Season-trend decomposition based significant wave height prediction method and system

Abstract

The application relates to the field of marine environment prediction, and discloses an effective wave height prediction method and system based on seasonal-trend decomposition, which comprises the following steps: obtaining an original wave height sequence and a zonal and meridional wind speed component, fusing channels after constructing multi-scale wave height features and wind field physical features, and outputting multi-scale wind wave coupling features through deep separable convolution; constructing multi-resolution time images for the multi-scale wind wave coupling features, respectively extracting seasonal structure features and trend structure features through a double-axis attention mechanism; outputting multi-resolution features through multi-scale fusion mechanism and main frequency attention weighted fusion operation of the seasonal structure features and the trend structure features; fusing the wind field physical features through construction of a two-dimensional periodic graph and multi-scale attention enhancement, and outputting an effective wave height prediction value through wind-induced gating operation and residual fusion of the multi-resolution features. The application realizes accurate and stable prediction of the effective wave height, and improves the model prediction robustness.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ocean prediction, and in particular to an effective wave height prediction method and system based on seasonal-trend decomposition. BACKGROUND

[0002] The effective wave height is a core parameter for describing the intensity of ocean wave motion, and its accurate prediction is of great significance for ensuring the safety of marine operations and optimizing the development of marine resources. Existing effective wave height prediction methods based on deep learning rely on models such as long short-term memory network (LSTM) and gated recurrent unit (GRU) to learn the time series rules of historical wave height sequences. However, there are significant drawbacks: the original wave height sequence has strong non-stationarity and high-frequency noise, and direct modeling can interfere with the network learning process; and the overall modeling of the wave height sequence cannot separate short-period disturbances (such as local wind waves) and long-period trends (such as seasonal changes), and the mixing of multi-scale features limits the prediction accuracy.

[0003] To solve the above problems, the existing "Ensemble Empirical Mode Decomposition (EEMD) + Long Short-Term Memory Network (LSTM)" combined model decomposes the original wave height sequence into multiple intrinsic mode functions (IMFs) corresponding to different frequency components and a residual term reflecting the overall baseline through EEMD, and then models each IMF component independently through a parallel LSTM network. Finally, the prediction results of each LSTM are added point by point to reconstruct the effective wave height prediction value. Although the above method can separate high and low frequency features and improve short-term prediction accuracy, it still has three technical bottlenecks that cannot be ignored: Lack of adaptive regulation and coordination mechanism for high and low frequency features: EEMD uses a fixed noise adding and decomposition strategy, which cannot dynamically adjust the decomposition scale and frequency division according to the sea conditions (such as calm sea conditions and extreme storms); LSTM only models each IMF independently, making it difficult to capture the nonlinear interaction between local high-frequency disturbances (such as short-term wind waves) and global low-frequency trends (such as seasonal changes), and "detail errors" or "trend deviations" may occur under extreme sea conditions.

[0004] Lack of explicit modeling of external physical driving factors such as wind field: wind is the core driving force for wave generation and evolution, and wind speed and direction directly determine the growth rate, energy accumulation, and propagation pattern of wave height. The existing EEMD-LSTM model only relies on historical wave height data for modeling, ignoring wind field information, which makes it difficult for the model to reflect the wind-wave coupling mechanism. For example, when wind speed suddenly increases, the model cannot perceive the external driving signal, which may lead to prediction lag or underestimation of wave peaks, and the model cannot distinguish the generation mechanism of the same wave height under different wind field backgrounds, resulting in poor physical consistency and generalization ability.

[0005] The feature fusion mode is simple and lacks deep interaction: a static linear fusion strategy of direct addition is adopted, only physical stacking of features is realized, and deep space-time dependence of high and low frequency features (such as nonlinear modulation of low frequency trend by high frequency fluctuation in strong storm) cannot be captured, resulting in local information blur, discontinuous edge transition, and insufficient global consistency of the prediction result.

[0006] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present application, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0007] To solve or at least alleviate one or more of the above problems, an effective wave height prediction method and system based on seasonal-trend decomposition are provided, multi-scale wind field coupling coding is used to extract multi-scale wind wave coupling features, a dual-axis attention mechanism is used to extract seasonal and trend components and perform multi-resolution fusion, and finally a wind-induced gating residual output effective wave height prediction value is introduced to realize accurate and stable prediction of the effective wave height and improve the prediction robustness of the model under extreme wind speed conditions.

[0008] To achieve the above-mentioned purpose, according to the first aspect of the present application, an effective wave height prediction method based on seasonal-trend decomposition is provided.

[0009] S1, multi-scale wind field coupling coding: The original wave height sequence is obtained to construct multi-scale wave height features, the zonal and meridional wind speed components are obtained to construct wind field physical features and embedded as high-dimensional wind field embedded features, the wind field embedded features are fused with the wave height features of the corresponding scale at the channel level, and the multi-scale wind wave coupling feature sequence is output through depth separable convolution; S2, seasonal-trend structure modeling: The multi-scale wind wave coupling feature sequence is constructed into a multi-resolution time image, and the seasonal structure features and the trend structure features are extracted through a dual-axis attention mechanism; the multi-scale fusion mechanism is used to obtain the main frequency feature map from the seasonal structure features and the trend structure features, and the multi-resolution features are output through the main frequency attention weighted fusion operation on the main frequency feature map; S3, residual fusion prediction of gating mechanism: The wind field physical features are embedded to obtain wind field embedded features, and a two-dimensional periodic graph is constructed based on the wind field embedded features; the wind field comprehensive representation is obtained through a multi-scale attention enhanced fusion operation, and after the wind field comprehensive representation is converted into a wind-induced adjustment signal, the wind-induced gating operation is performed on the wind-induced adjustment signal and the multi-resolution features to output the effective wave height prediction value.

[0010] Furthermore, the original wave height sequence is obtained to construct multi-scale wave height features, and the zonal and meridional wind speed components are obtained to construct wind field physical features and embedded into high-dimensional features to obtain wind field embedding features. These wind field embedding features are then fused with the corresponding scale wave height features at the channel level, and a multi-scale wind-wave coupled feature sequence is output through a depthwise separable convolution operation; including: Multi-scale wave height features are constructed by downsampling the original input wave height sequence layer by layer. The zonal and meridional wind speed components are obtained, and four types of wind field physical features, namely wind speed, wind direction, wind stress and wind shear, are constructed through numerical calculation to form a wind field feature vector. The wind field feature vector is then embedded into a high dimension through linear mapping to obtain the wind field embedding features. Wind field embedding features and wave height features at various scales are concatenated along the channel dimension to obtain fused features corresponding to each scale; these fused features are then mapped to a high-dimensional space to obtain the [missing information - likely a specific feature or parameter]. High-dimensional fusion features at the layer scale are sequentially subjected to channel-wise convolution and pointwise convolution to output the first... The corresponding wind and wave coupling feature sequence at the layer scale.

[0011] Furthermore, multi-resolution temporal images are constructed from the multi-scale wind and wave coupled feature sequences, including: The wind-wave coupling feature sequences at various scales are represented in the frequency domain by Fast Fourier Transform, and the sequence with the largest amplitude is extracted. For each dominant frequency, the corresponding period length is calculated. For the wind-wave coupling feature sequence at each scale, zero-padding is performed at each dominant frequency so that the length of the padded sequence is an integer multiple of the corresponding period length, thus obtaining the padded wind-wave coupling feature sequence. Based on the period length corresponding to each dominant frequency and the number of periods contained in the padded sequence length, the padded wind-wave coupling feature sequence is reconstructed into a two-dimensional periodic image. The two-dimensional periodic images corresponding to different dominant frequencies are used as time image representations of different resolution layers, and the hidden dimension of the two-dimensional periodic image of each resolution layer is expanded to obtain a multi-resolution time image.

[0012] Furthermore, seasonal and trend-based structural features are extracted separately using a dual-axis attention mechanism, including: Convolution operations are performed along the column axis on the multi-resolution temporal images of each resolution layer to obtain the query matrix, key matrix, and value matrix of the column axis, and attention calculation is performed to obtain seasonal structural features; Convolution operations are performed along the row axis on the multi-resolution temporal images of each resolution layer to obtain the query matrix, key matrix, and value matrix of the row axis, and attention calculation is performed to obtain trend structure features.

[0013] Furthermore, a multi-scale fusion mechanism is used to obtain the dominant frequency feature map from the seasonal and trend structural features. The dominant frequency feature map is then processed through a dominant frequency attention-weighted fusion operation to output multi-resolution features, including: To extract the seasonal structural features, starting from the first resolution layer, we proceed upwards layer by layer, adding the seasonal structural features of the current resolution layer to the seasonal structural features of the previous resolution layer after convolution. In response to the trend-based structural characteristics, with the first Starting from the resolution layer, layer by layer downwards, the trend structure features of the current resolution layer and the trend structure features of the next resolution layer are added together after deconvolution to obtain the enhanced trend structure features of the current resolution layer. Seasonal and trend structural features at the same resolution layer are fused into a two-dimensional representation at the corresponding scale, and then reshaped into a one-dimensional dominant frequency feature map through dimensionality reshaping. Each one-dimensional main frequency feature map is processed by global average pooling and linear mapping to calculate the main frequency attention score, and then normalized to obtain the main frequency weight. The main frequency weight is then weighted and fused with the corresponding main frequency feature map to output multi-resolution features.

[0014] Furthermore, wind field physical features are embedded to obtain wind field embedded features, and a two-dimensional periodic map is constructed based on these features. A comprehensive wind field representation is obtained through multi-scale attention-enhanced fusion. After the comprehensive wind field representation is converted into a wind-induced conditioning signal, it is fused with multi-resolution feature residuals through wind-induced gating to output the effective wave height prediction value, including: At each scale, four types of wind field physical features—wind speed, wind direction, wind stress, and wind shear—are embedded into a unified high-dimensional feature space through linear mapping to obtain four types of wind field embedding features. Fast Fourier Transform is performed on each of the four types of wind field embedding features, and the top K wind field dominant frequencies with the largest amplitudes are selected to calculate the period length corresponding to each wind field dominant frequency. Zero-padding is then applied to each of the four types of wind field embedding features at each wind field dominant frequency, ensuring that the length of the padded wind field embedding feature sequence is an integer multiple of the corresponding period length. Based on the period length corresponding to each wind field dominant frequency and the number of periods contained in the padded wind field embedding features, a two-dimensional periodic image of the four types of wind field embedding features is reconstructed. Each two-dimensional periodic image is mapped to a query vector, a key vector, and a value vector through a learnable query matrix, a key matrix, and a value matrix. The dot product between the query vector and the key vector is calculated and scaled. After Softmax normalization, attention enhancement features at each wind field's dominant frequency are obtained. The attention enhancement features of the same type of wind field embedding features at different wind field dominant frequencies are weighted and fused to obtain fused representations of wind speed, wind direction, wind stress, and wind shear at each scale. Channel splicing is performed on the fused representations at each scale to obtain the comprehensive wind field representation at each scale. After the comprehensive wind field representation is converted to one dimension, the wind-induced regulation signals of the wind field at each scale are obtained through linear mapping. Based on the wind-induced regulation signals, a dynamic gating factor is constructed. The contribution intensity of the wind field is weighted and adjusted by the dynamic gating factor and then fused with the multi-resolution feature residual to output the effective wave height prediction value.

[0015] To achieve the above objectives, according to a second aspect of the present invention, an effective wave height prediction system based on seasonal-trend decomposition is provided, comprising: The multi-scale wind field coupling coding module is used for multi-scale wind field coupling coding: it obtains the original wave height sequence to construct multi-scale wave height features, obtains the zonal and meridional wind speed components to construct wind field physical features and embeds them into high-dimensional wind field embedding features, performs channel-level fusion with the corresponding scale wave height features, and outputs a multi-scale wind and wave coupling feature sequence through depthwise separable convolution. The seasonal-trend structure modeling module is used for seasonal-trend structure modeling: it constructs multi-resolution time images from multi-scale wind and wave coupled feature sequences, and extracts seasonal and trend structure features respectively through a dual-axis attention mechanism; it uses a multi-scale fusion mechanism to obtain the dominant frequency feature map from the seasonal and trend structure features, and outputs multi-resolution features through a dominant frequency attention weighted fusion operation on the dominant frequency feature map; The residual fusion prediction module based on the gating mechanism is used for gated residual fusion prediction: the wind field physical features are embedded to obtain wind field embedded features, and a two-dimensional periodic map is constructed based on the wind field embedded features. The wind field comprehensive representation is obtained through multi-scale attention enhancement fusion operation. After the wind field comprehensive representation is converted into a wind-induced adjustment signal, it is residually fused with multi-resolution features through wind-induced gating operation to output the effective wave height prediction value.

[0016] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art: This invention explicitly constructs four types of wind field features—wind speed, wind direction, wind stress, and wind shear—and employs depthwise separable convolution for channel-level coupled modeling, addressing the shortcomings of traditional methods in modeling wind-driven factors and enhancing the model's ability to perceive the structural changes in wind-wave interaction. Then, it reconstructs a two-dimensional periodic image using the dominant frequency in the frequency domain and utilizes a dual-axis attention mechanism to explicitly decompose and independently enhance seasonal and trend-based structural features, effectively overcoming the shortcomings of high-low frequency aliasing and unclear trend representation in time-series modeling, and improving the model's adaptability to time-frequency structures and its cross-time-domain prediction performance. Finally, it introduces wind-gated residuals to output the predicted effective wave height. This achieves accurate and stable prediction of the effective wave height, improving the model's prediction robustness under extreme wind speed conditions. Attached Figure Description

[0017] The accompanying drawings, as part of this invention, are provided to further illustrate the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention, but do not constitute an undue limitation thereof. Clearly, the drawings described below are merely some embodiments; those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0018] In the attached diagram: Figure 1 This is a diagram illustrating the overall architecture of the effective wave height prediction system based on seasonal-trend decomposition in this specific embodiment.

[0019] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0021] In the following description, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

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

[0023] Example 1: Please see Figure 1 This invention provides an effective wave height prediction method based on seasonal-trend decomposition, comprising the following steps: S1. Multi-scale wind field coupled coding: The original wave height sequence is obtained to construct multi-scale wave height features. The zonal and meridional wind speed components are obtained to construct wind field physical features and embedded into high-dimensional wind field embedding features. The wind field embedding features are fused with the corresponding scale wave height features at the channel level and output a multi-scale wind and wave coupled feature sequence through depthwise separable convolution. S2, Seasonal-Trend Structure Modeling: Multi-resolution temporal images are constructed from multi-scale wind and wave coupling feature sequences, and seasonal and trend structural features are extracted separately using a dual-axis attention mechanism. A multi-scale fusion mechanism is used to obtain the dominant frequency feature map from the seasonal and trend structural features, and the dominant frequency feature map is then weighted and fused by dominant frequency attention to output multi-resolution features. S3, Gated Residual Fusion Prediction: Wind field embedding features are obtained by embedding the physical features of the wind field, and a two-dimensional periodic map is constructed based on the wind field embedding features. The wind field comprehensive representation is obtained by multi-scale attention enhancement fusion operation. After the wind field comprehensive representation is converted into a wind-induced adjustment signal, it is fused with multi-resolution features by wind-induced gating operation to output the effective wave height prediction value.

[0024] In some embodiments, the original wave height sequence is obtained to construct multi-scale wave height features, zonal and meridional wind speed components are obtained to construct wind field physical features and embedded into high-dimensional wind field embedding features, the wind field embedding features are fused with the corresponding scale wave height features at the channel level, and multi-scale wind-wave coupled feature sequences are output through depthwise separable convolution operations, including: Multi-scale wave height features are constructed by downsampling the original input wave height sequence layer by layer. The zonal and meridional wind speed components are obtained, and four types of wind field physical features, namely wind speed, wind direction, wind stress and wind shear, are constructed through numerical calculation to form a wind field feature vector. The wind field feature vector is then embedded into a high dimension through linear mapping to obtain the wind field embedding features. Wind field embedding features and wave height features at various scales are concatenated along the channel dimension to obtain fused features corresponding to each scale; these fused features are then mapped to a high-dimensional space to obtain the [missing information - likely a specific feature or parameter]. High-dimensional fusion features at the layer scale are sequentially subjected to channel-wise convolution and pointwise convolution to output the first... The corresponding wind and wave coupling feature sequence at the layer scale.

[0025] Specifically, input the original wave height sequence Multi-scale wave height features are constructed by downsampling through layer-by-layer one-dimensional convolution. , No. The wave height characteristics at the layer scale are expressed as: ; in, Indicates the first The time step at the scale is The wave height characteristics of Indicates scale index. Indicates the maximum scale layer number. Indicates the time step index. , indicating the first Wave height characteristics at the layer scale Indicates the kernel size. Indicates the internal index of the convolution kernel. Indicates the first The next-dimensional convolution kernel at the layer scale is in the first layer. The weight of each position, This indicates the wave height characteristics at the previous scale. Indicates the first Bias term at the layer scale.

[0026] Specifically, The total number of scale layers representing multi-scale wave height characteristics can be used for subsequent seasonal-trend structure modeling and residual fusion prediction based on gating mechanisms. Each scale layer performs the corresponding processing; see [link / reference] Figure 1 “×” "This indicates that seasonal-trend structure modeling and residual fusion prediction based on gating mechanisms are in..." Repeated execution on each scale layer.

[0027] Specifically, obtain the latitudinal data at a height of 10 meters from ERA5. , via The wind speed component is constructed through numerical calculation, forming a wind field feature vector composed of four types of wind field physical characteristics: wind speed, wind direction, wind stress, and wind shear. This vector is represented as follows: ; ; ; ; ; in, , , and These represent wind speed, wind direction, wind stress, and wind shear, respectively. and These represent the time steps. The zonal and meridional wind speed components at a height of 10m. Represents the arctangent function The extended form is used to determine the appropriate form based on the given information. The sign of the wind direction is used to determine the quadrant in which it is located, thereby calculating the correct wind direction angle; Indicates the time step. This represents the change in wind speed between adjacent time steps. This represents the gradient operator along the time dimension. Indicates time step The wind field characteristic vector consists of wind speed, wind direction, wind stress, and wind shear.

[0028] The wind field feature vector is embedded into a high-dimensional form through a linear mapping to obtain the wind field embedding feature, which is represented as: ; in, Indicates wind field embedding characteristics, and These are the learnable weight matrix and bias vector, respectively. This represents the dimension of the wind field embedding features.

[0029] The above linear mapping embedding can be represented by a physical embedding function: ; in, This represents the wind field embedding feature obtained by physical embedding mapping from the wind field feature vector.

[0030] Specifically, the wind field embedding features are high-order representations of the wind field. When they are subsequently fused with the wave height features at various scales obtained by downsampling and input mapping from the original wave height sequence in the channel dimension, they can retain the original physical meaning while providing the ability to model complex wind-wave interaction structures.

[0031] Specifically, to enhance the moderating effect of wind field on wave height modeling, multi-scale modeling is performed on the wind field embedding features: the wind field embedding features at each scale and the wave height features at each scale are concatenated along the channel dimension to obtain fused features, which are then mapped to a higher dimension to obtain high-dimensional fused features, represented as follows: ; ; ; in, Indicates the first The high-dimensional fusion feature is obtained by concatenating wind field embedding features and corresponding scale wave height features at the layer scale and then performing high-dimensional mapping. express Layer fusion features express Layer-scale wave height characteristics, express Layer-scale wind field embedding features. This indicates a splicing operation. express Time duration at the layer scale Indicates the channel dimension.

[0032] High-dimensional fusion features are output through channel-wise convolution and pointwise convolution using depthwise separable convolution. The corresponding wind and wave coupling feature sequence at the layer scale; wherein, channel-wise convolution performs one-dimensional convolution on each channel separately, without inter-channel mixing; point convolution is performed through... Convolution merges channels, represented as: ; in, Indicates the first Multi-scale wind-wave coupling characteristics at the layer scale This represents depthwise separable convolution, which consists of channel-wise convolution and pointwise convolution.

[0033] In some embodiments, constructing multi-resolution temporal images from multi-scale wind and wave coupled feature sequences includes: The wind-wave coupling feature sequences at various scales are represented in the frequency domain by Fast Fourier Transform (FFT) and the top K dominant frequencies with the largest amplitudes are extracted. The period length of each dominant frequency is calculated, and zero-padding is performed on the wind-wave coupling feature sequences at each scale at each dominant frequency. Zero-padding is performed on the wind-wave coupling feature sequences at each scale at each dominant frequency so that the length of the padded sequence is an integer multiple of the corresponding period length, resulting in the padded wind-wave coupling feature sequences. Based on the period length corresponding to each dominant frequency and the number of periods contained in the padded sequence length, the padded wind-wave coupling feature sequences are reconstructed into two-dimensional periodic images. The two-dimensional periodic images corresponding to different dominant frequencies are used as temporal image representations of different resolution layers, and the hidden dimensions of the two-dimensional periodic images of each resolution layer are expanded to obtain multi-resolution temporal images.

[0034] Specifically, the wind-wave coupling feature sequences at various scales are represented in the frequency domain by Fast Fourier Transform, as follows: ; in, Indicates the first Frequency domain representation of wind-wave coupling characteristics at the layer scale. Indicates Fast Fourier Transform; Extract the top K frequencies with the largest amplitudes and calculate the period length of each frequency, expressed as: ; Where K represents the total number of selected clock frequencies. Primary frequency index, For the first One main frequency value, Indicates the first The cycle length corresponding to each main frequency; For each scale of the wind-wave coupling characteristic sequence, zero-padding is performed at each dominant frequency, so that the first... The length of the wind-wave coupling feature sequence at the layer scale is the first... The length of the padded wind-wave coupling feature sequence is obtained by taking an integer multiple of the period length corresponding to each major frequency. The length of the padded wind-wave coupling feature sequence is expressed as: ; in, Indicates the first Layer scale The length of the filled wind-wave coupling characteristic sequence corresponding to each major frequency; Indicates the first The original length of the wind-wave coupling feature sequence at the layer scale. Indicates the first The cycle length corresponding to each main frequency. This indicates the rounding up operation; After obtaining the padded wind-wave coupling feature sequence, based on the period length corresponding to each dominant frequency and the number of periods contained in the padded sequence length, the padded wind-wave coupling feature sequence is reconstructed into a two-dimensional periodic image, represented as: ; ; in, This indicates a dimension reshaping operation. Indicates the first Layer scale The wind-wave coupling characteristic sequence after filling the main frequency. Indicates the first Layer scale Two-dimensional periodic image corresponding to each main frequency. This represents a time series segment within a period, i.e., the column axis length. The number of periods contained in the filled wind-wave coupling feature sequence, i.e., the row axis length, where d represents the feature dimension; The two-dimensional periodic images corresponding to different main frequencies are used as time images of different resolution layers. The hidden dimension of the two-dimensional periodic images of each resolution layer is expanded to obtain a multi-resolution time image, represented as follows: ; in, The first hidden dimension obtained by expanding the hidden dimension is... At the layer scale, the first Multi-resolution time images corresponding to each main frequency. Indicates the resolution layer index. Indicates the total number of resolution layers. Indicates the first The first resolution layer The row axis length corresponding to each main frequency.

[0035] In some embodiments, seasonal and trend structural features are extracted separately using a dual-axis attention mechanism, including: Convolution operations are performed along the column axis on the resolution time image of each resolution layer to obtain the query matrix, key matrix and value matrix of the column axis, and attention is calculated to obtain seasonal structural features; Convolution operations are performed along the row axis on the time-series images of each resolution layer to obtain the query matrix, key matrix, and value matrix of the row axis, and attention calculations are performed to obtain trend structure features.

[0036] Specifically, the resolution time image of each resolution layer is processed along the column axis. The convolution operation yields the query matrix, key matrix, and value matrix of the column axes, and attention calculation is performed to obtain the seasonal structural features, represented as: ; ; in, These represent the query matrix, key matrix, and value matrix of the column axes, respectively. Indicates the kernel size as column-axis convolution operation, This indicates column-axis attention calculation. Indicates the first Layer scale The seasonal structural features corresponding to the m-th resolution layer for each major frequency. Indicates the scale layer index. Indicates the primary frequency index. Indicates the resolution layer index; Perform resolution time images along the row axis for each resolution layer The convolution operation yields the query matrix, key matrix, and value matrix of the row axes, and attention calculation is performed to obtain trend structure features, represented as: ; ; in, These represent the query matrix, key matrix, and value matrix of the row axis, respectively. Indicates the kernel size as Axial convolution operation, This indicates row axis attention calculation. Indicates the first At the layer scale, the first Each main frequency corresponds to a trend structure feature at the m-th resolution layer.

[0037] In some embodiments, a multi-scale fusion mechanism is used to obtain a dominant frequency feature map from seasonal structural features and trend structural features. The dominant frequency feature map is then processed through a dominant frequency attention-weighted fusion operation to output multi-resolution features, including: For seasonal structural features, starting from the first resolution layer, layer by layer upwards, the seasonal structural features of the current resolution layer are added to the seasonal structural features of the previous resolution layer after convolution, to obtain the enhanced seasonal structural features of the current resolution layer. In response to the trend-based structural characteristics, with the first Starting from the resolution layer, layer by layer downwards, the trend structure features of the current resolution layer and the trend structure features of the next resolution layer are added together after deconvolution to obtain the enhanced trend structure features of the current resolution layer. Seasonal and trend structural features at the same resolution layer are fused into a two-dimensional representation at the corresponding scale, and then reshaped into a one-dimensional dominant frequency feature map through dimensionality reshaping. Each one-dimensional main frequency feature map is processed by global average pooling and linear mapping to calculate the main frequency attention score, and then normalized to obtain the main frequency weight. The main frequency weight is then weighted and fused with the corresponding main frequency feature map to output multi-resolution features.

[0038] Specifically, for seasonal structural features, the seasonal structural features of the current resolution layer are convolved with the seasonal structural features of the previous resolution layer to obtain the enhanced seasonal structural features of the current resolution layer, represented as: ; in, Indicates the first At the layer scale, the first The first main frequency, the first Seasonal structural features corresponding to each resolution layer Indicates the first At the layer scale, the first The first main frequency, the first Seasonal structural features corresponding to each resolution layer This represents the original seasonal structural features obtained by dual-axis attention extraction at the current resolution layer. This represents a two-dimensional convolution operation with a downsampling factor of 2; In response to the trend-based structural characteristics, with the first Starting with the current resolution layer, and proceeding downwards layer by layer, the trend structure features of the current resolution layer are added to the trend structure features of the next resolution layer after deconvolution, resulting in the enhanced trend structure features of the current resolution layer, represented as: ; in, Indicates the first At the layer scale, the first The first main frequency, the first Trend structural features corresponding to each resolution layer Indicates the first At the layer scale, the first The first main frequency, the first Trend structural features corresponding to each resolution layer This represents the original trend structure features obtained by dual-axis attention extraction at the current resolution layer. This indicates a deconvolution operation with an upsampling factor of 2; Specifically, seasonal and trend structural features at the same resolution layer are fused into a two-dimensional representation at the corresponding scale, and then a one-dimensional dominant frequency feature map is obtained through dimensional reshaping, as shown below: ; in, Indicates the first At the layer scale, the first The first main frequency, the first The one-dimensional dominant frequency feature map corresponding to the resolution layer. This indicates that the size is The image is flattened to a length of The sequence representation.

[0039] Specifically, each one-dimensional dominant frequency feature map is processed through global average pooling and linear mapping to calculate the dominant frequency attention score, and then normalized to obtain the dominant frequency weight. The dominant frequency weight is then weighted and fused with the corresponding dominant frequency feature map to output a multi-resolution feature, represented as follows: ; ; ; in, Indicates the first At the layer scale, the first The first main frequency, the first The resolution layer corresponds to a one-dimensional dominant frequency feature map. GAP(⋅) represents global average pooling, and FC(⋅) represents fully connected mapping. Indicates the first The frequency attention score corresponding to each frequency. This indicates a normalization operation. This represents the normalized main frequency weight. Indicates the first At the layer scale, the resolution features corresponding to the m-th resolution layer. This represents element-wise multiplication. Use the primary frequency index.

[0040] In other embodiments, wind field physical features are embedded to obtain wind field embedded features, and a two-dimensional periodic map is constructed based on the wind field embedded features. A comprehensive wind field representation is obtained through multi-scale attention enhancement fusion. After the comprehensive wind field representation is converted into a wind-induced adjustment signal, a predicted effective wave height is output through wind-induced gating and fusion with multi-resolution feature residuals, including: At each scale, four types of wind field physical features—wind speed, wind direction, wind stress, and wind shear—are embedded into a unified high-dimensional feature space through linear mapping to obtain four types of wind field embedding features. Fast Fourier Transform is then performed on each of the four types of wind field embedding features, and the top K wind field dominant frequencies with the largest amplitudes are selected to calculate the period length corresponding to each dominant frequency. Zero-padding is then applied to each of the four types of wind field embedding features at each dominant frequency, ensuring that the length of the padded wind field embedding feature sequence is an integer multiple of the corresponding period length. Based on the period length corresponding to each dominant frequency and the number of periods contained in the padded wind field embedding feature sequence, a two-dimensional periodic image of the four types of wind field embedding features is reconstructed. Each two-dimensional periodic image is mapped to a query vector, a key vector, and a value vector through a learnable query matrix, a key matrix, and a value matrix. The dot product between the query vector and the key vector is calculated and scaled. After Softmax normalization, attention enhancement features at each wind field's dominant frequency are obtained. The attention enhancement features of the same type of wind field embedding features at different wind field dominant frequencies are weighted and fused to obtain fused representations of wind speed, wind direction, wind stress, and wind shear at each scale. Channel splicing is performed on the fused representations at each scale to obtain the comprehensive wind field representation at each scale. After the comprehensive wind field representation is converted to one dimension, the wind-induced regulation signals of the wind field at each scale are obtained through linear mapping. Based on the wind-induced regulation signals, a dynamic gating factor is constructed. The contribution intensity of the wind field is weighted and adjusted by the dynamic gating factor and then fused with the multi-resolution feature residual to output the effective wave height prediction value.

[0041] Specifically, at each scale, the four types of wind field physical characteristics—wind speed, wind direction, wind stress, and wind shear—are embedded into a unified high-dimensional feature space through linear mapping, as follows: ; ; ; ; in, , , , Projection modules representing four types of wind field characteristics: wind speed, wind direction, wind stress, and wind shear. , , , These represent the learnable weight matrices for four types of wind field characteristics: wind speed, wind direction, wind stress, and wind shear. , , , These represent the components of the bias terms. , , , The wind field embedding features, representing the linear mappings of four types of wind field embedding features (wind speed, wind direction, wind stress, and wind shear), are all... 3D vector, adapted to the corresponding scale The length of time.

[0042] The four types of embedded features are subjected to Fast Fourier Transform (FFT) respectively, and the period length of each of the top K frequencies with the largest amplitudes is calculated, as follows: ; in, This represents the period length of each dominant frequency of the four types of embedded features. Represents the first of the four types of wind field embedding features One main frequency, This represents the clock frequency index, and K represents the total number of clock frequencies.

[0043] Zero-filling was performed on the embedding features of four types of wind fields—wind speed, wind direction, wind stress, and wind shear—at the dominant frequency of each wind field, so that the first... The sequence length of various wind field embedding features at the layer scale is the first... The period length corresponding to the dominant frequency of each wind field is an integer multiple of the period length, resulting in various filled wind field embedding feature sequences. Based on the period length corresponding to the dominant frequency of each wind field and the number of periods contained in the filled wind field embedding feature sequence, the filled wind field embedding feature sequences corresponding to wind speed, wind direction, wind stress, and wind shear are reconstructed into two-dimensional periodic images, represented as follows: ; ; ; ; ; ; in, Indicates the first Wind field embedding features at the layer scale in the first The sequence length after zero padding for each major frequency corresponding to the period length. Indicates the first Wind field embedding features at the layer scale in the first The number of cycles contained in each main frequency. , , , Two-dimensional periodic images representing wind speed, wind direction, wind stress, and wind shear, respectively.

[0044] Specifically, a self-attention mechanism is applied to each two-dimensional periodic image. A learnable query matrix, key matrix, and value matrix are mapped to query vectors, key vectors, and value vectors, respectively. The dot product between the query vector and the key vector is calculated and scaled. After Softmax normalization, the attention enhancement features of each two-dimensional periodic image are obtained, represented as follows: ; ; ; ; in, , , , These represent the attention enhancement features of the two-dimensional periodic images corresponding to wind speed, wind direction, wind stress, and wind shear, respectively. , , For learnable query, key, and value matrices, For feature dimensions.

[0045] Attention-enhanced features of the same type of wind field embedding features at different dominant wind field frequencies are weighted and fused to obtain fused representations of wind speed, wind direction, wind stress, and wind shear at various scales, as follows: ; ; ; ; in, Indicates wind speed Wind speed characteristics at the layer scale in the first Attention weight under the main frequency of each wind field , , , They represent the first The fused representation of wind speed, wind direction, wind stress, and wind shear at the layer scale is obtained by weighting and fusing the data with the dominant frequency. The dominant frequency index represents the embedded features of the wind field.

[0046] Specifically, the four types of fused representations at each scale are spliced ​​together to obtain a comprehensive wind field representation at each scale, as follows: ; in, This indicates a splicing operation. Indicates the first A comprehensive representation of wind fields at the strata scale; After the wind field is comprehensively represented and converted to one dimension, the wind-induced regulation signals at various scales are obtained through linear mapping, as follows: ; ; ; in, This represents the comprehensive wind field representation after conversion to one dimension. Indicates the first Layer-scale wind-induced regulation signals, , These represent the learnable weight matrix and bias vector used to generate the wind-induced regulation signal, respectively. To further enhance the physical consistency and generalization ability of the model, a dynamic gating factor is constructed based on the wind-induced adjustment signal to control the real-time adjustment intensity of the wind field information, thereby ensuring the model's adaptability under different wind speed conditions, expressed as: ; in, Indicates the dynamic gating factor. This represents the sigmoid activation function. and These are learnable parameters; Specifically, the intensity of real-time adjustment of wind field information is as follows: Local winds strengthen (such as near tropical cyclones or strong convective systems), resulting in strong wind energy input. The model fully preserves the contribution of wind-induced undulations; When the wind speed is weak It effectively suppresses gain signals without physical basis and avoids overfitting or spurious signal interference.

[0047] The effective wave height prediction value is output by weighting the wind field contribution intensity using a dynamic gating factor and fusing it with multi-resolution feature residuals. Specifically: ; in, Multi-resolution features output for modeling seasonal-trend structures. This is an element-wise multiplication.

[0048] Example 2: In some embodiments, to train the significant wave height prediction model based on seasonal-trend decomposition, a training dataset is constructed by acquiring wind field data and buoy significant wave height data. The wind field data uses ERA5 reanalysis data provided by the European Centre for Medium-Range Weather Forecasts (ECMWF), extracting the zonal wind component u and meridional wind component v at a height of 10 meters around the target buoy point; the buoy significant wave height data uses historical significant wave height data from buoy observation stations established by the National Oceanic and Atmospheric Administration (NOAA), with significant wave height referred to as Significant Wave Height (SWH).

[0049] In some embodiments, Santa Monica Bay buoy station numbered 46221 is selected as the target buoy point, located at 33.860°N, 118.641°W, with a water depth of 387 meters. In other embodiments, multiple buoy observation stations can be selected to construct a training dataset, for example, buoy observation stations numbered 46221 and 46251, and the historical significant wave height sequence and ERA5 wind field data covering the surrounding area of ​​each buoy observation station can be obtained respectively.

[0050] Specifically, ERA5 wind field data and buoy significant wave height observation data are matched according to timestamps to construct a joint wind field-wave height input sample. Each sample includes a historical significant wave height sequence within a preset time window, as well as the zonal wind component u and meridional wind component v of the same period. The sample label is the significant wave height of the target buoy point at a preset future time. In some embodiments, the significant wave height sequence of the past 10 days and the corresponding wind field sequence are used as model inputs to predict the significant wave height for the next 12 hours, 24 hours, and 48 hours, respectively.

[0051] Before inputting the model, the significant wave height sequence and the wind field input sequence are normalized and represented as follows: ; in, This represents the data values ​​to be normalized. This represents the maximum value in the corresponding dataset. This represents the minimum value in the corresponding dataset. This represents the normalized data value.

[0052] After the model outputs normalized predicted values, it performs inverse normalization using the corresponding normalization parameters, as shown below: ; in, This represents the normalized predicted value output by the model. This represents the predicted value after inverse normalization.

[0053] In some embodiments, the constructed samples are divided into training, validation, and test sets in chronological order. The training set is used to update model parameters, the validation set is used for model selection and hyperparameter tuning during training, and the test set is used to evaluate the predictive performance of the model after training. Dividing the dataset in chronological order can prevent data leakage from future time points into the training process.

[0054] In some embodiments, the ratio of the training set, validation set, and test set is 7:1:2; in other embodiments, the ratio can be adjusted according to the data scale and prediction task requirements.

[0055] In some embodiments, the model training employs the Adam optimizer with an initial learning rate of 0.001, a batch size of 64, and 300 training epochs. During training, training set samples are input into the model and sequentially undergo multi-scale wind field coupling encoding, seasonal-trend structure modeling, and gated residual fusion prediction to obtain the predicted significant wave height. The loss function is calculated based on the error between the predicted and actual significant wave heights, and the model parameters are updated via backpropagation. After each training epoch, the validation error is calculated using the validation set, and the model parameters corresponding to lower validation errors are saved.

[0056] In some embodiments, the model is implemented based on the PyTorch deep learning framework and trained on a server equipped with an NVIDIA 4090 GPU. It should be noted that the above training environment is only one optional implementation and is not intended to limit the scope of protection of this invention. This invention can also be trained or deployed on other computing devices capable of neural network training.

[0057] In some embodiments, Mean Squared Error (MSE) is used as the model training loss function to measure the difference between the predicted significant wave height and the true significant wave height, expressed as: ; in, Represents the total number of samples. Indicates the first The true effective wave height corresponding to each sample Indicates the first The predicted effective wave height corresponding to each sample.

[0058] Using MSE as the loss function can assign higher weights to larger prediction errors, allowing the model to prioritize reducing prediction biases in scenarios with high-energy wave events or drastic changes in sea state during training, thereby improving the model's prediction stability in complex sea conditions.

[0059] After the model is trained, the test set is used to make predictions on the trained model, and the model output is inversely normalized and compared with the true effective wave height to calculate the prediction error of the model on the test set.

[0060] Comparative experiment: To verify the technical effectiveness, this invention uses ERA5 reanalysis data provided by the European Centre for Medium-Range Weather Forecasts (ECMWF) to extract wind speed field data (zonal wind component at a height of 10 meters) covering the area around the buoy point between 2019 and 2023. With meridional wind component As an auxiliary input, historical significant wave height (SWH) data from buoy 46221, an observation station established by the National Oceanic and Atmospheric Administration (NOAA), was also collected as the primary input and prediction target. The buoy information used is as follows: Buoy 46221: (Right now ), the water depth is 387 meters.

[0061] To comprehensively evaluate the performance of the proposed model in the significant wave height (SWH) prediction task, root mean square error (RMSE), mean square error (MSE), and mean absolute error (MAE) were used as evaluation metrics. This invention selected ANN, AR, ARMA, SVR, MICN, and LightTS methods for comparative experiments. Detailed experimental results are shown in Table 1. Table 1: Comparison of the present invention with existing methods

[0062] In all the aforementioned effective wave height prediction tasks, this invention consistently demonstrates the lowest performance. The values ​​of MSE and MAE further demonstrate the reliability of the invention in improving the accuracy of effective wave height prediction, and prove the potential and advantages of the invention in practical applications.

[0063] Example 3: Based on the same inventive concept, this invention also provides an effective wave height prediction system based on seasonal-trend decomposition, comprising: The multi-scale wind field coupling coding module is used for multi-scale wind field coupling coding: it obtains the original wave height sequence to construct multi-scale wave height features, obtains the zonal and meridional wind speed components to construct wind field physical features and embeds them into high-dimensional wind field embedding features, performs channel-level fusion with the corresponding scale wave height features, and outputs a multi-scale wind and wave coupling feature sequence through depthwise separable convolution. The seasonal-trend structure modeling module is used for seasonal-trend structure modeling: it constructs multi-resolution time images from multi-scale wind and wave coupled feature sequences, and extracts seasonal structure features and trend structure features respectively through a dual-axis attention mechanism; it uses a multi-scale fusion mechanism to obtain the dominant frequency feature map from the seasonal structure features and trend structure features, and outputs multi-resolution features through a dominant frequency attention weighted fusion operation on the dominant frequency feature map; The residual fusion prediction module based on the gating mechanism is used for gated residual fusion prediction: the wind field physical features are embedded to obtain wind field embedded features, and a two-dimensional periodic map is constructed based on the wind field embedded features. The wind field comprehensive representation is obtained through multi-scale attention enhancement fusion operation. After the wind field comprehensive representation is converted into a wind-induced adjustment signal, the effective wave height prediction value is output through wind-induced gating operation and multi-resolution feature residual fusion.

[0064] Unless otherwise defined, the technical or scientific terms used in this invention shall have the general meaning understood by one of ordinary skill in the art to which this invention pertains. Words such as “a,” “an,” “an,” “the,” “the,” and “these” used in this invention do not indicate quantitative limitations and may be singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this invention are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or modules (units) is not limited to the listed steps or modules (units) but may include steps or modules (units) not listed, or may include other steps or modules (units) inherent to these processes, methods, products, or devices. Words such as “connected,” “linked,” and “coupled” used in this invention are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. “A plurality” used in this invention refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. Normally, the character " / " indicates that the objects before and after it are in an "or" relationship. In this invention, the terms "first," "second," "third," etc., are used only to distinguish similar objects and do not represent a specific order of objects.

[0065] Obviously, the accompanying drawings are merely some examples or embodiments of the present invention. Those skilled in the art can apply the present invention to other similar situations based on these drawings without any creative effort. Furthermore, it is understood that although the work done in this development process may be complex and lengthy, certain design, manufacturing, or production modifications made according to the technical content disclosed in the present invention are merely conventional technical means and should not be considered as insufficient disclosure of the present invention.

[0066] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-described technical content to create equivalent embodiments without departing from the scope of the present invention. The implementation schemes in the above embodiments can also be further combined or replaced. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An effective wave height prediction method based on seasonal-trend decomposition, characterized in that, Includes the following steps: S1. Multi-scale wind field coupled coding: The original wave height sequence is obtained to construct multi-scale wave height features. The zonal and meridional wind speed components are obtained to construct wind field physical features and embedded into high-dimensional wind field embedding features. The wind field embedding features are fused with the wave height features at the corresponding scale at the channel level and output as a multi-scale wind and wave coupled feature sequence through depthwise separable convolution. S2, Seasonal-Trend Structure Modeling: Multi-resolution temporal images are constructed from multi-scale wind and wave coupling feature sequences, and seasonal and trend structural features are extracted separately using a dual-axis attention mechanism. A multi-scale fusion mechanism is used to obtain the dominant frequency feature map from the seasonal and trend structural features, and the dominant frequency feature map is then weighted and fused by dominant frequency attention to output multi-resolution features. S3, Residual fusion prediction with gating mechanism: Wind field embedding features are obtained by embedding the physical features of the wind field, and a two-dimensional periodic map is constructed based on the wind field embedding features. The wind field comprehensive representation is obtained by multi-scale attention enhancement fusion operation. After the wind field comprehensive representation is converted into a wind-induced adjustment signal, the effective wave height prediction value is output by fusing it with the multi-resolution feature residual through wind-induced gating operation.

2. The method according to claim 1, characterized in that, The original wave height sequence is obtained to construct multi-scale wave height features. Zonal and meridional wind speed components are obtained to construct wind field physical features, which are then embedded into high-dimensional wind field embedding features. These wind field embedding features are fused with corresponding scale wave height features at the channel level, and after depthwise separable convolution operations, a multi-scale wind-wave coupled feature sequence is output, including: Multi-scale wave height features are constructed by downsampling the original input wave height sequence layer by layer. The zonal and meridional wind speed components are obtained, and four types of wind field physical features, namely wind speed, wind direction, wind stress and wind shear, are constructed through numerical calculation to form a wind field feature vector. The wind field feature vector is then embedded into a high dimension through linear mapping to obtain the wind field embedding features. Wind field embedding features and wave height features at various scales are concatenated along the channel dimension to obtain fused features corresponding to each scale; these fused features are then mapped to a high-dimensional space to obtain the [missing information - likely a specific feature or parameter]. High-dimensional fusion features at the layer scale are sequentially subjected to channel-wise convolution and pointwise convolution to output the first... The corresponding wind and wave coupling feature sequence at the layer scale.

3. The method according to claim 1, characterized in that, Constructing multi-resolution temporal images from multi-scale wind and wave coupled feature sequences, including: The wind-wave coupling feature sequences at various scales are represented in the frequency domain by Fast Fourier Transform, and the sequence with the largest amplitude is extracted. For each dominant frequency, the corresponding period length is calculated. For the wind-wave coupling feature sequence at each scale, zero-padding is performed at each dominant frequency so that the length of the padded sequence is an integer multiple of the corresponding period length, thus obtaining the padded wind-wave coupling feature sequence. Based on the period length corresponding to each dominant frequency and the number of periods contained in the padded sequence length, the padded wind-wave coupling feature sequence is reconstructed into a two-dimensional periodic image. The two-dimensional periodic images corresponding to different dominant frequencies are used as time image representations of different resolution layers, and the hidden dimension of the two-dimensional periodic image of each resolution layer is expanded to obtain a multi-resolution time image.

4. The method according to claim 3, characterized in that, Seasonal and trend-based structural features were extracted using a dual-axis attention mechanism, including: Convolution operations are performed along the column axis on the multi-resolution temporal images of each resolution layer to obtain the query matrix, key matrix, and value matrix of the column axis, and attention calculation is performed to obtain seasonal structural features; Convolution operations are performed along the row axis on the multi-resolution temporal images of each resolution layer to obtain the query matrix, key matrix, and value matrix of the row axis, and attention calculation is performed to obtain trend structure features.

5. The method according to claim 4, characterized in that, A multi-scale fusion mechanism is used to obtain the dominant frequency feature map from seasonal and trend structural features. The dominant frequency feature map is then subjected to a dominant frequency attention-weighted fusion operation to output multi-resolution features, including: To extract the seasonal structural features, starting from the first resolution layer, we proceed upwards layer by layer, adding the seasonal structural features of the current resolution layer to the seasonal structural features of the previous resolution layer after convolution. In response to the trend-based structural characteristics, with the first Starting from the resolution layer, layer by layer downwards, the trend structure features of the current resolution layer and the trend structure features of the next resolution layer are added together after deconvolution to obtain the enhanced trend structure features of the current resolution layer. Seasonal and trend structural features at the same resolution layer are fused into a two-dimensional representation at the corresponding scale, and then reshaped into a one-dimensional dominant frequency feature map through dimensionality reshaping. Each one-dimensional main frequency feature map is processed by global average pooling and linear mapping to calculate the main frequency attention score, and then normalized to obtain the main frequency weight. The main frequency weight is then weighted and fused with the corresponding main frequency feature map to output multi-resolution features.

6. The method according to claim 2, characterized in that, Wind field physical features are embedded to obtain wind field embedded features, and a two-dimensional periodic map is constructed based on these features. A comprehensive wind field representation is obtained through multi-scale attention-enhanced fusion. This comprehensive representation is then converted into a wind-induced conditioning signal, which is fused with multi-resolution feature residuals through wind-induced gating to output the predicted effective wave height, including: At each scale, four types of wind field physical features—wind speed, wind direction, wind stress, and wind shear—are embedded into a unified high-dimensional feature space through linear mapping to obtain four types of wind field embedding features. Fast Fourier Transform is performed on each of the four types of wind field embedding features, and the top K wind field dominant frequencies with the largest amplitudes are selected to calculate the period length corresponding to each wind field dominant frequency. Zero-padding is then applied to each of the four types of wind field embedding features at each wind field dominant frequency, ensuring that the length of the padded wind field embedding feature sequence is an integer multiple of the corresponding period length. Based on the period length corresponding to each wind field dominant frequency and the number of periods contained in the padded wind field embedding features, a two-dimensional periodic image of the four types of wind field embedding features is reconstructed. Each two-dimensional periodic image is mapped to a query vector, a key vector, and a value vector through a learnable query matrix, a key matrix, and a value matrix. The dot product between the query vector and the key vector is calculated and scaled. After Softmax normalization, attention enhancement features at each wind field's dominant frequency are obtained. The attention enhancement features of the same type of wind field embedding features at different wind field dominant frequencies are weighted and fused to obtain fused representations of wind speed, wind direction, wind stress, and wind shear at each scale. Channel splicing is performed on the fused representations at each scale to obtain the comprehensive wind field representation at each scale. After the comprehensive wind field representation is converted to one dimension, the wind-induced regulation signals of the wind field at each scale are obtained through linear mapping. Based on the wind-induced regulation signals, a dynamic gating factor is constructed. The contribution intensity of the wind field is weighted and adjusted by the dynamic gating factor and then fused with the multi-resolution feature residual to output the effective wave height prediction value.

7. An effective wave height prediction system based on seasonal-trend decomposition, characterized in that, include: The multi-scale wind field coupling coding module is used for multi-scale wind field coupling coding: it obtains the original wave height sequence to construct multi-scale wave height features, obtains the zonal and meridional wind speed components to construct wind field physical features and embeds them into high-dimensional wind field embedding features, performs channel-level fusion with the corresponding scale wave height features, and outputs a multi-scale wind and wave coupling feature sequence through depthwise separable convolution. The seasonal-trend structure modeling module is used for seasonal-trend structure modeling: it constructs multi-resolution time images from multi-scale wind and wave coupled feature sequences, and extracts seasonal and trend structure features respectively through a dual-axis attention mechanism; it uses a multi-scale fusion mechanism to obtain the dominant frequency feature map from the seasonal and trend structure features, and outputs multi-resolution features through a dominant frequency attention weighted fusion operation on the dominant frequency feature map; The residual fusion prediction module based on the gating mechanism is used for gated residual fusion prediction: the wind field physical features are embedded to obtain wind field embedded features, and a two-dimensional periodic map is constructed based on the wind field embedded features. The wind field comprehensive representation is obtained through multi-scale attention enhancement fusion operation. After the wind field comprehensive representation is converted into a wind-induced adjustment signal, it is residually fused with multi-resolution features through wind-induced gating operation to output the effective wave height prediction value.

Images