Regional extreme significant wave height space-time prediction method based on adaptive graph neural network

Abstract

The invention relates to the crossing field of marine environment information technology and artificial intelligence, and discloses a regional extreme significant wave height space-time prediction method based on an adaptive graph neural network, which comprises the following steps: acquiring input data, constructing input sequences, performing adaptive graph structure learning AGSL on each input sequence through a branch, and obtaining a prediction result; performing day sequence spatio-temporal correlation feature dynamic extraction STCFDE on the day sequence corresponding branches to obtain first output; comparing the prediction deviation of the day sequence with the prediction deviation of the month sequence, and generating a self-adaptive correction graph; performing year sequence periodic dependency mining PDM on the output of the branch corresponding to the year sequence to obtain a second output; and carrying out multi-modal fusion MMF on the first output, the correction graph and the second output, and finally outputting an effective wave height space-time prediction sequence in a future time period. According to the method, the dynamic space-time dependency relationship between the extreme significant wave height variables of the region can be automatically learned instead of depending on the preset static space relationship.

Description

Technical Field

[0001] This application belongs to the interdisciplinary field of marine environmental information technology and artificial intelligence, and specifically relates to a spatiotemporal prediction method for regional extreme effective wave height based on adaptive graph neural networks. Background Technology

[0002] Accurate prediction of the spatiotemporal evolution of extreme significant wave heights in a region is crucial for climate change research, disaster prevention and mitigation, and marine economic activities. Existing prediction methods are mainly divided into two categories: numerical models based on physical equations and data-driven machine learning models.

[0003] Numerical models (such as Wavewatch III and SWAN) offer physical interpretability but consume enormous computational resources and are sensitive to initial conditions and parameterization schemes, making them prone to introducing errors. Data-driven methods, such as the early ConvLSTM models, can learn spatiotemporal patterns from historical data, but they have significant limitations: 1. Static Spatial Assumption: Traditional models (such as those based on convolutional neural networks) assume that spatial relationships are translationally invariant, meaning that the same pattern produces the same response in different locations. This does not match the actual conditions of the marine environment. For example, the same wind field can produce drastically different wave responses in different sea areas due to factors such as topography and water depth.

[0004] 2. Euclidean Spatial Constraints: Most models deal with regular grid data, and their spatial dependencies are based on Euclidean distances. However, the true spatial dependencies between marine environmental variables are often complex non-Euclidean spatial relationships (for example, two ocean areas that are far apart but affected by the same ocean current system may be strongly correlated).

[0005] 3. Rigid Graph Structure: Although Graph Neural Networks (GNNs) have been introduced to handle non-Euclidean relationships, existing research largely relies on prior knowledge (such as geographical distance and Pearson correlation coefficients) to construct a globally static graph structure (adjacency matrix). The marine environment is dynamic, and this fixed graph structure cannot capture the characteristics of spatial dependencies evolving over time, resulting in limited prediction accuracy.

[0006] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] To address or at least alleviate one or more of the above problems, a spatiotemporal prediction method for regional extreme effective wave height based on an adaptive graph neural network is provided.

[0008] To achieve the above objectives, according to the first aspect of this application, a spatiotemporal prediction method for regional extreme effective wave height based on an adaptive graph neural network is provided, comprising: Acquire input data, which includes multivariate spatiotemporal sequences. and the corresponding graph structure; construct the input sequence, which includes daily sequences. , monthly sequence and year series ; Each input sequence is processed through a branch to perform Adaptive Graph Structure Learning (AGSL). The AGSL within one branch includes performing Specific Point Multivariate Spatiotemporal Feature Extraction (SSTC) on each input sequence and then obtaining an adaptive and optimal adjacency matrix A through graph construction (GC). The day sequence The corresponding branch performs dynamic extraction of daily series spatiotemporal correlation features using STCFDE to obtain the first output; Compare the recent daily sequences With the monthly sequence To calculate the prediction bias, an adaptive correction plot is generated. ; The annual sequence The output of the corresponding branch is subjected to periodicity mining (PDM) to obtain the second output; The first output and the correction diagram The second output is then subjected to multimodal fusion (MMF) to finally output the effective wave height spatiotemporal prediction sequence for future time periods.

[0009] By adopting the above technical solution, this application has the following beneficial effects compared with the prior art: Dynamic spatial dependency learning: Through the Adaptive Graph Structure Learning (AGSL) module, the assumption of global static spatial relationships is abandoned. It can automatically learn and update the dependencies between nodes at each time step or time period, which can better reflect the dynamic nature of the marine environment. Multi-scale feature fusion: By designing inputs and dedicated modules (STCFDE, BC, PDM) with multiple time scales (daily, monthly, and annual cycles), the short-term fluctuations, medium- and long-term evolution, and annual periodic patterns of marine environmental variables are captured simultaneously, resulting in more comprehensive predictions. Multivariate collaborative modeling: The SSTC module explicitly models the interaction between different marine environmental variables (such as wind field U10 / V10 and wave Hs), which improves the richness of feature representation and the physical rationality of prediction. Strong generalization ability: The model does not rely on a fixed sensor network or specific geographical prior knowledge. The graph structure obtained through learning has stronger generalization ability and can be applied to the prediction of different sea areas and different marine environmental variables.

[0010] The specific embodiments of this application will be described in further detail below with reference to the accompanying drawings. Attached Figure Description

[0011] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application. The illustrative embodiments and descriptions of the application are used to explain the application, but do not constitute an undue limitation of the application. Obviously, the drawings described below are merely some embodiments, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0012] In the attached diagram: Figure 1 This is a logical schematic diagram of the spatiotemporal prediction method for regional extreme effective wave height based on adaptive graph neural networks in this specific embodiment; Figure 2 This is a schematic diagram of specific point multivariate spatiotemporal feature extraction (SSTC) in this specific embodiment; Figure 3 This is a schematic diagram of the GC (GC) construction in this specific implementation. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application.

[0014] Please see Figure 1 This application provides a spatiotemporal prediction method for regional extreme effective wave height based on an adaptive graph neural network, comprising: Acquire input data, which includes multivariate spatiotemporal sequences. and the corresponding graph structure; construct the input sequence, which includes daily sequences. , monthly sequence and year series ; Each input sequence is processed through a branch to perform Adaptive Graph Structure Learning (AGSL). The AGSL within one branch includes performing Specific Point Multivariate Spatiotemporal Feature Extraction (SSTC) on each input sequence and then obtaining an adaptive and optimal adjacency matrix A through graph construction (GC). The day sequence The corresponding branch performs dynamic extraction of daily series spatiotemporal correlation features using STCFDE to obtain the first output; Compare the recent daily sequences With the monthly sequence To calculate the prediction bias, an adaptive correction plot is generated. ; The annual sequence The output of the corresponding branch is subjected to periodicity mining (PDM) to obtain the second output; The first output and the correction diagram The second output is then subjected to multimodal fusion (MMF) to finally output the effective wave height spatiotemporal prediction sequence for future time periods.

[0015] In some embodiments, input data is acquired, the input data including multivariate spatiotemporal sequences. and the corresponding graph structure; construct the input sequence, which includes daily sequences. , monthly sequence and year series .

[0016] S1: Spatiotemporal Sequence Dataset (STSD).

[0017] In this embodiment, we will define a module for generating graph spatiotemporal sequence data. This module mainly generates two types of graph spatiotemporal sequence data: (1) Graph spatiotemporal sequence data. The superscript 1-3 indicates that there are 3 channels, which are used to store data such as U10, significant wave height, and V10. These data are used to input the AGSL module to generate the optimal graph structure; (2) Graph spatiotemporal sequence data ,from Extract the effective wave height data of the second channel, and then combine it with the optimal graph structure generated in (1) to input the subsequent STCFDE, BC and PDM modules for wave variable prediction.

[0018] Assuming the reference time is The size of the prediction window is First, the graph node feature sequences input to the four modules are defined as daily sequences. , monthly sequence and year series The specific definition is as follows: day series From reference time Before Time step (1 day ≤ The graph node feature sequence (≤ 7 days) constitutes a historical time series segment directly adjacent to the prediction time. The length of this historical time series segment is equal to or slightly longer than the prediction time, reflecting the daily or weekly influence of the historical time series. Specifically, The structure is as follows: =( ) ; monthly series : By proximity of The graph consists of a sequence of graph node features within a time step. It is roughly located about a year before the recent time, reflecting the monthly to annual impact of historical time series. The specific number of days depends on Is it 2 days or 7 days? Specifically, The structure is as follows: =( ) ; Annual sequence : By proximity of The graph consists of a sequence of graph node features within a time step. It is a segment of historical time series, whose time attributes (such as season and month) and prediction time window Exactly the same. Specifically, It refers to the entire period of time during which the typhoon passes through. The structure is as follows: =( ) ; Secondly, the graph node feature sequences that need to be input into the AGSL module and the other three modules are divided into two types, represented as significant wave height sequence segments. and full participation wave data segment From this, we can infer that... , ,as well as These represent the daily, monthly, and yearly periodic data segments influencing the historical significant wave height time series, respectively; while , ,as well as These represent the complete historical data (including U). 10 V 10 The daily, monthly, and yearly cycles of the Hs time series affect the data segments.

[0019] In some embodiments, each input sequence is processed through a branch to perform Adaptive Graph Structure Learning (AGSL). The AGSL within one branch includes performing Specific Point Multivariate Spatiotemporal Feature Extraction (SSTC) on each input sequence and then obtaining an adaptive, optimal adjacency matrix A through graph construction (GC).

[0020] S2: Adaptive Graph Structure Learning (AGSL).

[0021] As can be seen from the knowledge of graph structure, Figure 1 Generally, it consists of graph nodes (in this article, geographic spatial points) and their corresponding adjacency matrices. The spatiotemporal sequence of the graph is stored in each graph node (spatial point), and the adjacency matrix represents the characteristic attributes (such as geographic attributes, relevance, etc.) between spatial points and their spatiotemporal sequences.

[0022] The graph structure learning process is as follows: Figure 2 and Figure 3 As shown, it mainly consists of a Specific Point Multivariate Spatiotemporal Feature Extraction (SSTC) module and a Graph Construction (GC) module. Specifically, SSTC is first used to obtain a fixed-point multivariate spatiotemporal relationship feature map, and then the obtained spatiotemporal relationship feature map is input into GC to obtain the adjacency matrix A.

[0023] S2.1: Specific point multivariate spatiotemporal feature extraction (SSTC).

[0024] Specifically, such as Figure 2 As shown, firstly, based on the channel dimension, the multivariable spatiotemporal sequence of shape (T,C,H,W) is... Separate into U 10 Spatiotemporal sequence (T,H,W), Spatiotemporal sequence of effective wave height (T,H,W) and V 10 The spatiotemporal sequence X³(T,H,W) is then obtained. , and Inputting "CNN block" will generate a feature vector with a corresponding shape of (T, H, W). , and Where T, H, and W represent the time length, height, and width of the sequence data, respectively. A "CNN block" consists of three cascaded convolutional layers, each followed by a ReLU activation function.

[0025] Next, , and Input into the "Attention Module" to generate the calculated attention map. , , ), and its shape is (Hᵢ×Hᵢ). These attention maps are generated from features from three channels respectively ( , , The generated query vector ( , , ), key vector ( , , ) and value vector ( , , The attention maps were calculated. Then, these attention maps ( , , ) and encoded features ( , , Perform element-wise multiplication (Hadamard product) to obtain the refined feature map. , and .

[0026] Subsequently, the learnable weight parameters λ¹, λ², and λ³ are each element-wise multiplied with the refined feature map, and the results are summed to obtain the refined multivariate spatiotemporal relationship feature map F. i .

[0027] Finally, a convolutional layer with a kernel size of 1, a stride of 1, and a ReLU activation function is applied to obtain the mixed multivariate spatiotemporal sequences. These mixed spatiotemporal sequences are then transformed into a shape of (C... i ×N i The graph spatiotemporal sequence X, where N i and C i These represent the number of nodes and the number of channels, respectively. Here, the number of channels is 1.

[0028] S2.2: Graph Construction (GC).

[0029] The graph structure learning process is as follows: Figure 3 As shown. Specifically, the mixed multivariate spatiotemporal sequence X obtained from the SSTC module is used to obtain the global static variable embedding E∈R. N×D , where each row vector ∈R D Representing variables The embedding is defined by D, where D is the embedding dimension. Subsequently, we define a novel variable proximity metric based on a two-layer fully connected neural network, which is based on... and Calculate the proximity of variables To assess and The interdependencies between them are determined. We calculate the proximity view of n interdependent variables by optimizing n weight vectors, and apply the Softmax function to their average as the final result. The formula for the graph learning function gl(·) is as follows: (1); (2);

[0030] in Let P represent the P-th viewpoint. In formula (1), || represents the tensor concatenation operation, and the training parameters... = ReLU(x) = max(0,x) is the modified linear unit activation function. , Let represent the learnable biases of the first and second fully connected layers from the p-th viewpoint. Softmax(x) is used to rescale the calculated variable proximity to the (0,1) range, thus obtaining the normalized adjacency matrix. Furthermore, we set a non-negative threshold τ. The formula for the normalized adjacency matrix A is as follows: ; otherwise (3);

[0031] In some embodiments, the daily sequence The corresponding branch performs dynamic extraction of daily series spatiotemporal correlation features using STCFDE to obtain the first output.

[0032] S3: Dynamic extraction of spatiotemporal relationship features of daily series (STCFDE).

[0033] like Figure 3 As shown, the STCFDE module consists of a Dynamic Spatiotemporal Convolutional Block (DSTGB) and a Graph Convolutional Layer (Gconv). Inside the DSTGB, there are two parallel convolutional layers (called AGconv layers) and a Temporal Gated-Conv layer. The two AGconv layers are used to capture spatial dependencies with the neighborhood, while the Temporal Gated-Conv layer is used to mine temporal dependencies with neighboring observations along the time dimension. Although both AGconv layers aim to capture spatial dependencies with adjacent intervals, they differ in certain aspects. The upper AGconv layer considers multiple factors (U... 10The AGconv layer below considers only the influence of Hs, while V10 and Hs are excluded from the AGconv layer. Specifically, the AGconv layer performs a spectral domain graph convolution operation, as shown in the following expression: (4); in These are the input values ​​of N Hs variables at time t. D is the identity matrix, and D is the degree matrix. Global static variable embedding. This can be viewed as node-specific parameters (also used for graph structure learning), weight pool These are shared parameters among nodes. This represents a learnable bias. The process EW can be viewed as learning node-specific patterns from a set of candidate patterns discovered in the Hs time series data. In this study, since the learned graph structure A is constructed from E learned from the time series data and automatically updated, and we have used the softmax(⋅) function to rescale the calculated variable proximity to the range (0,1) (Equation (3)), we can use the normalized adjacency matrix A as an adaptive adjacency matrix to represent spatial relationships and avoid computation. Therefore, we can further simplify the graph convolution in equation (4) to adaptive graph convolution (AGconv): (5); In our method, we use functions. This represents the AGconv layer. This layer performs different operations on the upper and lower branches: for the upper branch, A and E are denoted as... and For the lower branches, A and E are denoted as... and Therefore, the feature maps input to the "temporally gated convolution" layer can be obtained in the following way: (6); Where ⊙ represents the element-wise Hadamard product. It is a learnable parameter matrix and uses the Corrected Linear Unit (ReLU) activation function. The "temporally gated convolution" layer employs a full convolutional structure on the time axis to capture temporal dynamic features. Specifically, the "temporally gated convolution" layer contains a one-dimensional causal convolution (1-D Conv) of width Kt, followed by gated temporal units (GTUs) as non-linear activations, expressed as: (7); Among them, the function Used to represent the operation of a "temporally gated convolutional" layer. This represents the feature map input to the "temporally gated convolutional" layer. This represents the convolution operation. These are the parameters of the convolution kernel. and These are the hyperbolic tangent and sigmoid functions, respectively. Finally, an additional Gconv layer is implemented in the STCFDE module as the output layer, and its output is represented as: (8);

[0034] In some embodiments, the recent day sequence is compared With the monthly sequence To calculate the prediction bias, an adaptive correction plot is generated. .

[0035] S4: Monthly Series Bias Correction (BC).

[0036] To mitigate the impact of natural factors such as time, terrain, and season on the performance of prediction results, we designed a bias correction module, such as... Figure 3 As shown, this module is used to learn spatial differences and outputs an adaptive correction map accordingly to correct the prediction results. First, the differences will be calculated separately by... and The generated adjacency matrix and The mixing occurs before the "AGCN block" in the input BC module, and its mathematical expression can be expressed as: (9);

[0037] Short-term matrix More sensitive to recent, instantaneous changes in state (such as the immediate alteration of the wind and wave field relationship by an approaching typhoon); long-term matrix It places greater emphasis on stable, background climatological patterns (such as seasonal wind and wave relationships caused by large-scale circulation or fixed topography). Through fusion, the generated... As a spatiotemporally adaptive calibration baseline, it is fed into subsequent AGCN blocks. This baseline contains information on both "what is happening now" and "what usually happens in this season," enabling the model's final bias correction map to more accurately capture prediction biases caused by both fixed natural factors (topography, season) and transient effects, thus allowing for more effective correction. (Monthly series segments) Hs and These features are input into the "AGCN Block" to extract features containing medium- to long-term spatiotemporal patterns. This process can be represented as: (10); Among them, AGCN-Block typically consists of multiple stacked adaptive graph convolutional layers (AGconv) and activation functions (such as ReLU) to capture complex spatiotemporal dependencies.

[0038] Features output by AGCN Block With actual observation data of the daily cycle period By comparing the two and calculating the differences, preliminary deviation characteristics can be obtained: (11);

[0039] Here, ΔF quantifies the gap between the forecast based on the medium- to long-term background and the recent actual situation.

[0040] The calculated bias feature ΔF is fed into a graph convolutional layer (Gconv) to further mine its spatial distribution pattern, ultimately generating an adaptive bias correction map that is spatially related to location. : (12); in, Represents graph convolution operations. This calibration graph It includes corrections tailored to each spatial location. The final output of the BC module is a deviation correction plot. This correction plot will be used to correct the master prediction results generated by the STCFDE module. (i.e., the output of formula (8)). The core value of the BC module lies in its explicit learning and correction of systematic prediction biases caused by the combined effects of fixed natural factors (such as topography and climate) and different time scales. By comparing the "medium- to long-term background prediction" with the "recent reality", it can capture these systematic errors that the master prediction model (STCFDE) failed to learn and accurately compensate them through a spatially adaptive correction map, thereby significantly improving the accuracy and physical consistency of the final prediction results.

[0041] In some embodiments, the year sequence The output of the corresponding branch is subjected to periodicity mining (PDM) of the annual sequence to obtain the second output.

[0042] S5: Annual Series Periodicity Dependency Mining (PDM).

[0043] From the year sequence The system extracts cyclical characteristics from historical data of the same period to account for seasonal effects and improve the accuracy of long-term forecasts. The specific implementation mechanism is as follows: Figure 3 As shown, this module implements its function through the following process: and The data is input into a network consisting of multiple spatiotemporal convolutional blocks. This network is responsible for extracting typical, recurring spatiotemporal patterns, or "annual periodic features," from historical data from the same period. The formula is expressed as: (13); Then, a "temporally gated convolutional" layer is used on the time axis to capture temporal dynamic features, expressed as follows: (14); Among them, the function Used to represent the operation of a "temporally gated convolutional" layer. This represents the feature map input to the "temporally gated convolutional" layer. This represents the convolution operation. These are the parameters of the convolution kernel. and These are the hyperbolic tangent and sigmoid functions, respectively. Finally, an additional Gconv layer is implemented in the PDM module as the output layer, and its output is represented as: (15);

[0044] In some embodiments, the first output, the correction diagram The second output is then subjected to multimodal fusion (MMF) to finally output the effective wave height spatiotemporal prediction sequence for future time periods.

[0045] S6: Multimodal Fusion (MMF).

[0046] The output features of the STCFDE, BC, and PDM modules are fused. Fusion strategies can include weighted averaging, convolutional fusion, or attention-based fusion, ultimately outputting the future time period. Predicted sequences of marine environmental variables (such as significant wave height). First, the correction plot obtained from BC. It will be used to correct the master prediction results generated by the STCFDE module. (i.e., the output of formula (8)) (16); in, This is the final prediction result after bias correction. Then, a Gconv layer is used to process the corrected data. Map back to the predicted sequence: (17); Finally, through Fusion, and After weighted fusion, the final prediction result is obtained: (18); in, These are learnable weights.

[0047] The above are merely preferred embodiments of this application and are not intended to limit this application in any way. Although this application has disclosed preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make some modifications or alterations to the above-mentioned technical content to create equivalent embodiments without departing from the scope of the technical solution of this application. The implementation schemes in the above embodiments can also be further combined or replaced. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of this application without departing from the content of the technical solution of this application shall still fall within the scope of this application.

Claims

1. A spatiotemporal prediction method for regional extreme effective wave height based on adaptive graph neural networks, characterized in that, include: Acquire input data, which includes multivariate spatiotemporal sequences. And the corresponding graph structure; An input sequence is constructed based on the input data, the input sequence including daily sequences. , monthly sequence and year series ; Each input sequence is processed through a branch to perform Adaptive Graph Structure Learning (AGSL). The AGSL within one branch includes obtaining a fixed-point multivariate spatiotemporal relationship feature map by extracting specific point multivariate spatiotemporal features (SSTC) for each input sequence, and constructing a GC using the obtained spatiotemporal relationship feature map as input to obtain an adjacency matrix A. For the daily sequence The output of the corresponding branch is used to dynamically extract the daily series spatiotemporal correlation features using STCFDE to obtain the first output; The BC comparison with the nearest daily series is performed by correcting for monthly series deviation. With the monthly sequence To calculate the prediction bias, an adaptive correction plot is generated. ; For the year series The output of the corresponding branch is subjected to periodicity mining (PDM) to obtain the second output; The first output and the correction diagram The second output is then combined with a multimodal fusion (MMF) to finally output the effective wave height spatiotemporal prediction sequence for future time periods.

2. The method according to claim 1, characterized in that, The specific point multivariate spatiotemporal feature extraction (SSTC) method for obtaining a fixed-point multivariate spatiotemporal relationship feature map includes: First, based on the channel dimension, the multivariable spatiotemporal sequence is... Separate into U 10 Spatiotemporal sequence Effective wave height spatiotemporal sequence With V 10 Spatiotemporal sequence ; The obtained spatiotemporal sequence , and Input CNN blocks separately to generate feature vectors , and The CNN block consists of three cascaded convolutional layers, each followed by a ReLU activation function; eigenvectors , and The input is fed into the attention module to generate the calculated attention map. , , ; attention map , , With encoded features , , Perform element-wise multiplication to obtain the refined feature map. , , ; The learnable weight parameters λ¹, λ², and λ³ are respectively compared with the refined feature map. , , Perform element-wise multiplication, then sum the results to obtain the refined multivariate spatiotemporal relationship feature map F. i ; By applying a convolutional layer with a kernel size of 1, a stride of 1, and a ReLU activation function, a mixed multivariate spatiotemporal sequence is obtained. The resulting multivariate spatiotemporal sequence was then converted into a shape of (C). i ×N i The graph spatiotemporal sequence X, where N i and C i These represent the number of nodes and the number of channels, respectively. Here, the number of channels is 1.

3. The method according to claim 2, characterized in that, The step of constructing the adjacency matrix A from the obtained spatiotemporal relationship feature map input graph includes: The obtained spatiotemporal sequence X is used to obtain the global static variable embedding E; A novel variable proximity measurement method based on a two-layer fully connected neural network is proposed. This novel variable proximity measurement method is based on vectors. sum vector Calculate the proximity of variables To evaluate variables and variables The interdependencies between them; where vectors Representing variables Embedding; The proximity perspectives of n interdependent variables are calculated by optimizing n weight vectors, and the Softmax function is applied to the average of the proximity perspectives of the variables as the final result. The formula for the graph learning function gl(·) is as follows: (1)； ) (2)； in, Let || represent the variable proximity of nodes i and j from the P-th viewpoint; || represents the tensor concatenation operation. Represents the training parameters; ReLU is the modified linear unit activation function. , This represents the learnable weights of the first and second fully connected layers from the p-th viewpoint. , This represents the learnable bias of the first and second fully connected layers from the p-th viewpoint; Softmax is used to rescale the calculated neighborliness of variables to the range of (0,1), thereby obtaining a normalized adjacency matrix; Furthermore, by setting a non-negative threshold τ, the formula for normalizing the adjacency matrix A is as follows: ； otherwise (3); in, This represents the adjacency matrix of nodes i and j. The variable representing the proximity between nodes i and j.

4. The method according to claim 3, characterized in that, The daily sequence The output of the corresponding branch is used to dynamically extract the spatiotemporal correlation features of the daily series using STCFDE, and the first output includes: The dynamic extraction of daily sequence spatiotemporal correlation features STCFDE is achieved through a dynamic spatiotemporal convolutional block DSTGB and a graph convolutional layer Gconv. Inside the dynamic spatiotemporal convolutional block, there are two parallel AGconv layers and one temporally gated convolutional layer. The two AGconv layers are used to capture spatial dependencies with the neighborhood, and the temporally gated convolution is used to mine temporal dependencies with neighboring observations along the time dimension. The AGconv layer performs spectral domain graph convolution operations, as shown in the following expression: (4)； in, The input values ​​of the N effective wave heights Hs at time t are: It is the identity matrix, and D is the degree matrix; global static variable embedding. This can be viewed as a node-specific parameter, the weight pool. These are node-shared parameters; the EW process can be viewed as learning node-specific patterns from a set of candidate patterns discovered in Hs time series data. The normalized adjacency matrix A is used as an adaptive adjacency matrix to represent spatial relationships, thus avoiding computation. Furthermore, the graph convolution in formula (4) is simplified to adaptive graph convolution AGconv: (5)； Among them, the function This represents the AGconv layer. Indicates a learnable bias; The AGconv layer performs different operations on the upper and lower branches: for the upper branch, A and E are denoted as... and For the lower branches, A and E are denoted as... and ; Therefore, the feature maps input to the temporally gated convolutional layer can be obtained in the following way: (6)； Where ⊙ represents the element-wise Hadamard product. It is a learnable parameter matrix and uses the rectified linear unit ReLU activation function; The temporally gated convolutional layer consists of a one-dimensional causal convolution with a width of Kt, followed by gated temporal units (GTUs) as non-linear activations, expressed as: (7)； Among them, the function Used to represent the operations of time-gated convolutional layers. This represents the feature map input to the temporally gated convolutional layer. This represents the convolution operation. These are the parameters of the convolution kernel. and These are the hyperbolic tangent and the sigmoid function, respectively. Finally, a Gconv layer is used as the output layer, represented as: (8)； in, Represents graph convolution operations. This indicates that the final feature map is obtained.

5. The method according to claim 1, characterized in that, The BC comparison, which corrects for monthly series deviation, is closer to the daily series. With the monthly sequence To calculate the prediction bias, an adaptive correction plot is generated. include: will be respectively by and The generated adjacency matrix and Mixing is performed before inputting the AGCN block, as described below: (9)； Monthly sequence segments Effective wave height Hs and Together they are input into the AGCN block, represented as: (10)； The AGCN block consists of multiple stacked adaptive graph convolutional layers AGconv and the activation function ReLU; Features output by AGCN blocks With actual observation data of the daily cycle period By comparing the two and calculating the differences, preliminary deviation characteristics can be obtained: (11)； Among them, ΔF quantifies the gap between the forecast based on the medium- and long-term background and the recent actual situation; The calculated bias feature ΔF is fed into a graph convolutional layer Gconv to generate an adaptive, spatially relevant bias correction map. : (12)； in, This represents the graph convolution operation.

6. The method according to claim 5, characterized in that, The year sequence Perform periodic dependency mining (PDM) on the output of the corresponding branch to obtain the second output, which includes: Will and The input is fed into a network consisting of multiple adaptive graph convolutional layers (AGonv), and can be expressed as: (13)； Then, a temporally gated convolutional layer is used on the time axis to capture temporal dynamic features, expressed as: (14)； Among them, the function Used to represent the operations of time-gated convolutional layers. This represents the feature map input to the temporally gated convolutional layer. This represents the convolution operation. These are the parameters of the convolution kernel. and These are the hyperbolic tangent and the sigmoid function, respectively. Finally, a Gconv layer is used as the output layer, and the output is represented as: (15)； in, express, This represents the graph convolution operation.

7. The method according to claim 6, characterized in that, The first output and the correction diagram The second output is combined with a multimodal fusion (MMF) to finally output the spatiotemporal prediction sequence of effective wave height for future time periods, which includes: Correction plot obtained from monthly series deviation correction BC It will be used to correct the master prediction results generated by STCFDE. , is represented as: （16）； in, This is the final prediction result after bias correction; The corrected values ​​are then processed through a Gconv layer. Map back to predicted sequence : (17)； Through Fusion, and Weighted fusion is performed to obtain the final prediction result. : (18)； in, These are learnable weights.

Images