Effective wave height prediction method and system based on spatio-temporal evolution multi-scale feature extraction

Abstract

The invention relates to the technical field of sea wave height prediction, and discloses a significant wave height prediction method and system based on spatio-temporal evolution multi-scale feature extraction. The method comprises the following steps: performing significant wave height prediction on input significant wave height grid data of a sea area to be predicted by applying a trained hybrid heterogeneous parallel double-convolution dynamic network; the method further comprises the steps of obtaining the effective wave height grid data, inputting the effective wave height grid data into the hybrid heterogeneous parallel double-convolution dynamic network for effective wave height prediction, and then outputting an effective wave height prediction sequence of a future time step. According to the method, multiple features of SWH spatio-temporal dynamic evolution can be extracted, and hierarchical spatio-temporal features from a local scale to a global scale and local spatio-temporal dynamic crossing from a coarse scale to a fine scale can be mined at the same time; by constructing a heterogeneous parallel double-convolution framework, the internal irregularity of the SWH field and the irregular relation between SWH field data are overcome, and meanwhile the capacity of capturing the invariant relation in the SWH field is kept.

Description

Technical Field

[0001] This application belongs to the field of ocean wave height prediction technology, and specifically relates to an effective wave height prediction method and system based on spatiotemporal evolution multi-scale feature extraction. Background Technology

[0002] Significant wave height (SWH) is a very important parameter in coastal and marine engineering. However, the randomness and intermittency of ocean waves pose a great challenge to wave prediction. The formation and development of ocean waves are affected by a variety of factors, including wind speed, wind direction, and topography. The complex interaction of these factors makes ocean waves highly uncertain and variable, increasing the difficulty of wave prediction.

[0003] Traditional significant wave height prediction relies on numerical wave models, which suffers from high computational cost, sensitivity to boundary conditions, and insufficient near-shore accuracy. In recent years, deep learning-based spatiotemporal sequence models have been introduced, but existing technologies have the following main drawbacks: while simple CNNs or LSTMs can model time dependencies, their ability to extract spatial features is limited, making it difficult to simultaneously capture the spatial distribution and temporal evolution of waves; traditional models are prone to overfitting under extreme sea conditions, leading to performance degradation under normal weather conditions; significant wave height is influenced by multiple factors such as wind speed, air pressure, and topography, but traditional models often rely only on single-factor time series data, failing to fully utilize physical correlations; and with increasing prediction time, the prediction error of traditional models increases rapidly, especially under typhoon or high sea states, where the error becomes significant. Summary of the Invention

[0004] The technical problem to be solved by this application is to overcome the shortcomings of the existing technology. This application provides an effective wave height prediction method and system based on spatiotemporal evolution multi-scale feature extraction, which can extract multiple features of the spatiotemporal dynamic evolution of SWH, and at the same time mine the hierarchical spatiotemporal features from local to global scale and the local spatiotemporal dynamic leap from coarse to fine scale; at the same time, it overcomes the inherent irregularity of SWH field and the irregular relationship between SWH field data, while maintaining the ability to capture the invariant relationship within SWH field.

[0005] To achieve the above objectives, the first aspect of this application provides an effective wave height prediction method based on spatiotemporal evolution multi-scale feature extraction, including applying a trained hybrid heterogeneous parallel dual convolutional dynamic network to predict the effective wave height of the input sea area to be predicted grid data. The method includes: acquiring the effective wave height grid data, inputting the effective wave height grid data into the hybrid heterogeneous parallel dual convolutional dynamic network to predict the effective wave height, and then outputting the effective wave height prediction sequence for future time steps; The mixed heterogeneous parallel double convolution dynamic network predicts the effective wave height grid data input, including processing the effective wave height grid data through at least two channels processed in parallel, and then fusing the outputs of the two channels to output: The dynamic channel: the effective wave height grid data input is constructed into a dynamic graph sequence, and dynamic fine-scale spatiotemporal correlation feature extraction is performed on the dynamic graph sequence, wherein the dynamic fine-scale spatiotemporal correlation feature extraction is to extract local to global dynamic spatiotemporal features by using graph convolution and time gate convolution; The static channel: the effective wave height grid data input is constructed into a static grid sequence, and static coarse-scale spatiotemporal correlation feature extraction is performed on the static grid sequence, wherein the static coarse-scale spatiotemporal correlation feature extraction is to extract static stationary spatiotemporal features by using convolution LSTM; and Fusion output prediction: the dynamic spatiotemporal features and the static stationary spatiotemporal features output by the dynamic channel and the static channel are weighted and fused to generate an effective wave height prediction sequence of a future time step.

[0006] Optionally, the effective wave height grid data input is constructed into a dynamic graph sequence, including: In the mixed heterogeneous parallel double convolution dynamic network, at least one irregular graph construction submodule is constructed to construct the effective wave height grid data into a dynamic graph sequence; The dynamic graph sequence regards each grid point as a graph node, and constructs an undirected graph structure based on latitude and longitude coordinates , wherein a node set V contains N spatial positions, an edge set E generates a static adjacency matrix through a distance threshold rule, and the correlation between the center node and the adjacent node of the static adjacency matrix is 1, and the rest is 0.

[0007] Optionally, dynamic fine-scale spatiotemporal correlation feature extraction is performed on the dynamic graph sequence, and the dynamic fine-scale spatiotemporal correlation feature extraction is to extract local to global dynamic spatiotemporal features by using graph convolution and time gate convolution, including: In the mixed heterogeneous parallel double convolution dynamic network, at least two spatiotemporal convolution blocks are constructed to perform dynamic fine-scale spatiotemporal correlation feature extraction, wherein at least a spatial graph convolution layer and two time gate convolution layers are used for feature extraction in the spatiotemporal convolution block, and a residual connection and a bottleneck strategy are applied in the spatiotemporal convolution block; In any of the spatiotemporal convolution blocks, the spatial graph convolution layer captures the spatial dependence relationship with adjacent nodes, and the two time gate convolution layers at either end of the spatial graph convolution layer model the time dependence relationship of adjacent observations along the time dimension.

[0008] Optionally, the effective wave height grid data input is constructed into a static grid sequence, including: In the hybrid heterogeneous parallel dual-convolutional dynamic network, at least one regular grid construction submodule is constructed to construct the effective wave height grid data into a static grid sequence; Define reference time as The size of the prediction window is The static grid sequence is represented as , is represented as: ; in, Depend on Before Combination of regular grid feature sequences over a time step. , , They represent the first , and current time step Effective wave height grid data, This represents the dimensional structure of the entire input data, where H and W represent the grid height and width defined by the latitude and longitude coordinates of the SWH field.

[0009] Optionally, static coarse-scale spatiotemporal correlation feature extraction is performed on the static grid sequence. This extraction utilizes a convolutional LSTM to extract static stationary spatiotemporal features, including: In the hybrid heterogeneous parallel dual convolutional dynamic network, at least two SSTCFE modules are constructed to extract static coarse-scale spatiotemporal related features, and the spatiotemporal long short-term memory block is constructed to extract features by at least Conv layers and Conv-LSTM layers. The Conv layer is used to capture spatial dependencies with the neighborhood, and the Conv-LSTM layer is used to utilize temporal dependencies with nearby observations along the time dimension.

[0010] Optionally, the dynamic spatiotemporal features output from the dynamic channel and the static stationary spatiotemporal features are weighted and fused with the static stationary spatiotemporal features to generate an effective wave height prediction sequence for future time steps, including: The hybrid heterogeneous parallel dual convolutional dynamic network at least constructs an MMF module, which is configured to integrate the outputs of the dynamic channel and the static channel to provide a final prediction for the entire delivery cycle time series; the MMF module at least constructs a Gconv layer and a Conv layer to process the outputs of the dynamic channel and the static channel, respectively. First, the feature maps output by the dynamic and static channels are converted into prediction sequences through two parallel operations: the Gconv layer processes the dynamic channel output to generate a dynamic prediction sequence, and the Conv layer processes the static channel output to generate a static prediction sequence. Subsequently, by learning from historical data that the weights of these two predicted sequences are different, it can be represented as: ; in, This represents the final fused effective wave height prediction sequence. Indicates a dynamically predicted sequence. The learnable weights represent the dynamically predicted sequence. Represents a static prediction sequence. The learnable weights represent the static prediction sequence. This represents element-wise product.

[0011] Optionally, a hybrid heterogeneous parallel dual-convolutional dynamic network can be trained, with the loss function expressed as: ; in, It is real and valid wave height data. This represents the effective wave height sequence predicted by the model. The L2 norm is the square of the sum of squares of the element-wise differences, used to measure the error between the predicted and actual values. This represents the total number of time steps in the forecast. This indicates the index of the current time step.

[0012] To achieve the above objectives, a second aspect of this application provides an effective wave height prediction system based on spatiotemporal evolution multi-scale feature extraction, the effective wave height prediction system comprising: The acquisition unit is used to acquire effective wave height grid data for the sea area to be predicted. The data processing unit is equipped with a trained hybrid heterogeneous parallel dual convolutional dynamic network. The hybrid heterogeneous parallel dual convolutional dynamic network is used to receive the effective wave height grid data acquired by the acquisition unit, input the effective wave height grid data into the hybrid heterogeneous parallel dual convolutional dynamic network to predict the effective wave height, and then output the effective wave height prediction sequence for the future time step. The display unit receives the output of the data processing unit and then outputs the effective wave height prediction sequence. The hybrid heterogeneous parallel dual-convolutional dynamic network predicts the effective wave height of the input effective wave height grid data, including processing the effective wave height grid data separately through at least two channels in parallel, and then fusing the outputs of the two channels to output the final result. Dynamic channel: The input effective wave height grid data is constructed into a dynamic graph sequence, and dynamic fine-scale spatiotemporal correlation feature extraction is performed on the dynamic graph sequence. The dynamic fine-scale spatiotemporal correlation feature extraction is to extract local to global dynamic spatiotemporal features using graph convolution and time-gated convolution. Static Channel: The input effective wave height grid data is constructed into a static grid sequence. Static coarse-scale spatiotemporal correlation features are extracted from the static grid sequence. This extraction utilizes a convolutional LSTM to extract static stationary spatiotemporal features. Fusion output prediction: The dynamic spatiotemporal features output by the dynamic channel and the static channel are weighted and fused with the static stationary spatiotemporal features to generate an effective wave height prediction sequence for future time steps.

[0013] To achieve the above objectives, a third aspect of this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, is used to implement the effective wave height prediction method based on spatiotemporal evolution multi-scale feature extraction as described above.

[0014] By adopting the above technical solution, this application has the following beneficial effects compared with the prior art: In this application, a hybrid heterogeneous parallel dual-convolutional dynamic network is used to extract multiple features of the spatiotemporal dynamic evolution of the SWH field. Compared with existing technologies that mainly focus on single local scale, global scale, or static local spatial features, the proposed hybrid heterogeneous parallel dual-convolutional dynamic network can simultaneously mine hierarchical spatiotemporal features from local to global scale and local spatiotemporal dynamic transitions from coarse to fine scale. The hybrid heterogeneous parallel dual-convolutional dynamic network constructs a heterogeneous parallel dual-convolutional framework to overcome the inherent irregularity of the SWH field and the irregular relationship between SWH field data, while maintaining the ability to capture invariant relationships within the SWH field.

[0015] The specific embodiments of this application will be described in further detail below with reference to the accompanying drawings. Attached Figure Description

[0016] 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.

[0017] In the attached diagram: Figure 1 This is a flowchart illustrating the effective wave height prediction method based on spatiotemporal evolution multi-scale feature extraction in this specific embodiment. Figure 2 This is a schematic diagram showing how the grid data of SWH is mapped to the nodes of the graphic data in the effective wave height prediction method based on spatiotemporal evolution multi-scale feature extraction in this specific embodiment. Figure 3This is a schematic diagram of the network structure of the hybrid heterogeneous parallel dual convolutional dynamic network in the effective wave height prediction method based on spatiotemporal evolution multi-scale feature extraction in this specific embodiment. Figure 4 This is a schematic diagram of the effective wave height prediction system based on spatiotemporal evolution multi-scale feature extraction in this specific embodiment. Detailed Implementation

[0018] 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.

[0019] The application of deep learning-based spatiotemporal sequence models in wave forecasting has become a research hotspot. However, existing methods still have significant limitations in terms of accuracy and robustness. Specifically, traditional convolutional or recurrent network models often focus on learning time-dependent relationships, while failing to adequately characterize the complex spatial correlations of wave fields. They are prone to insufficient adaptability and a sharp drop in prediction performance when encountering extreme weather conditions. Most models only use historical sequences of significant wave heights as input, failing to organically integrate key physical driving factors such as wind speed, pressure field, and seabed topography into the prediction framework, resulting in a lack of understanding of wave generation mechanisms. As the forecast duration increases, existing methods generally face the challenge of error accumulation and amplification, especially in complex marine environments such as high sea states, where the reliability of prediction results decreases sharply with increasing forecast time.

[0020] Based on this, please see Figure 1 This application provides an effective wave height prediction method based on spatiotemporal evolution multi-scale feature extraction, which includes applying a trained hybrid heterogeneous parallel dual convolutional dynamic network to predict the effective wave height of the input sea area to be predicted grid data. The method includes: acquiring effective wave height grid data, inputting the effective wave height grid data into a hybrid heterogeneous parallel dual convolutional dynamic network for effective wave height prediction, and then outputting the effective wave height prediction sequence for future time steps; A hybrid heterogeneous parallel dual-convolutional dynamic network predicts the effective wave height of the input effective wave height grid data. This involves processing the effective wave height grid data separately through at least two channels in parallel, and then fusing the outputs of the two channels to produce the final output. Dynamic channel: The input effective wave height grid data is constructed into a dynamic graph sequence, and dynamic fine-scale spatiotemporal correlation feature extraction is performed on the dynamic graph sequence. Dynamic fine-scale spatiotemporal correlation feature extraction uses graph convolution and time-gated convolution to extract local to global dynamic spatiotemporal features. Static Channel: The input effective wave height grid data is constructed into a static grid sequence. Static coarse-scale spatiotemporal correlation features are extracted from this sequence using a convolutional LSTM to extract static stationary spatiotemporal features. Fusion output prediction: The dynamic spatiotemporal features output from the dynamic and static channels are weighted and fused with the static stationary spatiotemporal features to generate an effective wave height prediction sequence for future time steps.

[0021] It should be noted that the effective wave height prediction method based on spatiotemporal evolution multi-scale feature extraction in this embodiment is executed by an effective wave height prediction device based on spatiotemporal evolution multi-scale feature extraction. This device can be an electronic device, a component in an electronic device, an integrated circuit, or a chip. The electronic device can be a mobile electronic device or a non-mobile electronic device. For example, mobile electronic devices can be mobile phones, tablets, laptops, PDAs, in-vehicle electronic devices, wearable devices, etc., while non-mobile electronic devices can be servers and personal computers, etc., and this application does not make specific limitations. The following description uses a server as the execution subject to illustrate the effective wave height prediction method based on spatiotemporal evolution multi-scale feature extraction in this embodiment.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0023] Oceanographic variable datasets are typically organized in a regular grid format based on geographic latitude and longitude coordinates, which can be easily mapped to nodes in graphical data. Definition As an undirected graph A set of nodes; where each node Corresponding to fixed grid points determined by longitude and latitude coordinates, the base number This represents the total number of spatial grid points. Define the edges between spatial grid points. This represents a bidirectional connection between nodes. Furthermore, all connection relationships are stored in the adjacency matrix. In this matrix, the following properties are satisfied: ; in, Represents the elements in the adjacency matrix. These represent the nodes in the graph, i.e., the first node. The and the first A spatial location, Represents a node and The edge between, This represents the total number of nodes in the graph. Let each represent a set of all nodes in the graph, with each node corresponding to a spatial grid point. This represents the set of all edges in the graph, and indicates the connection relationships between nodes;

[0024] In this application's oceanographic variable SWH prediction, SWH is both the sole input and sole output variable. Let... Represents a node The SWH value at that location, then express The dynamic SWH sequence values ​​of N nodes within a time step; This represents the eigenvalue of a single SWH for N nodes at time t. Furthermore, if... and There is spatial correlation between them, we set ,otherwise ,in Defined nodes and The proximity. Then, the SWH prediction objective is to predict the following. The future sequence within the time step is represented as: ; in, Represents the entire input spatiotemporal sequence. They represent the first SWH values ​​of all spatial points at each time step Indicates the dimension of the input data. These represent the SWH values ​​of the 1st, 2nd, ..., Nth nodes at time t, respectively. This represents a vector consisting of the SWH values ​​of all nodes at a given moment. , Let represent the SWH values ​​of nodes i and j at a certain time, respectively. This represents an element in the adjacency matrix, indicating whether node i and node j are connected. This represents the spatial similarity between node i and node j. Indicates the number of time steps in the prediction. This represents a hybrid heterogeneous parallel dual-convolutional dynamic network. Represents learnable parameters, from time arrive The SWH prediction sequence is represented as , This represents the first future moment of the predicted output. Effective wave height vector, This represents the last future moment of the predicted output. The effective wave height vector, G represents the graph structure.

[0025] In one feasible implementation, the input valid wave height grid data is constructed into a dynamic graph sequence, including: In a hybrid heterogeneous parallel dual-convolutional dynamic network, at least one irregular graph construction submodule is constructed to construct a dynamic graph sequence from effective wave height grid data. The dynamic graph sequence treats each grid point as a graph node, constructing an undirected graph structure based on latitude and longitude coordinates. The node set V contains N spatial locations, and the edge set E generates a static adjacency matrix through a distance threshold rule. The central node of the static adjacency matrix has a correlation of 1 with its adjacent nodes, and the rest have a correlation of 0.

[0026] For details, please see Figure 2 and Figure 3 The overall framework of the hybrid heterogeneous parallel dual-convolutional dynamic network consists of the DFSTCFE module, the SCSTCFE module, and the MMF module, namely the Dynamic Fine-scale Spatial-Temporal Correlation Feature Extraction (DFSTCFE) module, the Static Coarse-scale Spatial-Temporal Correlation Feature Extraction (SCSTCFE) module, and the Multi-Module Fusion (MMF) module.

[0027] The DFSTCFE and SCSTCFE modules are two parallel double convolutional structures. The DFSTCFE module consists of two sub-modules: an Irregular Graph Construction (IGC) sub-module and a Dynamic Spatio-Temporal Correlation Feature Extraction (DSTCFE) sub-module. The output features generated by the IGC sub-module serve as the direct input to the DSTCFE sub-module, enabling the capture of the dynamic spatio-temporal evolution of the SWH field.

[0028] The IGC submodule first formalizes the spatiotemporal sequence of the irregular graph-based SWH field, which serves as the main input to the IGC module. Then, we detail the construction of the adjacency matrix, which, when combined with the aforementioned irregular spatiotemporal sequence, forms the input pipeline for the DSTCFE module.

[0029] Spatiotemporal Sequences of SWH Fields Based on Irregular Graphs: According to graph theory principles, a graph is typically composed of graph nodes and their corresponding adjacency matrices. Spatiotemporal sequences of a graph are stored within these nodes, while the adjacency matrices characterize the features between these nodes. Graph nodes are commonly used in traffic flow prediction studies to represent road intersections or monitoring stations, where their connection patterns simulate spatial dependencies within the traffic network. Correspondingly, in ocean variable prediction, the definition of graph nodes depends on the data source: observation-based systems utilize sensor locations as graph nodes; reanalysis systems (such as ERA5) use grid points as node representations.

[0030] This embodiment uses data from a reanalysis system, and defines the reference time as... The prediction window size is The feature sequence to be input into this module is represented as follows: The details are as follows: ; in, express Before The feature sequences of the nodes in the above figure at the time step. This represents a segment of a historical time series that is directly adjacent to the predicted time.

[0031] The construction of adjacency matrix A typically follows a geospatial attribute-based approach, which explicitly encodes spatial relationships by combining node connectivity, pairwise distance metrics, boundary geometry of influence zones, and relevant statistical parameters. For example, the spatial distance between nodes i and j... It is given by the following formula: ; in, and This indicates the relative latitude and longitude of node i. and This indicates the relative latitude and longitude of node j; Elements of adjacency matrix A The calculation is as follows: ; in, It is a threshold, located at a distance greater than Nodes are considered disconnected in the graph. Geospatial attribute-based methods rely on statically predefined node correlations, which have limited adaptability to the inherent spatiotemporal dynamics of ocean variable evolution. This embodiment, building upon geospatial attribute-based methods, designs a simpler adjacency matrix method: defining the correlation between the central node and its neighboring nodes as 1, while setting the correlation with all other nodes to 0. This constructs a simple adjacency matrix with only 0 / 1s, eliminating the need for additional adjustment of the distance decay function or Gaussian bandwidth, and ensuring that long-range noise is directly truncated. This confines node influence to a local region and generates a static adjacency matrix A. ; in, Represents the elements in the adjacency matrix. Represents a node The neighborhood of, that is, all those with Distance less than threshold The set of nodes, This represents a candidate neighbor node.

[0032] As a specific implementation method, dynamic fine-scale spatiotemporal correlation feature extraction is performed on dynamic graph sequences. This extraction utilizes graph convolution and temporally gated convolution to extract local to global dynamic spatiotemporal features, including: In a hybrid heterogeneous parallel dual-convolutional dynamic network, at least two spatiotemporal convolutional blocks are constructed to extract dynamic fine-scale spatiotemporal related features. In the spatiotemporal convolutional blocks, feature extraction is performed through at least a spatial graph convolutional layer and two temporally gated convolutional layers. Residual connections and bottleneck strategies are applied within the spatiotemporal convolutional blocks. In any spatiotemporal convolutional block, spatial dependencies with neighboring nodes are captured by a spatial graph convolutional layer. Two temporally gated convolutional layers located at either end of the spatial graph convolutional layer model the temporal dependencies of neighboring observations along the time dimension.

[0033] Specifically, in the DSTCFE submodule, during spatiotemporal sequence prediction, the DL model first extracts temporal and spatial dependencies from the input sequence along their respective dimensions, and then integrates these dependencies for predictive modeling. For temporal dependency extraction, the main DL methods involve RNN-based or CNN-based architectures, while spatial dependency extraction primarily employs CNN-based architectures or attention mechanisms. In the DSTCFE submodule, this embodiment, based on traditional geospatial methods, preliminarily designs a simpler method for constructing the adjacency matrix A, which can effectively capture the spatial relationships between nodes.

[0034] Furthermore, multiple spatiotemporal convolutional blocks in the DSTCFE submodule are designed to extract high-level features of spatiotemporally stationary and non-stationary changes. Specifically, the DSTCFE submodule consists of two spatiotemporal convolutional blocks. In each spatiotemporal convolutional block, a spatial graph convolutional (Spatio Gconv) layer captures spatial dependencies with neighboring nodes, while two temporally gated convolutional layers located at either end of the Spatio Gconv layer model the temporal dependencies of neighboring observations along the time dimension. Residual connectivity and bottleneck strategies are applied within each spatiotemporal convolutional block.

[0035] Temporally gated convolutional layers capture temporal dynamics by employing the entire convolutional structure along the time axis. A temporally gated convolutional layer contains a 1-D causal convolution (1-D convolution) with a kernel of width Kt, followed by a gated time unit GTU as the non-linearity, expressed as: ; Among them, the function Used to represent the operations of time-gated convolutional layers. This represents the input to the first gated convolution of the first spatiotemporal convolution block. This represents the convolution operation. These are the parameters of the convolution kernel. This represents the element-wise Hadamard product. It controls the current state. The sigmoid function, associated with discovering the compositional structure and dynamic variance in time series, and nonlinear gates, also facilitate the utilization of the complete input field through stacked time layers. yes , represents the input SWH graph sequence, where L is the sequence length, N is the number of nodes, and C is the feature dimension.

[0036] Spatial graph convolutional layers perform frequency domain graph convolution operations, represented as: ; in, This represents the convolution operation of the l-th spatiotemporal convolution block. It is the output of the first temporally gated convolution of the l-th spatiotemporal convolutional block; It is an identity matrix of size N×N, where D is the degree matrix and the weight pool is... It is a node-shared parameter that projects the input features onto In the delayed space of the dimension.

[0037] Therefore, the feature map output by the temporally gated convolutional layer can be obtained, as shown below: ; in, Indicates the first The spatial-temporal feature map output by each convolutional block after "spatial graph convolution + ReLU" is the representation of the current block that is finally fed into the next temporally gated convolution. Indicates the activation function; The Rectified Linear Unit (ReLU) activation function is used.

[0038] Similarity, the second temporally gated convolution of the first spatiotemporal convolution block can be calculated using the equation: ; in, This represents the output of the second temporally gated convolutional layer in the l-th spatiotemporal convolutional block.

[0039] In practical applications, the input effective wave height grid data is constructed into a static grid sequence, including: In a hybrid heterogeneous parallel dual-convolutional dynamic network, at least one regular grid construction submodule is constructed to construct the effective wave height grid data into a static grid sequence. Define reference time as The size of the prediction window is The static grid sequence is represented as , is represented as: ; in, Depend on Before Combination of regular grid feature sequences over a time step. , , They represent , and current time step Effective wave height grid data, This represents the dimensional structure of the entire input data, where H and W represent the grid height and width defined by the latitude and longitude coordinates of the SWH field.

[0040] Specifically, the SCSTCFE module consists of two dedicated sub-modules: the Regular Gridded Construction (RGC) sub-module and the Static Spatio-Temporal Correlation Feature Extraction (SSTCFE) sub-module. The output features generated by the RGC sub-module serve as the direct input to the SSTCFE sub-module, enabling the capture of the stationary spatio-temporal evolution of the SWH field.

[0041] RGC Submodule: The RGC submodule defines another spatiotemporal sequence for SWH—a regular grid spatiotemporal sequence—to serve as input to the SCSTCFE module. It also defines a reference time as... The size of the prediction window is The feature sequence to be input into the RGC submodule is represented as follows: The details are as follows: ; in, Depend on Before The combination of regular grid feature sequences over the time step, where H and W represent the grid height and width defined by the latitude and longitude coordinates of the SWH field.

[0042] As a specific implementation method, static coarse-scale spatiotemporal correlation feature extraction is performed on the static grid sequence. This extraction utilizes a convolutional LSTM to extract static stationary spatiotemporal features, including: In a hybrid heterogeneous parallel dual-convolutional dynamic network, at least two spatiotemporal long short-term memory blocks are constructed to extract static coarse-scale spatiotemporal related features. At least Conv layers and Conv-LSTM layers are constructed in the spatiotemporal long short-term memory blocks for feature extraction. Conv layers are used to capture spatial dependencies with the neighborhood, while Conv-LSTM layers are used to utilize temporal dependencies with nearby observations along the time dimension.

[0043] Specifically, the SSTCFE submodule: As shown above, since the adjacency matrix value is determined by the relationship between nodes, CNN outperforms GCN on spatial variation graph data. Therefore, this embodiment utilizes the inherent "translation invariance" of CNN to capture local static spatial correlations.

[0044] The SSTCFE submodule consists of two spatiotemporal long short-term memory blocks. Within each spatiotemporal long short-term memory block, there is a Conv layer for capturing spatial dependencies with the neighborhood, and a Conv-LSTM layer for utilizing temporal dependencies with nearby observations along the time dimension.

[0045] Specifically, the Conv layer, also known as the convolutional layer, first performs convolution operations using a 3×3 convolutional kernel, where... It's a convolution operation. We input a sequence of SWH grids to reduce the dimensionality of the data space. Then, we introduce the Rectified Linear Unit (ReLU) activation function to enhance the nonlinear representation and obtain spatial feature maps, such as: ; in, This represents the spatial feature map output by the first convolutional layer in the static channel. Indicates the activation function; Spatial feature maps are used as input to the Conv-LSTM layer, which consists of an input gate, a forget gate, an output gate, and storage units. The Conv-LSTM layer is specifically designed to capture long-term dependencies in time series data through its storage units and gating mechanism. The equations used to compute the input gate, forget gate, and output gate illustrate how the inputs of the hidden states at the current and previous time steps are fed into the gates to obtain their respective values: ; ; ; ; ; ; in, This represents calculating the feature map of the Conv-LSTM layer in the memory block when the input is at time t and the spatiotemporal length is k. It is the hidden state of the k-th ST-LSTM block at the previous time step. Specifically, the input gate, forget gate, and output gate of the k-th spatiotemporal long short-term memory block at time t, and also... These are the convolution kernel parameters. It is a bias term. These represent the hidden state, intermediate state, and storage unit in the k-th ST-LSTM block at time step t, respectively. This represents the convolution operation. This represents the element-wise Hadamard product. It is the sigmoid activation function.

[0046] In one feasible implementation, the dynamic spatiotemporal features output from the dynamic channel and the static channel are weighted and fused with the static stationary spatiotemporal features to generate an effective wave height prediction sequence for future time steps, including: At least an MMF module is constructed in the hybrid heterogeneous parallel dual convolutional dynamic network. The MMF module is configured to integrate the outputs of the dynamic channel and the static channel to provide the final prediction for the entire delivery cycle time series. The MMF module at least constructs a Gconv layer and a Conv layer to process the outputs of the dynamic channel and the static channel, respectively. First, the feature maps output by the dynamic and static channels are converted into prediction sequences through two parallel operations: the Gconv layer processes the dynamic channel output to generate a dynamic prediction sequence, and the Conv layer processes the static channel output to generate a static prediction sequence. Subsequently, by learning from historical data that the weights of these two predicted sequences are different, it can be represented as: ; in, This represents the final fused effective wave height prediction sequence. Indicates a dynamically predicted sequence. The learnable weights represent the dynamically predicted sequence. Represents a static prediction sequence. The learnable weights represent the static prediction sequence. This represents element-wise product.

[0047] Specifically, the MMF module is operated to integrate the outputs of DSTCFE and SSTCFE to provide a final forecast for the entire delivery cycle time series; First, the feature maps of the DFSTCFE and SCSTCFE modules are converted into prediction sequences through two parallel operations: the Gconv layer processes the DFSTCFE output to generate dynamic prediction sequences. The Conv layer processes the SCSTCFE output to produce a static prediction sequence. ; Subsequently, the weights of the two predicted sequences were learned from historical data.

[0048] Therefore, the final prediction result provided by the MMF module can be written as: ; in, This represents the final fused effective wave height prediction sequence. Indicates a dynamically predicted sequence. The learnable weights represent the dynamically predicted sequence. Represents a static prediction sequence. The learnable weights represent the static prediction sequence. This represents element-wise product.

[0049] In another feasible implementation, a hybrid heterogeneous parallel dual-convolutional dynamic network is trained, with the loss function expressed as: ; in, It is real and valid wave height data. This represents the effective wave height sequence predicted by the model. The L2 norm is the square of the sum of squares of the element-wise differences, used to measure the error between the predicted and actual values. This represents the total number of time steps in the forecast. This indicates the index of the current time step.

[0050] Please see Figure 4 Based on the same inventive concept, this application also provides an effective wave height prediction system based on spatiotemporal evolution multi-scale feature extraction. The effective wave height prediction system includes: The acquisition unit is used to acquire effective wave height grid data for the sea area to be predicted. The data processing unit is equipped with a trained hybrid heterogeneous parallel dual convolutional dynamic network. The hybrid heterogeneous parallel dual convolutional dynamic network is used to receive the effective wave height grid data acquired by the acquisition unit, input the effective wave height grid data into the hybrid heterogeneous parallel dual convolutional dynamic network to predict the effective wave height, and then output the effective wave height prediction sequence for the future time step. The display unit receives the output of the data processing unit and then outputs the effective wave height prediction sequence. A hybrid heterogeneous parallel dual-convolutional dynamic network predicts the effective wave height of the input effective wave height grid data. This involves processing the effective wave height grid data separately through at least two channels in parallel, and then fusing the outputs of the two channels to produce the final output. Dynamic channel: The input effective wave height grid data is constructed into a dynamic graph sequence, and dynamic fine-scale spatiotemporal correlation feature extraction is performed on the dynamic graph sequence. Dynamic fine-scale spatiotemporal correlation feature extraction uses graph convolution and time-gated convolution to extract local to global dynamic spatiotemporal features. Static Channel: The input effective wave height grid data is constructed into a static grid sequence. Static coarse-scale spatiotemporal correlation features are extracted from this sequence using a convolutional LSTM to extract static stationary spatiotemporal features. Fusion output prediction: The dynamic spatiotemporal features output from the dynamic and static channels are weighted and fused with the static stationary spatiotemporal features to generate an effective wave height prediction sequence for future time steps.

[0051] Based on the same inventive concept, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the effective wave height prediction method based on spatiotemporal evolution multi-scale feature extraction as described above.

[0052] The program product of this application for implementing the above method may employ a portable compact disk read-only memory and include program code, and may run on a terminal device, such as a personal computer. However, the program product of this application is not limited thereto. In this application, the readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, apparatus, or device.

[0053] It should be noted that a computer-readable storage medium may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0054] 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. An effective wave height prediction method based on spatiotemporal evolution multi-scale feature extraction, characterized in that, This includes using a trained hybrid heterogeneous parallel dual convolutional dynamic network to predict the effective wave height of the input sea area to be predicted from the grid data. The method includes: acquiring the effective wave height grid data, inputting the effective wave height grid data into the hybrid heterogeneous parallel dual convolutional dynamic network to predict the effective wave height, and then outputting the effective wave height prediction sequence for future time steps; The hybrid heterogeneous parallel dual-convolutional dynamic network predicts the effective wave height of the input effective wave height grid data, including processing the effective wave height grid data separately through at least two channels in parallel, and then fusing the outputs of the two channels to output the final result. Dynamic channel: The input effective wave height grid data is constructed into a dynamic graph sequence, and dynamic fine-scale spatiotemporal correlation feature extraction is performed on the dynamic graph sequence. The dynamic fine-scale spatiotemporal correlation feature extraction is to extract local to global dynamic spatiotemporal features using graph convolution and time-gated convolution. Static Channel: The input effective wave height grid data is constructed into a static grid sequence. Static coarse-scale spatiotemporal correlation features are extracted from the static grid sequence. This extraction utilizes a convolutional LSTM to extract static stationary spatiotemporal features. Fusion output prediction: The dynamic spatiotemporal features output by the dynamic channel and the static channel are weighted and fused with the static stationary spatiotemporal features to generate an effective wave height prediction sequence for future time steps.

2. The method according to claim 1, characterized in that, The input effective wave height grid data is constructed into a dynamic graph sequence, including: At least one irregular graph construction submodule is constructed in the hybrid heterogeneous parallel dual convolutional dynamic network to construct the effective wave height grid data into a dynamic graph sequence; The dynamic graph sequence treats each grid point as a graph node and constructs an undirected graph structure based on latitude and longitude coordinates. The node set V contains N spatial locations, and the edge set E generates a static adjacency matrix through a distance threshold rule. The central node of the static adjacency matrix has a correlation of 1 with its adjacent nodes, and the rest have a correlation of 0.

3. The method according to claim 1, characterized in that, Dynamic fine-scale spatiotemporal correlation feature extraction is performed on the dynamic graph sequence. This extraction utilizes graph convolution and temporally gated convolution to extract local-to-global dynamic spatiotemporal features, including: In the hybrid heterogeneous parallel dual-convolutional dynamic network, at least two spatiotemporal convolutional blocks are constructed to extract dynamic fine-scale spatiotemporal related features. The spatiotemporal convolutional blocks extract features through at least a spatial graph convolutional layer and two temporally gated convolutional layers. Residual connections and bottleneck strategies are applied in the spatiotemporal convolutional blocks. In any of the spatiotemporal convolutional blocks, the spatial dependencies with neighboring nodes are captured by the spatial graph convolutional layer, and the two temporally gated convolutional layers located at either end of the spatial graph convolutional layer model the temporal dependencies of neighboring observations along the time dimension.

4. The method according to claim 1, characterized in that, Constructing the input effective wave height grid data into a static grid sequence includes: In the hybrid heterogeneous parallel dual-convolutional dynamic network, at least one regular grid construction submodule is constructed to construct the effective wave height grid data into a static grid sequence; Define reference time as The size of the prediction window is The static grid sequence is represented as , is represented as: ； in, Depend on Before Combination of regular grid feature sequences over a time step. , , They represent the first , and current time step Effective wave height grid data, This represents the dimensional structure of the entire input data, where H and W represent the grid height and width defined by the latitude and longitude coordinates of the SWH field.

5. The method according to claim 1, characterized in that, Static coarse-scale spatiotemporal correlation feature extraction is performed on the static grid sequence. This extraction utilizes a convolutional LSTM to extract static stationary spatiotemporal features, including: In the hybrid heterogeneous parallel dual convolutional dynamic network, at least two SSTCFE modules are constructed to extract static coarse-scale spatiotemporal related features, and the spatiotemporal long short-term memory block is constructed to extract features by at least Conv layers and Conv-LSTM layers. The Conv layer is used to capture spatial dependencies with the neighborhood, and the Conv-LSTM layer is used to utilize temporal dependencies with nearby observations along the time dimension.

6. The method according to claim 1, characterized in that, The dynamic spatiotemporal features output from the dynamic channel and the static channel are weighted and fused with the static stationary spatiotemporal features to generate an effective wave height prediction sequence for future time steps, including: The hybrid heterogeneous parallel dual convolutional dynamic network at least constructs an MMF module, which is configured to integrate the outputs of the dynamic channel and the static channel to provide a final prediction for the entire delivery cycle time series; the MMF module at least constructs a Gconv layer and a Conv layer to process the outputs of the dynamic channel and the static channel respectively; First, the feature maps output by the dynamic channel and the static channel are converted into prediction sequences through two parallel operations: the Gconv layer processes the output of the dynamic channel to generate a dynamic prediction sequence, and the Conv layer processes the output of the static channel to generate a static prediction sequence; Subsequently, by learning from historical data that the weights of these two predicted sequences are different, it can be represented as: ； in, This represents the final fused effective wave height prediction sequence. Indicates a dynamically predicted sequence. The learnable weights represent the dynamically predicted sequence. Represents a static prediction sequence. The learnable weights represent the static prediction sequence. This represents element-wise product.

7. The method according to claim 1, characterized in that, The loss function for training a hybrid heterogeneous parallel dual-convolutional dynamic network is expressed as: ； in, It is real and valid wave height data. This represents the effective wave height data predicted by the model. The square of the L2 norm, which is the sum of squares of the element-wise differences, is used to measure the error between the predicted and actual values. This represents the total number of time steps in the forecast. This indicates the index of the current time step.

8. An effective wave height prediction system based on spatiotemporal evolution multi-scale feature extraction, characterized in that, The effective wave height prediction system includes: The acquisition unit is used to acquire effective wave height grid data for the sea area to be predicted. The data processing unit is equipped with a trained hybrid heterogeneous parallel dual convolutional dynamic network. The hybrid heterogeneous parallel dual convolutional dynamic network is used to receive the effective wave height grid data acquired by the acquisition unit, input the effective wave height grid data into the hybrid heterogeneous parallel dual convolutional dynamic network to predict the effective wave height, and then output the effective wave height prediction sequence for the future time step. The display unit receives the output of the data processing unit and then outputs the effective wave height prediction sequence. The hybrid heterogeneous parallel dual-convolutional dynamic network predicts the effective wave height of the input effective wave height grid data, including processing the effective wave height grid data separately through at least two channels in parallel, and then fusing the outputs of the two channels to output the final result. Dynamic channel: The input effective wave height grid data is constructed into a dynamic graph sequence, and dynamic fine-scale spatiotemporal correlation feature extraction is performed on the dynamic graph sequence. The dynamic fine-scale spatiotemporal correlation feature extraction is to extract local to global dynamic spatiotemporal features using graph convolution and time-gated convolution. Static Channel: The input effective wave height grid data is constructed into a static grid sequence. Static coarse-scale spatiotemporal correlation features are extracted from the static grid sequence. This extraction utilizes a convolutional LSTM to extract static stationary spatiotemporal features. Fusion output prediction: The dynamic spatiotemporal features output by the dynamic channel and the static channel are weighted and fused with the static stationary spatiotemporal features to generate an effective wave height prediction sequence for future time steps.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it is used to implement the effective wave height prediction method based on spatiotemporal evolution multi-scale feature extraction as described in any one of claims 1-7.

Images