Data-driven large-scale regional sea wave rapid prediction method and device

Abstract

The application discloses a data-driven large-scale regional sea wave rapid prediction method and device, wherein the method comprises the following steps: acquiring multi-source marine physical environment data of a target sea area, and generating a standard physical field data stream with unified space-time resolution; constructing a multi-channel space-time input tensor containing wind field driving information, topographic boundary information and historical wave state information; minimizing the difference between the predicted wave height and the real wave height through a back propagation mechanism, thereby obtaining a trained wave height prediction model; receiving the latest wind speed field data observed in real time or output by a numerical prediction model, and outputting a sea wave significant wave height prediction field at a future target time; and the device is used for realizing the above method. The method technical scheme provided by the application accurately captures the time delay of wind energy transmission to wave energy through a physical lag alignment mechanism, and improves the modeling accuracy; meanwhile, the deep learning is used to realize the rapid deduction of a large-scale sea wave field, and high accuracy and high timeliness are achieved.

Description

Technical Field

[0001] This invention relates to the technical field of marine weather forecasting, and in particular to a data-driven method and apparatus for rapid forecasting of large-scale regional ocean waves. Background Technology

[0002] With the expansion of global shipping and trade and the advancement of deep-sea engineering construction, the safety of the marine environment has increasingly become a focus of international attention. The accuracy of wave forecasting directly affects the selection of offshore operation windows, the optimization of shipping routes, and the effectiveness of coastal disaster prevention and mitigation systems. As a core indicator describing wave energy and intensity, accurate prediction of significant wave height (SWH) is crucial for assessing fatigue damage to marine engineering structures, guiding the development and utilization of wave energy resources, and responding to emergency rescue in extreme sea conditions. However, the marine dynamic environment is a typical strongly coupled, nonlinear, and complex system. Wave generation and evolution are not only directly driven by the current wind field but also have complex spatiotemporal dependencies on multiple physical factors such as seabed topography, historical wave propagation, and pressure fields.

[0003] Current wave forecasting primarily relies on physics-based numerical models, which, while theoretically rigorous, still have significant limitations in practical applications. Numerical models depend on solving complex systems of differential equations, resulting in enormous computational costs and making it difficult to meet the demands for real-time, rapid forecasting over large-scale regions. Significant physical lag exists in the transfer of wind energy to wave energy, making it difficult for traditional methods to accurately capture the temporal dislocation characteristics of wind and wave generation and the long-distance propagation patterns of cross-regional swells. Under extreme weather conditions such as typhoons, the nonlinear characteristics of ocean waves are dramatically enhanced, and traditional models often underestimate peak values ​​for high sea states due to limitations in their parameterization schemes. Therefore, effectively addressing these issues has become a pressing problem for the industry. Summary of the Invention

[0004] This invention provides a data-driven method and apparatus for rapid forecasting of large-scale regional ocean waves, which enables rapid extrapolation of large-scale ocean wave fields and combines high accuracy with high computational efficiency.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] A data-driven method for rapid forecasting of large-scale regional ocean waves is provided, including:

[0007] S1. Acquire multi-source marine physical environment data of the target sea area. The multi-source marine physical environment data includes historical wind speed vector sequences, historical wave height sequences, and static seabed topography data. Perform standardized preprocessing with physical constraints on the multi-source marine physical environment data. The standardized preprocessing performs quality control on wind speed, wave height, and topography data. Fill in time gaps and remove statistical outliers through spline interpolation. Construct a unified latitude and longitude reference grid. Fit the wind field distribution through spatial interpolation to ensure gradient smoothness. Introduce a land-sea binary mask matrix and perform logical operations with the physical data. Remove invalid land grids to generate a standard physical field data stream with unified spatiotemporal resolution.

[0008] S2, based on the physical hysteresis characteristics of wave excitation, the standard physical field data stream is misaligned in the time dimension, and combined with feature stacking in the channel dimension and sliding window sampling in the time dimension, a multi-channel spatiotemporal input tensor containing wind field driving information, terrain boundary information and historical wave state information is constructed.

[0009] S3, Construct a deep learning prediction model that includes a spatial feature encoding module, a spatiotemporal feature perception module and a decoder module. Iteratively train the deep learning prediction model using the multi-channel spatiotemporal input tensor. Minimize the difference between the predicted wave height and the actual wave height through the backpropagation mechanism to obtain the trained wave height prediction model.

[0010] S4, receive the latest wind speed field data output by real-time observation or numerical forecasting model, perform forward inference calculation on the latest wind speed field data through the trained wave height prediction model, and output the effective wave height prediction field of the ocean waves at the future target time.

[0011] Furthermore, in S1, the multi-source marine physical environment data undergoes standardization preprocessing that includes physical constraints, including:

[0012] S11, Perform data quality control on the historical wind speed vector sequence, historical wave height sequence and static seabed topography data. The data quality control includes at least numerical filling of missing time nodes using cubic spline interpolation function and removing abnormal values ​​based on statistical distribution patterns.

[0013] S12, construct a unified latitude and longitude reference grid, map the quality-controlled historical wind speed vector sequence to the latitude and longitude reference grid, and fit the spatial continuous distribution characteristics of the wind speed field through cubic spline interpolation to ensure that the physical quantities stored in the grid nodes have smooth gradient characteristics.

[0014] S13, introduce a binarized mask matrix representing the land-sea distribution, perform logical operations on the binarized mask matrix and the physical data on the reference grid, remove invalid grid points in the land area, and retain only the physical values ​​of the valid sea area, thereby generating the standard physical field data stream.

[0015] Furthermore, in step S2, the standard physical field data stream is subjected to a misalignment alignment operation in the time dimension, and combined with feature stacking in the channel dimension and sliding window sampling in the time dimension, a multi-channel spatiotemporal input tensor containing wind field driving information, terrain boundary information, and historical wave state information is constructed, including:

[0016] S21, Align the historical wind speed vector sequence and the historical wave height sequence in the standard physical field data stream in the spatial coordinate system, and then shift the historical wind speed vector sequence forward by n time steps relative to the historical wave height sequence in the time axis direction to match the time lag process of wind energy being transferred to the sea surface to form waves.

[0017] S22, Perform a copy broadcast operation on the static seabed topography data in the time dimension to expand it into a topography feature sequence with the same time length as the historical wind speed vector sequence, thereby giving the static topography data a dynamic time dimension attribute.

[0018] S23, the translated historical wind speed vector sequence, historical wave height sequence, and terrain feature sequence are deeply stacked along the channel dimension to generate a dimension of... The four-dimensional tensor is used to generate the multi-channel spatiotemporal input tensor by extracting a fixed-length time segment from the four-dimensional tensor through a sliding window.

[0019] Furthermore, in S3, a deep learning prediction model is constructed, comprising a spatial feature encoding module, a spatiotemporal feature perception module, and a decoder module, including:

[0020] S31, receive the multi-channel spatiotemporal input tensor, and use multiple sets of parallel depth-separable convolutional layers with different kernel sizes to extract local spatial texture features and global background distribution features of wind field, wave field and terrain at different physical scales, respectively. The depth-separable convolutional layers maintain the attribute isolation of each physical variable based on channel independence.

[0021] S32 performs stride convolution on the feature maps extracted by each parallel branch to perform spatial downsampling, in order to filter out high-frequency noise interference and compress the spatial resolution of the feature maps;

[0022] S33 concatenates and fuses the downsampled branch features along the channel dimension, and overlays a learnable geolocation encoding matrix to explicitly inject absolute location semantic information of the latitude and longitude grid, outputting initial embedding features with geolocation awareness.

[0023] Furthermore, in S3, the specific execution steps of the spatiotemporal feature perception module include:

[0024] S34, construct a dual-stream parallel processing architecture to process the initial embedded features, wherein the first branch extracts shallow texture features of wind waves directly driven by the local wind field through stacked residual convolutional layers, and the second branch extracts deep semantic features of swell waves with cross-regional long-distance propagation characteristics through continuous downsampling convolutional layers.

[0025] S35, the shallow texture features of the wind and waves are mapped into a query matrix through a linear projection transformation layer, and the deep semantic features of the surging waves are mapped into a key matrix and a value matrix;

[0026] S36, introduce a spatiotemporal propagation bias matrix that characterizes the physical laws of wave propagation, and incorporate the spatiotemporal propagation bias matrix into the attention weight calculation based on the query matrix and the key matrix. Output global spatiotemporal features that integrate local wind and wave information and long-distance swell information through an interactive retrieval mechanism.

[0027] Furthermore, it also includes:

[0028] S5, iteratively train the deep learning prediction model using an adaptive weighted loss function. The calculation steps of the adaptive weighted loss function include:

[0029] S51, obtain the true wave height value distribution at the current time step, set a high sea state judgment threshold, and identify high sea state sample points belonging to extreme sea states based on the comparison result between the true wave height value and the high sea state judgment threshold.

[0030] S52, construct a nonlinear penalty weight function for the high sea state sample points, and generate a sample weight mask matrix corresponding to the current training batch based on the nonlinear penalty weight function, so that the weight coefficients corresponding to the high sea state sample points are significantly higher than those of ordinary sea state sample points.

[0031] S53, perform element-wise multiplication of the sample weight mask matrix and the prediction error matrix to calculate the weighted total loss value, and use minimizing the weighted total loss value as the optimization objective of gradient descent to drive model parameter updates to preferentially fit the wave height characteristics of high sea state areas.

[0032] According to a second aspect of the present invention, a data-driven, large-scale regional rapid wave forecasting device is provided, comprising:

[0033] The acquisition module is used to acquire multi-source marine physical environment data of the target sea area. The multi-source marine physical environment data includes historical wind speed vector sequences, historical wave height sequences, and static seabed topography data. The module performs standardized preprocessing with physical constraints on the multi-source marine physical environment data. The standardized preprocessing performs quality control on wind speed, wave height, and topography data, fills in time gaps and removes statistical outliers through spline interpolation, constructs a unified latitude and longitude reference grid, fits the wind field distribution through spatial interpolation to ensure gradient smoothness, introduces a land-sea binary mask matrix and performs logical operations with the physical data, removes invalid land grids, and generates a standard physical field data stream with unified spatiotemporal resolution.

[0034] The construction module is used to perform misalignment alignment operation on the standard physical field data stream in the time dimension based on the physical hysteresis characteristics of wave excitation, and combine feature stacking in the channel dimension with sliding window sampling in the time dimension to construct a multi-channel spatiotemporal input tensor containing wind field driving information, terrain boundary information and historical wave state information.

[0035] The training module is used to construct a deep learning prediction model that includes a spatial feature encoding module, a spatiotemporal feature perception module, and a decoder module. The deep learning prediction model is iteratively trained using the multi-channel spatiotemporal input tensor. The difference between the predicted wave height and the actual wave height is minimized through the backpropagation mechanism, thereby obtaining the trained wave height prediction model.

[0036] The output module is used to receive the latest wind speed field data output by real-time observation or numerical prediction models, perform forward inference calculation on the latest wind speed field data through the trained wave height prediction model, and output the effective wave height prediction field of the ocean waves at the future target time.

[0037] According to a third aspect of the present invention, an electronic device is provided, comprising: a communication interface, a processor, and a memory;

[0038] The memory is used to store program instructions, which, when executed by the processor connected to the memory via the communication interface, implement any of the above-described data-driven large-scale regional rapid wave forecasting methods.

[0039] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, which, when executed by a computer (e.g., a processor in the computer), implement any of the above-described data-driven large-scale regional rapid wave forecasting methods.

[0040] Compared with the prior art, the advantages of the present invention are as follows:

[0041] By introducing the physical hysteresis characteristics of wave excitation and performing misalignment alignment on the data, the time delay mechanism of wind energy to wave energy transfer is effectively captured, improving the model's physical representation ability and prediction accuracy of the wave generation process. Replacing the traditional numerical solution process with a deep learning model significantly reduces computational complexity while ensuring the accuracy of large-scale spatiotemporal feature extraction, enabling rapid extrapolation of the effective wave height prediction field. Attached Figure Description

[0042] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0043] Figure 1 A flowchart of a data-driven method for rapid forecasting of large-scale regional ocean waves, provided as an embodiment of the present invention;

[0044] Figure 2 A flowchart of data preprocessing in a data-driven, large-scale regional rapid wave forecasting method provided for embodiments of the present invention;

[0045] Figure 3 A flowchart illustrating the construction of input samples in a data-driven, large-scale regional rapid wave forecasting method provided for embodiments of the present invention;

[0046] Figure 4 A flowchart of the spatial feature encoding module in a data-driven large-scale regional rapid wave forecasting method provided for embodiments of the present invention;

[0047] Figure 5 A flowchart of a spatiotemporal feature perception module in a data-driven large-scale regional rapid wave forecasting method is provided for embodiments of the present invention.

[0048] Figure 6 A flowchart of a data-driven large-scale regional rapid wave forecasting method with high sea state weighted penalty function prediction is provided for embodiments of the present invention;

[0049] Figure 7 A structural diagram of a data-driven large-scale regional rapid wave forecasting device provided for embodiments of the present invention;

[0050] Figure 8 A structural diagram of an electronic device provided as an embodiment of the present invention;

[0051] Figure 9A flowchart illustrating the overall structure and operation of a deep learning model for large-scale regional wave prediction provided in an embodiment of the present invention.

[0052] Figure 10 A schematic diagram of a spatial feature encoding module provided for an embodiment of the present invention;

[0053] Figure 11 A schematic diagram of a spatiotemporal feature perception module provided in an embodiment of the present invention;

[0054] Figure 12 A schematic diagram of the training loss curve provided for an embodiment of the present invention;

[0055] Figure 13 A schematic diagram illustrating the 1-hour prediction results for grid points at different water depths within a study area, provided as an embodiment of the present invention;

[0056] Figure 14 This is a schematic diagram of wave prediction results within a study area for the next 1 hour, provided as an embodiment of the present invention. Detailed Implementation

[0057] The present invention will be further described below with reference to the accompanying drawings and specific preferred embodiments, but this does not limit the scope of protection of the present invention.

[0058] The technical solution adopted in this embodiment is as follows: Figure 1 As shown, it includes the following steps:

[0059] S1. Acquire multi-source marine physical environment data of the target sea area. The multi-source marine physical environment data includes historical wind speed vector sequences, historical wave height sequences, and static seabed topography data. Perform standardized preprocessing with physical constraints on the multi-source marine physical environment data. The standardized preprocessing performs quality control on wind speed, wave height, and topography data. Fill in time gaps and remove statistical outliers through spline interpolation. Construct a unified latitude and longitude reference grid. Fit the wind field distribution through spatial interpolation to ensure gradient smoothness. Introduce a land-sea binary mask matrix and perform logical operations with the physical data. Remove invalid land grids to generate a standard physical field data stream with unified spatiotemporal resolution.

[0060] S2, based on the physical hysteresis characteristics of wave excitation, the standard physical field data stream is misaligned in the time dimension, and combined with feature stacking in the channel dimension and sliding window sampling in the time dimension, a multi-channel spatiotemporal input tensor containing wind field driving information, terrain boundary information and historical wave state information is constructed.

[0061] S3, Construct a deep learning prediction model that includes a spatial feature encoding module, a spatiotemporal feature perception module and a decoder module. Iteratively train the deep learning prediction model using the multi-channel spatiotemporal input tensor. Minimize the difference between the predicted wave height and the actual wave height through the backpropagation mechanism to obtain the trained wave height prediction model.

[0062] S4, receive the latest wind speed field data output by real-time observation or numerical forecasting model, perform forward inference calculation on the latest wind speed field data through the trained wave height prediction model, and output the effective wave height prediction field of the ocean waves at the future target time.

[0063] For step S1, this embodiment selects a specific sea area as the target sea area and acquires multi-source marine physical environment data covering the past five years. Specifically, the data sources include, but are not limited to, ERA5 reanalysis wind field data provided by the European Centre for Medium-Range Weather Forecasts (ECMWF). This ERA5 reanalysis wind field data serves as a historical wind speed vector sequence, a historical wave height sequence fused from buoy observation stations and satellite altimeters, and static seabed topography data provided by the ETOPO1 global relief model. In the data preprocessing stage, the system first performs strict data quality control, using cubic spline interpolation to fill in time series gaps caused by sensor malfunctions, with the filling window set to a maximum of 6 hours of continuous missing data; simultaneously, statistically significant outliers are removed based on the 3σ criterion. A unified latitude and longitude reference grid with a spatial resolution of 0.25°×0.25° is constructed, and wind speed and wave height data from different sources are mapped to this grid, ensuring a smooth gradient of physical quantities in spatial distribution through interpolation. To eliminate the interference of land on the calculation of ocean physical fields, this embodiment introduces a high-precision land-sea binarization mask matrix. The mask matrix is ​​then subjected to the Hadamard product operation with the physical data field, and the values ​​of land grid points are set to zero, retaining only the physical values ​​of the effective sea area. This generates a standard physical field data stream with uniform spatiotemporal resolution, such as one frame per hour with a 64×64 grid size.

[0064] Step S2 addresses the physical lag issue in the wind-wave coupling process. Unlike traditional pure data-driven models that directly input wind and wave data at the same moment into the network, this embodiment, based on physical oceanography principles, considers the time required for wind energy to be transferred to the sea surface and generate wave growth. It innovatively designs a time-shift alignment operation. Specifically, the system sets a lag time parameter n, for example, n=2, representing a 2-hour lag. The historical wind speed vector sequence is shifted forward n time steps relative to the wave sequence on the time axis, so that the wind field at time t in the input tensor corresponds to the wave field at time t+n, thus explicitly matching the causal logic of wind-wave generation at the data level. Simultaneously, for static seabed topography data, a time-dimensional broadcast copy operation is performed, expanding it into a dynamic tensor of the same length as the wind-wave sequence. In the channel dimension, the processed zonal wind speed u, meridional wind speed v, significant wave height Hs, and topographic depth d are depth-stacked to construct a dimension... A multi-channel spatiotemporal input tensor is used. Here, B is the batch size, T is the sliding window length (e.g., 24 hours), H and W are the grid length and width, and C is the number of feature channels. This tensor fully preserves wind field driving information, terrain boundary constraints, and the historical inertial information of the waves themselves.

[0065] In the stage of building the deep learning prediction model, this embodiment designs an end-to-end network architecture that includes spatial feature encoding, spatiotemporal feature awareness, and a decoder. For the spatial feature encoding module, to avoid premature confusion of different physical attribute features, the first layer of the model uses multiple sets of parallel depthwise separable convolutional layers. The depthwise separable convolutional layers extract features independently for the wind speed channel, wave height channel, and terrain channel, and use convolutional kernels of different sizes to capture local texture and global background. Spatial downsampling is performed through convolution operations with a stride of 2, which filters out high-frequency observation noise and compresses the feature map resolution to 1 / 2 or 1 / 4 of the original size to reduce memory consumption. Considering the differences in the Coriolis effect and propagation characteristics of ocean waves at different latitudes and longitudes, this embodiment superimposes a learnable geolocation encoding matrix on the stitched feature map, explicitly injecting the absolute positional semantics of each grid point, enabling the model to perceive the spatial differences between nearshore and deep sea areas and output initial embedded features with geolocation awareness.

[0066] The spatiotemporal feature perception module employs a dual-stream parallel processing architecture and an improved attention mechanism. Considering that ocean waves mainly consist of wind waves and swells, where wind waves are high-frequency and directly driven by local winds, while swells are low-frequency and propagate long distances across regions, this embodiment designs a dual-stream branch. The first branch utilizes stacked residual convolutional layers (ResNetBlocks) to extract shallow texture features of wind waves caused by drastic changes in the local wind field; the second branch extracts deep semantic features of swells with large-scale spatiotemporal correlations through continuous downsampling. To fuse these two types of features, the model introduces a self-attention mechanism, mapping wind wave features to a query matrix Q and swell features to a key matrix K and a value matrix V. A predefined spatiotemporal propagation bias matrix is ​​introduced when calculating the attention weights. This matrix, based on the group velocity theory of wave propagation, acts as a mask to restrict the attention mechanism to focus only on the physically possible neighborhood range, rather than blindly associating the entire image. Through this interactive retrieval mechanism, the model successfully outputs global spatiotemporal features that fuse local wind wave outbreak information and far-field swell background information.

[0067] In the training and optimization phases of the model, to address the issue that traditional mean squared error loss functions tend to lead to underestimation of extreme values, this embodiment employs an adaptive weighted loss function. In each training iteration, the system first acquires the true wave height distribution at the current time step and sets a high sea state threshold. If the wave height at a grid point exceeds this threshold, it is marked as a high sea state sample point. A nonlinear penalty weight function, such as an exponential growth function, is constructed to generate weight coefficients for these high sea state samples that are significantly higher than those for ordinary sea states, forming a sample weight mask matrix. During backpropagation gradient calculation, this mask matrix is ​​multiplied element-wise with the prediction error matrix, forcing the optimizer to prioritize the fitting error in areas with typhoon waves or giant waves. The trained model, while maintaining a low overall average error, improves its ability to capture the peak values ​​of extreme and hazardous waves, solving the problem of accurate average values ​​but poor extreme value predictions.

[0068] In a preferred embodiment, such as Figure 9 The method for predicting significant wave height of ocean waves in a large-scale region based on a deep learning model, as shown, specifically includes the following steps:

[0069] A1. Obtain historical wind speed data, historical wave height data, and corresponding static seabed topography data for the study area and perform preprocessing.

[0070] A2 concatenates processed wind speed, wave height, and static terrain data along the time dimension to construct a multi-channel tensor, and uses a sliding window approach to generate training samples for historical wind speed and future wave height.

[0071] A3. Construct a deep learning model for predicting the significant wave height of ocean waves in a large-scale region. Iteratively train the deep learning model using the training samples. The deep learning model for predicting the significant wave height of ocean waves in a large-scale region in A3 includes: a spatial feature encoding module, a spatiotemporal feature perception module, and a decoder module. During training, the model weights are updated by constructing a high sea state adaptive weighted loss function combined with a backpropagation mechanism to determine the optimal parameter configuration.

[0072] A4 receives the latest wind speed and wave height data from real-time external observations or numerical weather prediction models. First, it processes the wind speed data according to step A2. Then, while keeping the input time window length unchanged, it uses the processed data to slide and update the previous input time window to form a new input sample. This sample is then input into the trained deep learning model to output the effective wave height prediction result for the next target time.

[0073] Specifically, the data acquisition in step A1 includes the following steps:

[0074] A1.1 To ensure modeling accuracy and spatial coverage, the data used can come from multiple sources, and one or more of them can be fused depending on the actual situation: satellite remote sensing data, buoy observation data, reanalysis data, and numerical simulation data, etc.

[0075] A1.2, Data Preprocessing.

[0076] Specifically, step A1.2 includes the following steps:

[0077] A1.2.1 After acquiring the data, it is necessary to unify the projection coordinate system and spatial resolution, and to crop the data according to the target area and the research period.

[0078] A1.2.2, Use physical constraint methods to perform basic physical constraint checks on the original data and remove invalid data;

[0079] A1.2.3, different interpolation methods are used to fill in missing time steps or spatial locations. For missing time series data, cubic spline interpolation is used. For example, given a set of time series data points (X0, Y0), (X1, Y1), ..., (X... n Y n ), for the interval (X) i X i+1 Construct a cubic interpolation function:

[0080]

[0081] This ensures the continuity of function values, first derivatives, and second derivatives at each node, thus achieving a smooth transition. For missing spatial location data, an inverse distance weighted average is used for interpolation, calculated as follows. Assuming the target point is X0, the wind and wave data at that point is obtained by inverse distance weighted average of the surrounding N known points.

[0082]

[0083] Where d(X0, X i () is the distance from point X0 to point X i The Euclidean distance between them, where p is a power function of the distance;

[0084] A1.2.4 For outliers in the data, the Z-score detection method under the sliding window is adopted. The local mean and standard deviation are calculated within a sliding window, and then the outliers are filled in using the cubic spline interpolation method.

[0085] Section A1.2.5 performs masking on grid points containing land areas in the input data, including wind, wave, and topographic data. Based on the ETOPO1 global topographic elevation model, the land-sea boundary is determined, and a binary mask matrix is ​​constructed, retaining only the valid sea area grid. This mask can be used as input weights during training and inference phases to shield the model from interference from land areas, improving spatial prediction consistency.

[0086] Specifically, step A2 includes the following steps:

[0087] A2.1 aligns the historical wind speed vector field (U and V components) and the historical significant wave height field according to spatial coordinates. In the time dimension, the historical wind speed data and the historical significant wave height data are misaligned by n hours before being aligned. Then, they are stacked and spliced ​​to form a multi-channel tensor X∈C×T×H×W, where C, T, H, and W represent the number of input channels of the wind speed sequence, the time dimension of the wind speed sequence, and the height and width of the input tensor, respectively.

[0088] A2.2, the static seabed topography data is copied and expanded in the time dimension to a time series data with the same length as the wind speed sequence, and this data is added as an additional feature dimension to the multi-channel tensor X to obtain the input data that fuses the static topography information.

[0089] A2.3, the sliding window method is used to extract historical wind speed sequences of length Tin as input samples from the multi-channel tensor. The sliding window step size is 1 hour to generate continuous samples. The wave height data of Tout at the end of each window is the output label. After processing by the sliding window method, the multi-channel tensor is integrated into the input sample tensor X0∈N×C×T×H×W, where N is the number of samples, and the other variables are the same as in A2.1.

[0090] Specifically, such as Figure 10 As shown, the spatial feature encoding module in step A3 includes the following steps:

[0091] The input tensor X0∈N×C×T×H×W from step A2 is received and independently applied to each hour across T time steps using shared-weight 2D convolutions. To capture physical features at different scales, this case constructs three parallel depthwise separable convolution branches. Each branch uses a different kernel size to define a different physical receptive field: 3×3 for extracting high-frequency local wind and wave textures and terrain edge features, 5×5 for extracting medium-range wind field distribution features, and 7×7 for capturing large-scale background dependencies. An independent depthwise convolution is used for each branch, with the kernel depth matching the number of input channels. Each input physical variable has an independent kernel for feature extraction, avoiding the physical information chaos caused by the forced data fusion of traditional convolutions. The feature maps output from each branch are then subjected to stride convolution. In this embodiment, the convolution stride s is set to 2 to filter out high-frequency noise and compress the spatial dimension. Finally, the downsampled feature maps from each branch are concatenated along the channel dimension to obtain high-dimensional features that fuse multi-scale information. In addition, since convolutional neural networks have translation invariance, they cannot perceive the absolute geographical location of ocean waves. Therefore, in this embodiment, a learnable geolocation encoding matrix Ppos is introduced, which is automatically updated during network training to explicitly inject the absolute location semantic information of the latitude and longitude grid. Finally, the initial embedded feature U with time dimension information is output. The overall calculation formula is as follows:

[0092]

[0093] Where BN represents batch normalization operation, This represents the activation function. Down s `s` represents a downsampling operation with a step size of `s`, used for spatial aggregation. `Concat` represents the concatenation of information from different channel dimensions. (DW) k×k A channel-independent depthwise separable convolution with a kernel size of k is used to extract physical field features at a specific scale. P pos ⊕ represents the geolocation coding matrix, and ⊕ represents the element-wise addition operation at spatial locations.

[0094] Furthermore, such as Figure 11 As shown, the spatiotemporal feature perception module in step S3 specifically includes the following steps:

[0095] The initial embedded features U are processed using parallel processing branches. First, the first branch consists of three stacked residual convolutional blocks. Each residual block comprises a 3×3 convolutional layer, a normalization layer, and a ReLU activation function, with a convolution stride of 1. While maintaining the feature map size, this branch performs local convolution operations on the feature sequence U at each time step, outputting the shallow texture features Fs of the wind and waves driven by the current local wind field. Second, the second branch uses three consecutive downsampling stride convolutions, each with a 3×3 kernel and a stride of 2, progressively downsampling the feature sequence U. This compresses the feature map resolution to 1 / 8 of its original size, thereby expanding the receptive field to cover the entire study area and outputting the deep semantic features Fd of the swells, which exhibit long-distance propagation characteristics across regions. Subsequently, a linear projection transformation is used to map the shallow texture features Fs of the wind and waves at the current time t to a query matrix Q, and the deep semantic features Fd of the swells to a key matrix K and a value matrix V. Finally, based on this, a spatiotemporal propagation bias matrix of the same dimension as the attention mechanism is introduced to participate in the attention weight calculation. The attention score is calculated through the interactive retrieval of the query matrix Q and the key matrix K, outputting a global spatiotemporal feature F that integrates local wind and wave information with long-distance surge information. The overall calculation expression is as follows:

[0096]

[0097]

[0098] Among them, F res For residual feature extraction operations, H down For downsampling operations, W q W k、 W v These are the learnable linear projection matrices, d k β is the scaling factor. p This is the physical propagation bias matrix.

[0099] Furthermore, the decoder module in step S3 includes:

[0100] The decoder employs a cascaded transposed convolutional structure to progressively decode and refine the feature F. In each decoding stage, transposed convolution is used for upsampling to gradually recover the fine spatial structure of the wave field. Normalization and activation functions are used to eliminate inter-layer feature offsets, avoiding the spatial resolution loss and detail blurring caused by single projection in conventional convolution. Finally, the features are mapped to wave height values ​​through output layer convolution. Non-negative physical constraints are applied during the construction process to output the high-resolution effective wave height prediction result Y at the target time. The overall calculation expression is as follows:

[0101]

[0102] in, This represents the transposed convolution of the k-th layer, where n is the number of transposed convolution layers. conv represents a cascade operation. out This indicates the output layer convolution operation, and ReLU represents the non-negative physical constraint activation function.

[0103] Furthermore, the adaptive weighted loss function for high sea state conditions in step A3 is as follows:

[0104] During training, a sample weight matrix is ​​first constructed based on the numerical distribution of the actual wave height at time t to be predicted. This matrix uses a high sea state threshold τ. Since waves greater than 4m are generally considered giant waves, τ is set to 4m in this implementation. For sample points with actual wave heights exceeding this threshold, a non-linear penalty coefficient is generated, forming a weight mask for high sea states. Secondly, during model validation, the calculated absolute error matrix and the sample weight matrix are multiplied element-wise to correct the prediction error of high sea state samples, resulting in the final weighted total loss (Loss). Finally, minimizing this weighted total loss drives the model to prioritize optimizing the parameter weights in high sea state regions during gradient descent. The specific loss function calculation expression is as follows:

[0105]

[0106] Furthermore, step A4 specifically includes the following steps:

[0107] A4.1 receives the latest wind speed field data (u, v) and wave height observation data from real-time external observations or numerical weather prediction models; performs the same preprocessing operations as A1 on the wind speed field, including spatial resampling, temporal alignment, quality control, and normalization; and only extends static seafloor topography data in the time dimension.

[0108] A4.2, while keeping the input time window length L unchanged, perform a sliding update on the previous round of input window: discard the earliest frame of wind speed in the last input wind speed data, then extract the input data of the next L-1 hours and append the latest frame of wind speed; stack the latest data along the time channel to form a new input sample Xi+1 according to step A2.

[0109] A4.3 Input Xi+1 into the trained model and output the effective wave height prediction for the next target time.

[0110] As another embodiment of the present invention, a data-driven large-scale regional rapid wave forecasting system is also provided to implement the data-driven large-scale regional rapid wave forecasting method described above. The system includes a data acquisition and preprocessing module, a deep model inference module, and a rolling update and output module. The data acquisition and preprocessing module is used to acquire data and construct tensors. The deep model inference module is used to infer the trained model using new data. The rolling update and output module is used to maintain a time window queue of fixed length. Whenever the latest real-time wind field data is received, the old data at the head of the queue is automatically removed, the latest data is added, and the deep model module is triggered to perform inference, outputting the forecast result for the next moment.

[0111] In another embodiment, to ensure the representativeness of the training data, wind and wave data from the Northwest Pacific region were used as the training set. This data was derived from the European Centre for Medium-Range Weather Forecasts (ECMWF) Generation 5 Atmospheric Reanalysis Dataset (ERA5). The wind and wave data underwent preprocessing, including spatiotemporal interpolation, normalization, and landmass labeling. Spatiotemporal interpolation was performed on the wind and wave field, with a spatial range of 0°N to 80°N and 100°E to 180°E, a spatial resolution of 0.25° × 0.25°, a temporal resolution of 1 hour, and a grid size of 320 × 320. After spatial alignment of the data, historical wind speed and historical significant wave height data were misaligned in the temporal dimension to form a C×T×H×W structure, where C, T, H, and W represent the number of channels, the temporal dimension, height, and width, respectively. The seafloor topography data was copied and expanded, and added as additional feature dimension information to the multi-channel tensor X, resulting in input data X fused with static topography information. o .

[0112] After model construction, to improve prediction performance and generalization ability, the key hyperparameters need to be systematically tuned using the controlled variable method. During the tuning process, other variables are fixed while only one hyperparameter is changed for iterative training, thereby evaluating its specific impact on model performance. Specifically, to comprehensively consider the effects of input wind speed and historical wave height sequence length on model accuracy, different input time windows need to be set in the experiment to compare the model's fitting and generalization abilities at different time scales, and to select the optimal time length that minimizes prediction error. During model training, the mean absolute error is used as the basic loss function, and an adaptive weighted loss function for high sea states is defined, with a high sea state threshold of 4m and a penalty strength coefficient of 0.5. During training, the model applies additional gradient penalties to wave height samples exceeding 4m, forcing the network to preferentially fit extreme value regions. In addition, to improve training stability and optimize convergence speed, a cosine annealing learning rate strategy is introduced, which makes the learning rate decrease periodically with the training rounds. The total number of training rounds is set to 2000. In the early stage, a large learning step size is maintained to accelerate the convergence speed, and in the later stage, it is gradually reduced to refine the model weights, avoid oscillation and overfitting, thereby achieving further convergence of the overall error and performance improvement.

[0113] To comprehensively evaluate the data for the entire study area and the performance of subsequent models, the following spatial statistical indicators—MAE, RMSE, MAPE, and R²—were constructed. During data validation, these indicators were used only for point-by-point error calculation. Represents the true effective wave height on the coordinate system. This represents the predicted value output by the prediction model on the coordinate system. denoted by , where m and n represent the mean of the true effective wave height values ​​on the coordinates, m and n represent the dimensions of the range data, and m×n represents the number of wave height data points. During model training and prediction, the smaller the values ​​of MAE, RMSE, and MAPE, the lower the prediction error and the higher the accuracy of the model; while the larger the R2 value, the better the model's fit and the better its performance.

[0114]

[0115]

[0116]

[0117]

[0118] The experimental data were processed according to the method of this invention, and the experimental results were analyzed as follows:

[0119] like Figure 12As shown, the model's loss value drops rapidly and fluctuates significantly before 1000 iterations, but converges around 1500 iterations, with the final loss value converging to around 0.07m, demonstrating the effectiveness of the model's learning during training.

[0120] In the 1-hour forecast, the mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) remained at low levels, at 0.098m, 0.189m, and 7.90%, respectively. The correlation coefficient R² also remained within 0.982, indicating that the method provided by this invention can accurately predict future wave field changes.

[0121] like Figure 13 As shown, in order to verify the spatial effectiveness of the model, we extracted and analyzed data from three different water depth locations, P1-P3, which gradually spread from nearshore to offshore. As can be seen from the figure, the model exhibits excellent performance at different spatial locations, capturing the temporal and spatial coupling relationship well overall. Even when the wave height is greater than 5m, the error remains small, with a maximum error of approximately 0.35m.

[0122] like Figure 14 As shown in the figure, in order to verify the predictive performance of this model under extreme conditions, the prediction results when the typhoon passed through the study area were specifically selected for comparison. It can be seen from the figure that this model can well characterize the nonlinear characteristics between wind and waves, and also characterizes the structural details of the wind field relatively accurately. Even without the propagation of physical mechanisms, it still captures the propagation laws of wind, waves and swells very well.

[0123] In practical application, this embodiment deploys the trained wave height prediction model to a high-performance computing server. The system receives real-time wind speed field forecast data for the next 24 to 72 hours from numerical weather prediction models, using this as the latest wind speed field data, and inputs this latest wind speed field data into the model for forward inference. Because the model employs a fully convolutional architecture, the inference process involves only efficient matrix multiplication operations, eliminating the need for iterative solutions to differential equations. Therefore, calculations can be completed in milliseconds, directly outputting the predicted effective wave height field for the target future time. This output not only includes the wave height value for each grid point but also visually demonstrates the wave system's movement path and intensity distribution in the form of a field.

[0124] The technical solution in this embodiment introduces the physical hysteresis characteristics of wave excitation to perform misalignment alignment on the data, effectively capturing the time delay mechanism of wind energy to wave energy transfer, and improving the model's physical representation ability and forecast accuracy of the wave generation process. Based on a deep learning model to replace the traditional numerical solution process, the computational complexity is significantly reduced while ensuring the accuracy of large-scale spatiotemporal feature extraction, enabling rapid extrapolation of the effective wave height prediction field.

[0125] Preferably, such as Figure 2 As shown, in S1, the multi-source marine physical environment data undergoes a standardization preprocessing process that includes physical constraints, including:

[0126] S11, Perform data quality control on the historical wind speed vector sequence, historical wave height sequence and static seabed topography data. The data quality control includes at least numerical filling of missing time nodes using cubic spline interpolation function and removing abnormal values ​​based on statistical distribution patterns.

[0127] S12, construct a unified latitude and longitude reference grid, map the quality-controlled historical wind speed vector sequence to the latitude and longitude reference grid, and fit the spatial continuous distribution characteristics of the wind speed field through cubic spline interpolation to ensure that the physical quantities stored in the grid nodes have smooth gradient characteristics.

[0128] S13, introduce a binarized mask matrix representing the land-sea distribution, perform logical operations on the binarized mask matrix and the physical data on the reference grid, remove invalid grid points in the land area, and retain only the physical values ​​of the valid sea area, thereby generating the standard physical field data stream.

[0129] For step S11, the system acquires multi-source raw data of the target sea area from preset data interfaces. These preset data interfaces include the European Centre for Medium-Range Weather Forecasts (ERA5) database or the NOAA Buoy Data Center, covering historical wind speed vector sequences, historical significant wave height sequences, and high-precision static seabed topography data. The historical wind speed vector sequence includes a meridional component *u* and a zonal component *v*. Since raw observation data is often limited by sensor malfunctions or transmission packet loss, it exhibits temporal discontinuities. This embodiment uses a cubic spline interpolation function to numerically fill in the missing time points. Compared to linear interpolation, cubic spline interpolation ensures the continuity of the first and second derivatives at the filled points, thereby preserving the nonlinear trend and smoothness of the wind and wave data evolution over time to the greatest extent. Simultaneously, the system is based on... The statistical distribution law of the criteria is used to perform outlier detection on the data, automatically identify and remove physical values ​​that deviate significantly from the normal distribution range, and prevent noise data caused by sensor drift from polluting the subsequent model training process.

[0130] To address the issue of inconsistent spatial resolution among multi-source data—for example, wind field data might have a resolution of 0.25°, while wave data might be based on station observations—this embodiment executes step S12 to construct a unified latitude and longitude reference grid. This grid covers the target sea area and has a fixed spatial resolution, such as 0.1°. At 0.1°, the system maps the quality-controlled historical wind speed vector sequence onto the reference grid. During this process, the spatial continuity distribution characteristics of the wind speed field are again fitted using a cubic spline interpolation algorithm. Since the wind field, as a fluid, possesses spatial continuity and smoothness, cubic spline interpolation can effectively eliminate jagged artifacts that may occur during grid resampling, ensuring that the physical quantities stored in the grid nodes have smooth gradient characteristics.

[0131] For step S13, a binarized mask matrix representing the land-sea distribution is introduced. This binarized mask matrix has the same size as the reference grid, with sea area grid points assigned a value of 1 and land area grid points assigned a value of 0. The system performs element-wise logical multiplication of this binarized mask matrix with the wind speed, wave height, and topographic data on the reference grid, precisely forcing invalid grid point values ​​in the land area to zero, thus eliminating land data. This not only reduces the computational overhead of the model in processing invalid information, but more importantly, it imposes explicit geophysical constraints on the model, forcing it to focus only on the evolution of physical processes within the valid sea area, generating a standard physical field data stream with uniform spatiotemporal resolution, smooth numerical distribution, and conformity to the land-sea boundary physical constraints.

[0132] Preferably, such as Figure 3 As shown, in step S2, the standard physical field data stream is subjected to a misalignment alignment operation in the time dimension, and combined with feature stacking in the channel dimension and sliding window sampling in the time dimension, a multi-channel spatiotemporal input tensor containing wind field driving information, terrain boundary information, and historical wave state information is constructed, including:

[0133] S21, Align the historical wind speed vector sequence and the historical wave height sequence in the standard physical field data stream in the spatial coordinate system, and then shift the historical wind speed vector sequence forward by n time steps relative to the historical wave height sequence in the time axis direction to match the time lag process of wind energy being transferred to the sea surface to form waves.

[0134] S22, Perform a copy broadcast operation on the static seabed topography data in the time dimension to expand it into a topography feature sequence with the same time length as the historical wind speed vector sequence, thereby giving the static topography data a dynamic time dimension attribute.

[0135] S23, the translated historical wind speed vector sequence, historical wave height sequence, and terrain feature sequence are deeply stacked along the channel dimension to generate a dimension of... The four-dimensional tensor is used to generate the multi-channel spatiotemporal input tensor by extracting a fixed-length time segment from the four-dimensional tensor through a sliding window.

[0136] After completing the data standardization preprocessing, this embodiment proceeds to step S2, which involves constructing a multi-channel spatiotemporal input tensor containing wind field driving information, terrain boundary information, and historical wave state information. The core of this process lies in explicitly encoding the physical mechanism of wind-induced waves through data-level operations. Specifically, referring to step S21, the system ensures that the historical wind speed vector sequence and the historical wave height sequence are strictly aligned in a unified spatial coordinate system, where the historical wind speed vector sequence contains u and v components. Addressing the inherent time lag characteristic of wind energy transfer to the sea surface and the resulting wave growth, this embodiment creatively performs a time-axis shift operation. The system sets a lag time step parameter n, which can be empirically set based on the average water depth and wind zone scale of the target sea area, for example, n = 2 hours. During operation, the historical wind speed vector sequence is shifted forward by n time steps relative to the wave sequence on the time axis. The constructed data structure forcibly establishes a correspondence between the wind field data at time t and the wave data at time t+n. Misalignment allows subsequent deep learning models to directly capture the causal delay between driving force and response at the input, thereby reducing the difficulty for models to learn complex physical mappings.

[0137] To address the dimensionality mismatch between static data and dynamic sequences, this embodiment executes step S22. Static seabed topography data is physically an unchanging two-dimensional matrix that cannot be directly manipulated with temporal wind and wave data using tensors. Therefore, the system performs a copy-broadcast operation on this two-dimensional topography data along the time dimension. Specifically, the same topography matrix is ​​copied along the time axis, expanding it to have the exact same time length as the historical wind speed vector sequence, for example, T... total The three-dimensional terrain feature sequence is obtained. The above operation gives the static terrain data a pseudo-dynamic time dimension attribute, ensuring that the model can perceive the water depth and seabed friction dissipation boundary conditions at the current grid point on any time slice.

[0138] In step S23, the system performs depth stacking of the processed data components along the channel dimension. The translated historical wind speed vector sequence (occupying 2 channels), the historical wave height sequence (occupying 1 channel), and the expanded terrain feature sequence (1 channel) are concatenated to generate a dimension... (or set according to the framework convention) The system generates a four-dimensional total tensor, where C represents the total number of feature channels, T is the total time span, and H and W are the spatial grid sizes. A sliding window technique is used to extract fixed-length time segments from this four-dimensional tensor; for example, each segment contains data from the past 24 hours. This generates a series of multi-channel spatiotemporal input tensors. Each input tensor is not merely a simple data matrix, but a physically comprehensive package encapsulating the dynamic causes (wind), boundary constraints (topography), and its own evolutionary inertia (historical waves), serving as the standard input for subsequent deep learning models.

[0139] Preferably, such as Figure 4 As shown, in S3, a deep learning prediction model is constructed, comprising a spatial feature encoding module, a spatiotemporal feature perception module, and a decoder module, including:

[0140] S31, receive the multi-channel spatiotemporal input tensor, and use multiple sets of parallel depth-separable convolutional layers with different kernel sizes to extract local spatial texture features and global background distribution features of wind field, wave field and terrain at different physical scales, respectively. The depth-separable convolutional layers maintain the attribute isolation of each physical variable based on channel independence.

[0141] S32 performs stride convolution on the feature maps extracted by each parallel branch to perform spatial downsampling, in order to filter out high-frequency noise interference and compress the spatial resolution of the feature maps;

[0142] S33 concatenates and fuses the downsampled branch features along the channel dimension, and overlays a learnable geolocation encoding matrix to explicitly inject absolute location semantic information of the latitude and longitude grid, outputting initial embedding features with geolocation awareness.

[0143] After constructing the multi-channel spatiotemporal input tensor, this embodiment proceeds to step S3, which involves building the specific architecture of the deep learning prediction model. Regarding the spatial feature encoding module described in step S31, this module serves as the model's perceptual front end, responsible for extracting effective feature representations from the raw physical data. Specifically, the system receives data from step S2 with a dimension of... The model employs a multi-channel spatiotemporal input tensor. To fully capture the characteristics of wind field, wave field, and topographic data at different physical scales, this embodiment does not use a traditional single-size convolutional kernel, but instead designs multiple sets of parallel convolutional branches. Each branch uses a depthwise separable convolutional layer, which can group the input data based on channel independence. The model can independently extract the hydrodynamic features of the wind field, the wave characteristics of the ocean waves, and the geometric features of the seabed topography, avoiding premature mixing and interference of different physical variable attributes in the early stages of feature extraction. Simultaneously, each parallel branch is configured with convolutional kernels of different sizes (e.g., set to...). , and Small-sized convolutional kernels focus on capturing local spatial texture features, such as wind shear lines and abrupt changes in local topography; while large-sized convolutional kernels focus on extracting global background distribution features, such as large-scale cyclone structures, thereby achieving multi-scale comprehensive perception of the ocean physical field.

[0144] For step S32, the system performs spatial downsampling on the feature maps extracted from each parallel branch. In this embodiment, strided convolution with a stride of 2 is preferred over traditional max pooling. Pooling operations easily lose key positional information in high-frequency features, while strided convolution, as a learnable downsampling method, can retain more valuable structural information while compressing the spatial resolution of the feature map. Striked convolution can not only effectively filter out high-frequency random noise interference commonly found in the original observation data, but also compress the length and width of the feature map to half of the original input, for example, from... Dimensional reduction Spatial dimension compression significantly reduces the computational load and memory usage of subsequent deeper layers of the network, enabling the model to handle data with a larger time span, while also expanding the receptive field of subsequent network layers.

[0145] In step S33, the system performs feature fusion and geographic semantic injection. After multi-scale extraction and downsampling, the branch features are concatenated along the channel dimension to form a unified feature vector that integrates information from multiple physical variables. Considering that conventional convolutional neural networks have translation invariance (meaning the network cannot automatically perceive the specific geographical location of data points), and that wave generation is influenced by Coriolis force parameters and shoreline boundary effects, absolute location information is crucial. This embodiment innovatively superimposes a learnable geographic location encoding matrix. This matrix has the same spatial dimensions as the feature map, and by adding elements one by one, it explicitly injects the absolute location semantic information of the latitude and longitude grid into the feature map. The output initial embedded features not only contain meteorological and hydrological dynamic information but also implicitly indicate whether the data is located in nearshore shallow waters or deep sea.

[0146] Preferably, such as Figure 5 As shown, in S3, the specific execution steps of the spatiotemporal feature perception module include:

[0147] S34, construct a dual-stream parallel processing architecture to process the initial embedded features, wherein the first branch extracts shallow texture features of wind waves directly driven by the local wind field through stacked residual convolutional layers, and the second branch extracts deep semantic features of swell waves with cross-regional long-distance propagation characteristics through continuous downsampling convolutional layers.

[0148] S35, the shallow texture features of the wind and waves are mapped into a query matrix through a linear projection transformation layer, and the deep semantic features of the surging waves are mapped into a key matrix and a value matrix;

[0149] S36, introduce a spatiotemporal propagation bias matrix that characterizes the physical laws of wave propagation, and incorporate the spatiotemporal propagation bias matrix into the attention weight calculation based on the query matrix and the key matrix. Output global spatiotemporal features that integrate local wind and wave information and long-distance swell information through an interactive retrieval mechanism.

[0150] Following the initial feature extraction described above, this embodiment further executes step S34 to construct a dual-stream parallel processing architecture driven by physical mechanisms. This architecture can decouple the feature components of wind waves and swells from complex mixed wave fields. Considering that the physical composition of ocean waves mainly includes high-frequency wind waves directly driven by local winds and low-frequency swells transmitted from distant storms, a single network stream often struggles to accommodate these two drastically different physical processes. Therefore, this architecture designs two independent processing branches: the first branch focuses on extracting wind wave features and employs stacked residual convolutional layers. The design of residual connections allows the network to retain the original high-frequency texture information while enhancing its nonlinear expressive power by deepening the network layers, thereby outputting shallow texture features of wind waves that primarily reflect the driving effect of local wind fields. Simultaneously, the second branch focuses on extracting swell features and employs continuous downsampling convolutional layers. This branch expands the receptive field of neurons by progressively reducing the spatial resolution of the feature maps, enabling them to capture long-distance spatial associations spanning hundreds of kilometers, thereby outputting deep semantic features of swells that contain cross-regional propagation characteristics.

[0151] To achieve effective interaction and fusion of wind and wave information and swell information, this embodiment executes step S35, introducing a feature mapping operation based on an attention mechanism. The system uses a learnable linear projection transformation layer to map the shallow texture features of wind and waves output from the first branch into a query matrix (Q), representing the active driving factors at the current local level; and maps the deep semantic features of swell output from the second branch into a key matrix (K) and a value matrix (V), representing the wide-area background environment field. The physical meaning of this mapping strategy is that the model attempts to use the current local wind and wave state as a query condition to retrieve the swell components most relevant to the current sea state from the broad swell background field, thereby simulating the nonlinear superposition effect of wind and waves and swell in a specific sea area.

[0152] In step S36, addressing the lack of physical space constraints in traditional attention mechanisms during computation, this embodiment innovatively introduces a spatiotemporal propagation bias matrix characterizing the physical laws of wave propagation. This matrix, constructed based on wave group velocity theory, serves as a priori mask in the calculation of attention weights; that is, the bias term is superimposed on the dot product of Q and K. Its function is to suppress the weights of distant grid points that are physically impossible to reach in the current time step, strengthening the association of grid points conforming to the wave propagation path. Through this physically constrained interactive retrieval mechanism, the model deeply fuses the weighted swell information V with local wind and wave information, outputting global spatiotemporal features that combine local details with global background. These global spatiotemporal features encode the dynamic evolution of the wave field, providing a high-confidence potential feature representation for the decoder to accurately reconstruct the wave height field.

[0153] Preferably, such as Figure 6 As shown, it also includes:

[0154] S5, iteratively train the deep learning prediction model using an adaptive weighted loss function. The calculation steps of the adaptive weighted loss function include:

[0155] S51, obtain the true wave height value distribution at the current time step, set a high sea state judgment threshold, and identify high sea state sample points belonging to extreme sea states based on the comparison result between the true wave height value and the high sea state judgment threshold.

[0156] S52, construct a nonlinear penalty weight function for the high sea state sample points, and generate a sample weight mask matrix corresponding to the current training batch based on the nonlinear penalty weight function, so that the weight coefficients corresponding to the high sea state sample points are significantly higher than those of ordinary sea state sample points.

[0157] S53, perform element-wise multiplication of the sample weight mask matrix and the prediction error matrix to calculate the weighted total loss value, and use minimizing the weighted total loss value as the optimization objective of gradient descent to drive model parameter updates to preferentially fit the wave height characteristics of high sea state areas.

[0158] After completing the network architecture construction of the deep learning forecasting model, given the long-tail effect of ocean wave data distribution (i.e., the sea surface is in a low to medium wave state most of the time, with wave height less than 2 meters), and the extremely sparse spatiotemporal distribution of typhoon waves or giant waves that pose the greatest threat to shipping safety (with wave height greater than 4 meters), if the traditional mean squared error loss function is directly used, the model tends to fit the average sea state that occurs most frequently, resulting in a smoothing effect in the prediction of extreme values ​​and a serious underestimation of hazardous wave heights. Therefore, in step S51 of this embodiment, a sample selection mechanism based on a physical threshold is established. In each iteration of training, the system obtains the true wave height numerical distribution of the current time step and sets a high sea state judgment threshold, for example, 4.0 meters according to the International Meteorological Organization standard, or based on the 95th percentile value of the statistical characteristics of the training set. The system compares the value of each grid point in the true wave height field with this threshold point by point, marking all grid points with values ​​exceeding the threshold as high sea state sample points, and the rest as ordinary sea state sample points.

[0159] Regarding step S52, to give a larger gradient contribution to high sea state regions during the optimization process, this embodiment constructs a nonlinear penalty weighting function for high sea state sample points. Unlike simple fixed weighting, this function is designed to be a nonlinear form that monotonically increases with wave height, such as an exponential function:

[0160]

[0161] in, H represents the intensity of the punishment. true H represents the true significant wave height value of the corresponding grid point in the current training sample. threshold This represents the preset high sea state threshold. Based on this exponential function, the system generates a sample weight mask matrix that is exactly the same size as the output of the current training batch. In this sample weight mask matrix, the weight coefficients corresponding to ordinary sea state sample points remain at the base value. However, the weight coefficients corresponding to high sea state sample points are dynamically amplified to several times or even tens of times the base value, realizing spatially differentiated allocation of weights. Furthermore, through a non-linear growth mechanism, the higher the wave height, the greater the attention given to extreme areas, thus mathematically forcing the model to penalize prediction errors generated in the eye of the typhoon or the center of the storm.

[0162] In the parameter update step S53, the system performs element-wise multiplication of the generated sample weight mask matrix with the prediction error matrix output by the model. This weighting operation calculates the weighted total loss value. Because the error is amplified in high sea state regions, this total loss value becomes extremely sensitive to deviations from extreme values. The gradient of this weighted loss function relative to the model parameters is calculated using the backpropagation algorithm. Driven by the optimization objective of gradient descent, the update direction of the model parameters will be primarily dominated by the gradient of the high sea state sample points, forcing the network weights to be adjusted preferentially to fit the wave height characteristics of high sea state regions. After multiple rounds of iterative training, the obtained wave height prediction model can improve the peak value capture capability for extreme and hazardous sea states such as typhoon waves and cold waves while maintaining the accuracy of the overall background field, effectively solving the problem of traditional data-driven models being accurate on average but poor in extreme values.

[0163] In one embodiment, Figure 7 This is a block diagram illustrating a data-driven, large-scale regional rapid wave forecasting device according to an exemplary embodiment. Figure 7 As shown, the data-driven large-scale regional rapid wave forecasting device includes an acquisition module 71, a construction module 72, a training module 73, and an output module 74.

[0164] The acquisition module 71 is used to acquire multi-source marine physical environment data of the target sea area. The multi-source marine physical environment data includes historical wind speed vector sequences, historical wave height sequences, and static seabed topography data. The module performs standardized preprocessing with physical constraints on the multi-source marine physical environment data. The standardized preprocessing performs quality control on wind speed, wave height, and topography data, fills in time gaps and removes statistical outliers through spline interpolation, constructs a unified latitude and longitude reference grid, fits the wind field distribution through spatial interpolation to ensure gradient smoothness, introduces a land-sea binary mask matrix and performs logical operations with the physical data, removes invalid land grids, and generates a standard physical field data stream with unified spatiotemporal resolution.

[0165] Module 72 is used to perform misalignment alignment operation on the standard physical field data stream in the time dimension based on the physical hysteresis characteristics of wave excitation, and combine feature stacking in the channel dimension with sliding window sampling in the time dimension to construct a multi-channel spatiotemporal input tensor containing wind field driving information, terrain boundary information and historical wave state information.

[0166] Training module 73 constructs a deep learning prediction model that includes a spatial feature encoding module, a spatiotemporal feature perception module, and a decoder module. The deep learning prediction model is iteratively trained using the multi-channel spatiotemporal input tensor. The difference between the predicted wave height and the actual wave height is minimized through the backpropagation mechanism, thereby obtaining the trained wave height prediction model.

[0167] Output module 74 receives the latest wind speed field data output by real-time observation or numerical prediction model, performs forward inference calculation on the latest wind speed field data through the trained wave height prediction model, and outputs the effective wave height prediction field of the ocean waves at the future target time.

[0168] The data-driven large-scale regional rapid wave forecasting device block diagram includes an acquisition module 71, a construction module 72, a training module 73, and an output module 74, which are controlled to execute the data-driven large-scale regional rapid wave forecasting method described in any of the above embodiments.

[0169] like Figure 5 As shown, the present invention provides an electronic device 800, which includes: a communication interface, a processor 801, and a memory 802;

[0170] The memory 802 stores program instructions. When executed by the processor 801, which is connected to the memory 802 via the communication interface, the program instructions acquire multi-source marine physical environment data of the target sea area. This multi-source marine physical environment data includes historical wind speed vector sequences, historical wave height sequences, and static seabed topography data. The multi-source marine physical environment data undergoes standardized preprocessing with physical constraints. This standardized preprocessing performs quality control on wind speed, wave height, and topography data, fills in time gaps and removes statistical outliers through spline interpolation, constructs a unified latitude and longitude reference grid, fits the wind field distribution through spatial interpolation to ensure gradient smoothness, introduces a land-sea binary mask matrix for logical operations with the physical data, and removes invalid land grids to generate a standard physical field data stream with unified spatiotemporal resolution. The physical hysteresis characteristics of wave excitation are utilized to perform a misalignment alignment operation on the standard physical field data stream in the time dimension. Combined with feature stacking in the channel dimension and sliding window sampling in the time dimension, a multi-channel spatiotemporal input tensor containing wind field driving information, terrain boundary information, and historical wave state information is constructed. A deep learning prediction model, including a spatial feature encoding module, a spatiotemporal feature sensing module, and a decoder module, is constructed. The deep learning prediction model is iteratively trained using the multi-channel spatiotemporal input tensor, and the difference between the predicted wave height and the actual wave height is minimized through a backpropagation mechanism, thereby obtaining a trained wave height prediction model. The latest wind speed field data output from real-time observations or numerical weather prediction models is received, and forward inference calculations are performed on the latest wind speed field data using the trained wave height prediction model to output the effective wave height prediction field for the future target time.

[0171] Compared with the prior art, the advantages of this embodiment are:

[0172] By introducing the physical hysteresis characteristics of wave excitation and performing misalignment alignment on the data, the time delay mechanism of wind energy to wave energy transfer is effectively captured, improving the model's physical representation ability and prediction accuracy of the wave generation process. Replacing the traditional numerical solution process with a deep learning model significantly reduces computational complexity while ensuring the accuracy of large-scale spatiotemporal feature extraction, enabling rapid extrapolation of the effective wave height prediction field.

[0173] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0174] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A data-driven based fast wave prediction method for large-scale regions, characterized in that, The method comprises the following steps: S1, obtaining multi-source marine physical environment data of a target sea area, wherein the multi-source marine physical environment data comprises a historical wind speed vector sequence, a historical sea wave significant wave height sequence, and static seabed topography data, and performing standardized preprocessing with physical constraints on the multi-source marine physical environment data, wherein the standardized preprocessing performs quality control on wind speed, wave height, and topography data, fills in time gaps through spline interpolation, and removes statistical outliers, constructs a unified latitude and longitude grid, fits wind field distribution through spatial interpolation to ensure gradient smoothness, introduces a land-sea binary mask matrix for logical operation with physical data, and removes invalid land grids to generate a standard physical field data stream with unified spatiotemporal resolution; S2, based on the physical lag characteristics of wave excitation, performing a misalignment alignment operation on the standard physical field data stream in the time dimension, and combining feature stacking in the channel dimension with sliding window sampling in the time dimension to construct a multi-channel spatiotemporal input tensor containing wind field driving information, topographic boundary information, and historical wave state information; S3, constructing a deep learning prediction model comprising a spatial feature encoding module, a spatiotemporal feature perception module, and a decoder module, iteratively training the deep learning prediction model through the multi-channel spatiotemporal input tensor, minimizing the difference between the predicted wave height and the actual wave height through a backpropagation mechanism, and thereby obtaining a trained wave height prediction model; S4, receiving the latest wind speed field data observed in real time or output by a numerical prediction model, performing forward inference calculation on the latest wind speed field data through the trained wave height prediction model, and outputting a sea wave significant wave height prediction field at a future target time; In step S2, the misalignment alignment operation is performed on the standard physical field data stream in the time dimension, and the multi-channel spatiotemporal input tensor containing wind field driving information, topographic boundary information, and historical wave state information is constructed by combining feature stacking in the channel dimension with sliding window sampling in the time dimension, which comprises: S21, aligning the historical wind speed vector sequence and the historical sea wave significant wave height sequence in the standard physical field data stream in the spatial coordinate system, and then shifting the historical wind speed vector sequence forward relative to the historical sea wave significant wave height sequence by n time steps in the time axis direction to match the time lag process of wind field energy transmission to the sea surface to form waves; S22, performing a replication broadcast operation on the static seabed topography data in the time dimension to expand it into a topographic feature sequence with the same time length as the historical wind speed vector sequence, thereby giving the static topography data a dynamic time dimension attribute; S23, stack the translated historical wind speed vector sequence, the historical sea wave significant wave height sequence and the terrain feature sequence in the channel dimension to generate a four-dimensional tensor with the dimension of , wherein C is the number of input channels of the wind speed sequence, T is the time dimension of the wind speed sequence, H is the height of the input tensor, W is the width of the input tensor, and a fixed-length time segment is intercepted on the four-dimensional tensor through a sliding window to generate the multi-channel spatio-temporal input tensor; In step S3, the deep learning prediction model comprising a spatial feature encoding module, a spatiotemporal feature perception module, and a decoder module is constructed, which comprises: S31, receiving the multi-channel spatiotemporal input tensor, using multiple groups of parallel deep separable convolution layers with different convolution kernel sizes to extract local spatial texture features and global background distribution features of wind field, wave field, and topography at different physical scales, wherein the deep separable convolution layer maintains the attribute isolation of each physical variable based on channel independence; S32 performs stride convolution on the feature maps extracted by each parallel branch to perform spatial downsampling, in order to filter out high-frequency noise interference and compress the spatial resolution of the feature maps; S33 splices and fuses the downsampled branch features along the channel dimension and overlays a learnable geolocation encoding matrix to explicitly inject absolute location semantic information of the latitude and longitude grid, outputting initial embedding features with geolocation awareness. In step S3, the specific execution steps of the spatiotemporal feature perception module include: S34, construct a dual-stream parallel processing architecture to process the initial embedded features, wherein the first branch extracts shallow texture features of wind waves directly driven by the local wind field through stacked residual convolutional layers, and the second branch extracts deep semantic features of swell waves with cross-regional long-distance propagation characteristics through continuous downsampling convolutional layers. S35, the shallow texture features of the wind and waves are mapped into a query matrix through a linear projection transformation layer, and the deep semantic features of the surging waves are mapped into a key matrix and a value matrix; S36, introduce a spatiotemporal propagation bias matrix that characterizes the physical laws of wave propagation, and incorporate the spatiotemporal propagation bias matrix into the attention weight calculation based on the query matrix and the key matrix. Output global spatiotemporal features that integrate local wind and wave information and long-distance swell information through an interactive retrieval mechanism.

2. The data-driven based fast wave prediction method for large-scale region according to claim 1, wherein, In S1, the multi-source marine physical environment data undergoes a standardization preprocessing process that includes physical constraints, including: S11, Perform data quality control on the historical wind speed vector sequence, historical wave height sequence and static seabed topography data. The data quality control includes at least numerical filling of missing time nodes using cubic spline interpolation function and removing abnormal values ​​based on statistical distribution patterns. S12, construct a unified latitude and longitude reference grid, map the quality-controlled historical wind speed vector sequence to the latitude and longitude reference grid, and fit the spatial continuous distribution characteristics of the wind speed field through cubic spline interpolation to ensure that the physical quantities stored in the grid nodes have smooth gradient characteristics. S13, introduce a binarized mask matrix representing the land-sea distribution, perform logical operations on the binarized mask matrix and the physical data on the reference grid, remove invalid grid points in the land area, and retain only the physical values ​​of the valid sea area, thereby generating the standard physical field data stream.

3. The data-driven based fast wave prediction method for large-scale region according to claim 1, wherein, Also includes: S5, iteratively train the deep learning prediction model using an adaptive weighted loss function. The calculation steps of the adaptive weighted loss function include: S51, obtain the true wave height value distribution at the current time step, set a high sea state judgment threshold, and identify high sea state sample points belonging to extreme sea states based on the comparison result between the true wave height value and the high sea state judgment threshold. S52, construct a nonlinear penalty weight function for the high sea state sample points, and generate a sample weight mask matrix corresponding to the current training batch based on the nonlinear penalty weight function, so that the weight coefficients corresponding to the high sea state sample points are significantly higher than those of ordinary sea state sample points. S53, element-wise multiply the sample weight mask matrix with the prediction error matrix, calculate the weighted total loss value, and minimize the weighted total loss value as the optimization objective of gradient descent, drive model parameter update to preferentially fit the wave height characteristics of high sea state area.

4. A data-driven based fast forecasting device for ocean waves in large scale areas, characterized by, Comprise: An acquisition module is configured to acquire multi-source marine physical environment data of a target sea area, the multi-source marine physical environment data comprising a historical wind speed vector sequence, a historical sea wave significant wave height sequence, and static seabed topography data, and perform standardized preprocessing comprising physical constraints on the multi-source marine physical environment data, the standardized preprocessing performing quality control on wind speed, wave height, and topography data, filling in time gaps through spline interpolation and removing statistical outliers, constructing a unified latitude and longitude grid, fitting wind field distribution through spatial interpolation to ensure gradient smoothing, introducing a land-sea binary mask matrix for logical operation with physical data, and removing invalid land grids to generate a standardized physical field data stream with unified spatiotemporal resolution; A construction module is configured to perform a misalignment operation on the standardized physical field data stream in the time dimension based on the physical lag characteristics of wave excitation, and combine feature stacking in the channel dimension and sliding window sampling in the time dimension to construct a multi-channel spatiotemporal input tensor comprising wind field driving information, topographic boundary information, and historical wave state information, comprising: Aligning the historical wind speed vector sequence and the historical sea wave significant wave height sequence in the spatial coordinate system in the standardized physical field data stream, and then shifting the historical wind speed vector sequence forward by n time steps relative to the historical sea wave significant wave height sequence in the time axis direction to match the time lag process of wind field energy transmission to the sea surface to form waves; Performing a replication broadcast operation on the static seabed topography data in the time dimension to extend it to a topographic feature sequence with the same time length as the historical wind speed vector sequence, thereby giving the static topography data a dynamic time dimension attribute; The translated historical wind speed vector sequence, the historical sea wave significant wave height sequence and the terrain feature sequence after translation are stacked in the channel dimension to generate a four-dimensional tensor with a dimension of , wherein C is the number of input channels of the wind speed sequence, T is the time dimension of the wind speed sequence, H is the height of the input tensor, W is the width of the input tensor, and a fixed-length time segment is intercepted on the four-dimensional tensor through a sliding window to generate the multi-channel spatiotemporal input tensor; A training module is configured to construct a deep learning prediction model comprising a spatial feature encoding module, a spatiotemporal feature perception module, and a decoder module, iteratively train the deep learning prediction model using the multi-channel spatiotemporal input tensor, minimize the difference between the predicted wave height and the true wave height through a backpropagation mechanism, and thereby obtain a trained wave height prediction model. Wherein, constructing a deep learning prediction model comprising a spatial feature encoding module, a spatiotemporal feature perception module, and a decoder module comprises: Receiving the multi-channel spatiotemporal input tensor, using multiple groups of parallel deep separable convolution layers with different convolution kernel sizes to extract local spatial texture features and global background distribution features of wind field, wave field, and topography in different physical scales, respectively, wherein the deep separable convolution layer maintains the attribute isolation of each physical variable based on channel independence; Performing step convolution operation on the feature maps extracted by each parallel branch to perform spatial down-sampling to filter out high-frequency noise interference and compress the spatial resolution of the feature maps; The down-sampled branch features are spliced and fused in the channel dimension, and a learnable geographic position coding matrix is superimposed to explicitly inject the absolute position semantic information of the latitude and longitude grid, and an initial embedding feature with geographic position perception is output; The spatio-temporal feature perception module specifically includes: A double-flow parallel processing architecture is constructed to process the initial embedding feature, wherein a first branch extracts wind wave shallow texture features directly driven by local wind fields through stacked residual convolution layers, and a second branch extracts swell deep semantic features with cross-regional long-distance propagation characteristics through consecutive down-sampling convolution layers; The wind wave shallow texture features are mapped into a query matrix through a linear projection transformation layer, and the swell deep semantic features are mapped into a key matrix and a value matrix; A spatio-temporal propagation bias matrix representing the physical law of sea wave propagation is introduced, and the spatio-temporal propagation bias matrix is involved in the attention weight calculation based on the query matrix and the key matrix, and an interactive retrieval mechanism is used to output global spatio-temporal features fused with local wind wave information and long-distance swell information. An output module is configured to receive latest wind speed field data output by real-time observation or numerical prediction mode, perform forward inference calculation on the latest wind speed field data through the trained wave height prediction model, and output sea wave significant wave height prediction field at a target future time.

5. The data-driven based fast forecasting apparatus for ocean waves in large-scale areas according to claim 4, wherein: The acquisition module, the construction module, the training module, and the output module are controlled to perform the data-driven large-scale regional sea wave rapid prediction method according to any one of claims 1 to 3.

6. An electronic device, comprising: The method comprises the following steps: A communication interface, a processor, and a memory are included. The memory is configured to store program instructions, and the program instructions, when executed by the processor in communication with the memory through the communication interface, cause the electronic device to implement the data-driven large-scale regional sea wave rapid prediction method according to any one of claims 1 to 3.

7. A computer-readable storage medium having stored thereon program instructions, wherein, The program instructions, when executed by the computer, cause the computer to implement the data-driven large-scale regional sea wave rapid prediction method according to any one of claims 1 to 3.

Images