interaction module is used to: perform relationship interaction on the meteorological characteristics of each meteorological station to obtain the relationship characteristics of each meteorological station;

[0215] The rainfall prediction module is used to: concatenate the meteorological characteristics and relational characteristics of each meteorological station and input them into a feedforward neural network to predict rainfall, thereby obtaining the predicted rainfall for each meteorological station.

[0216] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0217] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0218] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0219] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0220] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0221] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A hierarchical hypergraph neural network method for rainfall prediction, characterized in that, include: Feature extraction was performed on the meteorological index data and meteorological text data of each meteorological station to obtain the meteorological characteristics of each meteorological station. The meteorological characteristics of each meteorological station are analyzed to obtain the relationship characteristics of each meteorological station. The meteorological characteristics and relational characteristics of each meteorological station are concatenated and input into a feedforward neural network to predict rainfall, thus obtaining the predicted rainfall for each meteorological station. The process of extracting features from the acquired meteorological index data and meteorological text data of each meteorological station to obtain the meteorological characteristics of each station includes: The temperature, humidity, wind speed and wind direction data of each meteorological station were normalized by maximum and minimum values, and the rainfall was normalized by logarithmic normalization to obtain the normalized meteorological index data of each meteorological station. The acquired meteorological text data from each meteorological station were normalized to obtain normalized meteorological text data for each meteorological station. The normalized meteorological index data and normalized meteorological text data of each meteorological station are input into the multilayer perceptron mixer for feature splicing and feature mixing operations to obtain the fused features of each meteorological station. The meteorological characteristics of each meteorological station are obtained by linear projection and multi-head attention calculation of the fusion features of each meteorological station using the time decay self-attention mechanism. The step of performing relationship interaction on the meteorological characteristics of each meteorological station to obtain the relationship characteristics of each meteorological station includes: A multi-relationship hypergraph is constructed based on the relationship types between the meteorological stations; wherein, the multi-relationship hypergraph includes a set of meteorological station nodes, a set of hyperedges, and an association matrix; Input the meteorological characteristics of each meteorological station into A multilayer perceptron was used to obtain initial feature representations of weather station nodes under different relation types; among them... Indicates the total number of relation types; The initial weather station node feature representation is aggregated based on the correlation matrix to obtain the initial hyperedge feature representation. Based on the initial weather station node feature representation and the initial hyperedge feature representation, the attention score between the weather station node and the hyperedge is calculated using the maskless hypergraph attention mechanism, and the correlation matrix is ​​updated using the attention score to obtain the updated correlation matrix; Based on the updated association matrix and the initial weather station node feature representation, the weather station node feature representation under different relation types is obtained by using hypergraph convolution based on hyperedge aggregator. Based on the node feature representation of the meteorological stations, the relationship features of each meteorological station are obtained using a hypergraph-level attention mechanism.

2. The hierarchical hypergraph neural network rainfall prediction method according to claim 1, characterized in that, The expression for the maximum and minimum value normalization process is as follows: ， in, This represents the data after normalization of the maximum and minimum values. This represents the data before normalization of the maximum and minimum values. express The minimum value in, express The maximum value in; The logarithmic normalization process is achieved through the following formula: ， in, This represents the data before logarithmic normalization. This represents the data after logarithmic normalization.

3. The hierarchical hypergraph neural network rainfall prediction method according to claim 1, characterized in that, The fusion characteristics of the various meteorological stations are obtained using the following formula: ， ， in, For weather station indexing, Indicates weather station The fusion characteristics Indicates transpose. Indicates splicing characteristics, This represents the trainable weights of the first fully connected layer. This represents the trainable weights of the second fully connected layer. express Activation function Presentation layer normalization operation, Indicates weather station Normalized meteorological index data, Indicates weather station Normalized meteorological text data, This represents a vector concatenation operation; The meteorological characteristics of each meteorological station are obtained using the following formula: ， ， ， in, Represents a set of meteorological characteristics. The total number of weather stations. Indicates weather station meteorological characteristics, Indicates weather station The fusion characteristics Represents the set of fused features. Represents the query vector. Represents the key vector. Represents the position mask matrix, Represents a value vector. Indicates the query weight. Indicates key weight. Indicates value weight, Indicates the attention head index, This indicates the total number of heads of attention. Represents a symbolic function. This represents the activation function. Represents absolute value. Scaling factor This represents the deviation matrix.

4. The hierarchical hypergraph neural network rainfall prediction method according to claim 1, characterized in that, The step of obtaining meteorological station node feature representations under different relation types using hypergraph convolution based on hyperedge aggregators, based on the updated association matrix and the initial meteorological station node feature representation, includes: Based on the updated correlation matrix and the initial weather station node feature representation, the embedding representation of the hyperedge is calculated; Calculate the comprehensive environmental characterization under different relationship types based on the initial meteorological station node feature representation; Based on the comprehensive environmental characterization and the embedding representation of hyperedges, a gating network is used to obtain the hyperedge weight matrix; The feature representations of weather station nodes under different relation types are calculated based on the hyperedge weight matrix, the updated correlation matrix, and the embedding representation of the hyperedge.

5. The hierarchical hypergraph neural network rainfall prediction method according to claim 1, characterized in that, The expression for the multi-relationship hypergraph is: ， ， ， ， ， in, This represents a multi-relationship hypergraph. For a set of relation types, For relation type indexing, Representing relation type The supergraph For the set of weather station nodes, Indicates the first Each weather station node, Indicates the total number of weather stations; Let be a hyperedge set, representing the connection relationships between weather station nodes; Representing relation type The set of superedges below, Representing relation type The correlation matrix below, , This represents the meteorological station node index. Representing relation type Next weather station node With super-edge The relationship between them .

6. The hierarchical hypergraph neural network rainfall prediction method according to claim 4, characterized in that, The initial weather station node feature representation is obtained through the following formula: ， in, Indicates the type used to handle relations. Multilayer perceptron, Indicates meteorological characteristics, Indicates in relation type Initial weather station node feature representation, , Indicates in relation type Next Initial weather station node characteristics representation for the day. , Indicates in relation type Next Weather station nodes of the day The initial feature representation; The initial hyperedge feature representation is obtained by the following formula: ， in, Indicates in relation type Next The first day A super edge, Indicates the superedge The initial feature representation, Indicates belonging to a superedge The set of all nodes , This represents the meteorological station node index. Indicates in relation type Next Weather station nodes of the day The initial feature representation, Indicates average operation; The attention score is obtained using the following formula: ， in, Indicates the first Weather Station Node With super-edge Attention score Indicates the first Weather Station Node With super-edge The correlation, , express Activation function It is a feedforward neural network. For learnable weight matrix, Indicates the first Weather Station Node meteorological characteristics, Indicates the first Weather Station Node With super-edge The correlation, As a filter factor, This represents the average correlation. ; The updated correlation matrix is ​​obtained using the following formula: ， in, Representing relation type Next The correlation matrix of days, Representing relation type Next Weather Station Node With super-edge The relationship between them; The embedding representation of the hyperedge is obtained by the following formula: ， in, Indicates in relation type Next Embedded representation of sky super edges Indicates in relation type The hyper-edge degree diagonal matrix below; The relation type hyper-edge degree diagonal matrix The expression is as follows: ， in, Indicates in relation type The number of bottom edges, Denotes the first element on the diagonal of the hyper-dimension diagonal matrix. One element; The comprehensive environmental characterization is obtained through the following formula: ， in, Indicates in relation type Next A comprehensive environmental characterization of the sky; The hyperedge weight matrix is ​​obtained by the following formula: ， in, Indicates in relation type The hyperedge weight matrix below, , Represents a diagonal matrix. Represents the hyperedge weight matrix The first diagonal One element, express and splicing, Represents the hyperbolic tangent function. This represents the activation function. Indicates the bias term. The weight matrix is ​​a learnable weight matrix; The feature representation of the meteorological station node is obtained by the following formula: ， in, Indicates in relation type Next The characteristics of weather station nodes on a given day are represented. Represents the degree matrix of weather station nodes; The relationship characteristics of the various meteorological stations are obtained through the following formula: ， in, Indicates the first Weather Station Node Relationship characteristics, Indicates the first Weather Station Node Relationship types The importance of , For learnable weight vectors, Indicates in relation type Next Weather station nodes of the day The characteristics are represented.

7. The hierarchical hypergraph neural network rainfall prediction method according to claim 1, characterized in that, The predicted rainfall amount is obtained using the following formula: ， in, Indicates weather station The predicted rainfall, This represents a dense layer in a feedforward neural network. Indicates weather station Corresponding meteorological station nodes Relationship characteristics, Indicates weather station Meteorological characteristics.

8. A hierarchical hypergraph neural network rainfall prediction system, characterized in that, include: The feature extraction module is used to extract features from the acquired meteorological index data and meteorological text data of each meteorological station to obtain the meteorological features of each meteorological station. The relationship interaction module is used to: perform relationship interaction on the meteorological characteristics of each meteorological station to obtain the relationship characteristics of each meteorological station; The rainfall prediction module is used to: concatenate the meteorological characteristics and relational characteristics of each meteorological station and input them into a feedforward neural network to predict rainfall, thereby obtaining the predicted rainfall for each meteorological station; The process of extracting features from the acquired meteorological index data and meteorological text data of each meteorological station to obtain the meteorological characteristics of each station includes: The temperature, humidity, wind speed and wind direction data of each meteorological station were normalized by maximum and minimum values, and the rainfall was normalized by logarithmic normalization to obtain the normalized meteorological index data of each meteorological station. The acquired meteorological text data from each meteorological station were normalized to obtain normalized meteorological text data for each meteorological station. The normalized meteorological index data and normalized meteorological text data of each meteorological station are input into the multilayer perceptron mixer for feature splicing and feature mixing operations to obtain the fused features of each meteorological station. The meteorological characteristics of each meteorological station are obtained by linear projection and multi-head attention calculation of the fusion features of each meteorological station using the time decay self-attention mechanism. The step of performing relationship interaction on the meteorological characteristics of each meteorological station to obtain the relationship characteristics of each meteorological station includes: A multi-relationship hypergraph is constructed based on the relationship types between the meteorological stations; wherein, the multi-relationship hypergraph includes a set of meteorological station nodes, a set of hyperedges, and an association matrix; Input the meteorological characteristics of each meteorological station into A multilayer perceptron was used to obtain initial feature representations of weather station nodes under different relation types; among them... Indicates the total number of relation types; The initial weather station node feature representation is aggregated based on the correlation matrix to obtain the initial hyperedge feature representation. Based on the initial weather station node feature representation and the initial hyperedge feature representation, the attention score between the weather station node and the hyperedge is calculated using the maskless hypergraph attention mechanism, and the correlation matrix is ​​updated using the attention score to obtain the updated correlation matrix; Based on the updated association matrix and the initial weather station node feature representation, the weather station node feature representation under different relation types is obtained by using hypergraph convolution based on hyperedge aggregator. Based on the node feature representation of the meteorological stations, the relationship features of each meteorological station are obtained using a hypergraph-level attention mechanism.

Images