Deep learning based method for constructing large ocean wave model

Abstract

The present application relates to the technical field of deep learning, and more particularly to a method for constructing a large ocean wave model based on deep learning. Multi-source ocean correlation monitoring data is obtained. Spatial interpolation mapping processing of ocean correlation monitoring features is performed on the multi-source ocean correlation monitoring data to generate ocean correlation monitoring feature spatial mapping data. Ocean wave representation feature analysis processing of ocean wave correlation heterogeneous feature channels is performed on the ocean correlation monitoring feature spatial mapping data to generate ocean wave representation feature data. A prediction mapping relationship of ocean wave representation and evolution is established based on a preset deep learning algorithm and the ocean wave representation feature data to generate an ocean wave evolution prediction model. Historical ocean wave representation evolution data is obtained. The ocean wave evolution prediction model is trained and subjected to ocean wave evolution constraint optimization processing using the historical ocean wave representation evolution data to generate an optimized ocean wave evolution prediction model. The present application realizes accurate prediction and analysis of ocean wave conditions.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a method for constructing a large ocean wave model based on deep learning. Background Technology

[0002] Ocean waves, as a core component of the marine dynamic environment, directly impact multiple important areas such as marine engineering construction, shipping safety, and marine disaster prevention and mitigation. Accurate prediction of wave generation, development, and evolution patterns provides crucial technical support for coastal infrastructure planning, offshore operation scheduling, and early warning of disasters such as typhoons and storm surges. With the intensification of global climate change and the increasing frequency of extreme marine weather events, wave evolution processes exhibit stronger nonlinearity, randomness, and spatiotemporal correlation, necessitating high-precision and high-timeliness prediction. However, existing wave prediction methods have shortcomings in processing multi-source marine monitoring data. They struggle to transform scattered and heterogeneous marine correlation monitoring data into unified feature space data that can be used for model training. In wave characterization feature analysis, they fail to systematically establish the influence relationship between multi-source heterogeneous feature channels and wave characterization, resulting in insufficient effectiveness and specificity of model input features. Furthermore, they often use a single data source for temporal or spatial feature modeling, ignoring the fact that ocean wave evolution is the result of the combined effects of multiple heterogeneous factors, making it difficult to comprehensively depict the complex correlation characteristics of wave evolution. They lack the ability to fuse spatiotemporal cross-scale features of wave characterization, failing to effectively capture wave evolution patterns at different spatial scales (such as nearshore and offshore) and different time scales. At the same time, they do not fully incorporate the physical constraints of wave evolution, causing model prediction results to easily deviate from actual physical laws, resulting in insufficient prediction stability and reliability under extreme sea conditions. Summary of the Invention

[0003] Based on this, the present invention provides a method for constructing a large ocean wave model based on deep learning to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a method for constructing a large ocean wave model based on deep learning includes the following steps:

[0005] Step S1: Acquire multi-source marine correlation monitoring data; perform spatial interpolation mapping processing on the multi-source marine correlation monitoring data to generate spatial mapping data of marine correlation monitoring features;

[0006] Step S2: Perform wave characterization feature analysis on the ocean-related monitoring feature spatial mapping data to generate wave characterization feature data;

[0007] Step S3: Based on the preset deep learning algorithm and wave characterization feature data, establish the prediction mapping relationship between wave characterization and evolution, and generate a wave evolution prediction model;

[0008] Step S4: Obtain historical wave characterization and evolution data; use the historical wave characterization and evolution data to train the wave evolution prediction model and optimize the wave evolution constraints to generate an optimized wave evolution prediction model.

[0009] The beneficial effects of this application are as follows: By acquiring multi-source marine correlation monitoring data covering multiple dimensions such as marine typhoon path intensity, sea surface wind field, sea surface air pressure, wave monitoring, marine astronomical tides, water depth, seabed topography, and nearshore line, the invention ensures comprehensive data coverage and lays the foundation for subsequent accurate modeling. Simultaneously, by eliminating temporal deviations in multi-source heterogeneous data through time-series synchronization benchmark preprocessing, and combining spatial interpolation mapping with a marine topography 3D spatial model constructed and optimized based on marine geographic information data, the invention not only solves the problem of integrating scattered heterogeneous data but also achieves accurate matching between monitoring features and real marine geographic space. The generated marine correlation monitoring feature spatial mapping data possesses both temporal consistency and spatial accuracy, providing high-quality and standardized data support for subsequent feature analysis. Furthermore, the adaptive mapping grid design and optimization process of the marine topography 3D spatial model further enhances the model's adaptability to marine geographic boundary features, ensuring the accuracy of spatial mapping of monitoring features in different sea areas (nearshore, offshore, etc.). This study analyzes the influence relationship between heterogeneous feature channels and wave characterization in marine correlation monitoring feature spatial mapping data, addressing the limitation of not being able to systematically establish the correlation between the two. By extracting data from different heterogeneous feature channels hierarchically, targeted analyses are conducted on key wave characterizations such as storm surges, tidal levels, and ocean-going tidal flats. These analyses are then integrated to generate influence relationship data and refine wave characterization features. This approach not only accurately captures the intrinsic correlation between different heterogeneous features and various wave characterizations but also achieves a comprehensive and detailed characterization of the core features of wave evolution. This hierarchical analysis and integrated refinement method ensures that the generated wave characterization feature data accurately reflects the key laws of wave evolution, effectively improving the effectiveness and relevance of subsequent model input features and providing core feature support for constructing a high-precision wave evolution prediction model. By constructing a deep learning large-scale model architecture and embedding a specially designed spatiotemporal cross-scale fusion layer for wave representation, the shortcomings of existing technologies in lacking spatiotemporal cross-scale feature fusion capabilities are effectively addressed. Specifically, through a fusion strategy combining spatial multi-scale analysis, temporal feature mining, and clustering attribute adaptation, the precise extraction of spatiotemporal cross-scale fusion features of wave representation is achieved, comprehensively capturing the wave evolution patterns at different spatial scales (nearshore and offshore) and different temporal scales (short-term swells and long-term evolution). Simultaneously, a multimodal adaptive attention strategy designed in conjunction with the influence analysis of single-state and superimposed-state evolution of wave representation allows the model to accurately focus on key features under different evolutionary scenarios, improving the model's adaptability to complex wave evolution scenarios. The constructed wave evolution prediction model possesses stronger feature representation and evolution mapping capabilities, laying an excellent model architecture foundation for high-precision prediction.By incorporating historical wave evolution data for model training and integrating analysis of the physical constraints of wave evolution, the shortcomings of inadequate integration of physical laws leading to predictions deviating from reality are effectively compensated for. By designing an auxiliary loss function based on physical constraints and optimizing model parameters, the model training process consistently follows the objective physical laws of wave evolution, significantly improving the rationality and reliability of the prediction results. Furthermore, this data-driven and physically constrained optimization approach not only improves the model's prediction accuracy under normal sea conditions but also enhances its prediction stability under extreme sea conditions, solving the problem of poor adaptability to extreme scenarios. The resulting optimized wave evolution prediction model possesses higher practical value and greater potential for wider application.

[0010] Therefore, the deep learning-based ocean wave prediction method of this invention effectively addresses the shortcoming of converting scattered heterogeneous ocean correlation monitoring data into unified feature space data by performing temporal synchronous preprocessing and spatial interpolation mapping on multi-source heterogeneous ocean correlation monitoring data such as typhoon path intensity, sea surface wind field, sea surface air pressure, wave monitoring, ocean astronomical tides, water depth, seabed topography, and nearshore lines. This provides a standardized and comprehensive data foundation for model training. By analyzing the influence relationship between multi-source heterogeneous feature channels and wave representations such as storm surge, tidal level, and floodplain, the method accurately establishes the correlation mapping between features and representations, significantly improving the effectiveness and relevance of the model input features and compensating for the shortcomings of traditional methods. This approach addresses the shortcomings of existing technologies in feature correlation analysis. By relying on multi-source fusion data for modeling, it fully considers the combined effects of multiple heterogeneous factors in ocean wave evolution, comprehensively depicting the complex correlation characteristics of wave evolution and overcoming the limitations of modeling from a single data source. Through the construction of a deep learning architecture that integrates a spatiotemporal cross-scale fusion layer and a multimodal adaptive attention mechanism, it can accurately capture the wave evolution patterns at different spatial scales, such as nearshore and offshore, and at different time scales, such as short-term swells and long-term evolution. At the same time, it optimizes model parameters by combining the physical constraints of wave evolution, ensuring that the prediction results strictly conform to actual physical laws, and significantly improving the stability and reliability of predictions under extreme sea conditions. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the steps of a deep learning-based method for constructing a large ocean wave model according to the present invention.

[0012] Figure 2 for Figure 1 A detailed flowchart illustrating the implementation steps of step S3.

[0013] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0014] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0015] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. Functional entities may be implemented in software, or in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods. The term "and / or" as used herein includes any and all combinations of one or more of the associated items listed.

[0016] To achieve the above objectives, please refer to Figures 1 to 2 This invention provides a method for constructing a large ocean wave model based on deep learning. In the embodiments of this invention, please refer to... Figure 1 The diagram shown illustrates the steps of a deep learning-based method for constructing a large ocean wave model according to the present invention. The deep learning-based method for constructing a large ocean wave model includes the following steps:

[0017] Step S1: Acquire multi-source marine correlation monitoring data; perform spatial interpolation mapping processing on the multi-source marine correlation monitoring data to generate spatial mapping data of marine correlation monitoring features;

[0018] In this embodiment of the invention, when acquiring multi-source marine correlation monitoring data, the typhoon path and intensity data are obtained through collaborative observation by polar-orbiting meteorological satellites and geostationary meteorological satellites. The data includes parameters such as the typhoon center latitude and longitude, maximum near-center wind speed, and minimum air pressure, with a sampling interval of 1 hour. Sea surface wind field data are acquired through a buoy array deployed in the target sea area and a shore-based Doppler radar. The buoy array is deployed at a density of one buoy per 100 square kilometers, and the radar observation range covers the entire target sea area. The data includes wind speed and direction parameters, with a sampling interval of 15 minutes. Sea surface air pressure data are acquired through a marine research vessel and coastal air pressure monitoring stations. The research vessel cruises once a month, and the monitoring station data sampling interval is set to 30 minutes. Wave monitoring data... Data was collected via wave buoys and seabed pressure sensors. The wave buoys recorded effective wave height, wave period, and wave direction parameters, while the seabed pressure sensors recorded wave energy variations. The sampling interval for both was set to 10 minutes. For marine geographic information data, astronomical tide data was obtained from coastal tide gauge stations, including tide level and time parameters, with a sampling interval of 1 hour. Ocean depth data was obtained through multibeam echo sounders, with a measurement accuracy controlled within ±0.5 meters, covering at least one measurement point per square kilometer of the target sea area. Ocean seabed topography data was acquired using side-scan sonar and 3D seismic exploration techniques, with a topography resolution controlled within 10 meters × 10 meters. Nearshore data was obtained through UAV aerial photography and satellite remote sensing image interpretation, updated quarterly. All of the above data was pre-stored in the marine monitoring system, and the data was retrieved to obtain multi-source marine correlation monitoring data. When establishing a three-dimensional spatial model of marine topography based on marine geographic information data, an irregular triangular mesh algorithm is first used to perform preliminary three-dimensional spatial modeling of marine topography based on water depth data and seabed topography data, generating a preliminary model. Then, nearshore coordinates and coordinates of water depth abrupt change areas are extracted from the marine geographic information data as marine geographic boundary feature data to determine the boundary range of the model and topographic abrupt change nodes. Based on the topographic undulation degree and boundary feature data of the preliminary model, an adaptive mapping mesh is designed. In areas with large topographic undulations (such as submarine mountains and trenches), the mesh resolution is set to 5 meters × 5 meters, and in areas with gentle topography (such as deep-sea plains), the mesh resolution is set to 20 meters × 20 meters, obtaining marine geographic spatial mapping mesh data. This mesh data is embedded into the preliminary model, and the topographic nodes in the model are re-matched and their coordinates are calibrated to complete the establishment and optimization of the mapping mesh of the three-dimensional spatial model, resulting in the three-dimensional spatial model of marine topography. When performing multi-source data time-series synchronization benchmark preprocessing on multi-source marine correlation monitoring data, a unified time benchmark is established, the original timestamps of each data source are extracted, the time deviation value is calculated by comparing with the atomic clock time, the deviation of the timestamp of each data sample is corrected, and the timestamp error of all data samples is ensured to be less than 1 second, thereby generating synchronized multi-source marine correlation monitoring data.When transmitting synchronous multi-source marine correlation monitoring data to a three-dimensional marine topographic model for spatial interpolation and mapping, the adaptive mapping grid nodes of the model are used as interpolation target points. The Kriging interpolation method is adopted to calculate the monitoring parameter values ​​corresponding to each grid node based on the parameter values ​​and spatial distance of the five nearest data samples around each grid node, so that each grid node corresponds to a complete set of multi-source monitoring data, generating marine correlation monitoring feature spatial mapping data.

[0019] Step S2: Perform wave characterization feature analysis on the ocean-related monitoring feature spatial mapping data to generate wave characterization feature data;

[0020] In this embodiment of the invention, when analyzing the influence relationship between heterogeneous wave-related feature channels and wave characterization based on marine correlation monitoring feature spatial mapping data, the heterogeneous wave-related feature channel data is first extracted from the marine correlation monitoring feature spatial mapping data. Specifically, the marine typhoon path intensity feature channel extracts the straight-line distance, maximum near-center wind speed, and minimum air pressure parameters corresponding to the typhoon center at each grid node; the sea surface wind field feature channel extracts the wind speed and direction parameters at each grid node; the sea surface air pressure feature channel extracts the air pressure value at each grid node; the marine water depth feature channel extracts the water depth value corresponding to each grid node; the marine seabed topography feature channel extracts the topographic slope and undulation parameters within a 100-meter radius around each grid node; the marine astronomical tide feature channel extracts the tide height and tidal range parameters corresponding to each grid node; the wave monitoring feature channel extracts the significant wave height and wave period parameters corresponding to each grid node; and the marine nearshore line feature channel extracts the straight-line distance parameter from each grid node to the nearshore line. After extracting feature spatial mapping data from marine typhoon track intensity data, sea surface wind field data, sea surface pressure data, ocean depth data, and seabed topography data, feature correlation analysis is performed on these data. The negative correlation coefficients between typhoon center distance and wind speed, wind speed and air pressure, and the obstruction coefficients of water depth and topographic slope on wave energy transmission are calculated. Based on these coefficients, the influence weights of each feature on ocean waves and storm surges are determined. Through weighted calculation, the storm surge height and influence range parameters corresponding to each grid node are obtained, generating ocean wave and storm surge characterization data. After extracting feature spatial mapping data from marine astronomical tide data, the storm surge height from the ocean wave and storm surge characterization data is superimposed with the tide height from the astronomical tide data. Simultaneously, combined with the tidal time parameters of the astronomical tide, the actual water level height of each grid node at different times is determined, generating ocean wave and tidal water level characterization data. After extracting the feature space mapping data from wave monitoring data and nearshore data, the storm surge impact range in the wave storm surge characterization data, the actual water level height in the wave tide level characterization data, and the significant wave height and wave period in the wave monitoring data are correlated to calculate the wave run-up value. Then, combined with the straight-line distance from each grid node to the nearshore line, it is determined whether the wave corresponding to each grid node can reach the nearshore area and floodplain range, generating floodplain characterization data. Based on the wave storm surge characterization data, wave tide level characterization data, and floodplain characterization data, the Pearson correlation coefficients between the heterogeneous feature channel parameters of each wave association and the three characterization data are calculated. Feature parameters with an absolute correlation coefficient greater than 0.6 are selected as key influencing features, and a correspondence table between key influencing features and each wave characterization is established, generating wave association heterogeneous feature-characterization influence relationship data.Based on the influence relationship data, intersection and union analyses were performed on the key influencing features of the three wave characterizations to identify common features that simultaneously affect two or more characterizations and specific features that only affect a single characterization. The weight of the common features was set to 1.5 times that of the specific features. Through weight allocation, the influence relationships of wave characterizations were integrated, generating integrated wave characterization influence relationship data. Linear and nonlinear fitting analyses were performed on the feature parameters and characterization data in this integrated data. The feature combination with the highest fitting degree was extracted as the core feature of the wave characterization. The core feature parameters were standardized (within the range of 0-1) to generate wave characterization feature data.

[0021] Step S3: Based on the preset deep learning algorithm and wave characterization feature data, establish the prediction mapping relationship between wave characterization and evolution, and generate a wave evolution prediction model;

[0022] In this embodiment of the invention, when establishing a large deep learning model architecture for wave representation and evolution based on a preset deep learning algorithm, a hybrid architecture combining a Transformer encoder and a Convolutional Neural Network (CNN) is adopted. The Transformer encoder has 8 layers, each with 12 multi-head attention heads, and the Feed-Forward network has a hidden layer dimension of 2048. The CNN part uses 3 convolutional layers with kernel sizes of 3×3, 5×5, and 3×3, respectively, a stride of 1, and SamePadding to ensure that the input and output feature map sizes are consistent, thus generating a wave evolution prediction model architecture. When performing spatiotemporal cross-scale fusion feature analysis of wave characterization based on wave characterization feature data, the wave characterization feature data is first divided into spatial multi-scale segments. The target sea area is divided into three scale levels: 10km×10km (macro scale), 1km×1km (meso scale), and 100m×100m (micro scale). Wave characterization feature parameters of grid nodes within each scale level are extracted, and the mean, variance, and extreme values ​​of feature parameters within each scale level are calculated to generate spatial multi-scale feature data of wave characterization. This spatial multi-scale feature data is then divided into time series segments with a time series window of 24 hours and a sliding step of 1 hour. The changing trends of feature parameters at each scale within each time window (such as rise rate, fall rate, and stationary duration) are extracted to generate spatiotemporal multi-scale feature data of wave characterization. Based on the wave characterization data, the K-means clustering algorithm was used to perform cluster analysis on the wave characterization. Five clusters were set, corresponding to calm sea states, slightly undulating sea states, moderately undulating sea states, heavily undulating sea states, and extreme sea states. The central feature parameters of each cluster were calculated to generate wave characterization cluster attribute data. Using this cluster attribute data, differentiated multi-scale fusion strategies for wave characterization space were designed for different cluster attributes: calm sea states and slightly undulating sea states were fused primarily using macro-scale features (60% weight); moderately undulating sea states were fused primarily using meso-scale features (50% weight); and heavily undulating sea states and extreme sea states were fused primarily using micro-scale features (60% weight). Based on this fusion strategy, the spatiotemporal multi-scale feature data of wave characterization were dynamically fused over time. Within each time window, the corresponding fusion strategy was called according to the current sea state cluster attribute, and the weighted sum of features at each scale was calculated to generate spatiotemporal cross-scale fused feature data of wave characterization.When embedding the spatiotemporal cross-scale fusion layer of wave representation into the wave evolution prediction model architecture using this spatiotemporal cross-scale fusion feature data, the fusion feature data is used as the input to the newly added fusion layer. The fusion layer is set between the Transformer encoder and the CNN, and the output dimension of the fusion layer is set to 1024, consistent with the output dimension of the Transformer encoder. A fully connected layer is used to concatenate the fusion features with the output features of the Transformer encoder, generating the embedded wave evolution prediction model architecture. When designing a multimodal adaptive attention strategy for wave evolution based on the wave representation feature data, the feature parameters corresponding to the single wave representation (storm surge, tidal level, floodplain) in the wave representation feature data are first extracted. The contribution value of each single representation feature parameter to the wave evolution in different time windows is calculated (the contribution value is calculated by the wave evolution amplitude caused by the change of the feature parameter), generating the single-state evolution influence data of wave representation. Spatial specificity analysis of wave representation superposition states (two or more single representations coexisting) was performed on spatiotemporal cross-scale fusion feature data. The difference between the feature parameters of each grid node in the superposition state and the corresponding parameters in the single state was calculated. Grid nodes with an absolute difference greater than 0.2 were selected as spatially specific regions. The variation law of superposition state feature parameters in these regions was extracted, and the contribution value of the superposition state to wave evolution was calculated to generate wave representation superposition state evolution influence data. Based on the wave representation single state evolution influence data and wave representation superposition state evolution influence data, a multimodal adaptive attention strategy was designed: in the single state scenario, the attention weight is allocated proportionally according to the contribution value of the single representation; in the superposition state scenario, the attention weight of the spatially specific regions is increased by 20%, and the weight ratio of each single representation is adjusted according to the contribution value of the superposition state, thus generating a multimodal adaptive attention strategy for wave evolution. When setting the attention mechanism for the superimposed state of the wave representation spatial scene in the embedded wave evolution prediction model architecture using this attention strategy, the attention strategy is transformed into a weight allocation rule that the model can recognize. This is then embedded into the multi-head attention layer of the Transformer encoder, enabling the model to automatically call the corresponding attention weight allocation rule when dealing with different sea state scenarios (single state, superimposed state) to obtain the wave evolution prediction model.

[0023] Step S4: Obtain historical wave characterization and evolution data; use the historical wave characterization and evolution data to train the wave evolution prediction model and optimize the wave evolution constraints to generate an optimized wave evolution prediction model.

[0024] In this embodiment of the invention, when acquiring historical wave characterization and evolution data, relevant data on wave evolution over the past 10 years are collected for the target sea area. Data sources include historical databases of marine observation stations, historical cruise records of marine research vessels, and interpretation results of historical satellite remote sensing images. The data content covers hourly wave and storm surge heights, wave and tide levels, and the extent of floodplains, as well as corresponding data such as typhoon path intensity, sea surface wind field, sea surface pressure, and marine geographic information. When analyzing the physical constraints of wave evolution in the historical wave characterization and evolution data, based on the principles of fluid mechanics and wave dynamics, the core physical constraints in the wave evolution process are extracted: firstly, the constraint of wave energy conservation, meaning that the energy loss rate of the wave during propagation must not exceed 30% of the initial energy (energy loss is calculated through wave height attenuation); secondly, the constraint of wave speed, meaning that the wave speed must be proportional to the square root of the water depth (the proportionality coefficient is determined based on seawater density); and thirdly, the constraint of floodplain extent, meaning that the floodplain extent must not exceed 5 kilometers from the nearshore line towards the land (adjusted according to the actual shoreline slope in special terrain areas). These physical constraints are transformed into quantitative indicators, such as energy loss rate thresholds, wave speed to water depth ratio ranges, and floodplain range thresholds, generating physical constraint characteristic data for wave evolution. Historical wave characterization evolution data are divided into training and validation sets in a 7:3 ratio, with the training set used for model parameter updates and the validation set used for model performance evaluation. The training set data is input into the wave evolution prediction model, with 100 iterations and an initial learning rate of 0.001. A cosine annealing learning rate scheduling strategy is used (the learning rate is halved every 20 iterations), and the mean squared error loss function is used. The model's weight parameters are updated using the backpropagation algorithm. After each training iteration, the model's prediction error (the mean absolute error between the predicted and actual values) is calculated using the validation set data. Training stops when the prediction error on the validation set decreases by less than 0.01 for five consecutive iterations, generating a trained wave evolution prediction model. When conducting training objective decision analysis for wave evolution constraints based on physical constraint characteristic data, the quantitative indicators of physical constraints are transformed into constraint objectives for model training: the predicted energy loss rate must be within ±5% of the true value, the predicted wave speed must be within ±3% of the true value, and the predicted floodplain range must be within ±10% of the true value, thus generating wave evolution constraint training objective decisions. Based on these training objective decisions, an auxiliary loss function for wave evolution constraints is designed. The auxiliary loss function adopts the hinge loss function. When the model's predicted value exceeds the constraint range, the loss value is calculated based on the magnitude of the exceedance (the larger the exceedance, the larger the loss value). The auxiliary loss function is then fused with the original mean squared error loss function at a weight ratio of 1:2 to generate the auxiliary loss function for wave evolution constraints.When optimizing the training parameters of the wave evolution prediction model by using wave evolution constraints to determine the training objectives and the wave evolution constraint auxiliary loss function, the model training is restarted with 50 iterations and an initial learning rate of 0.0005. The cosine annealing learning rate scheduling strategy is still used. During each training round, the fusion loss value is calculated simultaneously. Key parameters such as the attention weights, convolution kernel parameters, and Transformer encoder weights are adjusted through the backpropagation algorithm. After each training round, the model prediction values ​​are verified using validation set data to check whether they meet the constraint objectives. When the satisfaction rate of all constraint objectives reaches more than 95% and the prediction error of the validation set is stable, the optimization is stopped, and the optimized wave evolution prediction model is generated.

[0025] Furthermore, step S1 includes the following steps:

[0026] Step S11: Obtain multi-source marine correlation monitoring data, wherein the multi-source marine correlation monitoring data includes marine typhoon path intensity data, sea surface wind field data, sea surface air pressure data, ocean wave monitoring data, and marine geographic information data, wherein the marine geographic information data includes marine astronomical tide data, ocean water depth data, ocean seabed topography data, and ocean nearshore line data;

[0027] In this embodiment of the invention, pre-stored multi-source marine correlation monitoring data is acquired through a marine monitoring system. For typhoon path and intensity data, a collaborative observation mode using polar-orbiting meteorological satellites and geostationary meteorological satellites is employed to extract core parameters such as the typhoon's center latitude and longitude, maximum near-center wind speed, and minimum air pressure, ensuring the data fully reflects the typhoon's movement path and intensity changes. Sea surface wind field data is jointly acquired through an array of anchored buoys deployed in the target sea area and a shore-based high-frequency ground wave radar. The buoy array is distributed in a uniform grid to cover open sea areas, while the radar focuses on near-shore areas. The data from both are complementary to achieve comprehensive wind field data for the target sea area. The system covers a wide area, collecting parameters including wind speed and direction, with data accuracy ensured through hardware calibration mechanisms for wind speed and direction sensors. Sea surface pressure data relies on a three-dimensional monitoring network formed by fixed coastal pressure monitoring stations and mobile research vessels at sea. Monitoring stations are deployed at a density of one station every 50 kilometers to obtain continuous time-series data, while the research vessel cruises along a pre-set route to supplement data from the open ocean. The collected parameter is sea surface pressure, and the temperature compensation function of the barometer eliminates the interference of ambient temperature on the data. Wave monitoring data is collected using a combination of wave buoys and seabed pressure sensors. The wave buoys float on the sea surface and directly record... Effective wave height, wave period, and wave direction are measured. Seabed pressure sensors, deployed at different depths, invert wave energy transmission patterns through pressure changes. Combining this data enables comprehensive monitoring of ocean waves from the surface to the seabed. In marine geographic information data, astronomical tide data is acquired from long-operational tide gauge stations along the coast, including tidal height and time, and standardized by comparison with global sea level benchmarks. Ocean depth data is measured comprehensively across the target sea area using a multibeam echo sounder. Real-time dynamic positioning technology ensures the accuracy of measurement points, and the acquired parameter is the water depth value at the measurement point, further analyzed through sound velocity profiles. Surface correction eliminates the influence of changes in seawater sound velocity on the measurement results; ocean seabed topography data is acquired jointly through side-scan sonar and 3D seismic exploration technology. Side-scan sonar acquires seabed surface morphology data, while 3D seismic exploration acquires shallow seabed geological structure data. The two are fused to generate 3D topography data; ocean nearshore data is collected regularly through UAV aerial photography and high-resolution satellite remote sensing imagery. UAVs focus on nearshore detail areas and take images once a quarter, while satellite remote sensing covers a large nearshore area and acquires images once a month. Nearshore coordinates and morphological features are extracted through image interpretation to ensure that the data can reflect the dynamic changes of the nearshore line.

[0028] Step S12: Establish a three-dimensional spatial model of marine topography based on marine geographic information data;

[0029] In this embodiment of the invention, when establishing a three-dimensional spatial model of marine topography based on marine geographic information data, preliminary three-dimensional spatial modeling is carried out. Based on marine water depth data and seabed topography data, an irregular triangular mesh algorithm is used to construct a topographic surface model. During the modeling process, water depth measurement points are used as basic nodes, and triangular faces are formed by connecting them according to the rule that the spatial distance between nodes is less than a preset threshold. Each vertex of the triangular face corresponds to specific water depth and topographic parameters, thereby generating a preliminary three-dimensional spatial model of marine topography that can initially reflect the undulations of the seabed. Subsequently, marine geographic boundary feature analysis is performed. The coordinate sequences of nearshore lines, the coordinate range of water depth abrupt change areas (such as continental shelf edges and trench edges), and island outline coordinates are extracted from the marine geographic information data. This information is then processed to unify the coordinate systems of boundary data from different sources, converting them to a geodetic coordinate system. A boundary feature extraction algorithm is then used to identify key feature points such as inflection points and turning points of the boundaries, generating marine geographic boundary feature data containing the coordinates of key feature points and boundary types (such as shoreline boundaries and topographic abrupt change boundaries). Finally, adaptive... The mapping mesh design should combine the topographic relief and marine geographic boundary feature data of the preliminary marine topographic 3D spatial model. A mesh densification algorithm should be used to design the mapping mesh. In areas with large topographic relief (such as submarine mountains and trenches), the mesh node density should be increased to improve the model's ability to express topographic details. In areas with gentle topography (such as deep-sea plains), the mesh node density should be appropriately reduced to control the model complexity. Marine geospatial mapping mesh data containing mesh node coordinates, mesh size, and the corresponding topographic type of the mesh should be generated. The 3D spatial model should be optimized by embedding the marine geospatial mapping mesh data into the preliminary marine topographic 3D spatial model. The triangular faces in the model should be re-divided so that each mesh node corresponds to a topographic sampling point in the model. Then, the correspondence between mesh nodes and topographic sampling points should be adjusted by a coordinate calibration algorithm to ensure that the deviation between the topographic parameters of the mesh nodes and the actual measured data is less than a preset range. At the same time, topographic gaps and overlapping areas in the model should be repaired and merged. Finally, a marine topographic 3D spatial model that can accurately reflect the topographic features of the target sea area and adapt to the subsequent data mapping requirements should be generated.

[0030] Step S13: Perform multi-source data time-series synchronization benchmark preprocessing on the multi-source marine correlation monitoring data to generate synchronized multi-source marine correlation monitoring data;

[0031] In this embodiment of the invention, when performing multi-source data time-series synchronization benchmark preprocessing on multi-source marine correlation monitoring data, a unified time benchmark is first determined, and the original timestamps of each data source are extracted. From the original records of marine typhoon path intensity data, sea surface wind field data, sea surface air pressure data, wave monitoring data, and marine geographic information data, the collection timestamp corresponding to each data sample is extracted, and information such as the format and precision of the timestamps is recorded. Timestamps with inconsistent formats are converted. Next, the time deviation value is calculated. The original timestamps of each data sample are compared with the unified time benchmark, and a time difference calculation algorithm is used to calculate the deviation value between each timestamp and the benchmark time. Simultaneously, the deviation distribution of timestamps from each data source is statistically analyzed, and abnormal timestamps with large deviations are identified. Then, timestamp correction is performed. Based on the calculated time deviation value, the timestamp of each data sample is corrected. A linear correction algorithm is used in the correction process, adding the corresponding deviation value to the original timestamp to obtain the corrected timestamp. For abnormal timestamps with large deviations, the timetamp change pattern of adjacent data samples before and after the data sample is queried, and an interpolation algorithm is used to correct the abnormal timestamps to ensure that the corrected timestamps conform to the timing logic of data acquisition. All corrected data samples are sorted in timetamp order, and the time interval of adjacent data samples is checked to see if it conforms to the preset sampling interval of each data source. For cases where the time interval exceeds the preset range, the cause is analyzed and the timestamp is corrected again to obtain synchronous multi-source marine correlation monitoring data.

[0032] Step S14: Transmit the synchronous multi-source marine correlation monitoring data to the three-dimensional spatial model of the marine topography for spatial interpolation mapping of marine correlation monitoring features to generate spatial mapping data of marine correlation monitoring features.

[0033] In this embodiment of the invention, when transmitting synchronous multi-source marine correlation monitoring data to a three-dimensional marine topography model for spatial interpolation mapping of marine correlation monitoring features, the interpolation target and range are determined. The adaptive mapping grid nodes in the three-dimensional marine topography model are used as interpolation target points, and the monitoring feature parameters that need to be mapped for each target point are clearly defined. Simultaneously, the spatial range for interpolation processing is defined as the grid coverage area of ​​the entire target sea area. Then, an interpolation algorithm is selected and the interpolation rules are determined. Kriging interpolation is used for spatial interpolation. This algorithm establishes an interpolation model by analyzing the spatial correlation between sample points. During the interpolation process, a spatial search radius is defined with each grid node as the center. Within the search radius, a preset number of synchronous multi-source marine correlation monitoring data sample points are selected to ensure that the sample points can fully reflect the distribution pattern of monitoring features around the grid node. Next, interpolation calculations are performed. For each type of monitoring feature parameter of each grid node, the parameter values ​​of the searched sample points and the spatial distance from the sample point to the grid node are input... The Kriging interpolation model is used to calculate the predicted values ​​of monitoring feature parameters corresponding to grid nodes. The calculation process considers the spatial variability function between sample points to ensure the predicted values ​​reflect the spatial distribution trend of the monitoring features. Then, data mapping and integration are performed, associating the predicted values ​​of various monitoring feature parameters for each grid node with its coordinates and topographic parameters in the three-dimensional marine topography model. This establishes a correspondence between grid nodes and multi-source monitoring feature parameters, ensuring that each grid node in the model contains complete multi-source monitoring feature information. Finally, the mapping results are verified and corrected. A subset of grid nodes is randomly selected, and their interpolated monitoring feature parameter values ​​are compared with the parameter values ​​of sample points from synchronously collected multi-source marine correlation monitoring data around the node. The deviation is calculated; if the deviation exceeds a preset range, the search radius, number of sample points, and other parameters of the interpolation algorithm are adjusted, and the interpolation is recalculated until the accuracy of the mapping results meets the requirements, generating marine correlation monitoring feature spatial mapping data.

[0034] Furthermore, step S12 includes the following steps:

[0035] Step S121: Based on marine geographic information data, perform preliminary three-dimensional spatial modeling of marine topography to obtain a preliminary three-dimensional spatial model of marine topography;

[0036] In this embodiment of the invention, when performing preliminary three-dimensional spatial modeling of marine topography based on marine geographic information data, the core data source is first preprocessed. Ocean depth data and seabed topography data are selected from the marine geographic information data. Coordinate system unification processing is performed on both types of data, converting the coordinates of all data points to the geodetic coordinate system. Simultaneously, outliers are removed. Through statistical analysis of the data distribution range, data points exceeding three times the standard deviation of the mean are identified as outliers and removed, ensuring the reliability of the basic data. Subsequently, an irregular triangular mesh algorithm is used for modeling. The preprocessed depth data points and seabed topography data points are used as basic nodes. Neighborhood search and connection are performed according to the rule that the spatial distance between nodes is less than a preset threshold. Each node forms only a few triangular faces with its nearest neighbors, and adjacent triangular faces do not overlap, ensuring that the triangular mesh can completely cover the distribution range of all data points in the target sea area. During the modeling process, to ensure the realism of the terrain undulations, data points in areas with dramatic terrain undulations, such as submarine mountains and trenches, were encrypted. The refinement of triangular faces was increased by increasing node density, allowing for accurate depiction of the terrain features in these areas. For gently sloping areas such as deep-sea plains, the node connection density was appropriately reduced to control model complexity while maintaining the basic terrain trend. Ultimately, through this series of operations, a preliminary three-dimensional spatial model of the marine terrain was generated.

[0037] Step S122: Perform marine geographic boundary feature analysis based on marine geographic information data to obtain marine geographic boundary feature data;

[0038] In this embodiment of the invention, when performing marine geographic boundary feature analysis based on marine geographic information data, nearshore data, ocean depth data, seabed topography data, and island distribution data are extracted from the marine geographic information data as the basis for analysis. The extracted nearshore data undergoes contour refinement processing, and an edge detection algorithm is used to identify the continuous contour of the nearshore line, eliminating discrete breakpoints caused by image noise and generating a continuous nearshore coordinate sequence. Gradient analysis is performed on the ocean depth data to calculate the depth difference between adjacent data points. Areas where the depth difference exceeds a preset threshold are identified as abrupt depth changes, and the boundary coordinate range of these areas is extracted to clarify the spatial location of abrupt topographic boundaries such as continental shelf edges and trench edges. Overlay analysis is performed on the seabed topography data and island distribution data to extract the contour coordinates of the islands and determine the topographic boundary range around the islands. A feature point extraction algorithm is used to identify key feature points such as inflection points, turning points, and extreme points in various boundaries. The coordinates of each feature point, its boundary type (nearshore boundary, abrupt topographic boundary, island boundary), and corresponding topographic parameters are recorded to generate marine geographic boundary feature data.

[0039] Step S123: Based on the preliminary three-dimensional spatial model of marine topography and marine geographic boundary feature data, design an adaptive mapping grid for marine geospatial features to obtain marine geospatial mapping grid data;

[0040] In this embodiment of the invention, a quantitative evaluation standard for the degree of terrain undulation is established. The slope and curvature of each triangular facet in the preliminary three-dimensional marine topography model are calculated, and the product of slope and curvature is used as a quantitative indicator of terrain undulation. A larger value indicates more severe terrain undulation. Based on marine geographic boundary feature data, the core areas for grid design are delineated: the areas surrounding near-shore boundaries, areas with abrupt changes in water depth, areas surrounding islands, and areas where the quantitative indicator value of terrain undulation exceeds a preset threshold are designated as key refinement areas; the remaining areas are designated as regular areas. Differentiated grid density design strategies are adopted for different areas. High-density grids are used in key refinement areas to improve adaptability to boundary features and terrain details by reducing grid size; low-density grids are used in regular areas to reduce the computational load of the model while ensuring the representation of terrain trends. During the grid generation process, a structured grid generation algorithm is adopted to ensure that the grid nodes are arranged in a regular manner. At the same time, the grid boundary is precisely aligned with the key feature points in the marine geographic boundary feature data. The grid lines extend naturally along the boundary contour to avoid grid intersections or deviations from the boundary. This generates marine geospatial mapping grid data, which includes the coordinates of all grid nodes, the grid size of each region, the terrain type and boundary attributes corresponding to the grid, and the grid can completely cover the target sea area, achieving precise refinement of key areas and efficient representation of regular areas.

[0041] Step S124: Optimize the preliminary three-dimensional spatial model of marine topography by establishing a mapping grid for the three-dimensional spatial model using marine geospatial mapping grid data, and obtain the three-dimensional spatial model of marine topography.

[0042] In this embodiment of the invention, marine geospatial mapping grid data is embedded into a preliminary three-dimensional marine topographic spatial model. The triangular mesh structure of the preliminary model is re-divided based on the grid nodes, ensuring that each grid node corresponds to a topographic sampling point in the preliminary model, thus guaranteeing the correspondence between grid nodes and topographic sampling points. Subsequently, coordinate calibration is performed. A spatial registration algorithm is used to calculate the deviation between the coordinates of the grid nodes and the corresponding topographic sampling points. Grid nodes with deviations exceeding a preset range are adjusted to ensure that the adjusted grid nodes accurately match the actual spatial location of the topographic sampling points, guaranteeing spatial consistency between the grid and the topography. To address potential topographic gaps and overlapping triangular faces in the preliminary model, repairs are performed using the grid data: for areas with topographic gaps, missing topographic data is supplemented by interpolating the topographic parameters of the grid nodes, generating new triangular faces to fill the gaps; for areas with overlapping triangular faces, triangular faces are trimmed according to the grid boundaries, the overlapping parts are deleted, and adjacent nodes are reconnected, ensuring the integrity and rationality of the model structure and generating a three-dimensional marine topographic spatial model.

[0043] Furthermore, step S2 includes the following steps:

[0044] Step S21: Analyze the influence relationship between the heterogeneous feature channels of ocean wave association and the wave characterization based on the ocean-related monitoring feature spatial mapping data, and generate ocean wave association heterogeneous feature-characterization influence relationship data;

[0045] In this embodiment of the invention, when analyzing the influence relationship between heterogeneous wave-related feature channels and wave characterization based on marine correlation monitoring feature spatial mapping data, feature extraction of the heterogeneous wave-related feature channels is completed. From the marine correlation monitoring feature spatial mapping data, according to the preset feature channel division rules, eight types of heterogeneous feature channel data are separated, namely, marine typhoon path intensity feature channel, sea surface wind field feature channel, sea surface air pressure feature channel, ocean water depth feature channel, ocean seabed topography feature channel, ocean astronomical tide feature channel, wave monitoring feature channel, and ocean nearshore line feature channel. Each type of channel data contains continuous temporal feature parameters and spatial coordinate information of the corresponding grid node. Subsequently, a layered wave characterization analysis was conducted. First, spatial mapping data of five characteristic channels—typhoon path intensity, sea surface wind field, sea surface air pressure, ocean depth, and seabed topography—were extracted. Through feature correlation analysis, the intrinsic relationships between parameters of each channel were explored, and the influence of different parameters on wave energy transfer and water accumulation was calculated. Based on this, the contribution weight of each parameter to the wave and storm surge characterization was determined. Combined with the weights, the core storm surge parameters corresponding to each grid node were calculated to generate wave and storm surge characterization data. Next, spatial mapping data of the marine astronomical tide characteristic channel was extracted and superimposed with the storm surge height parameter in the wave and storm surge characterization data. Combined with the periodic variation of astronomical tides, the actual water level change trend at different times was derived to generate wave and tide water level characterization data. Then, spatial mapping data of two characteristic channels—wave monitoring and nearshore line—were extracted. The influence range parameter of the wave and storm surge characterization data and the water level height parameter of the wave and tide water level characterization data were integrated to calculate the correlation between wave rise and nearshore flooding, determine the flooding probability and range corresponding to each grid node, and generate wave and flooding characterization data. Finally, an influence relationship mapping is established. Based on three types of wave characterization data and eight types of heterogeneous feature channel data, the correlation strength between each feature channel parameter and each type of wave characterization is quantified through correlation analysis. Feature parameters with correlation strength reaching the preset standard are selected as key influence features. A correspondence matrix between key influence features and each wave characterization is established. The matrix contains information such as feature parameter name, correlation strength quantification value, and influence time series range. Finally, wave correlation heterogeneous feature-characterization influence relationship data is generated.

[0046] Step S22: Perform integrated analysis of wave characterization influence relationship based on the wave association heterogeneous characteristics-characterization influence relationship data to generate integrated wave characterization influence relationship data;

[0047] In this embodiment of the invention, when performing integrated analysis of the influence relationship of wave characterization based on the data of the influence relationship of heterogeneous features and characterizations of waves, feature association intersection and union analysis are carried out to extract three sets of key influence features corresponding to wave storm surge characterization, wave tide level characterization, and wave beach characterization in the data of the influence relationship of heterogeneous features and characterizations of waves. Through set operations, the intersection features (i.e., common features that simultaneously affect two or more wave characterizations) and union features (i.e., all features that affect any one wave characterization) of the three sets are obtained, clarifying the cross-characterization influence range of common features and the specific influence objects of specific features (features that only affect a single characterization). Subsequently, the influence weights were assigned. Based on the physical mechanism of wave evolution, the weight allocation rules for common and specific features were determined through statistical analysis. Common features, which play a dominant role in the overall process of wave evolution, were assigned higher weights than specific features. At the same time, differentiated sub-weights were assigned according to the differences in the influence intensity of common features on different representations. Specific features were assigned basic weights according to the influence intensity of their corresponding single representation. The weight allocation results were then integrated with the influence relationship data of the three types of wave representations to generate integrated data on the influence relationship of wave representations. This data can fully reflect the comprehensive influence effect of various heterogeneous features on different wave representations.

[0048] Step S23: Perform wave characterization feature analysis on the integrated data of wave characterization influence relationship to generate wave characterization feature data.

[0049] In this embodiment of the invention, based on the weight coefficients and temporal patterns of the integrated data on the influence relationships of wave representations, the key influence features of the integration are effectively screened. Weak influence features with weight coefficients below a preset threshold are eliminated, while core influence features with weight coefficients meeting the threshold and whose temporal patterns cover key stages of wave evolution are retained, ensuring that the screened features have significant explanatory power for wave representation. Subsequently, feature fitting and core combination extraction are performed. The screened core influence features are subjected to linear and nonlinear fitting analysis with the corresponding wave representation data. The fit degree is used to quantify the suitability between features and representations, and the feature combination with the highest fit degree is selected as the core feature set for wave representation. The core feature set contains feature parameters representing key stages of wave evolution, and there is no multicollinearity among the feature parameters. Through the above series of operations, wave representation feature data that can accurately depict the core patterns of wave representation and is suitable for the input requirements of deep learning models is generated.

[0050] Furthermore, step S21 includes the following steps:

[0051] Step S211: Extract feature spatial mapping data of marine typhoon track intensity data, sea surface wind field data, sea surface pressure data, ocean depth data and ocean seabed topography data from marine associated monitoring feature spatial mapping data, and perform wave and storm surge characterization analysis through the feature spatial mapping data of marine typhoon track intensity data, sea surface wind field data, sea surface pressure data, ocean depth data and ocean seabed topography data to generate wave and storm surge characterization data;

[0052] In this embodiment of the invention, when extracting feature spatial mapping data for five specified data categories based on marine correlation monitoring feature spatial mapping data, the process is implemented according to a preset feature channel division standard. The standard specifies a feature correlation threshold of 0.8, a spatial matching accuracy of 100-meter grid, and a temporal alignment accuracy of 1 hour. Through precise matching of feature attributes and channel labels, corresponding feature data is separated from the grid node information of the marine correlation monitoring feature spatial mapping data. Each data category contains continuous temporal parameters and spatial coordinate association information for each grid node. When extracting marine typhoon path intensity feature data, the relative distance between the grid node and the typhoon center is determined by spatial coordinate comparison in conjunction with typhoon movement trajectory data, and the associated typhoon intensity level data is converted to obtain near-term data. The impact intensity corresponding to the maximum wind speed at the center (e.g., an impact intensity of 0.4 for wind speeds of 10-17.1 m / s) is calculated based on the minimum pressure gradient using air pressure distribution data. Core parameters are directly extracted from the wind field time series corresponding to the grid nodes from the sea surface wind field characteristic data, and the wind speed time series change rate is obtained through time series curve fitting (calculation interval 30 minutes). Basic parameters and variation characteristics of the air pressure time series of the grid nodes are extracted from the sea surface air pressure characteristic data (variation amplitude calculation interval 1 hour). Ocean water depth characteristic data is directly correlated with the basic topographic elevation data corresponding to the grid nodes. Topographic parameters within a 500-meter radius of the surrounding ocean floor topographic characteristic data are calculated through interpolation of neighboring grid node data, and the topographic type is labeled using topographic classification standards. When conducting wave and storm surge characterization analysis using five types of data, we first carried out multi-feature correlation mining and used Pearson correlation analysis technology to analyze the coupling relationship between typhoon path intensity and sea surface wind field and sea surface air pressure. For example, when the typhoon moves towards the target sea area, we track the changing trend of wind speed at grid nodes as the distance from the typhoon center decreases, and simultaneously correlate the law of air pressure decrease to clarify the driving law that every 5 m / s increase in wind speed corresponds to a 15% increase in wave energy. At the same time, we correlate ocean depth and seabed topography data and use a topography influence quantification model (topography resistance coefficient ranges from 0 to 1, with 0.8 for trench areas and 0.2 for plain areas) to analyze the impact of different topography on water flow and determine the degree of influence of topographic factors on wave energy transfer and deposition. A feature-weighted fusion method was employed to quantify the contribution weights of various feature parameters to storm surge formation. The weights were assigned based on wave dynamics principles and statistical patterns from historical storm surge samples over the past 10 years. Specifically, the weights were: 0.35 for the maximum wind speed near the typhoon center, 0.25 for sea surface wind speed, 0.2 for seabed slope, 0.15 for sea surface pressure variation, and 0.05 for water depth. Through weighted calculations, the core characterization parameters of each grid node at different time series were obtained, ultimately generating comprehensive storm surge characterization data covering the target sea area and incorporating both temporal and spatial information.

[0053] Step S212: Extract feature spatial mapping data of marine astronomical tide data from marine correlation monitoring feature spatial mapping data, and perform ocean wave and storm surge characterization analysis by combining ocean wave and storm surge characterization data with feature spatial mapping data of marine astronomical tide data to generate ocean wave and storm surge characterization data.

[0054] In this embodiment of the invention, when extracting feature spatial mapping data of marine astronomical tide data based on marine correlation monitoring feature spatial mapping data, feature information of corresponding marine astronomical tide channels is selected from the marine correlation monitoring feature spatial mapping data. The selection criteria are the correlation attributes between feature parameters and astronomical tide evolution, and the channel selection threshold is that the contribution ratio of astronomical tide features is ≥0.7. The extracted content includes the core tide level parameters and periodic variation parameters of astronomical tides corresponding to each grid node. The periodic variation parameters are obtained through time series data regularity mining. For example, by analyzing the fluctuation period of long-term time series tide level data, the periodic duration of semi-diurnal tides (12.42 hours) and diurnal tides (24.84 hours) is identified. Phase features within a range of ±5° are obtained through phase fitting. All extracted data retain complete time series records (sampling interval of 1 hour) to ensure that the continuous evolution process of astronomical tides can be reflected. When performing ocean wave and storm surge characterization analysis by combining storm surge and astronomical tide data, a superposition model of storm surge and astronomical tide is first established. This model is based on the principle of water statics. The model configuration parameters clearly define the input features as the time series data of storm surge height and astronomical tide height, and the output as the superimposed tidal level. The storm surge height parameter in the storm surge characterization data and the tide height parameter in the astronomical tide data are aligned using the same time series reference. At each time series node, the values ​​of both are superimposed. Simultaneously, the phase difference in the superposition process of storm surge and astronomical tide is corrected. For example, when there is a discrepancy between the astronomical high tide time series and the storm surge peak time series, the phase correction submodule adjusts the temporal correspondence of the numerical superposition to avoid superposition errors caused by time series misalignment. Then, combining the periodic variation law of astronomical tides, the temporal evolution trend of the superimposed tidal level is derived, clarifying the occurrence time of the highest and lowest water levels at different times (accuracy 5 minutes). At the same time, combining the nearshore topographic features in the marine geographic information data, the tidal level calculation results for areas with a nearshore slope > 5° are corrected by ±0.2m to eliminate the interference of topographic undulations on water level observation. This yields ocean wave and tidal level characterization data, which accurately reflects the sea level variation law under the combined influence of storm surge and astronomical tide.

[0055] Step S213: Extract the feature spatial mapping data of wave monitoring data and nearshore data based on the marine correlation monitoring feature spatial mapping data, and perform wave-shoal characterization analysis by combining wave monitoring data and nearshore data with wave storm surge characterization data and wave tide level characterization data to generate wave-shoal characterization data.

[0056] In this embodiment of the invention, when extracting feature spatial mapping data of two specified types of data based on marine correlation monitoring feature spatial mapping data, the process is implemented according to the feature channel division standard. By matching feature dimensions and channel attributes, feature channel data corresponding to wave monitoring and the nearshore line are separated from the marine correlation monitoring feature spatial mapping data: core wave parameters of each grid node are extracted from the wave monitoring feature data, and core parameters such as effective wave height, wave period, wave direction, and wave energy density are obtained through feature extraction of wave time series data, with a time series sampling interval of 10 minutes; the nearshore line feature data obtains the straight-line distance from each grid node to the nearshore line through spatial geometric calculation, and spatial feature parameters such as nearshore slope (quantized from 0-90°) and nearshore area terrain type are extracted through nearshore topographic data. Both types of data are strictly correlated with grid node coordinates and time series information. When performing wave and beach characterization analysis using three types of data, first integrate the storm surge impact range of wave and storm surge characterization data and the tidal water level height of wave and tidal water level characterization data. For example, superimpose the spatial boundary of the storm surge impact range with the threshold range of tidal water level height ≥1.5m to determine the basic water level and coverage area that waves may reach nearshore areas (within 3km of the nearshore line being the key area). The integrated water level and range parameters are correlated with the significant wave height and wave period parameters from the wave monitoring data. A wave run-up calculation model (Goda formula) is used to calculate the wave run-up value. The model configuration parameters explicitly define the input features as significant wave height, wave period, wave direction, and the angle between the wave and the nearshore line, and the output is the wave run-up value. It includes a built-in wave energy loss calculation submodule (energy loss coefficient 0.85) and an angle correction submodule, with an iterative convergence accuracy of 0.01m. The calculation of the wave run-up value takes into account the energy loss during wave energy transmission and corrects the calculation results based on the relationship between the wave direction and the nearshore line. For example, the correction coefficient is 1.2 when the wave direction is perpendicular to the nearshore line and 0.8 when the angle is 0°. At the same time, combined with the distance from the grid nodes to the nearshore line in the nearshore line data (within 1km is the nearshore core area) and nearshore slope parameters (slope >15° is steep slope, <5° is gentle slope), it is determined whether the wave run-up can exceed the nearshore line elevation (elevation threshold is set at the local coastal datum +0.5m), thereby determining the probability of flooding. For areas identified as having experienced flooding, the flooding extent is further calculated by combining the slope and elevation distribution of the nearshore topography. The flooding depth is calculated based on the superposition of tidal level and wave rise (with the starting standard being 0.1m above the nearshore elevation). The duration of flooding is determined by combining time series data, clarifying the evolution of flooding under different time series. For example, the flooding extent expands and the depth increases during the rising tidal phase, while it gradually shrinks during the receding tidal phase. This generates flooding characterization data that fully reflects the entire process characteristics of nearshore flooding under the action of ocean waves.

[0057] Step S214: Based on the wave storm surge characterization data, wave tide level characterization data, and wave beach characterization data, analyze the influence relationship between wave-related heterogeneous feature channels and wave characterization, and generate wave-related heterogeneous feature-characterization influence relationship data.

[0058] In this embodiment of the invention, when analyzing the influence relationship between heterogeneous feature channels and wave characterization based on three types of wave characterization data, the analysis object and scope are clearly defined. The heterogeneous feature channels involved in the analysis are identified as eight types, including marine typhoon path intensity and sea surface wind field. Wave characterization is categorized into three types. The analysis scope covers all 100-meter precision grid nodes and a complete one-year time series (365 days, 1-hour sampling interval) of the target sea area. Subsequently, a quantitative correlation analysis between features and characterization is conducted using Pearson correlation analysis (95% confidence level, significance test P-value ≤ 0.05, sample size ≥ 1200 groups). The original data of the eight heterogeneous feature channels and the three types of wave characterization data are mapped according to grid nodes and time series. The correlation strength between each parameter in each feature channel and each wave characterization parameter is quantified, with the correlation strength represented by a 0-1 quantification value. Based on the quantification results, a correlation strength threshold of 0.6 was set. Feature parameters with correlation strength exceeding the threshold were selected as key influencing features. The wave characterization type corresponding to each key influencing feature was clarified. For example, the correlation strength of the sea surface wind speed parameter was 0.75, corresponding to two types of characterization: storm surge and seabed lava flow. Thus, it was identified as a key influencing feature and its corresponding characterization type was labeled. Next, an influence relationship mapping system was established, integrating information such as key influencing features, quantified correlation strength values, corresponding characterization types, and influence time series ranges (e.g., the influence time series of typhoon path intensity features is from 24 hours before the typhoon passes to 12 hours after the typhoon passes). This constructed a multi-dimensional influence relationship matrix and generated wave correlation heterogeneous feature-characterization influence relationship data.

[0059] Furthermore, as an embodiment of the present invention, reference is made to... Figure 2 As shown, Figure 1 A detailed flowchart illustrating the implementation steps of step S3 is provided in this embodiment. Step S3 includes:

[0060] Step S31: Establish a deep learning large model architecture for wave representation and evolution based on a preset deep learning algorithm, and generate a wave evolution prediction model architecture.

[0061] In this embodiment of the invention, when establishing a large-scale deep learning model architecture for wave representation and evolution based on a preset deep learning algorithm, the preset algorithm adopts a hybrid architecture combining a Transformer encoder and a Convolutional Neural Network (CNN). This architecture can take into account both the temporal correlation of wave evolution and the extraction requirements of spatial distribution features. The Transformer encoder has 8 layers, each containing a multi-head attention mechanism and a feedforward network. The number of multi-head attention heads is set to 12 to ensure deep mining capabilities of temporal features. The feedforward network's hidden layer dimension is set to 2048 to ensure sufficient information carrying capacity during feature transformation. The CNN part uses 3 convolutional layers with kernel sizes of 3×3, 5×5, and 3×3, configured in descending order to adapt to spatial feature extraction at different scales. The stride of each convolution is set to 1, and SamePadding is used to ensure that the feature map size remains consistent with the input after each convolutional operation, avoiding loss of spatial information. A Transformer encoder and a CNN are connected via a feature concatenation layer, with the output of the Transformer encoder serving as the input to the CNN, forming a progressive architecture logic of "temporal feature extraction - spatial feature enhancement". Simultaneously, an output layer is set at the end of the architecture, with a dimension of 6 (corresponding to 6 core wave evolution prediction parameters such as storm surge height, tidal level, and floodplain extent), used to output wave evolution prediction results. Through the parameter settings and connection logic design of each module in the above architecture, a wave evolution prediction model architecture capable of adapting to the needs of wave representation and evolution prediction is generated. This architecture has the ability to simultaneously process temporal and spatial features, laying the foundation for subsequent fusion of cross-scale features and attention mechanisms.

[0062] Step S32: Perform spatiotemporal cross-scale fusion feature analysis of wave characterization based on wave characterization feature data to generate spatiotemporal cross-scale fusion feature data of wave characterization;

[0063] In this embodiment of the invention, when performing spatiotemporal cross-scale fusion feature analysis on wave characterization data, spatial multi-scale feature extraction is first carried out. According to a preset spatial scale division standard, the target sea area is divided into three levels: 10km×10km (macroscale, covering a large open sea area), 1km×1km (mesoscale, focusing on the transition area between nearshore and offshore), and 100m×100m (microscale, targeting nearshore shoals and areas with complex topography). The wave characterization data within each scale level is then regionalized, and statistical features such as the mean, variance, and extreme values ​​of feature parameters within each region, as well as distribution features such as spatial distribution density and gradient changes, are extracted to form spatial multi-scale feature data corresponding to each scale. Subsequently, temporal feature analysis is performed. A 24-hour time series window is set, and a 1-hour sliding window is used to temporally segment the spatial multi-scale feature data. The rising rate, falling rate, and duration of stable operation of feature parameters at each scale within each window are extracted to generate spatiotemporal multi-scale feature data of wave characterization containing spatiotemporal dimension information. Next, we conducted a clustering attribute analysis of wave characteristics. The K-means clustering algorithm was used to classify the wave characteristic data, with five clusters corresponding to calm sea states, slightly undulating sea states, moderately undulating sea states, heavily undulating sea states, and extreme sea states. The central characteristic parameters (mean and median of each characteristic parameter) of each sea state were calculated to generate wave characteristic clustering attribute data. Based on the clustering attribute data, a spatial multi-scale fusion strategy was designed, setting differentiated scale weight allocation rules for different sea state types: calm sea states and slightly undulating sea states were fused primarily using macro-scale features (macro-scale weight 60%, meso-scale weight 30%, micro-scale weight 10%); moderately undulating sea states were fused primarily using meso-scale features (meso-scale weight 50%, macro-scale weight 30%, micro-scale weight 20%); and heavily undulating sea states and extreme sea states were fused primarily using micro-scale features (micro-scale weight 60%, meso-scale weight 30%, macro-scale weight 10%). Finally, based on this fusion strategy, the spatiotemporal multi-scale feature data of wave representation are dynamically fused in time series. Within each time window, the corresponding weight rules are called according to the current sea state clustering attributes to perform weighted summation calculation on the features of each scale, generating spatiotemporal cross-scale fused feature data of wave representation that can simultaneously reflect the evolution law of different spatial scales and time dimensions.

[0064] Step S33: The wave evolution prediction model architecture is embedded by performing a wave characterization spatiotemporal cross-scale fusion layer embedding process on the wave characterization spatiotemporal cross-scale fusion feature data to generate an embedded wave evolution prediction model architecture.

[0065] In this embodiment of the invention, when embedding the wave evolution prediction model architecture with the spatiotemporal cross-scale fusion feature data of wave representation, the embedding position of the fusion layer is determined. Combined with the feature transfer logic of the model architecture, the fusion layer is placed between the Transformer encoder and the CNN. This position allows the cross-scale spatiotemporal features to be fused after temporal extraction and before spatial enhancement, improving the coherence of feature transfer. Subsequently, the core parameters of the fusion layer are designed. The fusion layer adopts a fully connected layer structure, with an input dimension of 512 (matching the dimension of the spatiotemporal cross-scale fusion feature data of wave representation) and an output dimension of 1024 (consistent with the output dimension of the Transformer encoder), ensuring that the output features of the fusion layer can adapt to the input feature dimension of the subsequent CNN. Using the spatiotemporal cross-scale fusion feature data of wave representation as the input of the fusion layer, the fully connected layer performs linear transformation and information integration on the cross-scale spatiotemporal features, so that the fused features simultaneously contain spatial information of different scales and multi-temporal evolution information. Next, the connection between the fusion layer and the original modules of the model architecture is established. The input of the fusion layer directly receives the output features of the Transformer encoder, and the output is connected to the input layer of the CNN through feature concatenation (the concatenated feature dimension is 2048), realizing the deep fusion of cross-scale spatiotemporal features with the features of the original architecture. Through the above-mentioned setting of the fusion layer's position, parameter design, and connection logic construction, the embedding process of the wave evolution prediction model architecture is completed, generating an embedded wave evolution prediction model architecture. This architecture has the ability to process spatiotemporal cross-scale fusion features and can more comprehensively capture the complex laws of wave evolution.

[0066] Step S34: Design a multimodal adaptive attention strategy for wave evolution based on wave characterization feature data;

[0067] In this embodiment of the invention, when designing a multimodal adaptive attention strategy for wave evolution based on wave characterization feature data, the following steps are first taken: Analysis of the impact of single-state wave characterization evolution is conducted. Characteristic parameters corresponding to three types of single wave characterizations—storm surge, tidal level, and floodplain—are separated from the wave characterization feature data. For each single characterization, the influence of its characteristic parameters on the overall wave evolution at different time stages is analyzed. The evolutionary contribution value of each characteristic parameter is obtained by calculating the correlation coefficient between the change in characteristic parameters and the wave evolution amplitude, generating single-state wave characterization evolution impact data. Subsequently, superposition analysis of wave characterization is performed. Spatial specificity detection is conducted on the spatiotemporal cross-scale fusion feature data of wave characterization to identify superposition regions where two or more single characterizations coexist. The difference between the characteristic parameters in this region and the corresponding parameters in the single-state region is calculated. Regions with an absolute difference greater than 0.2 are selected as spatially specific regions. The evolution law of the superposition characteristic parameters in this region is extracted, and the contribution value of the superposition state to wave evolution is quantified, generating superposition wave characterization evolution impact data. A multimodal adaptive attention strategy was designed based on the evolutionary impact data of both single-state and superposition-state scenarios. Attention weight allocation rules were defined: In the single-state scenario, attention weights were allocated proportionally based on the evolutionary contribution value of each feature parameter, with higher contribution values ​​resulting in larger weight proportions; in the superposition-state scenario, the attention weight for feature parameters in spatially specific regions was increased by 20%, while the weight proportions of each individual representation feature were adjusted according to the superposition-state evolutionary contribution value to ensure focused attention on key impact features under superposition-state conditions. Through these rule designs, a multimodal adaptive attention strategy for wave evolution that is adaptable to both single-state and superposition-state scenarios was generated.

[0068] Step S35: The wave evolution prediction model is obtained by setting the superposition state attention mechanism of the wave representation space scene in the embedded wave evolution prediction model architecture through the multimodal adaptive attention strategy of wave evolution.

[0069] In this embodiment of the invention, when setting the attention mechanism for the superimposed state of the spatial scene of the wave representation using a multimodal adaptive attention strategy for wave evolution embedded in the wave evolution prediction model architecture, the multimodal adaptive attention strategy is first transformed into a weight allocation logic that the model can recognize. The weight allocation rules in the strategy are broken down into specific calculation logic, clarifying the calculation method of weights being calculated according to the contribution value ratio in the single state and weights being allocated according to the contribution value of the superimposed state after first increasing the proportion of the spatially specific region. Subsequently, the embedding module of the attention mechanism is determined. Combined with the structure of the embedded model architecture, the attention mechanism is embedded in the multi-head attention layer of the Transformer encoder (embedded in each of the 8-layer encoder multi-head attention layers). This module can directly act on the temporal feature extraction process, improving the ability to focus on key spatiotemporal features. The transformed weight allocation logic is integrated into the calculation process of the multi-head attention layer, enabling the attention layer to automatically determine whether the current sea state scene belongs to a single state or a superposition state when processing feature data: if only one type of single representation parameter is detected to be non-zero, it is determined to be a single state, and attention is allocated according to the single state weight rule; if two or more types of single representation parameters are detected to be non-zero, it is determined to be a superposition state, and the spatial specific region weight enhancement logic is automatically activated to adjust the attention allocation according to the superposition state weight rule, thereby completing the automatic identification and switching of the scene and obtaining a wave evolution prediction model with spatiotemporal cross-scale feature processing capabilities and a multimodal adaptive attention mechanism.

[0070] Furthermore, step S32 includes the following steps:

[0071] Step S321: Perform spatial multi-scale feature analysis of wave representation based on wave representation feature data to generate spatial multi-scale feature data of wave representation;

[0072] In this embodiment of the invention, when performing spatial multi-scale feature analysis of wave characterization based on wave characterization feature data, a three-level spatial scale quantification standard is first established: macro-scale, meso-scale boundary, and micro-scale. According to this standard, a grid partitioning method is used to divide the wave characterization feature data into spatial regions with a partitioning accuracy of 100m × 100m, ensuring that the boundaries of each scale region are accurately defined by grid coordinates (e.g., the macro-scale boundary is grid coordinates X1-Y1 to X2-Y2), and completely covering the target sea area without overlap or omission. For the wave characterization feature data within each scale region, statistical and distributional features of the feature parameters are extracted: statistical features include mean, extreme values, and variation range; distributional features include the spatial distribution density of the feature parameters (the number of grids meeting the parameter standard per unit area) and gradient variation trend (the mean of the parameter difference between adjacent grids). For macro-scale data, the focus is on extracting the large-scale wave propagation direction and the mean and extreme values ​​of the overall wave height distribution. For meso-scale data, the focus is on extracting the wave height attenuation rate and waveform distortion coefficient (wave crest width to wavelength ratio) from offshore to nearshore. For micro-scale data, the focus is on extracting fine features such as wave refraction angle and diffraction coefficient caused by local topography. The feature data extracted at each scale are correlated with the grid coordinates of the corresponding spatial region to generate multi-scale feature data representing the space of ocean waves.

[0073] Step S322: Perform temporal feature analysis on the spatial multi-scale feature data of ocean waves to generate spatiotemporal multi-scale feature data of ocean waves;

[0074] In this embodiment of the invention, when performing temporal feature analysis on the spatial multi-scale feature data representing ocean waves, the quantification parameters of the time series window and sliding step size are first set: the window duration is set to 24 hours (covering the basic cycle of semi-diurnal and diurnal tides in ocean wave evolution, ensuring the capture of the complete short-term evolution process); the window sliding step size is set to 1 hour (less than the window duration, ensuring the continuity and density of temporal features). The spatial multi-scale feature data representing ocean waves is temporally segmented using a 1-hour sliding step size to obtain continuous time window datasets. Each window dataset contains complete feature parameters for three spatial scales and corresponding spatial grid coordinate information. For the spatial multi-scale feature data within each time window, the rate of change, duration of stationary stability, time of extreme value occurrence, and magnitude of extreme value change are extracted. The extraction of the duration of stationary stability is achieved by traversing the parameter changes of each hourly time series node within the window, determining time series segments that continuously satisfy "change magnitude < 0.05", and calculating their duration. The extracted temporal evolution features are associated with the spatial multi-scale feature data of the corresponding window through the window number and spatial grid coordinates, so as to clarify the change pattern of each spatial scale feature in different temporal stages and generate spatiotemporal multi-scale feature data of ocean waves.

[0075] Step S323: Perform wave characterization clustering attribute analysis based on wave characterization feature data to generate wave characterization clustering attribute data;

[0076] In this embodiment of the invention, the K-means clustering algorithm is used to classify the wave characterization feature data. The clustering configuration parameters are: 5 clusters, ≥100 iterations, convergence accuracy 1e-3, and Euclidean distance is used for similarity determination (the preset distance threshold is 0.2, and samples with a distance <0.2 are grouped into one cluster). During the clustering process, the combination difference of the core parameters of the wave characterization is used as the classification criterion. By calculating the Euclidean distance between different data samples, samples with a distance below the threshold are grouped into one cluster. After clustering, each cluster corresponds to a specific sea state type, specifically divided as follows: calm sea state (significant wave height <0.5, wave period <3), slightly undulating sea state (0.5≤significant wave height <1.5, 3≤wave period <5), moderately undulating sea state (1.5≤significant wave height <3, 5≤wave period <8), heavily undulating sea state (3≤significant wave height <5, 8≤wave period <12), and extreme sea state (significant wave height ≥5, wave period ≥12). For each sea state type, the mean, median, and other statistical measures of each core parameter are calculated and used as the central characteristic parameters of that sea state type. Simultaneously, the quantity distribution, temporal distribution, and spatial distribution information of each sea state type are recorded to clarify the occurrence frequency and main affected areas of different sea states, generating wave characterization clustering attribute data. This provides a core basis for the subsequent design of differentiated spatial multi-scale fusion strategies.

[0077] Step S324: Design a multi-scale fusion strategy for wave representation space based on the differences in clustering attributes using wave representation clustering attribute data;

[0078] In this embodiment of the invention, based on the sea state type in the clustered attribute data, fusion weights for features at different spatial scales are specifically assigned to ensure that the fused features can accurately match the core evolutionary patterns of the current sea state. For extreme sea states and heavily undulating sea states, since these sea states are significantly affected by local topography, wave evolution exhibits strong nonlinearity and local abrupt changes. Therefore, the weight of micro-scale features is set to be the highest, followed by meso-scale features, and the weight of macro-scale features is the lowest, focusing on capturing the impact of local fine features on wave evolution. For moderately undulating sea states, which exhibit both overall propagation trends and local changes, the weight of meso-scale features is set to be the highest, with a balanced distribution of macro- and micro-scale feature weights to achieve a balanced capture of overall and local features. For calm sea states and slightly undulating sea states, which are dominated by large-scale propagation trends with weak local changes, the weight of macro-scale features is set to be the highest, with lower weights for meso- and micro-scale features, focusing on capturing the overall evolutionary patterns. A clear weight allocation table is established for each sea state type, including the weight percentage at each scale and the applicable time range for each weight. Simultaneously, trigger conditions for weight adjustments are set, automatically updating the weights when the sea state type changes. Through this rule design, a multi-scale fusion strategy for the wave representation space is generated, ensuring that subsequent fusion processing can selectively extract key features under different sea states.

[0079] Step S325: Based on the multi-scale fusion strategy of wave representation space, perform time-series dynamic multi-scale fusion feature analysis of wave representation space based on the spatiotemporal multi-scale feature data, and generate spatiotemporal cross-scale fusion feature data of wave representation.

[0080] In this embodiment of the invention, the spatiotemporal multi-scale feature data of ocean waves are traversed temporally according to the time series window and sliding step size set in step S322, ensuring that the data analysis of each time window is independent and coherent. For the spatiotemporal multi-scale feature data within each time window, the core parameters of the ocean waves corresponding to that window are first extracted and compared with the central feature parameters of various sea states in the clustering attribute data of ocean waves. The sea state type corresponding to the current window is determined by calculating the feature similarity. Based on the determined sea state type, the corresponding weight allocation rules in the multi-scale fusion strategy of ocean waves are called to clarify the weight ratio of macroscopic, mesoscopic, and microscopic scale features. The spatiotemporal feature data of each scale within the window are weighted and fused according to the corresponding weights. During the fusion process, the dimensions of the feature data of each scale are first unified to ensure that the data dimensions are consistent before the weighted summation operation is performed. The weighted result is the fused feature data of that time window. After the fusion calculation is completed, the validity of the result is verified to determine whether the fused features can completely retain the key evolutionary information of each scale. If key information is missing, the weight ratio is readjusted and fused again. After completing the dynamic fusion processing of all time windows according to the above procedure, the fusion feature data of each window are connected in chronological order to generate spatiotemporal cross-scale fusion feature data of ocean wave characterization.

[0081] Furthermore, step S34 includes the following steps:

[0082] Step S341: Analyze the evolution impact of the single state of wave characterization based on the wave characterization feature data, and generate wave characterization single state evolution impact data;

[0083] In this embodiment of the invention, when analyzing the evolutionary impact of single-state wave characterization based on wave characterization feature data, the criteria for classifying single-state wave characterization are first clarified. Three scenarios—those with only storm surge characterization, those with only tidal level characterization, and those with only floodplain characterization—are defined as single states, ensuring that each single-state scenario accurately corresponds to the independent core process of wave evolution. Feature data corresponding to the three single states are separated from the wave characterization feature data, based on the existence of core characterization parameters. Scenario where only the storm surge height parameter is significant and other characterization parameters are negligible is classified as a storm surge single state; scenario where only the tidal level parameter exhibits periodic changes and other characterization parameters remain stable is classified as a tidal level single state; and scenario where only the floodplain extent parameter is non-zero and other characterization parameters show no significant fluctuations is classified as a floodplain single state. For the feature data of each single state, the influence of its core characterization parameters on the overall wave evolution within a complete time series is analyzed. This is specifically achieved by calculating the correlation between changes in core parameters and changes in total wave energy, and the matching degree between extreme values ​​of core parameters and the intensity of wave evolution. The correlation and matching degree are quantified to obtain the evolutionary contribution value of each feature parameter under each single state. The quantified contribution value directly reflects the driving ability of the parameter on wave evolution. At the same time, the temporal distribution pattern and spatial influence range of each single state are recorded to clarify the temporal stage and main area of ​​action of the single state, and generate wave characterization single state evolution influence data, laying the foundation for subsequent scene adaptation of multimodal attention strategies.

[0084] Step S342: Perform spatial specificity analysis on the spatiotemporal cross-scale fusion feature data of wave characterization to generate spatial specificity data of wave characterization superposition state, and perform evolution impact analysis on wave characterization superposition state based on the spatial specificity data of wave characterization superposition state to generate evolution impact data of wave characterization superposition state.

[0085] In this embodiment of the invention, when performing spatial specificity analysis of the superposition state of wave characterization across spatiotemporal scales using spatiotemporal cross-scale fusion feature data, the superposition state of wave characterization is first defined as a scenario where two or more single states coexist, including four types: storm surge-tidal level superposition state, storm surge-floodplain superposition state, tidal level-floodplain superposition state, and a full superposition state of all three. Superposition state region identification is achieved through feature parameter combination detection; that is, regions where the core parameters corresponding to two or more single states are all significantly non-zero are identified as superposition state regions. For the identified superposition state regions, the core parameters of the spatiotemporal cross-scale fusion feature data of wave characterization within these regions are extracted and compared with the core parameters of the corresponding single states in the same spatial region and at the same temporal stage. The difference value between the superposition state and the single state is obtained. A difference value threshold is set, and regions with difference values ​​exceeding the threshold are selected as spatially specific regions. These regions can reflect the unique laws of wave evolution under superposition states (such as energy superposition, mutual inhibition, etc.). Based on the characteristic data of spatially specific regions, the influence of superposition states on wave evolution is analyzed, and the evolution contribution value of the core parameter combination under superposition states is calculated. The calculation of contribution value needs to consider the coupling effect between parameters (such as the amplification effect of the water level rise after the superposition of storm surge height and tidal water level on the floodplain range). The spatially specific data of wave characterization superposition states and the data of wave characterization superposition state evolution influence are generated to fully capture the specific characteristics and influence mechanism of wave evolution under superposition states.

[0086] Step S343: Design an adaptive attention strategy for multimodal wave evolution by using wave characterization single-state evolution impact data and wave characterization superimposed state evolution impact data, and generate a multimodal adaptive attention strategy for wave evolution.

[0087] In this embodiment of the invention, based on single-state evolution impact data, a single-state attention weight rule is designed: for each single-state type, attention weights are allocated according to the proportion of evolutionary contribution values ​​of each feature parameter. The feature parameter with the higher the contribution value, the greater the attention weight obtained by the corresponding feature channel, ensuring that the model prioritizes the extraction of core driving features in single-state scenarios. Based on superposition state evolution impact data and spatially specific data, a superposition state attention weight rule is designed: the attention weight of feature channels in spatially specific regions is increased to a preset high value, ensuring that the model focuses on the unique evolutionary regions of the superposition state; secondly, basic attention weights are allocated according to the proportion of evolutionary contribution values ​​of the core parameters of each single state under the superposition state, and additionally, the weights of feature channels with significant coupling effects are increased according to the strength of parameter coupling effects. Scene detection trigger conditions are set, and by monitoring the combination state of core representation parameters in real time, automatic identification and strategy switching between single-state and superposition states are realized. The aforementioned weight allocation rules, scene detection logic, and switching conditions are integrated into a complete strategy framework. The execution order and calculation logic of each step are clarified, and finally, a multimodal adaptive attention strategy for wave evolution is generated. This strategy can accurately adapt to different wave evolution scenarios and improve the model's efficiency and accuracy in extracting core features.

[0088] Furthermore, step S4 includes the following steps:

[0089] Step S41: Obtain historical wave characterization evolution data;

[0090] In this embodiment of the invention, historical wave characterization evolution data is obtained, and the time range of data collection is defined as continuous time series data of the target sea area for the past 10 years. This time range can cover the wave evolution process under different seasons and different climatic conditions, ensuring that the data includes various typical scenarios such as calm sea conditions and extreme sea conditions.

[0091] Step S42: Analyze the physical constraints of wave evolution on historical wave characterization evolution data to generate physical constraint data of wave evolution.

[0092] In this embodiment of the invention, when analyzing the physical constraints of wave evolution based on historical wave characterization evolution data, the core basis of the analysis is first established as the fundamental principles of fluid mechanics and wave dynamics, ensuring that the extracted physical constraints conform to the objective natural laws of wave evolution. Typical data samples under different sea state scenarios are selected from the historical wave characterization evolution data, covering calm sea states, slightly undulating sea states, moderately undulating sea states, heavily undulating sea states, and extreme sea states, ensuring that the analysis results are adaptable to various wave evolution scenarios. Based on the selected sample data and dynamic principles, three types of core physical constraints are extracted: The first type is the energy conservation constraint of ocean waves. During the propagation of ocean waves, energy is lost due to factors such as friction and refraction, but there is an upper limit to the loss. By analyzing the initial value of ocean wave energy and the attenuation value after propagation in historical data, a quantitative threshold for the energy loss rate is determined, that is, the energy loss rate during the propagation of ocean waves must not exceed a fixed proportion of the initial energy. The second type is the wave speed constraint. There is a clear dynamic relationship between wave speed and seawater depth. By fitting the relationship between wave speed and corresponding water depth in historical data, a quantitative relationship that wave speed must satisfy is proportional to the square root of water depth is determined, and a reasonable range of the proportionality coefficient is clarified. The third type is the floodplain range constraint. The floodplain range is limited by geographical conditions such as nearshore topography and shoreline slope. By analyzing the correlation between historical floodplain data and nearshore distance and shoreline slope, a fixed distance threshold from the nearshore to the land is determined. For special topographic areas, the threshold is adjusted according to the topographic slope. Finally, physical constraint characteristic data of ocean wave evolution are generated, providing a constraint basis that conforms to physical laws for subsequent model parameter optimization.

[0093] Step S43: Transmit the historical wave characterization evolution data to the wave evolution prediction model for training, and generate a training wave evolution prediction model;

[0094] In this embodiment of the invention, when transmitting historical wave characterization and evolution data to the wave evolution prediction model for training, the historical wave characterization and evolution data are first divided into a training set and a validation set according to a fixed ratio. The training set is used for iterative updates of model parameters, and the validation set is used for real-time evaluation of the model's predictive performance. During the division process, it is ensured that the distribution of sea state types in the two datasets is consistent to avoid insufficient model generalization ability due to data distribution deviations. Subsequently, the core parameters for model training are set, with the number of iterations set to a fixed value to ensure sufficient training iteration space for the model; the initial learning rate is set to a fixed value, and a cosine annealing learning rate scheduling strategy is adopted, dynamically adjusting the learning rate by halving it every fixed round to balance the convergence speed and convergence accuracy of model training; the mean squared error loss function is used to quantify the deviation between the model's predicted values ​​and the true values ​​of historical data. The training process is carried out according to a preset iterative process. The training set data is input into the wave evolution prediction model in batches according to time sequence. The model obtains the prediction results through forward propagation. Based on the loss function, the deviation between the prediction results and the true values ​​is calculated. The parameters of each module are updated layer by layer along the model architecture through the backpropagation algorithm, including the attention weights of the Transformer encoder, the convolution kernel parameters of the CNN, the weights of the fully connected layers, etc., to generate the trained wave evolution prediction model.

[0095] Step S44: Optimize the training parameters of the trained wave evolution prediction model by using the physical constraint characteristic data of wave evolution, and generate an optimized wave evolution prediction model.

[0096] In this embodiment of the invention, the quantitative indicators in the physical constraint characteristics data of wave evolution are transformed into explicit model training constraint objectives: the predicted wave energy loss rate must be controlled within a fixed error range of the true value, the predicted wave speed must conform to the quantitative proportional relationship with water depth, and the predicted floodplain range must be within a fixed error range of the true value. Based on these constraint objectives, an auxiliary loss function for wave evolution constraints is designed, using a hinge loss function. When the model's predicted value exceeds the constraint objective range, a corresponding loss value is calculated based on the magnitude of the exceedance; the larger the exceedance, the larger the loss value. When the predicted value is within the constraint range, the loss value is zero. This design strengthens the model's adherence to physical constraints. The auxiliary loss function is then fused with the original mean squared error loss function according to a fixed weight ratio to form a new comprehensive loss function. The fusion weight is determined based on the importance of the constraint objectives, ensuring the guiding role of physical constraints in model training. Subsequently, a second training optimization of the model is initiated, with the number of iterations set to half of the initial training iterations and the initial learning rate set to one-tenth of the final learning rate of the initial training iterations. The cosine annealing learning rate scheduling strategy is still used. During each training round, the loss value is calculated based on the comprehensive loss function, and the key parameters of the model are adjusted using the backpropagation algorithm, with a focus on optimizing parameters related to spatiotemporal feature extraction and attention weight allocation. After each round of optimization, validation set data is used to verify whether the model's predicted values ​​meet the constraint objectives, and the prediction error is calculated. When the satisfaction rate of all constraint objectives reaches a fixed percentage or higher and the prediction error on the validation set stabilizes at a low level, optimization stops, and an optimized wave evolution prediction model is generated.

[0097] Furthermore, step S44 includes the following steps:

[0098] Based on the physical constraint characteristics data of wave evolution, the training objective decision analysis of wave evolution constraint is carried out to generate wave evolution constraint training objective decision, and based on the wave evolution constraint training objective decision, the auxiliary loss function of wave evolution constraint is designed to generate wave evolution constraint auxiliary loss function.

[0099] By utilizing wave evolution constraints to train the target decision and wave evolution constraint auxiliary loss function, the training parameters of the trained wave evolution prediction model are optimized and adjusted to generate an optimized wave evolution prediction model.

[0100] In this embodiment of the invention, when performing training objective decision analysis for wave evolution constraints based on wave evolution physical constraint characteristic data, the core quantitative indicators in the wave evolution physical constraint characteristic data are first systematically reviewed to clarify the verifiable target directions corresponding to three types of core physical constraints: wave energy conservation constraints, wave speed constraints, and floodplain range constraints. For wave energy conservation constraints, combined with the energy loss patterns under different sea conditions in historical data, the constraint quantitative indicators are transformed into specific training objectives: the deviation between the predicted and actual wave energy loss rate must be controlled within a fixed percentage range to ensure that the energy evolution predicted by the model conforms to the basic law of energy conservation; for wave speed constraints, based on the dynamic correlation between wave speed and water depth, the training objective is determined to be that the predicted wave speed value must strictly fall within a quantitative interval proportional to the square root of the water depth, and the interval boundary is determined through historical data fitting results; for floodplain range constraints, combined with the limiting law of nearshore topography on the floodplain, the training objective is set to ensure that the deviation between the predicted and actual floodplain range values ​​does not exceed a fixed distance threshold, and the threshold for special topographic areas is adjusted according to the topographic slope correction rule. Simultaneously, the priorities and applicable scenarios of each training objective are clearly defined. Energy conservation constraints and wave speed constraints are global priorities, applicable to all sea state scenarios, while floodplain range constraints are specific to nearshore areas. This ensures accurate matching between training objectives and actual wave evolution scenarios, generating wave evolution constraint training objective decisions. When designing the wave evolution constraint auxiliary loss function based on this decision, a hinge loss function is adopted. This form can impose a clear penalty on predicted values ​​exceeding the constraint range, while not generating additional loss for predicted values ​​conforming to the constraints. Differentiated loss calculation logic is designed for different training objectives: when the predicted energy loss rate exceeds the deviation range, the loss value increases linearly with the magnitude of the exceedance; when the predicted wave speed deviates from the quantization range, the loss value increases squarely with the degree of deviation, strengthening adherence to core physical laws; when the predicted floodplain range exceeds the threshold, the loss value is positively correlated with the exceedance distance. The loss calculation logic corresponding to the three types of constraints is integrated to form the wave evolution constraint auxiliary loss function. The wave evolution constraint auxiliary loss function is fused with the mean squared error loss function used in the original model training at a fixed weight ratio. The fusion weight is set according to the priority in the training objective decision. The weight ratio of the auxiliary loss function ensures that the physical constraints can effectively guide the optimization of model parameters, while not masking the original loss function's requirement for prediction accuracy, thus generating a comprehensive loss function. Subsequently, the training parameters for the optimization phase are set, with the number of iterations set to half of the initial training iterations to ensure fine-tuning of parameters and avoid overtraining. The initial learning rate is set to one-tenth of the final learning rate of the initial training, and a cosine annealing learning rate scheduling strategy is adopted, decreasing by a fixed proportion every fixed round to balance the convergence speed of the optimization process and the precision of parameter fine-tuning.The optimization training process proceeds according to a pre-defined procedure. The training set of historical wave evolution data is input into the wave evolution prediction model in batches over time. Forward propagation yields prediction results, and the total loss for deviations from the actual values ​​and violations of physical constraints is calculated based on the comprehensive loss function. Key parameters are adjusted layer by layer along the model architecture using the backpropagation algorithm, with a focus on optimizing the attention weights of the Transformer encoder, the convolutional kernel parameters of the CNN, and the weight parameters of the spatiotemporal cross-scale fusion layer, ensuring that parameter adjustments specifically reduce the comprehensive loss value. After each round of optimization, the model performance is validated using validation set data, and the satisfaction rate of each physical constraint training objective is checked, calculating the percentage of samples whose predicted values ​​meet the constraint requirements. When the comprehensive loss value of the validation set stabilizes at a low level for several consecutive rounds, and the satisfaction rate of all physical constraint training objectives reaches a fixed proportion or higher, parameter optimization is stopped, generating an optimized wave evolution prediction model. This model's prediction results possess both high accuracy and strict adherence to the objective physical laws of wave evolution, with significantly improved prediction stability under extreme sea conditions.

[0101] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0102] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for constructing a large ocean wave model based on deep learning, characterized in that, Includes the following steps: Step S1: Acquire multi-source marine correlation monitoring data; perform spatial interpolation mapping processing on the multi-source marine correlation monitoring data to generate spatial mapping data of marine correlation monitoring features; Step S1 includes the following steps: Step S11: Obtain multi-source marine correlation monitoring data, wherein the multi-source marine correlation monitoring data includes marine typhoon path intensity data, sea surface wind field data, sea surface air pressure data, ocean wave monitoring data, and marine geographic information data, wherein the marine geographic information data includes marine astronomical tide data, ocean water depth data, ocean seabed topography data, and ocean nearshore line data; Step S12: Based on marine geographic information data, perform preliminary three-dimensional spatial modeling of marine topography to obtain a preliminary three-dimensional spatial model of marine topography; perform marine geographic boundary feature analysis based on marine geographic information data to obtain marine geographic boundary feature data; design an adaptive mapping mesh for marine geographic spatial features based on the preliminary three-dimensional spatial model of marine topography and marine geographic boundary feature data to obtain marine geographic spatial mapping mesh data; optimize the mapping mesh of the preliminary three-dimensional spatial model of marine topography using the marine geographic spatial mapping mesh data to obtain a three-dimensional spatial model of marine topography. Step S13: Perform multi-source data time-series synchronization benchmark preprocessing on the multi-source marine correlation monitoring data to generate synchronized multi-source marine correlation monitoring data; Step S14: Transmit synchronous multi-source marine correlation monitoring data to a three-dimensional marine topography model for spatial interpolation mapping of marine correlation monitoring features to generate marine correlation monitoring feature spatial mapping data; Step S2: Perform wave characterization feature analysis on the marine correlation monitoring feature spatial mapping data using heterogeneous wave correlation feature channels to generate wave characterization feature data. Step S3: Based on the preset deep learning algorithm and wave characterization feature data, establish the prediction mapping relationship between wave characterization and evolution, and generate a wave evolution prediction model; Step S3 includes the following steps: Step S31: Establish a deep learning large model architecture for wave representation and evolution based on a preset deep learning algorithm, and generate a wave evolution prediction model architecture. The preset deep learning algorithm adopts a hybrid architecture combining a Transformer encoder and a convolutional neural network. Step S32: Perform spatial multi-scale feature analysis of wave representation based on wave representation feature data to generate spatial multi-scale feature data of wave representation; perform temporal feature analysis on the spatial multi-scale feature data of wave representation to generate spatiotemporal multi-scale feature data of wave representation; perform clustering attribute analysis of wave representation based on wave representation feature data to generate wave representation clustering attribute data; design a spatial multi-scale fusion strategy for wave representation based on clustering attribute differences using the wave representation clustering attribute data; perform temporal dynamic spatial multi-scale fusion feature analysis of wave representation based on the spatial multi-scale fusion strategy of wave representation to generate spatiotemporal cross-scale fusion feature data of wave representation; Step S33: The wave evolution prediction model architecture is embedded by performing a wave characterization spatiotemporal cross-scale fusion layer embedding process on the wave characterization spatiotemporal cross-scale fusion feature data to generate an embedded wave evolution prediction model architecture. Step S34: Analyze the evolutionary impact of a single state of wave representation based on wave representation feature data, generating wave representation single-state evolution impact data; perform spatial specificity analysis on wave representation superposition state based on spatiotemporal cross-scale fusion feature data, generating wave representation superposition state spatial specificity data, and analyze the evolutionary impact of wave representation superposition state based on the wave representation superposition state spatial specificity data, generating wave representation superposition state evolution impact data; design an adaptive attention strategy for multimodal wave evolution using wave representation single-state evolution impact data and wave representation superposition state evolution impact data, generating a multimodal adaptive attention strategy for wave evolution. Step S35: The wave evolution prediction model is obtained by setting the superposition state attention mechanism of the wave representation space scene through the wave evolution multimodal adaptive attention strategy. Step S4: Obtain historical wave characterization and evolution data; use the historical wave characterization and evolution data to train the wave evolution prediction model and optimize the wave evolution constraints to generate an optimized wave evolution prediction model. Step S4 includes the following steps: Step S41: Obtain historical wave characterization evolution data; Step S42: Analyze the physical constraints of wave evolution on historical wave characterization evolution data to generate physical constraint data of wave evolution. Step S43: Transmit the historical wave characterization evolution data to the wave evolution prediction model for training, and generate a training wave evolution prediction model; Step S44: Based on the physical constraint characteristic data of wave evolution, perform training objective decision analysis for wave evolution constraints to generate wave evolution constraint training objective decisions, and design auxiliary loss functions for wave evolution constraints based on the wave evolution constraint training objective decisions to generate auxiliary loss functions for wave evolution constraints; use the wave evolution constraint training objective decisions and auxiliary loss functions to optimize and adjust the training parameters of the trained wave evolution prediction model to generate an optimized wave evolution prediction model.

2. The method for constructing a large ocean wave model based on deep learning according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Analyze the influence relationship between the heterogeneous feature channels of ocean wave association and the wave characterization based on the ocean-related monitoring feature spatial mapping data, and generate ocean wave association heterogeneous feature-characterization influence relationship data; Step S22: Perform integrated analysis of wave characterization influence relationship based on the wave association heterogeneous characteristics-characterization influence relationship data to generate integrated wave characterization influence relationship data; Step S23: Perform wave characterization feature analysis on the integrated data of wave characterization influence relationship to generate wave characterization feature data.

3. The method for constructing a large ocean wave model based on deep learning according to claim 2, characterized in that, Step S21 includes the following steps: Step S211: Extract feature spatial mapping data of marine typhoon track intensity data, sea surface wind field data, sea surface pressure data, ocean depth data and ocean seabed topography data from marine associated monitoring feature spatial mapping data, and perform wave and storm surge characterization analysis through the feature spatial mapping data of marine typhoon track intensity data, sea surface wind field data, sea surface pressure data, ocean depth data and ocean seabed topography data to generate wave and storm surge characterization data; Step S212: Extract feature spatial mapping data of marine astronomical tide data from marine correlation monitoring feature spatial mapping data, and perform ocean wave and storm surge characterization analysis by combining ocean wave and storm surge characterization data with feature spatial mapping data of marine astronomical tide data to generate ocean wave and storm surge characterization data. Step S213: Extract the feature spatial mapping data of wave monitoring data and nearshore data based on the marine correlation monitoring feature spatial mapping data, and perform wave-shoal characterization analysis by combining wave monitoring data and nearshore data with wave storm surge characterization data and wave tide level characterization data to generate wave-shoal characterization data. Step S214: Based on the wave storm surge characterization data, wave tide level characterization data, and wave beach characterization data, analyze the influence relationship between wave-related heterogeneous feature channels and wave characterization, and generate wave-related heterogeneous feature-characterization influence relationship data.

Images