Short-term wave forecasting method based on node association features and dynamic trajectory features
By using an adaptive adjacency matrix and a node graph convolutional model with globally shared parameters, combined with dynamic trajectory feature extraction, the problem of insufficient forecast accuracy and temporal stability in traditional short-term wave forecasting methods is solved, achieving high accuracy and stability in wave forecasting.
Patent Information
- Application Number
- CN202611584706.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-14
- Publication Date
- 2026-11-17
AI Technical Summary
Existing short-term wave forecasting methods suffer from insufficient forecast accuracy and temporal stability. Traditional graph convolutional models struggle to adaptively learn the relationships between ocean grid nodes, and fixed convolutional kernels cannot dynamically adjust sampling positions, lacking a continuously progressive joint modeling mechanism.
Learnable node embedding is used to generate an adaptive adjacency matrix. The relationship between ocean grid nodes is constructed by combining mask constraints, self-loop addition and Top-K sparsification. Node graph convolution parameters are generated by a globally shared weight pool and a globally shared bias pool. Feature aggregation and mapping are performed using the node graph convolution parameters. Dynamic trajectory features are extracted by dynamic sampling of the trajectory flow field and gating update.
It improves the ability to represent the spatial correlation features of ocean waves, enhances the model's ability to express the heterogeneous features of nodes at different spatial locations, realizes progressive joint modeling of node spatial correlation information and dynamic migration trajectory information, and improves forecast accuracy and temporal stability.
Smart Images

Figure CN123115698A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of ocean wave height prediction technology, and in particular to a short-term ocean wave forecasting method based on node correlation characteristics and dynamic trajectory characteristics. Background Technology
[0002] Significant wave height is a crucial ocean dynamic parameter characterizing wave intensity. Its accurate forecasting plays a vital role in maritime safety, port shipping, marine engineering construction, and the utilization of marine renewable energy. Wave evolution is influenced by a combination of factors, including wind fields, ocean currents, and seabed topography, exhibiting strong nonlinearity, non-stationarity, and spatiotemporal correlation. While traditional numerical wave models can describe wave propagation and evolution, they suffer from high computational costs, strong dependence on boundary conditions, and insufficient forecast accuracy in nearshore areas. With the development of deep learning technology, graph convolutional networks and convolutional recurrent networks are increasingly being applied to significant wave height prediction, providing a new technical approach for short-term wave forecasting.
[0003] However, current deep learning-based wave prediction methods often rely on graph convolutional models to construct adjacency matrices based on human experience or pre-defined rules. The connections between nodes are fixed before model training, making it difficult to adaptively learn effective associations between different ocean grid nodes based on changes in wave data. Traditional graph convolutional models typically use shared convolution weights and bias parameters for different nodes, failing to fully reflect the differences in feature distribution and association patterns at different spatial locations. Furthermore, for situations where wave features spatially migrate over time, spatiotemporal prediction models based on fixed convolution kernels usually sample and aggregate historical features at preset locations. These sampling locations cannot be dynamically adjusted according to the migration direction of wave features, making it difficult to effectively align historical features and extract migration trajectory features during continuous evolution. Finally, existing methods often separate the spatial association modeling between ocean grid nodes from the dynamic migration modeling of wave features, lacking a continuously progressive joint modeling mechanism. This affects the accuracy and temporal stability of short-term wave forecasts. Therefore, a short-term wave forecasting method based on node association features and dynamic trajectory features is needed at this stage. Summary of the Invention
[0004] To address the technical problems of insufficient forecast accuracy and temporal stability in existing short-term wave forecasting methods, this invention provides a short-term wave forecasting method based on node association features and dynamic trajectory features.
[0005] Firstly, the short-term wave forecasting method based on node association features and dynamic trajectory features provided by this invention adopts the following technical solution: Short-term wave forecasting methods based on node association features and dynamic trajectory features include: Obtain the effective wave height grid sequence of the target sea area for continuous historical periods, map the effective ocean grid points to graph nodes according to the preset spatial index, and construct the node temporal feature tensor; Using the node temporal feature tensor as input, an adaptive adjacency matrix is generated through learnable node embeddings, and node features are aggregated and mapped using node graph convolution parameters to obtain node association features. Based on node association features, the node association features are reconstructed into regular grid features according to the rule spatial location corresponding to the node; Local spatial features are extracted based on regular grid features. Multiple trajectory flow fields are generated using current features and historical hidden states. The historical hidden states are dynamically sampled and gatedly updated along the trajectory flow fields to obtain dynamic trajectory features. By utilizing dynamic trajectory features, predictive channel mapping and predictive time dimension recovery are performed, outputting an effective wave height grid sequence for multiple future time steps.
[0006] Furthermore, the step of generating an adaptive adjacency matrix through learnable node embeddings includes constructing a learnable node embedding matrix and initializing the elements in the learnable node embedding matrix through Kaiming uniform initialization. Construct a mask constraint matrix and calculate a node similarity matrix based on the learnable node embedding matrix. Use the mask constraint matrix to set the node association score corresponding to invalid grid points to a preset minimum value. Add self-loops to the node similarity matrix after mask constraints, perform Top-K sparsification along the node dimensions to obtain a sparse similarity matrix, and then perform row-wise normalization on the sparse similarity matrix to obtain an adaptive adjacency matrix. The expression for the row-wise normalization is as follows: , in, This represents the normalized association weight between node j and node i in the adaptive adjacency matrix. Let be the element in the i-th row and j-th column of the sparse similarity matrix. The total number of nodes in the graph. It is the sum of all elements in the i-th row of the sparse similarity matrix.
[0007] Furthermore, the construction of the node graph convolution parameters includes initializing a global shared weight pool and a global shared bias pool, mapping the node embedding vector of each graph node to the global shared weight pool and the global shared bias pool respectively, generating graph convolution weights and graph convolution biases corresponding to each graph node, and using the generated graph convolution weights and graph convolution biases as the node graph convolution parameters of the corresponding graph node.
[0008] Furthermore, the aggregation and mapping of node features using node graph convolution parameters includes: The node temporal feature tensor is input into a two-level concatenated ST-Conv block. In each ST-Conv block, a one-dimensional causal convolution and gating operation are performed on the node features along the time dimension through the first time-gated convolutional layer to obtain temporal gated features. The temporal gated features are aggregated based on an adaptive adjacency matrix, and the aggregated features are differentially mapped using node graph convolution parameters to obtain intermediate spatiotemporal features. Temporal modeling is performed on the intermediate spatiotemporal features through a second time-gated convolutional layer, and the output of the second-level ST-Conv block is used as the spatiotemporal fusion feature. The expression of the intermediate spatiotemporal feature is as follows: , in, For the first Intermediate spatiotemporal characteristics of the ST-Conv block output, For the hierarchical index of ST-Conv blocks, To enable first-time gated convolutional layers at nodes j The timing gating characteristics of the output, and These are the node graph convolution weights and node graph convolution biases corresponding to node i, respectively. This refers to the feature reorganization operation performed based on the node index. It is a non-linear activation function.
[0009] Furthermore, the aggregation and mapping of node features using node graph convolution parameters also includes inputting spatiotemporal fusion features into an additional time block, performing temporal convolution on the spatiotemporal fusion features through multiple parallel one-dimensional temporal convolution branches, fusing and nonlinearly activating the outputs of the multiple temporal convolution branches to obtain temporally refined features, flattening them, and mapping the flattened features to the target prediction dimension through a fully connected layer to obtain node association features. The expression for the node association features is: , in, For node association features, For additional time blocks, the input features are... , and These are three parallel one-dimensional temporal convolution operations. For gated activation functions, It is a non-linear activation function. For feature flattening operation, For fully connected linear mapping operations, To predict the time step, The number of feature channels, This represents the total number of nodes in the graph.
[0010] Furthermore, the step of reconstructing node-related features into regular grid features includes establishing a spatial positional correspondence between graph nodes and regular latitude and longitude grid points based on a preset spatial index, splitting the node dimensions of the node-related features using the correspondence, rearranging the features corresponding to each graph node to the height and width dimensions of the regular latitude and longitude grid points, fixing the prediction time dimension and feature channel dimension of the node-related features, and performing dimensional reconstruction on the rearranged features to obtain regular grid features.
[0011] Furthermore, the process of generating multiple trajectory flow fields using current features and historical hidden states includes performing spatial convolution and nonlinear activation on regular grid features, extracting local spatial features between adjacent regular grid points, and using the local spatial features as the current features; Convolutional mapping is performed on the current feature and the historical hidden state of the previous time step respectively to obtain the current feature mapping component and the historical hidden state mapping component. The two are then fused and nonlinearly activated to obtain the flow field generation feature. Convolutional mapping is performed on the generated flow field features, and the features are segmented along the feature channel dimension to obtain flow field offset information corresponding to multiple trajectories. The generation process of the flow field offset information satisfies the following: , in, This is a collection of flow fields corresponding to multiple trajectories. This refers to the segmentation operation performed along the feature channel dimension. For convolution operations, Features for flow field generation.
[0012] Furthermore, the dynamic sampling and gating update of the historical hidden state along the trajectory flow field includes determining the continuous historical sampling coordinates for each of the multiple trajectory flow fields based on the offset between the current spatial position and the trajectory flow field. Based on continuous historical sampling coordinates, the historical hidden state of the previous time step is sampled in the opposite direction of the trajectory flow field, and the historical trajectory features corresponding to each trajectory are obtained by bilinear interpolation. The historical trajectory features corresponding to multiple trajectories are concatenated along the feature channel dimension to obtain multi-trajectory concatenated features. Gated updates are then performed using these multi-trajectory concatenated features and the current features. An update gate adaptively fuses the historical hidden states and new memory features to obtain dynamic trajectory features. The historical trajectory features satisfy the following conditions: , in, For the first The historical trajectory characteristics of the trajectory at the current position p, For the first The historical sampling coordinates corresponding to each trajectory For the set of adjacent grid points participating in bilinear interpolation, For adjacent grid points, Let be the eigenvalue of the historical hidden state of the k-th ST-TrajGRU block at grid point q. Let grid point q be the coordinates of continuous historical sampling. Distance in the x-direction, Let grid point q be the coordinates of continuous historical sampling. Distance in the y-direction.
[0013] Furthermore, the output of the effective wave height grid sequence for multiple future time steps includes inputting dynamic trajectory features into the prediction output structure, converting the dynamic trajectory features to the target prediction channel through convolutional mapping to obtain prediction features, then performing dimensional transformation on the prediction features to restore the prediction time dimension, and generating the effective wave height grid sequence for multiple future time steps according to the spatial organization of the regular grid.
[0014] Furthermore, the step of outputting the effective wave height grid sequence for multiple future time steps also includes obtaining the predicted effective wave height grid sequence and the corresponding true effective wave height grid sequence, calculating the squared error between the predicted effective wave height grid sequence and the true effective wave height grid sequence at each prediction time step, summing the squared errors at each prediction time step to obtain the model training loss, and updating the model parameters based on the model training loss.
[0015] In summary, the present invention has the following beneficial technical effects: 1. This invention generates an adaptive adjacency matrix by embedding learnable nodes, and constructs the association relationship between ocean grid nodes by combining mask constraints, self-loop addition, Top-K sparsification and row-by-row normalization. This realizes the transformation of node connection relationship from manual preset to data-driven learning, which can eliminate the interference of invalid grid points on node association modeling, reduce the computational complexity of graph structure, and improve the representation ability of ocean wave spatial association features.
[0016] 2. This invention generates corresponding node graph convolution parameters for different ocean grid nodes by mapping a globally shared weight pool, a globally shared bias pool, and node embedding vectors. This enables different nodes to be aggregated and mapped differently according to their own characteristics, enhancing the model's ability to express the heterogeneous features of nodes in different spatial locations and improving the accuracy of node association feature extraction.
[0017] 3. This invention reconstructs node-related features into regular grid features and generates multiple trajectory flow fields based on current features and historical hidden states. By dynamically sampling the historical hidden states along the trajectory flow fields, it can extract the migration direction and spatial displacement information of wave features at continuous moments, thereby enhancing the model's ability to characterize the dynamic evolution trajectory of waves.
[0018] 4. This invention achieves progressive joint modeling of node spatial association information and dynamic migration trajectory information by concatenating node association feature extraction, regular grid feature reconstruction, dynamic trajectory feature extraction and prediction output in the order of feature transmission. This enables the final features used for prediction to simultaneously include the spatial association characteristics and continuous evolution characteristics of the wave field. Attached Figure Description
[0019] Figure 1 This is a schematic diagram of the overall process of the short-term wave forecasting method based on node association features and dynamic trajectory features according to an embodiment of the present invention.
[0020] Figure 2 This is a schematic diagram of the overall network structure of the NAT-HCNN model in an embodiment of the present invention. Detailed Implementation
[0021] The present invention will be further described in detail below with reference to the accompanying drawings.
[0022] Example 1 Reference Figure 1 The short-term wave forecasting method based on node association features and dynamic trajectory features in this embodiment includes: S1. Obtain the effective wave height grid sequence of the target sea area for continuous historical periods, map the effective ocean grid points to graph nodes according to the preset spatial index, and construct the node temporal feature tensor; S2. Using the node temporal feature tensor as input, an adaptive adjacency matrix is generated through learnable node embeddings, and node features are aggregated and mapped using node graph convolution parameters to obtain node association features. S3. Based on the node association features, reconstruct the node association features into regular grid features according to the rule spatial location corresponding to the node; S4. Extract local spatial features based on regular grid features, generate multiple trajectory flow fields using current features and historical hidden states, and dynamically sample and gate-update the historical hidden states along the trajectory flow fields to obtain dynamic trajectory features; S5. Utilize dynamic trajectory features to perform prediction channel mapping and prediction time dimension recovery, and output the effective wave height grid sequence for multiple future time steps.
[0023] Specifically, the short-term wave forecasting method based on node association features and dynamic trajectory features includes the following: like Figure 1 As shown in Figure S1, in this embodiment, the sea area to be predicted is selected as the target area, and the effective wave height grid sequence of the target sea area within a continuous historical period is obtained. The effective wave height grid sequence can use ERA5 ocean reanalysis data, whose original data is organized according to a regular spatial grid. Let the height of the regular spatial grid be H and the width be W, and the spatial position of any grid point be represented as... Where h represents the row coordinate of the grid point and w represents the column coordinate of the grid point.
[0024] To avoid invalid grid points such as land and islands from participating in subsequent graph structure modeling, the regular spatial grid is filtered according to a predefined ocean region mask to determine the set of valid ocean grid points. The ocean region mask satisfies: , Based on the ocean region mask, the effective set of ocean grid points is obtained: , in, For an effective set of ocean grid points, Spatial location The mask value at that location, and These represent the height and width of the regular spatial grid, respectively.
[0025] Then, the valid ocean grid points are mapped to graph nodes one by one according to the preset spatial index. The preset spatial index is used to establish the correspondence between graph nodes and regular spatial grid points. Through the above mapping, each graph node corresponds to a specific valid ocean grid point.
[0026] After completing the graph node mapping, the effective wave height values of each valid ocean grid point at different historical times are extracted according to the graph node index, and the effective wave height values at the same time are organized into node feature vectors. Let the current reference time be... The duration of historical input is Then from time At that time The continuous historical significant wave height data constitute the input time window. The node features at each moment within the input time window are arranged sequentially along the time dimension to construct the node temporal feature tensor: , in, For node temporal feature tensors, For a moment Corresponding node features The total number of graph nodes obtained by effectively mapping ocean grid points. The number of feature channels, The historical input time step. When only the effective wave height is used as the input variable, the number of feature channels... The value is 1.
[0027] S2, Node Temporal Feature Tensor As input, such as Figure 2 As shown, the node adaptive correlation feature extraction module performs spatial correlation modeling and temporal feature extraction on the effective wave height historical sequence. The node correlation feature extraction submodule includes two-level ST-Conv blocks, an additional time block, and a fully connected layer.
[0028] S21. Generate an adaptive adjacency matrix; First, construct a learnable node embedding matrix. ,in, Embedding dimensions for nodes, each graph node corresponds to one The node embedding vectors are updated end-to-end during model training, and then a predefined two-dimensional mask matrix is introduced. The two-dimensional mask is flattened into a one-dimensional vector according to fixed spatial indices. The flattening operation satisfies: , in, Let i be the i-th element in the one-dimensional mask vector. Let be the element in the h-th row and w-th column of the two-dimensional mask matrix. To index the flattened nodes, after flattening, a mask constraint matrix is constructed based on the one-dimensional mask vector. Each element in the mask constraint matrix is obtained by the outer product operation of the flattened vectors. , in, Let be the mask constraint matrix. The mask constraint matrix satisfies: if and only if the nodes i and nodes j When all are effective ocean grid points, ;otherwise, By using a mask constraint matrix, the node association score corresponding to invalid grid points can be set to a minimum value, ensuring that the graph structure is established only between valid grid points in the ocean, thus avoiding interference from land or island grid points on node association learning.
[0029] Based on the above node embedding matrix and mask constraint matrix Calculate the similarity matrix between node embedding vectors. This matrix quantifies the similarity between any two nodes, and the formula is: , in, The node similarity matrix is used to quantify the similarity between any two graph nodes. The more similar the node embedding vectors are, the higher the corresponding similarity value. Next, we will analyze the original similarity matrix... conduct Activation process: , in, This is the node similarity matrix after ReLU activation. This operation sets all negative similarity values to zero, retaining only non-negative node feature associations. Then, a mask constraint is applied: , in, This is the node similarity matrix after mask constraints. Let be the element in the i-th row and j-th column of the mask constraint matrix. The minimum value is set as the preset value. After masking constraints, the similarity values between valid ocean nodes are preserved, while the similarity values corresponding to non-ocean nodes are set to the minimum value to ensure that invalid nodes are not selected in subsequent Top-K sparsification.
[0030] To improve node feature stability and gradient propagation, a self-loop is forcibly added to each node: , in, for The identity matrix is of order 1. After adding self-loops, each node retains its own feature information during neighborhood aggregation, preventing node features from completely relying on neighboring nodes during the aggregation process. The similarity matrix after adding self-loops is sorted in descending order by row, and the top K highest-scoring neighboring nodes in each row are selected. The similarity of the remaining nodes is set to zero, resulting in a sparse similarity matrix. , in, Represents a node Highest similarity A set of neighborhood indexes Given the element in the i-th row and j-th column of the sparse similarity matrix, after obtaining the Top-K sparse similarity matrix, this invention further proposes a row sum constraint and normalization method for sparse node association relationships. First, the sum of elements in each row of the sparse similarity matrix is calculated. To avoid division by zero errors, the minimum value of the row sum is constrained to... : , in, The sum of the constraint rows corresponding to node i. Based on the constrained node association weights, the Top-K sparse similarity matrix is normalized row by row to generate an adaptive adjacency matrix for node association feature extraction. , in, To determine the normalized association weights of node j with respect to node i in the adaptive adjacency matrix, the normalized adjacency matrix is... Key spatial connections between ocean grids were preserved.
[0031] S22. Generate node-specific graph convolution parameters; Spatial graph convolutional layers based on adaptive adjacency matrix Perform graph convolution and generate unique weights and biases for each ocean grid node using a shared weight pool and a shared bias pool. First, define the node embedding matrix. Initialize the global shared weight pool and initializing the global shared bias pool ,in The spatial feature dimension of the output of the graph convolutional layer.
[0032] By using the Einstein summation convention, node embeddings are mapped to node-specific weight matrices: , In the formula, This is the node-specific weight matrix corresponding to node i; Let i be the node embedding vector of node i; Let k be the k-th element in the node embedding vector of node i; For the first in the globally shared weight pool k The weight parameters corresponding to each dimension, and the exclusive weight matrices of all nodes together form a weight tensor with dimensions N×C×F.
[0033] Map the node embedding to a node-specific bias vector: , In the formula, For nodes i The corresponding node-specific bias vector; For the first in the globally shared bias pool k The bias parameters corresponding to each dimension. The individual bias vectors of all nodes collectively constitute a bias tensor of dimension N×F.
[0034] Because each node has a different node embedding vector, the node-specific weights and biases generated by the mapping are also different, thus enabling different ocean grid nodes to use graph convolution parameters adapted to their own characteristics for spatial feature mapping.
[0035] S23. Perform ST-Conv block feature extraction; The node association feature extraction submodule comprises two levels of ST-Conv blocks. Each level of ST-Conv block includes a first temporally gated convolutional layer, a spatial graph convolutional layer, and a second temporally gated convolutional layer. The first temporally gated convolutional layer is used to extract temporal variation information of node features, the spatial graph convolutional layer is used to extract spatial association information between nodes, and the second temporally gated convolutional layer is used to further perform temporal modeling on the spatial association features.
[0036] The first time-gated convolutional layer applies one-dimensional causal convolution along the time dimension and combines it with gated time units to capture the evolution characteristics of the wave field. Its operation is represented as follows: , In the formula, This refers to the operations of time-gated convolutional layers. For the first The input of the first time-gated convolutional layer in each ST-Conv block For temporal convolution operations, For convolution kernel parameters, For element-wise Hadamard product; The Sigmoid function is used to control the correlation of the current state. Input the time length. is the temporal convolution kernel size.
[0037] After obtaining the node-specific graph convolution parameters, the adaptive adjacency matrix is used to aggregate the node features output by the first-time gated convolutional layer to obtain neighborhood information. For the i-th node, its adaptive neighborhood aggregation feature is represented as: , In the formula, For node i in the th Adaptive neighborhood aggregation features in each ST-Conv block For the first The temporal features output by the first time-gated convolutional layer at node j in each ST-Conv block; N is the total number of nodes.
[0038] Through weighted aggregation, each node can gather information from other ocean grid nodes with strong correlations with it based on the relationships learned through data-driven learning.
[0039] By inputting the aggregated features of node i into the dedicated parameter mapping process generated by node embedding, the spatial association features of that node are obtained: , In the formula, For node i in the th Spatial association features in each ST-Conv block The node-specific weight corresponding to node i. Assign a node-specific bias to node i; reorganize and nonlinearly activate the spatial correlation features obtained from each node to obtain the i-th node. Intermediate spatiotemporal characteristics of each ST-Conv block: , In the formula, For the first The intermediate spatiotemporal features output by each ST-Conv block; Stack(⋅) is the operation of recombining the spatially related features of each node according to the node index to form a node feature tensor; ReLU(⋅) is a nonlinear activation function.
[0040] Will Enter the first The second temporally gated convolutional layer of each ST-Conv block continues to perform temporal modeling on spatially correlated features: , In the formula, For the first The output of the second temporally gated convolutional layer in each ST-Conv block; after processing by two ST-Conv blocks, the output of the second ST-Conv block is the spatiotemporal fusion feature. In the formula, Input features for additional time blocks; The spatiotemporal fusion features output by the second-level ST-Conv block.
[0041] S24. Perform time-series purification and dimension mapping. After the two-level ST-Conv blocks output spatiotemporal fusion features, an independent time block is set up to refine the temporal features of the high-dimensional spatiotemporal fusion features. The time block no longer performs spatial graph convolution, but processes the high-dimensional spatiotemporal features along the time dimension to enhance the model's ability to capture long-term temporal dependencies across time steps.
[0042] Time blocks As input, a multi-branch one-dimensional temporal convolution structure is adopted, and temporal convolution is performed independently on the high-dimensional spatiotemporal features of each node. The temporal representation capability is improved by multi-branch fusion. Its core operation is as follows: , The temporal refined features output from the time block are first flattened, merging the time step dimension with the feature channel dimension: , Subsequently, a fully connected layer is used to linearly map the flattened node features to obtain the node association features: , In the formula, , and These are three parallel one-dimensional temporal convolution operations. For gated activation functions, This represents the number of remaining effective time steps after temporal convolution. It is a non-linear activation function. For feature flattening operation, For fully connected linear mapping operations, To predict the time step, The number of feature channels, This represents the total number of nodes in the graph.
[0043] S3. The node association features output in step S2 are transformed into their spatial organization form through the grid feature reconstruction submodule. The node association features are still organized according to the graph node dimension. Although the spatial association information between different ocean grid nodes has been integrated, the two-dimensional spatial arrangement of the regular latitude and longitude grid has not yet been restored. To facilitate the execution of spatial convolution, flow field generation, and trajectory sampling under a regular grid, the node association features need to be rearranged to their original regular spatial positions.
[0044] Based on the fixed spatial index established in step S1, the spatial correspondence between graph nodes and regular latitude and longitude grid points is determined. For any graph node, the corresponding regular latitude and longitude grid point position is determined according to its node index, and the features corresponding to the graph node are rearranged to the height and width dimensions of the regular latitude and longitude grid points. During the rearrangement process, only the node dimensions are split into the height and width dimensions of the regular grid, without changing the feature values already learned by each graph node, nor the information in the prediction time dimension and feature channel dimension.
[0045] After the node features are rearranged, dimensionality reconstruction is performed on the rearranged features to obtain regular grid features: , in, To reconstruct the regular grid features, For the dimension reconstruction operation, after dimension reconstruction, the node association features are transformed from being organized according to the graph node index to being organized according to the rule latitude and longitude grid.
[0046] S4, Regular mesh features reconstructed in step S3 As input, the spatial migration process of ocean wave features is modeled through the trajectory-aware spatiotemporal feature extraction submodule. This submodule comprises multiple ST-TrajGRU blocks, each including a spatial convolutional layer and a TrajGRU layer. The spatial convolutional layer extracts local spatial features between adjacent regular grid points, while the TrajGRU layer dynamically samples and gates the historical hidden states based on the trajectory flow field.
[0047] Spatial convolutional layers perform convolution operations on regular grid features, extracting the spatial dependencies between adjacent regular grid points, and introduce non-linear expressive power through the ReLU activation function. The output of the spatial convolutional layer satisfies: , in, For spatial convolution operations, The structured spatial features output by the spatial convolutional layer are preserved in this feature map. The regular grid structure ensures that TrajGRU can perform flow field generation and trajectory sampling operations for each grid feature point.
[0048] The feature maps output from the spatial convolutional layer are fed into the TrajGRU network. This network, based on flow field generation and multi-trajectory sampling mechanisms, replaces the static convolution operation of traditional gated recurrent units, explicitly models the temporal evolution trajectory of each grid feature point, and adaptively fuses current spatial features and historical temporal features sampled along the trajectory through a gating mechanism to capture the continuous motion and evolution of ocean waves. The TrajGRU layer contains reset gates, update gates, and new memory gates. Combined with the motion offset information generated by the flow field, it achieves efficient modeling of the long-term temporal dependence of ocean waves. Its core computation process is defined as follows: The first step is flow field generation. This involves generating the input feature map for each time step. Hidden state from the previous moment By fusing current spatiotemporal features and historical time-series features through convolutional networks, a new system is generated. The flow field migration information of the trajectory is used to characterize the feature points. Spatial motion direction in the grid: , , , in, Generate components for the flow field corresponding to the current feature. Let be the current feature of the k-th ST-TrajGRU block at time step t. For convolution operations, Generate components for the flow field corresponding to the historical hidden state. Let k be the historical hidden state of the k-th ST-TrajGRU block at time step t−1. Features are generated for the flow field. These features simultaneously contain spatial state information from the current time step and historical state information from the previous time step, and are used to predict the spatial migration direction and offset of ocean wave features.
[0049] The flow field generation features are convolutionally mapped and then segmented along the feature channel dimension to obtain the flow field corresponding to multiple trajectories: , in, For the first Flow field of the trajectory Representing grid feature points in Directional offset of motion The preset number of trajectories is used to adapt to the multi-directional characteristics of wave evolution; For convolution operations, This is a feature segmentation operation for the channel dimension.
[0050] Next is multi-track sampling. Based on the generated... The flow field hides the state from the previous moment. Perform reverse flow field mesh sampling operation to extract historical time-series features along the motion trajectory of feature points, thereby aligning the historical features with the current features' trajectories: , , in, This is a grid sampling operation based on flow field migration. The negative sign indicates that historical features are extracted against the direction of flow field motion. By concatenating features along the channel dimension and fusing historical temporal information from multiple trajectories, the output is... .
[0051] For the mesh sampling operation based on flow field migration, for the th The trajectory at the current position At that point, based on the flow field Determine the sampling location in the hidden state of the previous time step. Since historical features need to be traced back along the reverse direction of the flow field motion, the sampling coordinates are represented as follows: , in, For the first The trajectory is at the current position p along x The offset in direction. For the first The offset of the trajectory along the y-direction at the current position p. Since the offset is a continuous value, the sampling position... They typically do not completely overlap with regular grid nodes, therefore bilinear interpolation is used to extract the hidden state from the previous time step. Obtaining historical features: , in, for The continuous sampling coordinates corresponding to the trajectory This represents the set of neighboring points around the sampling location that participate in bilinear interpolation. Indicates one of the adjacent grid points, This indicates the hidden state at the grid point in the previous time step. eigenvalues at that location For the first The historical trajectory features extracted at the current position p of the trajectory are used. Through the above weighted interpolation, the historical trajectory features along the trajectory can be obtained. Historical features of dynamic alignment between the trajectory and the current position .
[0052] To each After performing the above grid sampling on the flow field, the obtained historical trajectory features are stitched together along the channel dimension: , in, This represents the feature concatenation operation along the channel dimension. As a result, historical motion information described by different flow fields is uniformly aggregated into multi-track historical features. This provides historical time-series information that is dynamically aligned with the current spatial location for subsequent gating calculations and hidden state updates.
[0053] Finally, there's gating computation and hidden state update. Convolution maps multi-trajectory concatenated features into gated linear components. Simultaneously, convolution is performed on the current input features to obtain gated linear components. The reset gate, update gate, and new memory gate are calculated respectively. Finally, the historical hidden state and the newly generated memory are adaptively fused through the gating mechanism to obtain the hidden state at the current time step, thus completing the update of spatiotemporal features.
[0054] , , , , , , , in These are the linear components of the reset gate, update gate, and new memory gate obtained from the multi-trajectory historical feature mapping. The linear components of the reset gate, update gate, and new memory gate obtained from the current input feature map. Indicates the first Each ST-TrajGRU block in time The reset door, Indicates the first Each ST-TrajGRU block in time The update gate, This represents the element-wise Hadamard product. Indicates the first Each ST-TrajGRU block in time New memory characteristics, Indicates the first Each ST-TrajGRU block in time The final hidden state.
[0055] S5. After obtaining the predicted features, perform a dimensional transformation on the predicted features to restore the prediction time dimension, and rearrange the predicted features according to the spatial organization of a regular grid, thereby generating a valid wave height grid sequence corresponding to multiple future time steps. The generated valid wave height grid sequence maintains the same spatial organization as the input valid wave height historical grid sequence, and each prediction time step corresponds to a valid wave height prediction result in a regular grid format.
[0056] During model training, the predicted significant wave height grid sequence and the corresponding true significant wave height grid sequence are obtained. For each prediction time step, the squared error between the predicted and true significant wave height grid sequences is calculated. The squared errors from all prediction time steps are then summed to obtain the model training loss. The formula for calculating the model training loss is: , in, To predict the time step, To predict the effective wave height grid sequence, To obtain the true effective wave height grid sequence corresponding to the predicted effective wave height grid sequence, the model training loss is obtained, and the model parameters are updated based on the model training loss so that the predicted effective wave height grid sequence gradually approaches the true effective wave height grid sequence.
[0057] Example 2 The difference between this embodiment and Embodiment 1 is that this embodiment provides a simulation experiment of a short-term wave forecasting method based on node association features and dynamic trajectory features; In this embodiment, significant wave height data from the ERA5 ocean reanalysis dataset were used to construct experimental samples, and tests were conducted on a 24-hour significant wave height prediction task. The node-adaptive trajectory-aware hybrid convolutional neural network model proposed in this application is denoted as NAT-HCNN, and ConvLSTM, ConvGRU, STGCN, and GCNGRU were selected as comparative models. Mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) were used as evaluation metrics to compare the prediction results of different models. The smaller the values of MAE, RMSE, and MAPE, the smaller the deviation between the model's prediction result and the actual significant wave height data.
[0058] The experimental results of different models in the 24-hour significant wave height prediction task are shown in the table below: Table 1 Evaluation Indicators for Each Model
[0059] For the 24-hour SWH prediction task, the NAT-HCNN model proposed in this application was compared with existing mainstream spatiotemporal prediction models such as ConvLSTM, ConvGRU, STGCN, and GCNGRU. The proposed model achieved the best results in the three core evaluation indicators: MAE, RMSE, and MAPE. The MAE was reduced to 0.1208m, which is about 13.1% lower than the second-best model; the RMSE was reduced to 0.1987m, which is 18.2% lower than the second-best model; and the MAPE was reduced to 9.80%, which is about 12.7% lower than the second-best model.
[0060] Experimental results show that the NAT-HCNN model proposed in this application significantly reduces prediction error and improves prediction accuracy and stability compared with existing technologies in the effective wave height prediction task.
[0061] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.
Claims
1. A short-term wave forecasting method based on node association features and dynamic trajectory features, characterized in that, include: Obtain the effective wave height grid sequence of the target sea area for continuous historical periods, map the effective ocean grid points to graph nodes according to the preset spatial index, and construct the node temporal feature tensor; Using the node temporal feature tensor as input, an adaptive adjacency matrix is generated through learnable node embeddings, and node features are aggregated and mapped using node graph convolution parameters to obtain node association features. Based on node association features, the node association features are reconstructed into regular grid features according to the rule spatial location corresponding to the node; Local spatial features are extracted based on regular grid features. Multiple trajectory flow fields are generated using current features and historical hidden states. The historical hidden states are dynamically sampled and gatedly updated along the trajectory flow fields to obtain dynamic trajectory features. By utilizing dynamic trajectory features, predictive channel mapping and predictive time dimension recovery are performed, outputting an effective wave height grid sequence for multiple future time steps.
2. The short-term wave forecasting method based on node association features and dynamic trajectory features according to claim 1, characterized in that, The process of generating an adaptive adjacency matrix through learnable node embeddings includes constructing a learnable node embedding matrix and initializing the elements in the learnable node embedding matrix through Kaiming uniform initialization. Construct a mask constraint matrix and calculate a node similarity matrix based on the learnable node embedding matrix. Use the mask constraint matrix to set the node association score corresponding to invalid grid points to a preset minimum value. Add self-loops to the node similarity matrix after mask constraints, perform Top-K sparsification along the node dimensions to obtain a sparse similarity matrix, and then perform row-wise normalization on the sparse similarity matrix to obtain an adaptive adjacency matrix. The expression for the row-wise normalization is as follows: , in, This represents the normalized association weight between node j and node i in the adaptive adjacency matrix. Let be the element in the i-th row and j-th column of the sparse similarity matrix. The total number of nodes in the graph. It is the sum of all elements in the i-th row of the sparse similarity matrix.
3. The short-term wave forecasting method based on node association features and dynamic trajectory features according to claim 2, characterized in that, The construction of the node graph convolution parameters includes initializing the global shared weight pool and the global shared bias pool, mapping the node embedding vector of each graph node to the global shared weight pool and the global shared bias pool respectively, generating graph convolution weights and graph convolution biases corresponding to each graph node, and using the generated graph convolution weights and graph convolution biases as the node graph convolution parameters of the corresponding graph node.
4. The short-term wave forecasting method based on node association features and dynamic trajectory features according to claim 3, characterized in that, The aggregation and mapping of node features using node graph convolution parameters includes: The node temporal feature tensor is input into a two-level concatenated ST-Conv block. In each ST-Conv block, a one-dimensional causal convolution and gating operation are performed on the node features along the time dimension through the first time-gated convolutional layer to obtain temporal gated features. The temporal gated features are aggregated based on an adaptive adjacency matrix, and the aggregated features are differentially mapped using node graph convolution parameters to obtain intermediate spatiotemporal features. Temporal modeling is performed on the intermediate spatiotemporal features through a second time-gated convolutional layer, and the output of the second-level ST-Conv block is used as the spatiotemporal fusion feature. The expression of the intermediate spatiotemporal feature is as follows: , in, For the first Intermediate spatiotemporal characteristics of the ST-Conv block output, For the hierarchical index of ST-Conv blocks, This represents the temporal gating feature output by the first-time gated convolutional layer at node j. and These are the node graph convolution weights and node graph convolution biases corresponding to node i, respectively. This refers to the feature reorganization operation performed based on the node index. It is a non-linear activation function.
5. The short-term wave forecasting method based on node association features and dynamic trajectory features according to claim 4, characterized in that, The method of aggregating and mapping node features using node graph convolution parameters also includes inputting spatiotemporal fusion features into an additional time block, performing temporal convolution on the spatiotemporal fusion features through multiple parallel one-dimensional temporal convolution branches, fusing and nonlinearly activating the outputs of the multiple temporal convolution branches to obtain temporal purified features, flattening them, and mapping the flattened features to the target prediction dimension through a fully connected layer to obtain node association features. The expression for the node association features is as follows: , in, For node association features, For additional time blocks, the input features are... , and These are three parallel one-dimensional temporal convolution operations. For gated activation functions, It is a non-linear activation function. For feature flattening operation, For fully connected linear mapping operations, To predict the time step, The number of feature channels, This represents the total number of nodes in the graph.
6. The short-term wave forecasting method based on node association features and dynamic trajectory features according to claim 1, characterized in that, The step of reconstructing node-related features into regular grid features includes establishing a spatial position correspondence between graph nodes and regular latitude and longitude grid points based on a preset spatial index, splitting the node dimensions of the node-related features using the correspondence, rearranging the features corresponding to each graph node to the height and width dimensions of the regular latitude and longitude grid points, fixing the prediction time dimension and feature channel dimension of the node-related features, and performing dimensional reconstruction on the rearranged features to obtain regular grid features.
7. The short-term wave forecasting method based on node association features and dynamic trajectory features according to claim 1, characterized in that, The process of generating multiple trajectory flow fields using current features and historical hidden states includes performing spatial convolution and nonlinear activation on regular grid features, extracting local spatial features between adjacent regular grid points, and using the local spatial features as the current features. Convolutional mapping is performed on the current feature and the historical hidden state of the previous time step respectively to obtain the current feature mapping component and the historical hidden state mapping component. The two are then fused and nonlinearly activated to obtain the flow field generation feature. Convolutional mapping is performed on the generated flow field features, and the features are segmented along the feature channel dimension to obtain flow field offset information corresponding to multiple trajectories. The generation process of the flow field offset information satisfies the following: , in, This is a collection of flow fields corresponding to multiple trajectories. This refers to the segmentation operation performed along the feature channel dimension. For convolution operations, Features for flow field generation.
8. The short-term wave forecasting method based on node association features and dynamic trajectory features according to claim 7, characterized in that, The dynamic sampling and gating update of the historical hidden state along the trajectory flow field includes determining the continuous historical sampling coordinates for each of the multiple trajectory flow fields based on the offset between the current spatial position and the trajectory flow field. Based on continuous historical sampling coordinates, the historical hidden state of the previous time step is sampled in the opposite direction of the trajectory flow field, and the historical trajectory features corresponding to each trajectory are obtained by bilinear interpolation. The historical trajectory features corresponding to multiple trajectories are concatenated along the feature channel dimension to obtain multi-trajectory concatenated features. Gated updates are then performed using these multi-trajectory concatenated features and the current features. An update gate adaptively fuses the historical hidden states and new memory features to obtain dynamic trajectory features. The historical trajectory features satisfy the following conditions: , in, For the first The historical trajectory characteristics of the trajectory at the current position p, For the first The historical sampling coordinates corresponding to each trajectory For the set of adjacent grid points participating in bilinear interpolation, For adjacent grid points, Let be the eigenvalue of the historical hidden state of the k-th ST-TrajGRU block at grid point q. Let grid point q be the coordinates of continuous historical sampling. Distance in the x-direction, Let grid point q be the coordinates of continuous historical sampling. Distance in the y-direction.
9. The short-term wave forecasting method based on node association features and dynamic trajectory features according to claim 1, characterized in that, The process of outputting an effective wave height grid sequence for multiple future time steps includes inputting dynamic trajectory features into the prediction output structure, converting the dynamic trajectory features to the target prediction channel through convolutional mapping to obtain prediction features, performing dimensional transformation on the prediction features to restore the prediction time dimension, and generating an effective wave height grid sequence for multiple future time steps according to the spatial organization of a regular grid.
10. The short-term wave forecasting method based on node association features and dynamic trajectory features according to claim 1, characterized in that, The output of the effective wave height grid sequence for multiple future time steps also includes obtaining the predicted effective wave height grid sequence and the corresponding true effective wave height grid sequence, calculating the squared error between the predicted effective wave height grid sequence and the true effective wave height grid sequence at each prediction time step, summing the squared errors at each prediction time step to obtain the model training loss, and updating the model parameters based on the model training loss.
