A method for spatiotemporal prediction of wind speed in typhoon wind fields

Abstract

This invention discloses a spatiotemporal prediction method for typhoon wind field and wind speed. Specifically, it involves: acquiring multi-source meteorological elements based on ground meteorological observation stations, satellite remote sensing platforms, and reanalysis datasets; preprocessing the multi-source data using a multi-source heterogeneous data fusion mechanism to obtain an irregular spatiotemporal observation set; designating all ground meteorological observation stations, satellite remote sensing platforms, and typhoon observation systems corresponding to the reanalysis datasets as nodes, totaling N nodes; constructing a feature matrix X for the N nodes, and simultaneously constructing temporal and spatial hyperedges; constructing a spatiotemporal hypergraph based on the feature matrix X and the two types of hyperedges; extracting features from the spatiotemporal hypergraph to obtain higher-order features; and inputting the higher-order features into an improved Transformer encoder to obtain the predicted typhoon wind field and wind speed values ​​for future times. This invention can serve applications such as wind energy assessment, extreme weather warnings, and dynamic wind field reconstruction.

Description

Technical Field

[0001] This invention belongs to the field of atmospheric science and technology, and in particular relates to a method for spatiotemporal prediction of typhoon wind field wind speed. Background Technology

[0002] Wind speed forecasting is a crucial component of meteorological operations, wind power grid connection, and typhoon warnings. However, existing methods suffer from two significant shortcomings in complex local terrain, typhoon paths, and abrupt wind field changes: First, insufficient forecast accuracy, particularly in scenarios involving typhoon landfall or abrupt wind field changes induced by complex terrain, making it difficult to accurately characterize peak values ​​and phases; second, insufficient global coupling representation, failing to capture higher-order relationships such as upstream-downstream transmission, long-distance transmission, and multi-site group co-driving, resulting in an inadequate grasp of the organizational structure of key weather processes. In recent years, deep learning methods have been introduced into the field of wind speed forecasting, such as convolutional neural networks for extracting spatial features and recurrent neural networks for capturing temporal dependent structures. However, these models generally assume that the relationships between spatial nodes are static and uniform, neglecting the non-Euclidean structural features of meteorological elements in the spatiotemporal dimensions.

[0003] Traditional numerical models are prone to parameterization errors under complex terrain and strong convection triggering conditions. Data-driven methods "normalize" multi-source data, often performing spatiotemporal interpolation and smoothing, which smooths out peaks and introduces phase shifts, making it difficult to reconstruct extreme wind speeds and rapid abrupt changes. Multi-source data (ground stations, reanalysis, remote sensing textures) differ in time steps and spatial resolution. Forcibly fusing them at a single scale amplifies noise and bias propagation. Furthermore, the lack of explicit modeling of long-range dependence and directional propagation limits the model's sensitivity to extreme processes and its extrapolation ability, thus manifesting as an overall accuracy bottleneck.

[0004] In real atmospheric systems, wind speed variations are not only controlled by local temperature, pressure, and humidity, but also influenced by the coupling effects of neighboring stations and upstream wind fields. Spatial dependence exhibits nonlinear, multi-scale, and dynamic characteristics. Traditional graph neural networks (GNNs) struggle to represent these complex spatial relationships, while relying solely on long short-term memory (LSTM) networks to process time-series data fails to reflect geographic topological features. Therefore, recent research has begun to explore combining graph neural networks with time-series models for non-Euclidean spatiotemporal modeling. However, most existing methods still treat edge relationships as second-order connections and static edges, failing to express higher-order interactions involving simultaneous action from multiple stations, thus weakening global coupling. Summary of the Invention

[0005] Purpose of the invention: In order to solve the problems existing in the prior art, the present invention provides a method for spatiotemporal prediction of typhoon wind field wind speed.

[0006] Technical solution: This invention discloses a spatiotemporal prediction method for typhoon wind field wind speed, specifically as follows:

[0007] Multi-source meteorological elements are acquired based on ground meteorological observation stations, satellite remote sensing platforms, and reanalysis datasets.

[0008] A multi-source heterogeneous data fusion mechanism was used to preprocess multi-source meteorological elements to obtain an irregular spatiotemporal observation set;

[0009] All ground meteorological observation stations, satellite remote sensing platforms, and typhoon observation systems corresponding to the reanalysis dataset are treated as nodes, with a total of N nodes; a feature matrix X is constructed for the N nodes, and temporal hyperedges and spatial hyperedges are also constructed. A spatiotemporal hypergraph is constructed based on the feature matrix X and the two types of hyperedges.

[0010] High-order features are obtained by extracting features from the spatiotemporal hypergraph;

[0011] By inputting high-order features into an improved Transformer encoder, the predicted wind speed of the typhoon wind field at future times can be obtained.

[0012] Furthermore, based on wind direction and time observed by ground meteorological stations, sine and cosine transforms are introduced into the multi-source meteorological elements to extract the sine and cosine features of wind direction and time.

[0013] Furthermore, the multi-source meteorological elements were preprocessed using the following method:

[0014] Data with a missing rate below a preset threshold are directly removed, while data with a missing rate above a preset threshold are labeled using an explicit masking mechanism.

[0015] Identify and parse the timestamp format of numerical or string data, and unify the time zone and granularity of all data;

[0016] Using (station_id, timestamp) as the joint key, perform an outer join between the data observed by the ground meteorological observation station and the typhoon's optimal trajectory data, where station_id is the ID of the ground meteorological observation station and timestamp is the timestamp;

[0017] Using timestamps as anchors, for each unique timestamp, the first remote sensing feature observed by the satellite remote sensing platform is broadcast and copied to all ground meteorological observation stations to achieve lightweight fusion; the first remote sensing feature is remote sensing information other than remote sensing texture information.

[0018] When the timestamp of the remote sensing image matches the ground observation time, the remote sensing texture information observed by the satellite remote sensing platform is broadcast and copied to all ground meteorological observation stations; when the difference between the remote sensing image and the ground observation time is less than the preset time threshold, the remote sensing image is considered to match the ground observation time.

[0019] For the typhoon trajectory data in the reanalysis dataset, a distance-based spatiotemporal nearest neighbor matching algorithm is used to construct an index bucket and select the typhoon center observation data that is closest to the corresponding ground meteorological observation station within a preset time window;

[0020] Finally, all the data is standardized.

[0021] Furthermore, the construction of the spatiotemporal hypergraph specifically involves:

[0022] The feature matrix ,in For the real number field, Where T is the batch size and T is the time. For feature dimension; based on Heterogeneous graphs are constructed using heterogeneous graph neural networks, and node features in the heterogeneous graphs are fused:

[0023] ;

[0024] in, This represents the node features after fusion. This represents the node features before fusion. For nodes The neighborhood set, For a general linear mapping matrix, For specific mapping matrices whose data source type is from ground meteorological observation stations, satellite remote sensing platforms, or reanalysis datasets, The fusion coefficient is... The expression is as follows:

[0025] ;

[0026] in, For temperature parameters, For nodes Confidence weights For nodes and Feature similarity;

[0027] After fusion Reassemble the new feature matrix ;

[0028] Construct temporal and spatial hyperedges, and combine the heterogeneous graph output by the heterogeneous graph neural network with... The input is fed into a hypergraph neural network to obtain the final spatiotemporal hypergraph.

[0029] Furthermore, the construction of temporal and spatial hyperedges is as follows: Temporal hyperedge: Connecting all ground meteorological observation stations at the same observation time into a hyperedge yields the temporal hyperedge;

[0030] Spatial hyperedge: Taking a single ground meteorological observation station as the axis, a hyperedge is formed by connecting all the observation nodes of the ground meteorological observation station throughout the day.

[0031] Furthermore, feature extraction is performed on the spatiotemporal hypergraph, specifically as follows:

[0032] Employing a hypergraph convolutional propagation structure to analyze the feature matrix in the spatiotemporal hypergraph The following processing is performed to obtain the matrix. :

[0033] ;

[0034] in, The time hyperedge incidence matrix, The spatial hyperedge incidence matrix, Let be the degree matrix corresponding to the time hyperedge. Let be the degree matrix corresponding to the hyperedge in space. The weight of the time-bound edge. The weight of the spatial hyperedge, For activation function, Indicates transpose; Introducing wind direction correction:

[0035] ;

[0036] in: Based on the weight matrix, For wind direction dot squaring, The angle between wind directions;

[0037] Attention weights for computation time hyperedge :

[0038] ;

[0039] in, For time-exceeding edge and node A set of connected nodes. The linear transformation matrix for the time channel. For the time channels, there are learnable attention vectors. node With nodes Geographical differences between them; and They are nodes With nodes eigenvalues; For activation functions;

[0040] Attention weights of computation space hyperedges :

[0041] ;

[0042] in, For spatial hyperedges and nodes A set of connected nodes. Let be the linear transformation matrix of the spatial channel. This represents the learnable attention vector for the spatial channel. Represents a node With nodes The time difference between them;

[0043] based on and Obtain higher-order features :

[0044] ;

[0045] in, for The matrix formed for The matrix formed As weight, It is a diagonal matrix function.

[0046] Furthermore, higher-order features The ground meteorological observation stations are rearranged chronologically and at the same time according to their order, and then packaged into tensors in batches. These tensors are then fed into an improved Transformer encoder, which processes the input features as follows:

[0047] Perform a linear transformation on the input features to generate a query Q, key K, and value V vector, and then segment these three vectors into... Each attention head is assigned a weight based on a temporal mask. :

[0048] ;

[0049] in, Indicates the first One point of attention, ; For the first A query vector with attention heads. For the first The key vector of each attention head. For the mask matrix, This means replacing all zeros in the mask matrix with negative infinity; For the dimension of attention head, Indicates transpose;

[0050] Perform matrix multiplication on each weight to obtain a weighted value vector. :

[0051] ;

[0052] in, For the first A vector of values ​​for each attention head;

[0053] All of Concatenate the vectors to obtain the concatenated vector. ;right Perform a linear transformation to obtain the transformed vector. ;

[0054] The final improved Transformer encoder output is obtained. :

[0055] ;

[0056] in, This indicates a normalization operation.

[0057] Furthermore, the objective function used during training is: :

[0058] ;

[0059] in, For all learning parameters, For loss function, The regularization coefficient is used. The expression is as follows:

[0060] ;

[0061] in, Represents the set of predicted durations. and All are related to the predicted duration The corresponding weights Indicates the mean absolute error. This represents the mean square error. and The possible values ​​are as follows:

[0062] ;

[0063] in, For the observed density:

[0064] ;

[0065] in, Indicates the prediction duration The number of valid observations within, This represents the theoretical maximum number of observations.

[0066] Furthermore, during training, an Adam optimizer combined with a cosine scheduler learning rate scheduling strategy is employed.

[0067] A computer device includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the steps of the spatiotemporal prediction method for typhoon wind field speed.

[0068] Beneficial Effects: This invention addresses the differences in resolution, confidence level, and timeliness among ground-based observations, satellite remote sensing, and reanalysis datasets. It constructs a heterogeneous graph structure that includes both homogeneous and heterogeneous edges, introducing priority-driven information propagation. This prioritizes the fusion of high-confidence data and provides controlled supplementation of low-confidence data, reducing information distortion caused by interpolation smoothing and improving cross-source utilization and spatiotemporal consistency from the source. Simultaneously, through joint modeling of time-adaptive topology updates and frequency embedding, the hypergraph structure can self-organize and evolve with wind field conditions, precisely depicting local abrupt changes while maintaining a globally coupled representation of large-scale circulation. The synergistic effect of these two mechanisms enables this method to achieve high-precision, interpretable predictions of wind speed evolution under non-Euclidean space and irregular sampling conditions, serving applications such as wind energy assessment, extreme weather warnings, and dynamic wind field reconstruction. Attached Figure Description

[0069] Figure 1 This is a flowchart of the present invention;

[0070] Figure 2 This is a schematic diagram of the heterogeneous graph structure of the present invention;

[0071] Figure 3 A schematic diagram for modeling spatial relationships;

[0072] Figure 4 This is a schematic diagram of the structure of the hypergraph convolutional neural network and encoder of the present invention;

[0073] Figure 5 This is a diagram illustrating the dynamic evolution of a hypergraph.

[0074] Figure 6 Figure (a) shows the predicted point distribution map obtained by the conventional method and Figure (b) shows the predicted point distribution map obtained by the method of the present invention.

[0075] Figure 7 This is a comparison chart of the errors between the method of the present invention and the conventional method;

[0076] Figure 8 This is a comparison diagram of the prediction time steps of the method of the present invention and the conventional method;

[0077] Figure 9 This is a diagram of the predictive head structure of the present invention;

[0078] Figure 10 This represents the error distribution of the method of the present invention over a 24-hour prediction time series. Detailed Implementation

[0079] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0080] like Figure 1 The method of the present invention shown is specifically as follows:

[0081] S1: Obtain multi-source meteorological elements based on ground meteorological observation stations, satellite remote sensing platforms, and reanalysis datasets;

[0082] S2: A multi-source heterogeneous data fusion mechanism is used to preprocess multi-source data to obtain an irregular spatiotemporal observation set;

[0083] S3: Take all ground meteorological observation stations, satellite remote sensing platforms and typhoon observation systems corresponding to the reanalysis dataset as nodes, for a total of N nodes; construct a feature matrix X for the N nodes, and construct temporal hyperedges and spatial hyperedges at the same time; construct a spatiotemporal hypergraph based on the feature matrix X and the constructed hyperedges.

[0084] S4: Extract features from the spatiotemporal hypergraph to obtain higher-order features;

[0085] S5: Input the high-order features into the improved Transformer encoder to obtain the wind speed prediction value of the typhoon wind field at future time.

[0086] S1 specifically refers to: In this embodiment, multi-source meteorological elements are obtained from ground meteorological observation stations, satellite remote sensing platforms, and reanalysis datasets (China Meteorological Administration, CMA) to form time series data containing multi-dimensional spatiotemporal characteristics.

[0087] Based on wind direction and time observed by ground meteorological stations, multi-source meteorological elements are analyzed. Sine and cosine transformations are introduced to extract the sine and cosine features of wind direction and time.

[0088] In the end, this embodiment has a total of 35 valid numerical features, as shown in Table 1 below:

[0089] Table 1

[0090]

[0091] S2 specifically refers to the following: To ensure the quality and physical consistency of the input data for the wind speed prediction model, this invention introduces a multi-source heterogeneous data fusion mechanism in the data processing flow. First, in the data cleaning stage, multi-source data such as ground meteorological stations, remote sensing images, and typhoon trajectories are processed separately: observation samples with a missing rate of less than 5% are directly removed, while those with a higher missing rate are labeled using an explicit masking mechanism instead of interpolation filling, in order to preserve the original distribution characteristics of the data to the greatest extent. Outlier detection combines statistical methods and machine learning methods, identifying extreme observation points based on the variable distribution characteristics. This embodiment designs a unified function to automatically identify and parse numerical or string-type timestamp formats, and unifies the time zone and granularity of all data sources to ensure accurate synchronization in the multi-source time dimension. Subsequently, an outer join is performed on the ground meteorological station data and the typhoon's optimal path data using (station_id, timestamp) as the joint key. This maximizes the retention of effective observations and achieves precise coupling between stations and typhoons in the time dimension. Then, using the timestamp as the anchor point, for each unique timestamp, the first remote sensing feature observed by the satellite remote sensing platform (the first remote sensing feature is remote sensing information other than remote sensing texture information) is broadcast and copied to all station records, achieving a lightweight fusion that prioritizes time consistency and simplifies spatial matching. Following this, the system performs horizontal stitching of multi-source data based on station ID and timestamp, and executes the feature engineering module. For the spatial dimension, a broadcast fusion strategy for remote sensing features is proposed: when the remote sensing image matches the ground observation time, the remote sensing texture information observed by the satellite remote sensing platform is broadcast and copied to all ground meteorological observation stations; when the interpolation between the remote sensing image and the ground observation time is less than a preset time threshold, the remote sensing image is considered to match the ground observation time. This avoids physical distortion caused by complex spatial interpolation operations. For the typhoon trajectory data in the reanalysis dataset, a spatiotemporal nearest neighbor matching algorithm based on the semi-sine Haversine distance is used to construct an index bucket. Within a preset time window (a window of ±3 hours is set in this embodiment), the observation data of the typhoon center closest to the corresponding ground meteorological observation station is selected to achieve a highly robust matching mechanism. Finally, Min-Max standardization is performed on all numerical features, and only the minimum and maximum values ​​are statistically analyzed using the training set and mapped to the [0,1] interval.

[0092] This data fusion mechanism effectively solves problems such as inconsistent spatiotemporal resolution and uneven observation density of multi-source meteorological data through data quality control, spatiotemporal intelligent matching, and periodic feature extraction. It greatly enhances the stability and physical consistency of model input, thereby improving the accuracy and robustness of wind speed prediction tasks from the source.

[0093] S3 specifically refers to: S2, through the process of "cleaning—alignment—key external connection and timestamp broadcasting—periodic expansion—standardization", outputs an irregular spatiotemporal observation set that can be directly used for modeling: the feature matrix. Missing test mask Time Index (Records the timestamp corresponding to each node) and spatial index (It records the spatial location information corresponding to each node), and also saves the inverse normalization parameters; based on this, S3 adopts a two-level structured modeling path of "heterogeneous first, then hypergraph": first, in the heterogeneous layer, a priority matrix is ​​used to... Constraining the ordered information propagation of homogeneous / heterogeneous edges enables the directional and controlled fusion of ground observations, remote sensing satellites, and reanalysis datasets. Subsequently, the hyperlayer organizes multiple stations at the same time and multiple time points at the same station into temporal / spatial hyperedges, explicitly characterizing higher-order spatiotemporal coupling relationships. The collaborative output of the two-level structure serves as the direct input for subsequent hypergraph feature extraction and dynamic hypergraph learning mechanisms. The feature matrix in this embodiment... ,in For the real number field, Where T is the batch size and T is the time. For feature dimensions.

[0094] The input consists of multi-source heterogeneous meteorological data, covering ground meteorological observation stations, satellite remote sensing platforms, and meteorological reanalysis datasets. These data sources differ in spatial resolution, temporal frequency, and confidence level. To ensure the reliability and interpretability of the fusion process, the priority order is set according to confidence level as follows: ground meteorological observation stations > satellite remote sensing platforms > reanalysis datasets.

[0095] Let time set Site collection .

[0096] The heterogeneous graph output by the heterogeneous graph neural network:

[0097] in, For a set of nodes, Let be the set of edges. It is an incidence matrix. The priority matrix is ​​shown below; a schematic diagram of the heterogeneous graph is shown below. Figure 2 As shown.

[0098] ;

[0099] in, Denotes a subset of edges with the same origin. It is a heterogeneous edge subset; and This represents any two nodes in a heterogeneous graph. The type of node is indicated, including ground meteorological observation stations, remote satellite sensing platforms, and reanalysis datasets.

[0100] when When (two nodes of the same type), (Original edge);

[0101] when When the two nodes are of different types, (Heterogeneous edge);

[0102] This definition ensures that heterogeneous graphs can accurately distinguish the connection properties between nodes from different data sources, providing a logical foundation for subsequent differentiated message passing and feature aggregation. Capture consistency within the same source. This captures complementary features across sources. By constructing heterogeneous graphs, joint modeling of meteorological features from different sources can be achieved at the topological level.

[0103] For any node in the heterogeneous graph, calculate the node's fusion feature:

[0104] ;

[0105] in, Indicates the characteristics after fusion. For nodes The neighborhood set, For a general linear mapping matrix, For specific mapping matrices whose data source type is from ground meteorological observation stations, satellite remote sensing platforms, or reanalysis datasets, Represents a node Characteristics before fusion The attention coefficient is expressed as follows:

[0106] ;

[0107] Temperature parameter controls the smoothness of attention distribution; the larger the value, the more the model focuses on similar nodes.

[0108] :node The confidence weight reflects the credibility and priority of its data source;

[0109] Feature similarity is used to measure the consistency of feature orientation between nodes.

[0110] based on Reconstruct the feature matrix to obtain a new feature matrix. .

[0111] This mechanism combines confidence level and feature similarity, enabling high-confidence and similar nodes to dominate information propagation during the fusion process, while the influence of low-confidence nodes is naturally suppressed, thus ensuring the robustness and interpretability of the fusion process.

[0112] Data from different sources exhibit systematic differences in temporal granularity, spatial resolution, noise levels, and timeliness. Direct splicing or simple interpolation can easily introduce cross-source bias and phase drift. To address this, this invention combines "priority-driven ordered message passing" with "source type-specific mapping" within a unified topology framework to achieve differentiated fusion: When multiple sources exist simultaneously at the same "site-timestamp" node, they are first sorted according to "confidence level × timeliness" (ground observations take priority, followed by remote sensing, and then analysis is performed last), with high-confidence sources as the dominant information, while attenuation weights are applied to the second-best sources to limit the influence domain of cross-source migration; in the message aggregation stage, attention weights modulated by priority are used to assign weights to homogeneous and heterogeneous edges respectively, so that high-confidence information is propagated first, and low-confidence information only intervenes as controlled compensation when there is a lack of measurement or inconsistency; a source type-specific linear transformation is introduced into the feature space to complete modal alignment together with the general mapping, alleviating the excessive smoothing caused by differences in dimensionality and statistical distribution; and time alignment and geographical proximity are used as priors to limit the effective propagation path of cross-source information, avoiding the mis-diffusion of local mutations to areas with no physical connection. The above design achieves comprehensive modeling of reliability, timeliness and physical adjacency without changing the consistency of the global structure, significantly reducing the distortion and lag caused by cross-source fusion, and ensuring high-quality unified representation at the input level.

[0113] Traditional methods often fail to effectively backfill missing samples or utilize modal complementarity information in a structured manner due to reliance on a single source or simple weighting of multiple sources. To improve information utilization, this invention explicitly constructs cross-source heterogeneous edges in heterogeneous graphs, establishing one-to-many / many-to-one structural connections between different source nodes at the same spatiotemporal location. This allows large-scale circulation background data from remote sensing textures, real-time ground station records, and reanalysis to form a mutually corroborating and corrective closed loop at the topological level. In message passing, confidence-modulated attention allocation and residual-based master-slave channels are employed, with high-confidence sources as the master channel and low-confidence sources entering with restricted weights. This provides robust backfilling when the master source is missing and provides directionally consistent correction signals when the master source has systematic biases. For spatiotemporally sparse regions, alternative observations from neighboring times / nearby stations are injected after physical prior filtering through heterogeneous edges, effectively improving the effective sample density and information coverage. Simultaneously, under the joint constraints of the correlation matrix and the missing measurement mask, the loss is calculated only for effective observation positions and backpropagated, avoiding interference from spurious filling on parameter learning. This mechanism achieves the dual benefits of "intra-source consistent modeling + inter-source complementary collaboration", which reduces over-reliance on a single source and transforms multimodal differences into learnable structural redundancy and reliable backfilling, thereby significantly improving data utilization efficiency and model robustness in complex observation ecosystems.

[0114] At the innovation level, this invention introduces the ordered message passing mechanism of ordered graph neural networks into the fusion of meteorological multi-source data for the first time. The propagation order of information flow is determined according to the reliability and time sensitivity of the data source, which significantly reduces noise interference and information distortion between different sources. A priority matrix P (which gives the weight distribution of nodes on each source) is proposed. During the aggregation process, P drives the weighted update of features, so that high-confidence observations dominate the learning process, while low-confidence information is retained as a prudent supplement, which enhances the robustness and interpretability of the fusion. At the same time, a heterogeneous graph structure with both homogeneous and heterogeneous edges is constructed, which makes the information propagation directional and hierarchical, closer to the real meteorological process: high-level nodes such as ground stations transmit observation signals downward, and low-level nodes such as analysis grids provide background constraints upward, forming a bidirectional flow from top to bottom and bottom to top. The three work together to achieve end-to-end optimization from "data source confidence modeling" to "topologically ordered propagation", providing high-confidence, directional and quantifiable fusion input for subsequent dynamic hypergraph spatiotemporal learning, and further improving the global consistency expression ability and multi-scale fusion accuracy.

[0115] To accommodate the heterogeneous ordered fusion results, after obtaining a unified and aligned node-level input, we will explicitly construct a spatiotemporal hypergraph topology on the node set. This will characterize the spatial collaboration of "multiple sites at the same moment" and the temporal evolution of "multiple moments at the same site" in the form of high-order connections, thus providing a structural foundation for subsequent hypergraph convolution and dynamic propagation.

[0116] Observation node set: The combination of (site, timestamp) for each observation sample is regarded as an observation node, and the feature of the observation node is the meteorological element vector at the corresponding time.

[0117] like Figure 3 The temporal and spatial hyperedges are constructed as shown:

[0118] Temporal hyperedge: This connects all station nodes at the same moment into a hyperedge, for example, including every meteorological station at 08:00 on July 1, 2023. The weight of the edge is determined by two parts: first, the mutual information of the wind fields of each station under real-time observation, which measures the similarity of their instantaneous wind arrows; and second, the temporal alignment, ensuring that all data are strictly aligned to the same moment. The hyperedge formed in this way specifically captures the spatial correlation between different geographical locations at the same time point, enabling the network to explicitly express the coordinated changes of instantaneous weather patterns in a region.

[0119] Spatial hyperedge: Centered on a single station, a hyperedge is formed by connecting all observation nodes of that station throughout the day—for example, 96 time points from 00:00 to 24:00 for station A are connected at once. The weight of the edge is determined by both spatial and physical factors: first, the geographical proximity between the station and its nearest neighbor is measured using Euclidean distance; then, a topographic slope correction term is introduced to give additional penalties or gains to stations located on hillsides, valleys, or plateaus, reflecting the diffusion and deposition effects of the real atmosphere constrained by the underlying surface. Hyperedges constructed in this way no longer focus on who collaborates with whom "at this moment," but rather on the temporal evolution of "this place" itself within a day, enabling the network to explicitly characterize the continuity of weather processes at the same geographical location and the unique rhythms modulated by topography.

[0120] This separate hyperedge design effectively solves the interference problem caused by irregular sampling, and can model temporal and spatial dependencies separately, thus improving the model's ability to express complex spatiotemporal patterns.

[0121] To overcome the bottleneck of "insufficient global coupling expression" in weather forecasting, this invention introduces a spatiotemporal hypergraph topology on a unified and aligned node input: on the one hand, time hyperedges explicitly connect multiple stations at the same moment, directly characterizing the instantaneous propagation of upstream and downstream and the coordinated influence of large-scale systems such as typhoons in the region; on the other hand, spatial (station) hyperedges connect multiple moments of the same station, and combined with physical priors such as geographical proximity and topography, constrain the modeling of long-distance indirect associations and multi-station group driving forces, thereby expressing "spatial consistency of multiple stations at the same moment" and "evolutionary continuity of multiple moments of the same station" within a unified framework.

[0122] Building upon this foundation, a triple hypergraph architecture and attention-based message passing are proposed to enhance expressiveness and robustness. Spatiotemporal dependencies are decomposed into "temporal hyperedges (capturing temporal dependencies) + variable / modal hyperedges (characterizing coupling between elements)," while spatial (site) hyperedges are retained to form higher-order many-to-many connections, enabling a single hyperedge to aggregate the interactions of multiple observation nodes. Message passing is divided into three stages: "node → hyperedge, hyperedge → hyperedge, and hyperedge → node." Irregularity-aware attention and adjacency masks are added in the hyperedge → hyperedge stage, propagating information only within temporally aligned and physically relevant ranges, suppressing noise and bias caused by irregular sampling. Ultimately, observation nodes, temporal hyperedges, and variable / modal hyperedges work collaboratively, achieving effective information aggregation and write-back through hypergraph convolution. This strengthens the global coupling expression of upstream-downstream transmission and group-driven processes while avoiding sampling interference caused by spatiotemporal hybrid modeling, significantly improving prediction accuracy and global consistency under extreme conditions.

[0123] Based on heterogeneous graphs Spatiotemporal hyperedge, using hypergraph neural networks to construct a spatiotemporal hypergraph. Spacetime Hypergraph Including hyperedge sets and static hypergraph association matrix ; Includes time hyperedge With spatial hyperedge .

[0124] S4 specifically refers to: (e.g., ...) Figure 4 As shown, in order to build upon the heterogeneous ordered fusion and spatiotemporal hypergraph construction results of S3, S4 extracts hypergraph features and completes the joint representation of temporal coordination and spatial evolution within the same framework: specifically, the input is the node feature matrix in the spatiotemporal hypergraph. And a hypergraph containing both temporal and spatial hyperedges. Based on this, S4's separate "temporal-spatial channel" hypergraph convolution and attention aggregation synchronously model cross-site collaboration at the same moment and cross-temporal evolution within the same site, ultimately outputting a high-order spatiotemporal feature representation. This provides a unified and interpretable input representation for subsequent time series coding and multi-scale prediction.

[0125] To model high-order coupling between meteorological elements, this module employs a dual-channel hypergraph convolutional propagation structure to separate temporal and spatial information. The propagation formula is defined as:

[0126] ;

[0127] in, The time hyperedge incidence matrix, The spatial hyperedge incidence matrix, Let be the degree matrix corresponding to the time hyperedge. , Let be the temporal hyperedge correlation matrix between node i and node j. Let be the degree matrix of the hyperedge in space. , Let be the spatial hyperedge incidence matrix between node i and node j. The weight of the time-bound edge. The weight of the spatial hyperedge, It is a non-linear activation function.

[0128] In time-channel propagation, to enhance physical consistency, the weight matrix... A wind direction correction term has been introduced:

[0129] ;

[0130] in: Basic weight matrix; Wind direction dot product, used to introduce physical constraints. The angle between wind directions.

[0131] This design allows the model to better follow physical laws during the learning process, improving the accuracy and interpretability of predictions.

[0132] To improve the flexibility of feature aggregation, this module introduces a hypergraph attention mechanism, assigning learnable weights to each hyperedge in both the temporal and spatial channels:

[0133] Attention weights of temporal superedge :

[0134] ;

[0135] in, For time-exceeding edge and node A set of contiguous elements The linear transformation matrix for the time channel. For the time channels, there are learnable attention vectors. Represents a node With nodes Geographical differences between them; and They are respectively Middle node With nodes eigenvalues; For activation functions; Indicates transpose;

[0136] Attention weights of computation space hyperedges :

[0137] ;

[0138] in, For spatial hyperedges and nodes A set of contiguous elements Let be the linear transformation matrix of the spatial channel. This represents the learnable attention vector for the spatial channel. Represents a node With nodes The time difference between them;

[0139] based on and Obtain higher-order features :

[0140] ;

[0141] in, for The matrix formed for The matrix formed As weight, It is a diagonal matrix function.

[0142] The attention mechanism in this embodiment introduces spatial geographic distance simultaneously. Time interval It achieves adaptive modeling of multi-scale correlations, making the model more robust when dealing with complex terrain, typhoon paths and sudden wind fields.

[0143] This module addresses the issue of insufficient prediction accuracy by employing a joint aggregation of hypergraph convolution and attention. Nodes no longer merely exchange information with paired neighbors; instead, they aggregate multi-order signals across the entire hyperedge using learnable weights, simultaneously absorbing instantaneous collaborative information from "simultaneous multi-station" spatial hyperedges and continuous evolutionary memory from "simultaneous multi-time" temporal hyperedges. As a result, the network can achieve end-to-end characterization of wind field peaks and phases under extreme scenarios such as typhoon landfall and abrupt changes in the leeward slope of complex terrain. When any subset of the hyperedge undergoes drastic changes, the convolutional kernel leverages the global receptive field at the hyperedge level to rapidly propagate the disturbance to the relevant spatiotemporal neighborhood, achieving a mapping from local jumps to global responses. To accurately reflect real atmospheric motion, physical constraints are explicitly introduced during the message transmission phase: the wind direction vectors of connected nodes are multiplied by a dot product and normalized before being multiplied into the aggregate weight. Higher contributions are only given when the directions are consistent or opposite (which conforms to ideal mass conservation and topographic flow). This suppresses false gradients that violate the geostrophic deflection or are blocked by mountains, while retaining physically reasonable high-energy structures such as typhoon vortices and canyon acceleration. Thus, while maintaining the flexibility of data-driven approaches, interpretability and extrapolation robustness are significantly improved.

[0144] To address the issue of insufficient global coupling representation, a hypergraph attention mechanism is introduced to assign learnable dynamic weights to each hyperedge. During training, the importance of higher-order relationships such as upstream-downstream transmission, long-distance energy dispersion, and multi-station collaborative mutations is automatically identified. The attention score is jointly determined by node features, hyperedge priors (such as geographical distance, elevation difference, and historical consistency of wind direction), and task context: when a spiral rainband or cold front causes upstream disturbances to be rapidly replicated downstream, the corresponding spatiotemporal hyperedge weights are increased; conversely, local noise or isolated convection is suppressed, ensuring that only physically meaningful driving patterns are amplified and participate in updates. To avoid irregular sampling interference, the model adopts a separate temporal / spatial dual-channel structure: the temporal channel performs convolution and attention only on the "same-station, multiple-time" hypergraph, unaffected by spatial sparsity; the spatial channel transmits messages only on the "same-moment, multiple-station" hypergraph, avoiding the coupling of noise due to temporal missing measurements. The two-channel outputs are adaptively fused using learnable gating, which retains the advantages of independent modeling in terms of temporal continuity and spatial synergy, while eliminating the contamination of spatiotemporal interaction terms by sampling bias, thus achieving an accurate and robust representation of complex high-order meteorological processes.

[0145] Within this framework, three key innovations are achieved: First, a message passing mechanism with physical constraints, which directly embeds the wind direction dot product into the aggregation weights, enabling the network to follow the basic laws of atmospheric dynamics; second, a temporal / spatial dual-channel hypergraph convolution, which processes temporal and spatial hyperedges separately and fuses them at a high level, significantly enhancing the ability to represent complex spatiotemporal patterns; and third, a hyperedge adaptive attention mechanism, which adopts differentiated computational strategies for different hyperedge types—the temporal channel focuses on geographical distance and terrain correction, while the spatial channel focuses on temporal alignment—enabling the information aggregation strategy to self-adjust according to the scenario.

[0146] S5 specifically refers to: (e.g., ...) Figure 9 As shown, high-order features are obtained after fusion of dual-channel hypergraph convolution and adaptive attention. In this step, the first step is to process the data according to the rule of "time sequence for the same station, and station sequence for the same time". The data is rearranged and batch-packed into tensors. Then, a learnable position embedding and validity mask are constructed by combining the time and variable indices from S1 (these two can be combined to filter out combinations of "valid time points + valid variable types," generating a mask to avoid interference from invalid time points and ensure temporal alignment and physical consistency in irregular sampling and multivariate scenarios). This enables the encoder to maintain temporal alignment and physical consistency in irregular sampling and multivariate mixed scenarios. The rearranged features are fed into the Transformer encoder, where multi-head self-attention explicitly models cross-delay dependencies and cross-station collaboration in the sequence dimension, outputting a spatiotemporal representation after feedforward networking and residual normalization. Simultaneously, attention weights generated during the encoding process... These features are retained for subsequent decoding and interpretability analysis, thereby achieving a unified representation of temporal enhancement and cross-scale dependence of the high-order features extracted by S4.

[0147] After spatial features are aggregated by a hypergraph convolutional layer, the node feature sequence is input into the time series encoding layer. A multi-head attention mechanism captures the dependencies between different time steps, and temporal alignment is introduced to enhance robustness to misaligned observations. An improved Transformer encoder is employed, and its computation process is as follows:

[0148] Linear mapping applies a linear transformation to the input features to generate a vector of query, key, and value elements:

[0149] ;

[0150] in Let D be the learnable parameter matrix, and D represent the feature dimension.

[0151] Multi-head split

[0152] The transformed vector is divided into There are 1 attention head, and each head has 1 dimension. .

[0153] Attention weights are calculated by performing a scaled dot product attention score for each head and applying a temporal mask. To avoid invalid time points:

[0154] ;

[0155] in, This represents the weight of the h-th attention head. This represents the h1th attention head. ; Let h1 be the query vectors of attention heads. Let h1 be the key vectors of attention heads. For the mask matrix, This means replacing all zeros in the mask matrix with negative infinity; For the dimension of attention head, Indicates transpose; This is a binary adjacency mask tensor used to mark "valid observation node pairs" (value 1) and "invalid / irrelevant node pairs" (value 0); in this embodiment, masked_fill is used to... The attention score for location is set as This ensures that the weights of these positions are 0 after the softmax function, thereby preventing information propagation between invalid node pairs and retaining only the attention interactions of physically related nodes.

[0156] Perform matrix multiplication on each weight to obtain a weighted value vector. :

[0157] ;

[0158] in, This is a vector of h1 attention heads;

[0159] All of Concatenate the vectors to obtain the concatenated vector. ;right Perform a linear transformation to obtain the transformed vector. ;

[0160] ;

[0161] in, This is the projection matrix.

[0162] The final, improved Transformer encoder output is obtained:

[0163] ;

[0164] in, This indicates a normalization operation.

[0165] Traditional time-series encoders struggle to simultaneously capture short-term abrupt changes and long-term dependencies when processing non-stationary wind speed sequences, and also fail to express high-order global relationships such as upstream-downstream transmission and long-distance collaboration. This invention introduces multi-head attention and time alignment constraints into the time-series encoding layer: multi-head attention learns the correlations between different time steps in parallel across multiple subspaces, effectively identifying peak and phase changes in extreme events such as typhoon landfall and abrupt wind field changes induced by complex terrain. Simultaneously, by constructing time index matrices and variable index matrices, time alignment is explicitly incorporated into the attention weights, ensuring the learning process follows physical temporal laws, thereby improving the accuracy and stability of predictions. Based on this mechanism, the model is no longer limited to a local sliding window, enabling it to model long-range dependencies and global couplings across the entire sequence, accurately depicting the transmission of upstream disturbances downstream, collaborative driving between long-distance nodes, and multi-scale evolution processes. This invention also introduces an irregularity perception mechanism in attention computation, achieving dynamic aggregation of global dependencies through mask matrices and adjacency constraints, ensuring that the model propagates attention only between physically related node pairs. In terms of time series encoding layer design, this embodiment adopts an innovative encoding strategy that integrates a multi-head attention mechanism with a hypergraph structure, effectively improving the model's ability to express complex time patterns. Through a dual-scale encoding structure, the model can simultaneously capture short-term changes and long-term dependencies, handling abrupt wind speed changes and seasonal patterns respectively. This design is implemented through a multi-head attention mechanism, where the query, key, and value are obtained through linear transformations, then partitioned head-by-head for parallel computation, and finally expressed using a formula. Calculate the attention score, where This refers to the dimension of each attention head. Regarding physical consistency constraints, the model deeply integrates a masking mechanism into the attention mechanism. By constructing a binary mask tensor to explicitly identify effective observations, it ensures the model's robustness to misalignment between satellite and ground station observations, enhancing the model's physical consistency. Specifically, the masking mechanism is implemented through the formula... To achieve this, the attention scores for invalid locations are set to negative infinity, ensuring that these locations have zero weights after softmax. For the rate of change constraint mechanism, the model dynamically adjusts attention weights during attention calculation by merging coefficients, prioritizing nodes with high temporal alignment, enabling the model to better handle irregular sampling data. Furthermore, the model introduces an adjacency mask constraint mechanism. During hyperedge-to-hyperedge message passing, the adjacency mask limits the attention range, ensuring the model only focuses on relevant variables and time points, further enhancing the model's adaptability to irregular sampling patterns. These innovative designs enable the model to effectively handle complex spatiotemporal dependencies in multi-source heterogeneous meteorological data, improving the accuracy and robustness of wind speed prediction.

[0166] Multi-scale prediction and loss function

[0167] Based on the spatiotemporal feature representation output by the S5 time series coding layer, parallel multi-scale decoding is performed for different prediction step sizes. Historical observation sparsity features are introduced. (Indicates step size) (effective observation density within), of which .

[0168] This is the multi-task dynamic weighted prediction structure used in this embodiment, which simultaneously generates wind speed prediction results for the next 6 hours, 12 hours, and 24 hours.

[0169] The output layer of the model consists of a decoder function. Configuration, used to determine the hidden state based on the encoder output. Predicting step size and observation sparsity :

[0170] ;

[0171] Where t represents the current time; this design enables the model to learn short-term abrupt changes and long-term trends of wind speed in parallel at different time scales, achieving multi-scale spatiotemporal information fusion prediction.

[0172] To balance accuracy and stability, this module employs a dynamic weighted loss mechanism, which adaptively adjusts the relative weights of mean square error (MSE) and mean absolute error (MAE) based on the observation density at each time scale.

[0173] The overall loss function is defined as:

[0174] ;

[0175] in, and For time step The adaptive weights satisfy Dynamic weights are based on observation density. Automatic adjustment, defined as follows:

[0176] ;

[0177] The formula for calculating the observation density is:

[0178] ;

[0179] in Indicates the prediction window The number of valid observations within, This represents the theoretical maximum number of observations. When the observation density is high ( When the observation density is low, the model focuses more on the overall stability of the error distribution (primarily MAE); when the observation density is low ( When the model prioritizes detail accuracy and local consistency (mainly MSE), it achieves dynamic adaptation of prediction accuracy in sparse observation environments.

[0180] In multi-scale prediction and loss function design, this embodiment effectively solves the problem of insufficient prediction accuracy in wind speed forecasting through an innovative dynamic loss mechanism and multi-scale prediction structure. To address this issue, the model simultaneously predicts results at different time steps through a multi-scale prediction structure, improving its ability to capture short-term abrupt changes and long-term trends. This design allows the model to output prediction results for multiple time steps simultaneously, adapting to the needs of different application scenarios. Simultaneously, the model incorporates observation sparsity features to perceive the reliability differences across different prediction periods, enhancing its adaptability to data quality. Specifically, the model assesses data integrity by calculating observation density and integrates this feature into the prediction process, enabling the model to adjust its prediction strategy based on the data quality at different time periods. This mechanism is achieved through a dynamic loss function, adaptively adjusting the weight ratio of mean squared error (MSE) and mean absolute error (MAE) based on observation density. When observation density is high, the MSE weight is increased to better capture data details; when observation density is low, the MAE weight is increased to improve robustness to noise.

[0181] This embodiment also effectively addresses the problem of insufficient global coupling representation. To solve this problem, the model utilizes a dynamic weighting mechanism to adaptively adjust the loss function, enabling the model to optimize prediction strategies based on different observation conditions. This dynamic weighting mechanism is implemented through observation density information, reflecting the differences in observation integrity between sites and enhancing the model's ability to express complex spatiotemporal patterns. Specifically, during training, the model calculates the overall observation density and uses it as a confidence index to dynamically adjust the weights of the loss function. When the observation density is high (>0.7), the model increases the weight of MSE (Mean Sequence of Observations), enabling it to better capture detailed changes in the data and be sensitive to outliers; when the observation density is low (≤0.7), the model increases the weight of MAE (Magnitude of Observations), improving robustness to missing data and noise. Furthermore, the model explicitly handles missing data through a masking mechanism, ensuring that only valid observations are considered in the loss calculation, further enhancing the model's adaptability to irregularly sampled data. In this way, the model can adaptively adjust the prediction strategy according to different observation conditions, ensuring reliable prediction results under various data quality conditions and effectively improving the model's global coupling representation capability.

[0182] This embodiment employs an innovative dynamic loss function mechanism and feature fusion decoding strategy, enabling adaptive adjustment of the prediction strategy based on data reliability to ensure good predictive performance under various observation conditions. Specifically, this embodiment uses observation density as a confidence index to dynamically adjust the weight ratio of mean squared error (MSE) and mean absolute error (MAE) to achieve confidence estimation. When the observation density is high, the model increases the weight of MSE to better capture detailed changes in data and is more sensitive to outliers; when the observation density is low, the weight of MAE is increased to improve robustness to missing data and noise. In the decoder design, this embodiment uses a triple feature fusion mechanism, concatenating the features of observation nodes, temporal hyperedges, and variable hyperedges and then performing a linear transformation to obtain the final prediction result. This design fully utilizes the information from different components in the hypergraph structure. For the calculation of observation density, the model uses mask tensors to statistically determine the proportion of effective observations, ensuring the accuracy of confidence estimation. This physical consistency constraint mechanism allows the model to adaptively adjust the prediction strategy according to different observation conditions, ensuring reliable prediction results under various data quality conditions. Through this innovative loss function design and decoder structure, the model can not only improve the accuracy and stability of long-term predictions, but also maintain good robustness when faced with irregular sampling and missing data.

[0183] Model training and evaluation

[0184] After completing the multi-scale prediction objective and dynamic loss design, this embodiment presents the training and evaluation settings for the model. During the training phase, a dynamically weighted objective is used as the optimization criterion to perform end-to-end learning on the encoded spatiotemporal feature representation. The model parameters are optimized by minimizing the following objective function:

[0185] ;

[0186] in For all learnable parameters of the model, for Regularization coefficient, The loss is a combined MAE / MSE loss based on observation density. The training process uses the Adam optimizer (initial learning rate set to...). This is combined with a cosine learning rate decay strategy; an early stopping mechanism is triggered if the validation set loss does not decrease for two consecutive rounds to avoid model overfitting. During the evaluation phase, MAE, RMSE, and [other metrics] are reported simultaneously on both the validation and test sets. Three metrics were used to compare the model with classic baseline models such as LSTM (Long Short-Term Memory), Informer (Beyond Efficient Transformer for Long Sequence Time-Series Forecasting), DCRNN (Diffusion Convolutional Recurrent Neural Network), and STGCN (Spatio-Temporal Graph Convolutional Network) under the same experimental settings. After training, the optimal weight file and performance report (including...) were exported. The sub-results and overall performance form a complete closed loop from the definition of the prediction target to the model convergence verification and then to the quantitative evaluation of the effect.

[0187] This module uses an end-to-end training approach, employing the Adam optimizer for model training, with an initial learning rate set to... It employs a cosine scheduler strategy for learning rate decay. This setup effectively balances the convergence speed in the early stages of training with the accuracy improvement in the later stages.

[0188] An early stopping mechanism is introduced during training. Training terminates when the validation set loss fails to decrease for two consecutive rounds, preventing overfitting. This mechanism ensures the model stops training in its optimal state, improving its generalization ability.

[0189] The final model performance was measured by MAE, RMSE, and The model was evaluated using three metrics: (coefficient of determination) and compared with baseline models such as LSTM, Informer, DCRNN, and STGCN. These metrics measure the model's predictive accuracy and stability from different perspectives, ensuring the comprehensiveness of the evaluation results.

[0190] In the model training and evaluation design, this embodiment effectively solves the problem of insufficient prediction accuracy in wind speed forecasting through innovative optimization strategies and training mechanisms. To address this issue, the model employs the Adam optimizer combined with a cosinescheduler learning rate scheduling strategy to achieve more stable gradient updates and learning rate adjustments, thereby improving model convergence speed and final accuracy. The Adam optimizer combines the advantages of momentum methods and RMSprop (Root Mean Square Propagation), dynamically adjusting the learning rate of each parameter by calculating the first and second moment estimates of the gradient, as shown in the formula: and ,in and These are the first and second moment estimates of the gradient, respectively. This is the current gradient. and It is the decay rate. The cosine scheduler learning rate scheduling strategy adjusts the learning rate through cosine annealing, as shown in the formula: ,in and These are the upper and lower bounds of the learning rate, respectively. and These represent the current round and the total number of rounds, respectively. This combined strategy allows the learning rate to decrease rapidly in the early stages of training and converge slowly in the later stages, helping the model escape local optima and steadily converge to the global optimum. Simultaneously, the model introduces an early stopping mechanism to prevent overfitting. By monitoring the validation set loss, training is terminated early if the validation loss does not decrease for several consecutive rounds, ensuring the model achieves optimal performance on the validation set and improving its generalization ability.

[0191] In the model training and evaluation design, this embodiment effectively addresses the problem of insufficient global coupling representation. To solve this problem, this embodiment employs a multi-scale prediction and adaptive feature aggregation mechanism, enabling the model to dynamically adjust based on the weight contributions of different observation stations, thus enhancing its ability to express complex spatiotemporal patterns. This mechanism is implemented through the attention mechanism in the hypergraph structure, where temporal and variable hyperedges capture the dependencies of the time and variable dimensions, respectively, and the importance of different nodes is dynamically adjusted through a multi-head attention mechanism. Simultaneously, this embodiment utilizes the visualization results of attention weights to reveal the main controlling factors of regional wind speed, providing a scientific basis for wind energy layout and disaster prevention. By analyzing the trained attention weight matrix, the observation stations and time steps that contribute the most to the prediction results can be identified, thereby understanding the main driving factors of wind speed changes. This visualization mechanism not only improves the interpretability of the model but also provides valuable insights for practical applications, enabling the model to not only provide accurate prediction results but also reveal the key factors affecting wind speed changes, providing scientific support for applications such as wind energy resource assessment, wind farm site selection, and extreme weather early warning.

[0192] In terms of model training and evaluation, this embodiment achieves efficient modeling of complex spatiotemporal dependencies through a hypergraph neural network architecture, possessing powerful high-order spatial correlation modeling capabilities. The model employs a triple hypergraph structure design, capable of simultaneously representing group interactions between multiple observation stations, rather than being limited to pairwise edge connections. This high-order connection mechanism effectively captures the regional collaborative variation characteristics of the wind field. By constructing a joint structure of observation nodes, temporal hyperedges, and variable hyperedges, the model achieves cross-dimensional information fusion. The temporal hyperedge captures the evolutionary patterns of the same variable at different time points, while the variable hyperedge models the correlation patterns between different variables at the same time point. Regarding temporal dependency modeling, although this embodiment does not directly use a gated recurrent unit structure, it achieves effective capture of complex time-series patterns through the fusion of a multi-head attention mechanism and hypergraph convolution, combining memory capacity and flexibility, and can simultaneously characterize periodic wind speed changes and sudden wind speed anomalies. The model supports multi-scale prediction output, providing prediction results at multiple time steps according to different application needs, and achieves adaptive feature aggregation through a hyperedge attention mechanism, dynamically adjusting its weights based on the contribution of different observation stations. In terms of generalization and interpretability, the model maintains high accuracy under different geographical regions and climate backgrounds through irregularity perception attention mechanism and adjacency mask constraint. Its attention weight provides insights into the main controlling factors of regional wind speed, providing a scientific basis for wind energy layout and disaster prevention and control.

[0193] Dynamic Hypergraph Spatiotemporal Learning Mechanism

[0194] like Figure 5As shown, this training phase will simultaneously enable the dynamic hypergraph spatiotemporal learning mechanism as a topology adaptation strategy: each training step updates the hypergraph connections and hyperedge weights online based on the latest node representations, and adds learnable temporal frequency embeddings to characterize day-night and seasonal variations; the updated representations are directly fed into the multi-scale prediction head in step S5, and participate in backpropagation with dynamically weighted loss, forming a closed-loop optimization of "adaptive topology - feature propagation - multi-scale prediction - evaluation", thereby maintaining robust convergence and higher generalization performance under non-stationary wind fields and irregular sampling conditions.

[0195] This mechanism aims to break through the fixed topology assumption of traditional static hypergraph convolution and achieve adaptive modeling of the spatiotemporal evolution of wind speed fields. Traditional hypergraph neural networks typically assume a fixed topology. While remaining constant throughout the sequence, the spatial dependence of the wind field changes significantly over time due to the influence of typhoon path, frontal movement, and complex terrain, making it difficult for the model to effectively capture the dynamic coupling features between nodes. To address this, this invention proposes a dynamic hypergraph spatiotemporal learning mechanism, introducing a learnable topology evolution function in the time dimension, enabling the hypergraph structure to adaptively update with each time step. At each time step, the system jointly determines the current topology update form based on the previous time step's topology structure and node feature states, defined as:

[0196] ;

[0197] in: Time step The correlation matrix reflects the connection relationship between nodes and hyperedges; The topology of the previous time step; : Node feature matrix of adjacent time steps; Learnable nonlinear mapping functions integrate historical topology, feature differences, and temporal evolution trends to generate new time-varying hypergraph structures.

[0198] This update mechanism enables the model to automatically reconstruct its topology based on the wind field conditions at each time step, exhibiting self-organizing properties. After the structure update is complete, the model synchronously performs feature propagation. The dynamic hypergraph convolution propagation formula is defined as:

[0199] ;

[0200] in: : Node degree matrix, representing the hyperedge connection strength of each node; Hyperedge degree matrix, representing the number of nodes connected by each hyperedge; : A time-dependent learnable parameter matrix used to capture temporal differences; : Input feature matrix at the current time step; : Non-linear activation functions (such as ReLU or GELU); Time step The higher-order feature output.

[0201] This structure considers both higher-order connections between nodes and the temporal changes in hyperedge weights during propagation, enabling dynamic modeling of the coupling relationship between local and global wind fields. To further capture the periodic characteristics of the wind field (such as diurnal cycles, seasonal fluctuations, and monsoon rhythms), a time-frequency embedding term is introduced into the feature space, giving the model multi-scale time awareness. It is defined as follows:

[0202] ;

[0203] in: : A learnable frequency coefficient vector used to control the embedding period at different time scales; , : Represent the sine and cosine time modulation terms, respectively, used to parameterize the periodic variation of wind speed over time. This design enables the model to distinguish between short-period fluctuations (such as gusts) and long-term trends (such as seasonal wind speed changes) at the feature level. Combining time-frequency embedding and structural evolution mechanisms, the final dynamic propagation equation is expressed as:

[0204] ;

[0205] This equation simultaneously achieves: the dynamic evolution of the topology layer (through... The time-varying update reflects changes in the wind field structure; the periodic adaptation of the feature layer (through...) (Frequency embedding reinforces time pattern modeling); thereby completing the joint optimization of structural evolution and feature propagation.

[0206] Traditional static hypergraphs assume fixed topological relationships between nodes, making it difficult to reflect the rapid evolution of typhoon paths, frontal movement, and boundary layer wind field structures. This static assumption leads to models only capturing local or time-averaged dependency patterns during training, failing to respond instantly to sudden changes in wind speed and spatial transport processes. This invention introduces a time-dependent dynamic topology update mechanism, updating the topology at each time step... Adaptive reconstruction of the relationships between nodes: that is, based on the topology of the previous time step. Node status Compared with current observations The similarity between nodes and the weights of hyperedge connections are recalculated to achieve real-time capture of the time-varying structure of the wind field. This mechanism enables the model to spontaneously generate dynamic responses to local abrupt changes such as convection, frontogenesis, and shear lines, thereby achieving a significant improvement in the accuracy of predictions for extreme wind speed surges, gust propagation, and typhoon spiral rainbands.

[0207] Wind field systems generally exhibit cross-scale, multi-layered coupling relationships: local abrupt events and large-scale circulation interact, while small-scale disturbances propagate and amplify downstream along topography or circulation channels. Static maps or fixed topologies struggle to uniformly characterize this "multi-scale-multi-regional" interaction. This mechanism introduces temporal frequency embedding into the feature space, explicitly encoding changes at different time scales as learnable frequency components, and achieving co-modeling of short-period and long-period components in the hyperedge update function. The short-period component captures rapid oscillations and gust behavior, while the long-period component corresponds to seasonal or circulation pattern changes. Both are learned synchronously in the same topology, achieving globally consistent modeling across the spatiotemporal dimensions. This enables the model not only to identify local abrupt events but also to reveal their propagation paths and energy transfer mechanisms at the regional scale, thus achieving cross-scale co-expression in the global dynamic reconstruction of the wind field.

[0208] This invention proposes a time-adaptive topology update mechanism for dynamic meteorological processes, providing a hyperedge update function that unifies structural evolution and feature propagation within the same optimization framework. At each time step, the model re-estimates connectivity relationships and weights online based on the previous time step's topology and the current / previous time step's node states. This enables the hypergraph topology to "self-organize" and reconstruct itself according to the wind field state, allowing feature learning and structural updates to adapt synchronously. This results in a faster and more accurate response to sudden wind speed surges such as typhoon landfall and shear enhancement.

[0209] Introducing multi-scale frequency embedding at the representation level It simultaneously characterizes two types of time scales—short-term oscillations (such as gusts and local shear) and long-term cycles (day and night, seasonal / monsoon)—using continuous parameterization, achieving explicit decoupling and joint modeling within a single model. The synergistic effect of dynamic topology and frequency enhancement transforms the learning paradigm from a static graph with fixed relationships to a dynamic hypergraph with time-varying relationships, significantly improving the accuracy and timeliness of spatial pattern reconstruction and peak capture.

[0210] Figure 6 Figure 1 shows the distribution of predicted points obtained by the method of the present invention and the traditional method (basic model). Figure 2 shows the distribution of predicted points obtained by the traditional method and Figure 3 shows the distribution of predicted points obtained by the method of the present invention. In Figure 4, the MAE of the traditional method is 0.1187 and the MAE of the method of the present invention is 0.0785. It can be seen that the method of the present invention is 33.9% better than the traditional method.

[0211] Depend on Figure 7 It can be seen that the standard deviation of the traditional method (std) is 0.0749, while the standard deviation of the method of this invention (std) is 0.0607. This shows that the error distribution of the method of this invention is more concentrated. Figure 8 It can be seen that the present invention achieves higher accuracy when the measurement step length is 16 hours; Figure 7 and8 It can be seen that the method of the present invention significantly improves the prediction accuracy, which proves the effectiveness of the dynamic loss weight mechanism of the present invention.

[0212] Figure 10 The error distribution of the method of the present invention on a 24-hour prediction time series is shown, demonstrating the stability of the method of the present invention in long-term prediction, identifying time periods with large errors to facilitate further optimization, and providing comprehensive performance evaluation indicators.

[0213] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

Claims

1. A method for spatiotemporal prediction of typhoon wind speed, characterized in that, Specifically: Multi-source meteorological elements are acquired based on ground meteorological observation stations, satellite remote sensing platforms, and reanalysis datasets. A multi-source heterogeneous data fusion mechanism was used to preprocess multi-source meteorological elements to obtain an irregular spatiotemporal observation set; All ground meteorological observation stations, satellite remote sensing platforms, and typhoon observation systems corresponding to the reanalysis dataset are treated as nodes, with a total of N nodes; a feature matrix X is constructed for the N nodes, and temporal hyperedges and spatial hyperedges are also constructed. A spatiotemporal hypergraph is constructed based on the feature matrix X and the two types of hyperedges. High-order features are obtained by extracting features from the spatiotemporal hypergraph; The higher-order features are input into the improved Transformer encoder to obtain the predicted wind speed of the typhoon wind field at future time. The construction of temporal and spatial hyperedges is as follows: Temporal hyperedge: Connect all ground meteorological observation stations at the same observation time into a hyperedge to obtain the temporal hyperedge; The weight of the edge in the temporal hyperedge is determined by two parts: one is the mutual information of the wind field of each station under real-time observation, which is used to measure the similarity of the instantaneous wind vector between them; the other is the temporal alignment, which ensures that all data are strictly aligned to the same moment. Spatial hyperedge: Taking a single ground meteorological station as the axis, a hyperedge is formed by connecting all the observation nodes of the ground meteorological station throughout the day. The weight of the spatial hyperedge is determined by both spatial and physical factors: first, the geographical proximity between the station and its nearest neighbor is measured by Euclidean distance, and then a topographic slope correction term is introduced to give extra penalties or gains to stations located on hillsides, valleys or plateaus to reflect the diffusion and deposition effects of the real atmosphere constrained by the underlying surface.

2. The method for spatiotemporal prediction of typhoon wind field wind speed according to claim 1, characterized in that, Based on wind direction and time observed by ground meteorological stations, multi-source meteorological elements are analyzed. Sine and cosine transformations are introduced to extract the sine and cosine features of wind direction and time.

3. The method for spatiotemporal prediction of typhoon wind field wind speed according to claim 1, characterized in that, The following method was used to preprocess multi-source meteorological elements: Data with a missing rate below a preset threshold are directly removed, while data with a missing rate above a preset threshold are labeled using an explicit masking mechanism. Identify and parse the timestamp format of numerical or string data, and unify the time zone and granularity of all data; Using (station_id, timestamp) as the joint key, perform an outer join between the data observed by the ground meteorological observation station and the typhoon's optimal trajectory data, where station_id is the ID of the ground meteorological observation station and timestamp is the timestamp; Using timestamps as anchors, for each unique timestamp, the first remote sensing feature observed by the satellite remote sensing platform is broadcast and copied to all ground meteorological observation stations to achieve lightweight fusion; the first remote sensing feature is remote sensing information other than remote sensing texture information. When the timestamp of the remote sensing image matches the ground observation time, the remote sensing texture information observed by the satellite remote sensing platform is broadcast and copied to all ground meteorological observation stations; when the difference between the remote sensing image and the ground observation time is less than the preset time threshold, the remote sensing image is considered to match the ground observation time. For the typhoon trajectory data in the reanalysis dataset, a distance-based spatiotemporal nearest neighbor matching algorithm is used to construct an index bucket and select the typhoon center observation data that is closest to the corresponding ground meteorological observation station within a preset time window; Finally, all the data is standardized.

4. The method for spatiotemporal prediction of typhoon wind speed according to claim 1, characterized in that, The construction of the spatiotemporal hypergraph specifically involves: The feature matrix ,in For the real number field, Where T is the batch size and T is the time. For feature dimension; based on Heterogeneous graphs are constructed using heterogeneous graph neural networks, and node features in the heterogeneous graphs are fused: ； in, This represents the node features after fusion. This represents the node features before fusion. For nodes The neighborhood set, For a general linear mapping matrix, For specific mapping matrices whose data source type is from ground meteorological observation stations, satellite remote sensing platforms, or reanalysis datasets, The fusion coefficient is... The expression is as follows: ; in, For temperature parameters, For nodes Confidence weights For nodes and Feature similarity; After fusion Reassemble the new feature matrix ; Construct temporal and spatial hyperedges, and combine the heterogeneous graph output by the heterogeneous graph neural network with... The input is fed into a hypergraph neural network to obtain the final spatiotemporal hypergraph.

5. The method for spatiotemporal prediction of typhoon wind speed according to claim 1, characterized in that, Feature extraction of the spatiotemporal hypergraph is performed as follows: Employing a hypergraph convolutional propagation structure to analyze the feature matrix in the spatiotemporal hypergraph The following processing is performed to obtain the matrix. : ； in, The time hyperedge incidence matrix, The spatial hyperedge incidence matrix, Let be the degree matrix corresponding to the time hyperedge. Let be the degree matrix corresponding to the hyperedge in space. The weight of the time-bound edge. The weight of the spatial hyperedge, For activation function, Indicates transpose; Introducing wind direction correction: ； in: Based on the weight matrix, For wind direction dot squaring, The angle between wind directions; Attention weights for computation time hyperedge : ; in, For time-exceeding edge and node A set of connected nodes. The linear transformation matrix for the time channel. For the time channels, there are learnable attention vectors. node With nodes Geographical differences between them; and They are nodes With nodes eigenvalues; For activation functions; Attention weights of computation space hyperedges : ; in, For spatial hyperedges and nodes A set of connected nodes. Let be the linear transformation matrix of the spatial channel. This represents the learnable attention vector for the spatial channel. Represents a node With nodes The time difference between them; based on and Obtain higher-order features : ； in, for The matrix formed for The matrix formed As weight, It is a diagonal matrix function.

6. The method for spatiotemporal prediction of typhoon wind field wind speed according to claim 1, characterized in that, High-order features The ground meteorological observation stations are rearranged chronologically and at the same time according to their order, and then packaged into tensors in batches. These tensors are then fed into an improved Transformer encoder, which processes the input features as follows: Perform a linear transformation on the input features to generate a query Q, key K, and value V vector, and then segment these three vectors into... One point of attention; The weights of each attention head are calculated based on the temporal mask. : ; in, Indicates the first One point of attention, ; For the first A query vector with attention heads. For the first The key vector of each attention head. For the mask matrix, This means replacing all zeros in the mask matrix with negative infinity; For the dimension of attention head, Indicates transpose; Perform matrix multiplication on each weight to obtain a weighted value vector. : ; in, For the first A vector of values ​​for each attention head; All of Concatenate the vectors to obtain the concatenated vector. ;right Perform a linear transformation to obtain the transformed ; The final improved Transformer encoder output is obtained. : ； in, This indicates a normalization operation.

7. The method for spatiotemporal prediction of typhoon wind field wind speed according to claim 1, characterized in that, The objective function used during training is : ； in, For all learning parameters, For loss function, The regularization coefficient is used. The expression is as follows: ； in, Represents the set of predicted durations. and All are related to the predicted duration The corresponding weights Indicates the mean absolute error. This represents the mean square error. and The possible values ​​are as follows: ； in, For the observed density: ； in, Indicates the prediction duration The number of valid observations within, This represents the theoretical maximum number of observations.

8. The method for spatiotemporal prediction of typhoon wind speed according to claim 1, characterized in that, During training, a learning rate scheduling strategy combining the Adam optimizer and cosine scheduler is employed.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of a spatiotemporal prediction method for typhoon wind field speed as described in any one of claims 1 to 8.

Images