Wave height prediction system and method integrating dynamic features and physical gating

Abstract

This invention discloses a wave height prediction system and method integrating dynamic features and physical gating, belonging to the field of artificial intelligence and machine learning technology, for marine environmental monitoring. The system includes a data acquisition and preprocessing module, a feature selection module, and a model prediction module. The method includes acquiring and preprocessing marine meteorological variable data, constructing a Shapley additive interpretation lightweight gradient booster regression model to select the core variable set, constructing a physically guided gating feature-aware Transformer prediction model, constructing a hybrid loss function-constrained prediction model, and outputting significant wave height prediction values ​​after the prediction model is trained. This invention significantly improves the utilization efficiency and scenario adaptability of multi-source marine meteorological features through feature selection and dynamic weight allocation mechanisms. It can be flexibly extended to wave prediction tasks at different sea areas and spatiotemporal scales, providing high-precision and highly reliable technical support for marine engineering, shipping safety, and disaster early warning.

Description

Technical Field

[0001] This invention discloses a wave height prediction system and method that integrates dynamic features and physical gating, belonging to the field of artificial intelligence and machine learning technology. Background Technology

[0002] Significant wave height, as a core parameter describing the marine environment, is crucial for accurate prediction of marine disasters, ensuring safety in offshore operations, and protecting coastal engineering projects. Disastrous waves are often triggered by weather systems such as tropical cyclones and extratropical cyclones, which can capsize ships, destroy coastal engineering projects, and cause serious casualties and economic losses. Therefore, the need for high-precision and timely wave prediction is extremely urgent.

[0003] Currently, significant wave height prediction techniques are mainly divided into three categories: numerical prediction methods, traditional machine learning methods, and deep learning methods. Numerical prediction methods are based on ocean dynamic equations to build models. Although they have a physical basis, the model structure is complex and the computational load is huge. Moreover, the longer the prediction time, the more obvious the error accumulation becomes, making it difficult to meet the needs of real-time early warning. Traditional machine learning methods (such as support vector machines and random forests) rely on manual feature engineering and cannot capture the temporal correlation and dynamic feature changes of marine meteorological data, resulting in limited generalization ability. Among deep learning methods, LSTM models capture temporal dependencies through gating mechanisms and have become a commonly used scheme for wave prediction. However, this type of model is subject to the risk of gradient vanishing and does not consider the dynamic adaptation of feature importance, resulting in insufficient prediction accuracy for extreme wave heights. In recent years, the Informer model, an improvement on the Transformer, has shown advantages in long-series time-series prediction due to its efficient probabilistic attention mechanism. However, it still has significant shortcomings in wave prediction scenarios: First, the attention mechanism uses a random sampling strategy, which easily focuses on invalid time steps, reducing prediction efficiency. Second, it does not incorporate the physical constraints of wave generation, such as the lag effect of wave height on wind speed and wave period, resulting in prediction results lacking physical rationality. Third, when dealing with multivariate input tasks, the input features are not specifically screened, and many marine meteorological variables contain low-correlation redundant information, increasing the model training burden. Fourth, the feature weights are fixed and cannot adapt to the dynamic changes in feature importance under different marine environments such as storms and calm, resulting in large prediction errors for extreme wave heights.

[0004] Furthermore, existing technologies generally employ a single loss function (such as MSE), which is sensitive to outliers like extreme wave heights, easily leading to overfitting of models to regular samples and neglecting critical disaster scenarios. Simultaneously, feature selection relies heavily on empirical judgment, lacking quantitative evaluation and intelligent selection mechanisms, further limiting the improvement of prediction accuracy. Therefore, developing a significant wave height prediction technology that integrates physical constraints, dynamic feature weights, and intelligent feature selection to address the problems of weak physical interpretability, insufficient dynamic adaptation, and low accuracy in extreme value prediction of existing methods has become a key technological gap in the field of intelligent marine environmental forecasting. Summary of the Invention

[0005] The purpose of this invention is to provide a wave height prediction system and method that integrates dynamic features and physical gating, so as to solve the problems of weak physical interpretability, insufficient dynamic feature adaptation, and low prediction accuracy of extreme sea state caused by the reliance on a single loss function and manual feature selection in the existing wave height prediction methods.

[0006] A wave height prediction system that integrates dynamic features and physical gating includes: a data acquisition and preprocessing module, a feature selection module, and a model prediction module;

[0007] The data acquisition and preprocessing module collects and preprocesses marine meteorological variable data. The preprocessed marine meteorological variable data is then input into the feature selection module. The feature selection module constructs a Shapley additive interpretation lightweight gradient booster regression model to select the core variable set. The preprocessed marine meteorological variable data and the core variable set are then input into the model prediction module. The model prediction module constructs a physically guided gated feature-aware Transformer prediction model for training. The trained physically guided gated feature-aware Transformer prediction model outputs the predicted values ​​of significant wave heights.

[0008] A wave height prediction method integrating dynamic features and physical gating, using a wave height prediction system integrating dynamic features and physical gating, including:

[0009] S1. Through the data acquisition and preprocessing module, the monitoring area is determined, marine meteorological variable data are acquired and preprocessed;

[0010] S2. The feature selection module constructs a Shapley additive interpretation lightweight gradient lift regression model based on the measured values ​​of significant wave heights. The measured values ​​of significant wave heights are used as the target values ​​for prediction, and marine meteorological variables other than the measured values ​​of significant wave heights are used as feature variables. The Shapley additive interpretation lightweight gradient lift regression model is trained, and the importance of the feature variables is analyzed based on the trained model to select the core variable set.

[0011] S3. The model prediction module constructs a physically guided gated feature-aware Transformer prediction model. It then trains the model using a neural network, employing a hybrid loss function to constrain the model. When the hybrid loss function reaches its minimum value, the trained model is output; otherwise, neural network training continues. The hybrid loss function includes Hubel loss, mean squared error, and mean absolute error. Based on the trained model and preprocessed marine meteorological variable data, the module outputs the optimal predicted wave height.

[0012] S1 includes determining the monitoring area, collecting marine meteorological variable data, and normalizing the marine meteorological variable data:

[0013] ;

[0014] In the formula, For normalized marine meteorological variable data, The range of values ​​is , For marine meteorological variable data, This represents the minimum value of the marine meteorological variable data. This represents the maximum value of the marine meteorological variable data.

[0015] S2 includes S2.1, constructing a Shapley additive interpretation lightweight gradient boosting machine regression model, including a data input layer, a core layer, an evaluation layer, an interpretation layer, and a result output layer; the Shapley additive interpretation lightweight gradient boosting machine regression model uses the measured value of significant wave height as the prediction target value, and uses marine meteorological variable data other than the measured value of significant wave height as feature variables to train the Shapley additive interpretation lightweight gradient boosting machine regression model;

[0016] The data input layer inputs feature variables into the core layer for tree model ensemble training. After training, it outputs the prediction results of significant wave heights and model parameters. The prediction results are input into the evaluation layer to calculate the gain importance score of each feature variable. The model parameters are input into the interpretation layer, and the SHAP value of each feature variable is calculated using a tree model interpreter. ;Will and As the analysis results, the input result output layer calculates all samples for each feature variable. absolute value average ,set up Weights and Weights, calculate the first The combined score of each feature variable , Index of the feature variable;

[0017] S2 includes, S2.2, based on Sort all feature variables in descending order and select the highest-ranking one. These 10 characteristic variables serve as core features, forming a core variable set.

[0018] S3 includes a physically guided gated feature-aware Transformer prediction model consisting of two dynamic weight layers, two data embedding layers, two encoder layers, one decoder layer, and a linear prediction layer.

[0019] The measured values ​​of significant ocean wave heights and the Pearson correlation coefficients of core variables are input into the first dynamic weighting layer to obtain the weighted features. and time stamp matrix ,Will and Input the first data embedding layer and generate the encoder embedding vector using the fusion embedding formula. ,Will Input the encoder and get the encoder output. Construct a prediction sequence of the target length. ,Will After the part to be predicted is padded with 0, it is input into the second data embedding layer, and the decoder embedding vector is generated using the fusion embedding formula. ,Will and Input to decoder, obtain decoder output ,Will Input linear prediction layer outputs preliminary significant wave height prediction values ;

[0020] The encoder consists of a gated feature-aware self-attention layer, a LayerNorm layer, a convolutional layer, and a LayerNorm layer in sequence; the decoder consists of a gated feature-aware self-attention layer, a LayerNorm layer, a cross-attention layer, a LayerNorm layer, a convolutional layer, and a LayerNorm layer in sequence.

[0021] S3 includes S3.1, inputting the measured values ​​of significant wave heights and the Pearson correlation coefficients of core variables into the first dynamic weighting layer to initialize the feature weight vector. , , As the core variable, Index for core variables, , For the first Prior weights of the core variables, The value of is the absolute value of the corresponding Pearson correlation coefficient;

[0022] Selected from marine meteorological variable data 1 feature, construct the input feature matrix :

[0023] ;

[0024] In the formula, for The measured value of the significant wave height at a given time, let's assume... An index for selecting features from marine meteorological variable data. , For the first A feature selected from marine meteorological variable data;

[0025] Calculate the time-step adaptive weight matrix :

[0026] ;

[0027] In the formula, For normalization function, It is a 1×1 convolutional layer. It is the ReLU activation function. For batch size, The time step length, Number of core variables Element-by-element multiplication It is the transpose symbol;

[0028] S3 includes S3.2, calculation :

[0029] ;

[0030] In the formula, Use the Sigmoid activation function;

[0031] Build :

[0032] ;

[0033] ;

[0034] In the formula, For time step index, For the first The time feature vector at each time step For the first The month after time step normalization For the first The normalized day after the time step For the first The time step is normalized to an hour.

[0035] S3 includes S3.3, the first data embedding layer generated using a fusion embedding formula. The calculation process is as follows:

[0036] ;

[0037] In the formula, For time feature embedding function, For numerical feature embedding functions, For the final embedding vector, , for The dimension;

[0038] S3 includes, S3.4, and will Input encoder, first of all Input-gated feature-aware self-attention layer performs non-random sampling of feature importance:

[0039] ;

[0040] ;

[0041] In the formula, The key tensor in the self-attention layer is used for gating feature perception. , For key In the variance of the last dimension, For the number of attention heads, For each dimension of attention head, for index, For the first The first batch, the first The first attention head, the first The time step, the first A dimensional key tensor For the first The first batch, the first The first attention head, the first Mean of the key tensor at each time step The sampling coefficients are... Let be the time step length of the key tensor. It is a rounding function. The number of samples;

[0042] according to Sort the time steps from largest to smallest, and select the first... Sampling is performed at each time step to obtain a set of sampling indices. ,according to from Extract the corresponding subtensor to obtain ;

[0043] Calculate the initial attention score :

[0044] , ;

[0045] In the formula, To query a tensor, for Time step length, Scaling factor It is the transpose symbol;

[0046] Will Perform hysteresis gating:

[0047] ;

[0048] In the formula, and For the weights and biases of the gated linear layer, This is the gate strength coefficient. The attention score after gating;

[0049] based on By fusing local and global attention through learnable parameters, the training process is adaptively optimized.

[0050] ;

[0051] In the formula, The fused attention matrix, For learnable parameters, ;

[0052] Calculate the context vector :

[0053] ;

[0054] In the formula, For value tensors, , for Dimensions ;

[0055] Output :

[0056] ;

[0057] In the formula, This is a LayerNorm layer.

[0058] S3 includes S3.5, setting the known historical sequence length of the decoder. and the number of future time steps predicted by the decoder ,Will The former The significant measured wave height values ​​and corresponding marine meteorological variables at each time step are: , After a time step The significant measured wave height values ​​and corresponding marine meteorological variables for each prediction step are: ,set up ,Will and By piecing them together in sequence ;

[0059] Will Input the second dynamic weight layer and generate the decoder time stamp matrix according to step S3.1. Weighted features of decoder The data embedding layer uses the fusion embedding formula to generate the decoder embedding vector. :

[0060] ;

[0061] Will The gated feature-aware self-attention layer of the input decoder is used as input, and according to step S3.4, the output of the gated feature-aware self-attention layer of the decoder is obtained. .

[0062] S3 includes, S3.6, and will The input is fed into the cross-attention layer, where the decoder attention score matrix is ​​first calculated. :

[0063] ;

[0064] ;

[0065] In the formula, For the decoder to query the tensor, , For decoder key tensors, , and The learnable parameter matrix generated during model training. The feature dimension of a single attention head;

[0066] Compute decoder multi-head context representation :

[0067] ;

[0068] ;

[0069] In the formula, For decoder value tensors, , The learnable parameter matrix generated during model training. For each attention head of the decoder, the context vector is... To swap the first and second dimensions, For tensor reshaping;

[0070] Calculate the output of the cross-attention layer:

[0071] ;

[0072] In the formula, Linear layer;

[0073] S3 includes S3.7, and the output of the compute decoder. :

[0074] ;

[0075] S3 includes, S3.8, and will The input projection layer is mapped to the prediction dimension to obtain preliminary significant wave height predictions:

[0076] ;

[0077] In the formula, These are preliminary predictions of significant wave heights. , To extract the last Features of each time step This refers to slicing operations, including taking all elements in the first dimension and starting from the last element in the second dimension. Starting from the first element and moving to the last element, take all elements in the third dimension; The number of input feature dimensions is 1.

[0078] S3 includes S3.9, which calculates the hybrid loss function based on the predicted values ​​of significant ocean wave heights and the corresponding measured values ​​of significant ocean wave heights. :

[0079] ;

[0080] In the formula, For Huber's loss, Mean square error, The mean absolute error, , , For the weighting coefficients, satisfying ;

[0081] S3 includes S3.10, inputting the preprocessed marine meteorological data into the trained physical-guided gated feature-aware Transformer prediction model to obtain the optimal salient wave height prediction value.

[0082] Compared with existing technologies, this invention has the following advantages: It significantly improves the utilization efficiency and scenario adaptability of multi-source marine meteorological features by interpreting the feature selection and dynamic weight allocation mechanism of the lightweight gradient booster regression model through Shapley additive interpretation; the gated feature-aware self-attention layer enhances the physical rationality of the results while ensuring prediction accuracy; the hybrid loss function and stabilization architecture design effectively improve the prediction robustness and engineering practicality under extreme sea conditions; and the overall modular design endows the model with strong generalization ability, which can be flexibly extended to wave prediction tasks at different sea areas and spatiotemporal scales, providing high-precision and high-reliability technical support for marine engineering, shipping safety, and disaster early warning. Attached Figure Description

[0083] Figure 1 This is the flowchart of the present invention;

[0084] Figure 2 This is a diagram of the structure of a physically guided, gated feature-aware Transformer prediction model.

[0085] Figure 3 This is a structure diagram of the LightGBM regression model;

[0086] Figure 4 This is a comparison chart of the Informer model's 24-hour predicted values ​​and actual values;

[0087] Figure 5 This is a comparison chart of the LSTM model's 24-hour predicted values ​​and the actual values;

[0088] Figure 6 This is a comparison chart of the predicted and actual values ​​of the model in this invention over 24 hours;

[0089] Figure 7 This is a comparison chart of the Informer model's 48-hour predicted values ​​and the actual values;

[0090] Figure 8 This is a comparison chart of the LSTM model's predicted values ​​and the actual values ​​after 48 hours.

[0091] Figure 9This is a comparison chart of the predicted and actual values ​​of the model of this invention over 48 hours;

[0092] Figure 10 This is a comparison chart of the Informer model's 72-hour predicted values ​​and the actual values;

[0093] Figure 11 This is a comparison chart of the LSTM model's 72-hour predicted values ​​and the actual values;

[0094] Figure 12 This is a comparison chart of the predicted and actual values ​​of the model of this invention over 72 hours. Detailed Implementation

[0095] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0096] A wave height prediction system that integrates dynamic features and physical gating includes: a data acquisition and preprocessing module, a feature selection module, and a model prediction module;

[0097] The data acquisition and preprocessing module collects and preprocesses marine meteorological variable data. The preprocessed marine meteorological variable data is then input into the feature selection module. The feature selection module constructs a Shapley additive interpretation lightweight gradient booster regression model to select the core variable set. The preprocessed marine meteorological variable data and the core variable set are then input into the model prediction module. The model prediction module constructs a physically guided gated feature-aware Transformer prediction model for training. The trained physically guided gated feature-aware Transformer prediction model outputs the predicted values ​​of significant wave heights.

[0098] A wave height prediction method integrating dynamic features and physical gating, using a wave height prediction system integrating dynamic features and physical gating, including:

[0099] S1. Through the data acquisition and preprocessing module, the monitoring area is determined, marine meteorological variable data are acquired and preprocessed;

[0100] S2. The feature selection module constructs a Shapley additive interpretation lightweight gradient lift regression model based on the measured values ​​of significant wave heights. The measured values ​​of significant wave heights are used as the target values ​​for prediction, and marine meteorological variables other than the measured values ​​of significant wave heights are used as feature variables. The Shapley additive interpretation lightweight gradient lift regression model is trained, and the importance of the feature variables is analyzed based on the trained model to select the core variable set.

[0101] S3. The model prediction module constructs a physically guided gated feature-aware Transformer prediction model. It then trains the model using a neural network, employing a hybrid loss function to constrain the model. When the hybrid loss function reaches its minimum value, the trained model is output; otherwise, neural network training continues. The hybrid loss function includes Hubel loss, mean squared error, and mean absolute error. Based on the trained model and preprocessed marine meteorological variable data, the module outputs the optimal predicted wave height.

[0102] S1 includes determining the monitoring area, collecting marine meteorological variable data, and normalizing the marine meteorological variable data:

[0103] ;

[0104] In the formula, For normalized marine meteorological variable data, The range of values ​​is , For marine meteorological variable data, This represents the minimum value of the marine meteorological variable data. This represents the maximum value of the marine meteorological variable data.

[0105] S2 includes S2.1, constructing a Shapley additive interpretation lightweight gradient boosting machine regression model, including a data input layer, a core layer, an evaluation layer, an interpretation layer, and a result output layer; the Shapley additive interpretation lightweight gradient boosting machine regression model uses the measured value of significant wave height as the prediction target value, and uses marine meteorological variable data other than the measured value of significant wave height as feature variables to train the Shapley additive interpretation lightweight gradient boosting machine regression model;

[0106] The data input layer inputs feature variables into the core layer for tree model ensemble training. After training, it outputs the prediction results of significant wave heights and model parameters. The prediction results are input into the evaluation layer to calculate the gain importance score of each feature variable. The model parameters are input into the interpretation layer, and the SHAP value of each feature variable is calculated using a tree model interpreter. ;Will and As the analysis results, the input result output layer calculates all samples for each feature variable. absolute value average ,set up Weights and Weights, calculate the first The combined score of each feature variable , Index of the feature variable;

[0107] S2 includes, S2.2, based on Sort all feature variables in descending order and select the highest-ranking one. These 10 characteristic variables serve as core features, forming a core variable set.

[0108] S3 includes a physically guided gated feature-aware Transformer prediction model consisting of two dynamic weight layers, two data embedding layers, two encoder layers, one decoder layer, and a linear prediction layer.

[0109] The measured values ​​of significant ocean wave heights and the Pearson correlation coefficients of core variables are input into the first dynamic weighting layer to obtain the weighted features. and time stamp matrix ,Will and Input the first data embedding layer and generate the encoder embedding vector using the fusion embedding formula. ,Will Input the encoder and get the encoder output. Construct a prediction sequence of the target length. ,Will After the part to be predicted is padded with 0, it is input into the second data embedding layer, and the decoder embedding vector is generated using the fusion embedding formula. ,Will and Input to decoder, obtain decoder output ,Will Input linear prediction layer outputs preliminary significant wave height prediction values ;

[0110] The encoder consists of a gated feature-aware self-attention layer, a LayerNorm layer, a convolutional layer, and a LayerNorm layer in sequence; the decoder consists of a gated feature-aware self-attention layer, a LayerNorm layer, a cross-attention layer, a LayerNorm layer, a convolutional layer, and a LayerNorm layer in sequence.

[0111] S3 includes S3.1, inputting the measured values ​​of significant wave heights and the Pearson correlation coefficients of core variables into the first dynamic weighting layer to initialize the feature weight vector. , , As the core variable, Index for core variables, , For the first Prior weights of the core variables, The value of is the absolute value of the corresponding Pearson correlation coefficient;

[0112] Selected from marine meteorological variable data 1 feature, construct the input feature matrix :

[0113] ;

[0114] In the formula, for The measured value of the significant wave height at a given time, let's assume... An index for selecting features from marine meteorological variable data. , For the first A feature selected from marine meteorological variable data;

[0115] Calculate the time-step adaptive weight matrix :

[0116] ;

[0117] In the formula, For normalization function, It is a 1×1 convolutional layer. It is the ReLU activation function. For batch size, The time step length, Number of core variables Element-by-element multiplication It is the transpose symbol;

[0118] S3 includes S3.2, calculation :

[0119] ;

[0120] In the formula, Use the Sigmoid activation function;

[0121] Build :

[0122] ;

[0123] ;

[0124] In the formula, For time step index, For the first The time feature vector at each time step For the first The month after time step normalization For the first The normalized day after the time step For the first The time step is normalized to an hour.

[0125] S3 includes S3.3, the first data embedding layer generated using a fusion embedding formula. The calculation process is as follows:

[0126] ;

[0127] In the formula, For time feature embedding function, For numerical feature embedding functions, For the final embedding vector, , for The dimension;

[0128] S3 includes, S3.4, and will Input encoder, first of all Input-gated feature-aware self-attention layer performs non-random sampling of feature importance:

[0129] ;

[0130] ;

[0131] In the formula, The key tensor in the self-attention layer is used for gating feature perception. , For key In the variance of the last dimension, For the number of attention heads, For each dimension of attention head, for index, For the first The first batch, the first The first attention head, the first The time step, the first A dimensional key tensor For the first The first batch, the first The first attention head, the first Mean of the key tensor at each time step The sampling coefficients are... Let be the time step length of the key tensor. It is a rounding function. The number of samples;

[0132] according to Sort the time steps from largest to smallest, and select the first... Sampling is performed at each time step to obtain a set of sampling indices. ,according to from Extract the corresponding subtensor to obtain ;

[0133] Calculate the initial attention score :

[0134] , ;

[0135] In the formula, To query a tensor, for Time step length, Scaling factor It is the transpose symbol;

[0136] Will Perform hysteresis gating:

[0137] ;

[0138] In the formula, and For the weights and biases of the gated linear layer, This is the gate strength coefficient. The attention score after gating;

[0139] based on By fusing local and global attention through learnable parameters, the training process is adaptively optimized.

[0140] ;

[0141] In the formula, The fused attention matrix, For learnable parameters, ;

[0142] Calculate the context vector :

[0143] ;

[0144] In the formula, For value tensors, , for Dimensions ;

[0145] Output :

[0146] ;

[0147] In the formula, This is a LayerNorm layer.

[0148] S3 includes S3.5, setting the known historical sequence length of the decoder. and the number of future time steps predicted by the decoder ,Will The former The significant measured wave height values ​​and corresponding marine meteorological variables at each time step are: , After a time step The significant measured wave height values ​​and corresponding marine meteorological variables for each prediction step are: ,set up ,Will and By piecing them together in sequence ;

[0149] Will Input the second dynamic weight layer and generate the decoder time stamp matrix according to step S3.1. Weighted features of decoder The data embedding layer uses the fusion embedding formula to generate the decoder embedding vector. :

[0150] ;

[0151] Will The gated feature-aware self-attention layer of the input decoder is used as input, and according to step S3.4, the output of the gated feature-aware self-attention layer of the decoder is obtained. .

[0152] S3 includes, S3.6, and will The input is fed into the cross-attention layer, where the decoder attention score matrix is ​​first calculated. :

[0153] ;

[0154] ;

[0155] In the formula, For the decoder to query the tensor, , For decoder key tensors, , and The learnable parameter matrix generated during model training. The feature dimension of a single attention head;

[0156] Compute decoder multi-head context representation :

[0157] ;

[0158] ;

[0159] In the formula, For decoder value tensors, , The learnable parameter matrix generated during model training. For each attention head of the decoder, the context vector is... To swap the first and second dimensions, For tensor reshaping;

[0160] Calculate the output of the cross-attention layer:

[0161] ;

[0162] In the formula, Linear layer;

[0163] S3 includes S3.7, and the output of the compute decoder. :

[0164] ;

[0165] S3 includes, S3.8, and will The input projection layer is mapped to the prediction dimension to obtain preliminary significant wave height predictions:

[0166] ;

[0167] In the formula, These are preliminary predictions of significant wave heights. , To extract the last Features of each time step This refers to slicing operations, including taking all elements in the first dimension and starting from the last element in the second dimension. Starting from the first element and moving to the last element, take all elements in the third dimension; The number of input feature dimensions is 1.

[0168] S3 includes S3.9, which calculates the hybrid loss function based on the predicted values ​​of significant ocean wave heights and the corresponding measured values ​​of significant ocean wave heights. :

[0169] ;

[0170] In the formula, For Huber's loss, Mean square error, The mean absolute error, , , For the weighting coefficients, satisfying ;

[0171] S3 includes S3.10, inputting the preprocessed marine meteorological data into the trained physical-guided gated feature-aware Transformer prediction model to obtain the optimal salient wave height prediction value.

[0172] The calculation process of K, Q, and V in the gated feature perception self-attention layer includes the following: the calculation process of K is as follows: first, the encoder fuses the embedded output. The initial input is used; then, a projection transformation is performed through independent linear layers: First, maintain the feature dimension as the global dimension of the model; finally, split the dimensions according to the number of attention heads H: Remodeling Then transpose Finally, we obtain the key tensor K (ensuring that each attention head independently captures features).

[0173] The calculation process for the query tensor Q is as follows: the encoder fuses the output of the embedding layer. As input, feature transformation is performed through an independent learnable linear layer to obtain a linear projection result:

[0174] ;

[0175] Then, a multi-head split will be performed, which is... Split by the number of attention heads H to obtain the single-head query tensor Q: .

[0176] The value tensor V and the key tensor K are of the same origin. The core principles are linear projection and multi-head splitting, ensuring that the dimensions adapt to the attention weight fusion requirements. The input is also the encoder fusion embedding layer output. Linear projection is performed through independently learnable linear layers:

[0177] ;

[0178] Then, a multi-head split is performed to obtain the value tensor V. The splitting formula is as follows:

[0179] .

[0180] The following description, in conjunction with the accompanying drawings and embodiments, further illustrates the process of this invention. Figure 1As shown, the first step is to determine the monitoring area, acquire marine meteorological variable data, and preprocess it; the second step is to construct the SHAP-LightGBM model (Shapley additive interpretation lightweight gradient booster regression model) to screen the core predictive variable set; the third step is to construct a physically guided gated feature-aware Transformer prediction model, train it, and use a hybrid loss function for constraints; finally, based on the trained physically guided gated feature-aware Transformer prediction model, the significant wave height prediction results are output.

[0181] The structure of the physically guided gated feature-aware Transformer prediction model of this invention is as follows: Figure 2 As shown, the source sequence data is sequentially input into the dynamic weight layer, data embedding layer, encoder, and decoder; the predicted sequence of the target length, with the predicted portion padded with 0s, is sequentially input into the dynamic weight layer, data embedding layer, and decoder; the decoder's processing result is input into the linear prediction layer to obtain the output sequence. The encoder sequentially includes a gated feature-aware self-attention layer, a LayerNorm layer, a convolutional layer, and a LayerNorm layer; the decoder sequentially includes a gated feature-aware self-attention layer, a LayerNorm layer, a cross-attention layer, a LayerNorm layer, a convolutional layer, and a LayerNorm layer.

[0182] The data used in this embodiment of the invention employs a global climate reanalysis dataset. Marine meteorological variables include significant wave height, wind speed, gust speed, average wind and wave direction, average wind and wave period, air temperature, average sea-level pressure, sea-level temperature, total swell wave height, surface pressure, free convection velocity over the ocean, and air density over the ocean. The temporal resolution reaches the hourly level, accurately capturing short-term climate fluctuations. Spatially, a regular latitude and longitude grid projection is used, with a core spatial resolution of 0.25° × 0.25°, extending vertically from the Earth's surface to an altitude of 80 kilometers, subdivided into 137 levels, systematically presenting atmospheric and oceanic structure information at different altitudes. Nine representative observation points are selected to comprehensively cover wave characteristics in different sea areas.

[0183] The structure of the lightweight gradient booster regression model, as explained by Shapley's additive interpretation in this invention, is as follows: Figure 3 As shown, it includes a data input layer (original features / target variables), which inputs the feature data stream into the LightGBM core layer (tree model ensemble training), the core layer inputs the prediction results stream into the evaluation layer (prediction + index calculation), the core layer inputs the model parameter stream into the SHAP interpretation layer (SHAP value calculation + feature interpretation), and the two converge into an analysis result stream input into the result output layer (feature importance + model interpretation report).

[0184] The construction and training process includes loading marine meteorological variable data, parsing the dates to sort the variable data in ascending order by time, and processing NaN values ​​through forward imputation. Specifically, for the sorted time series dataset, the observed values ​​of each marine meteorological characteristic variable are traversed. When the variable value of a certain time node is missing, the valid observed value of the same variable from the next adjacent time node is used to fill the missing value until there are no missing values ​​in the entire dataset.

[0185] Significant wave height measurements were extracted from marine meteorological variable data. Specifically, the significant wave height column in the dataset was used as the sole prediction target value, while all remaining columns were selected as marine meteorological feature variables, thus separating the features from the target value. Temporal cross-validation was employed to divide the training and validation sets: the complete marine meteorological data, ordered chronologically, was divided into five consecutive, equal-length temporal subsets. Each subset consisted of a training set and a validation set, with the training set comprising approximately 80% and the validation set approximately 20%. Each subset followed the rule that the training set consisted of historical data from preceding time periods, and the validation set consisted of data from immediately following time periods. The training and validation sets had no temporal overlap or intersection. This process was repeated for all temporal subsets.

[0186] In each predefined time-series subset, a SHAP-LightGBM regression model is constructed based on the training set (setting the regression objective (objective=regression), the gradient boosting decision tree boosting method (boosting_type=gbdt), and setting the maximum number of training rounds to 1000; controlling the number of iteration rounds through an early stopping mechanism; if the model performance on the validation set shows no improvement for 50 consecutive rounds, the model iteration training is immediately terminated). Based on the trained SHAP-LightGBM regression model, the model performance is evaluated using the validation set. Significant wave height prediction is performed on the feature data of the validation set to generate validation set prediction values. Combining the measured significant wave height values ​​of the validation set, the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²)—three core evaluation indicators for regression tasks—are calculated to quantify and evaluate the model's predictive performance.

[0187] The gain of marine meteorological variable data is recorded as an importance indicator. Using training set sampling as the background and validation set sampling as the sample, the SHAP value is calculated through a tree model interpreter. After all time-series subsets are trained, the average importance of marine meteorological variable data in different subsets is calculated and ranked. Based on this, a specified number of features are selected. The specific number of features selected can be set by the user; this implementation uses the following settings. The SHAP value is the value used to explain the prediction results of the machine learning model for a single sample. This method selects the top 5 features based on their importance, which can retain the core variables with strong correlations and avoid the introduction of redundant variables that would lead to model complexity, slower computation, overfitting, and poor generalization ability.

[0188] Let the set of measured values ​​of significant wave heights be . , , Given the number of measured significant wave height samples, output the predicted significant wave height. The loss function uses mean squared error:

[0189] ;

[0190] In the formula, The mean squared error of the LightGBM regression model. For the first Measured values ​​of significant wave height for each sample. ;

[0191] The SHAP value of each feature variable is calculated using a tree model interpreter, and the sample... Chinese characteristics SHAP value for:

[0192] ;

[0193] In the formula, A mapping function for SHAP values. It is the first Explanation functions for each characteristic variable, For the index of the feature variable, For the set of feature variables, Indicates the first The set of feature variables for each sample Take the first value for the set of feature variables The value of each sample For the first Remove the th sample from the set of feature variables. The set of remaining feature variables after each feature variable; In the first The average predicted value of significant wave height when the feature variables of each sample are input into the SHAP-LightGBM regression model; To make the first The feature variable set of the nth sample is removed. The average predicted value of significant wave height when the set of remaining feature variables after the first feature variable is input into the SHAP-LightGBM regression model.

[0194] Each input feature variable Calculate the absolute mean of the SHAP values. for:

[0195] ;

[0196] Combination and The weighted average score is obtained as follows:

[0197] ;

[0198] In the formula, Features The overall score, The LightGBM gain importance score for a feature is defined as the sum of the decrease in loss function brought about by that feature during all decision tree splits, and its calculation formula is as follows:

[0199] ;

[0200] In the formula, Let T be the LightGBM gain importance score for the nth feature variable, and T be the total number of decision trees in the LightGBM model. Let be the set of all nodes in the nth decision tree that use the nth feature as the splitting feature. Let $\mathbf{n}$ be the decrease in the loss function brought about by the $\math splitting node in the $\math tree$. This is the training loss function for the LightGBM regression model (mean squared error in this invention).

[0201] Compared to traditional wave prediction models, deep learning-based prediction methods have significant advantages in terms of high computational efficiency and low application cost. To ensure the effectiveness of model training and the reliability of testing, this invention selects wave observation data from 00:00 on January 1, 2015 to 23:00 on December 31, 2024 for 10 consecutive years for model training and testing, as shown in Table 1:

[0202] Table 1. Training and Testing Datasets

[0203] ;

[0204] To verify the model's predictive performance and generalization ability, significant wave height data and related feature components (total swell wave height, average wave period, 10-meter gust, air density over the ocean, and free convection velocity over the ocean) from nine observation points were input into the model of this invention for training. The average value of the nine points was taken as the final error result of the model. The input sequence length was fixed at 72 hours (3 days), and wave height data for the next 24, 48, and 72 hours were output. To fairly compare the model's wave height prediction performance, this invention sets approximately the same total number of parameters for each experiment, ensuring that the model's computational efficiency is roughly the same each time. Specifically, the model's encoder layer was set to 2 layers, the decoder layer to 1 layer, the Adam optimizer was used during training, the learning rate was set to 0.0001, the weight decay was set to 0.00001, the batch size was 32, the number of training epochs was 30, and the early stopping value was 8. Training was stopped when the validation set loss did not decrease for 8 consecutive epochs to avoid model overfitting.

[0205] Table 2 shows the average model error results at nine different locations under the training of the model of this invention, and compares them with mainstream time series prediction models LSTM and Informer:

[0206] Table 2. Average Model Error Results

[0207] ;

[0208] The three models used the same input data and input sequence length, and the LSTM and Informer models were tuned using a grid method to maintain optimal prediction performance and ensure the objectivity of the comparison. The experimental results, mean absolute percentage error (MAPE) and coefficient of determination (R²), objectively reflect the adaptability and accuracy advantages of each model in nearshore environmental time series prediction tasks.

[0209] From the perspective of the core indicator of prediction accuracy, MAPE, the PGFAT model outperforms the Informer model in the entire prediction time range. It also shows a slight advantage over LSTM in short-term prediction (24 hours). Although the absolute value of MAPE is slightly higher than that of LSTM in medium- and long-term prediction (48 hours and 72 hours), it can be seen that the R² index shows better fitting stability. Specifically, for 24-hour predictions, PGFAT's MAPE is 20.66, a decrease of 1.76% compared to LSTM's 21.03, and a significant decrease of 85.03% compared to Informer's 137.96. This indicates that in short-term nearshore environmental prediction tasks, PGFAT's accuracy in capturing local temporal features is slightly better than traditional LSTM, and far superior to the Informer model. For 48-hour predictions, PGFAT's MAPE is 45.37, which is higher than LSTM's 37.73, but still a decrease of 76.04% compared to Informer's 188.48. For 72-hour predictions, PGFAT's MAPE is 59.14, a decrease of 73.57% compared to Informer's 223.40, further validating PGFAT's significant accuracy advantage over Informer in medium- to long-term prediction scenarios. It is worth noting that the MAPE of the Informer model is abnormally high across all prediction durations, far exceeding that of LSTM and PGFAT. This may be because the Informer's long-sequence attention mechanism is not well adapted to the complex nonlinear fluctuations in Chinese coastal data, leading to a surge in prediction bias. In contrast, PGFAT may have improved its adaptability to complex coastal time-series data through targeted feature extraction or structural optimization.

[0210] From the model fit index R², PGFAT showed the best fit effect across the entire prediction time range, fully demonstrating that it can better capture the inherent variation patterns of time series data from China's coastal waters. For 24-hour forecasts, PGFAT's R² is 0.77, an improvement of 10.00% and 8.45% compared to LSTM's 0.70 and Informer's 0.71, respectively, indicating that PGFAT has a stronger ability to interpret data trends in short-term forecasts. For 48-hour forecasts, PGFAT's R² is 0.45, an improvement of 15.38% compared to LSTM's 0.39 and 21.62% compared to Informer's 0.37, further expanding its fitting advantage in medium-term forecast scenarios. For 72-hour forecasts, PGFAT's R² is 0.34, an improvement of 61.90% compared to LSTM's 0.21 and 36.00% compared to Informer's 0.25. Even though the R² of all models shows a decreasing trend as the forecast duration increases, PGFAT's goodness of fit decays relatively less and it always maintains a leading advantage over the other two models, highlighting its structural stability in long-term series forecasting.

[0211] From the perspective of the impact of prediction duration on model performance, all three models exhibit the common characteristic that the longer the prediction duration, the higher the MAPE and the lower the R². However, the performance degradation rates of different models differ significantly: PGFAT's MAPE increased from 20.66 after 24 hours to 59.14 after 72 hours, an increase of 186.34%; LSTM's MAPE increased from 21.03 to 50.35, an increase of 139.42%; and Informer's MAPE increased from 137.96 to 223.40, an increase of 61.93%. Although Informer's MAPE increase was the smallest, its absolute error value was too high, making it unsuitable for practical application. While PGFAT's increase was higher than LSTM's, considering the R² degradation, its R² decreased from 0.77 to 0.34, a degradation of 55.84%, which is lower than LSTM's R² degradation. This indicates that PGFAT has a stronger ability to preserve data trends in long-term predictions and exhibits better performance stability.

[0212] To clearly reveal the differences in the wave height prediction performance of the three networks for the next 24, 48, and 72 hours in different regions, a line graph was plotted using the prediction values ​​for 400 consecutive hours at point D as an example. This visually presents the fitting effect between the predicted values ​​and the actual values ​​of the LSTM, Informer, and models of this invention under different prediction durations.

[0213] Predicted length 24 hours Figure 4 , Figure 5 and Figure 6 As shown, although the LSTM prediction can roughly follow the overall trend of the true value, there is a slight deviation from the true value at local peaks and valleys; the Informer prediction is significantly less consistent with the true value, and the fluctuation trend deviates significantly from the true value in some periods; the prediction of the model of this invention almost coincides with the fluctuation trajectory of the true value. Whether it is a sudden rise in peak value or a gradual decline in valley value, it can accurately capture the detailed changes of the true value, and the consistency is significantly better than the other two models.

[0214] Predicted length 48 hours Figure 7 , Figure 8 and Figure 9 As shown, with the extension of the prediction period, the deviation between each model and the true value increases, but the differences between the models remain clear: the deviation between the LSTM prediction and the true value intensifies, and the error in some periods (such as the peak around 150 hours) is more obvious than that of the 24-hour prediction; the "lag" or "deviation" of the Informer prediction to the fluctuation of the true value is further amplified, making it difficult to match the local changes of the true value; although the model of this invention has a slight deviation from the 24-hour prediction, it can still closely follow the core fluctuation trend of the true value, and its matching degree of peaks and troughs is far superior to LSTM and Informer.

[0215] Predicted length 72 hours Figure 10 , Figure 11 and Figure 12 As shown, the difficulty of inferring data trends increases under long-term forecasting, and the fitting effect of each model declines. However, the advantages of the model of this invention are still clear: the trajectory of the LSTM predicted value deviates significantly from the true value, and it can hardly match the fluctuation rhythm of the true value at the trough; the Informer predicted value has the worst fit with the true value, and the deviation of the overall trend has affected the effective judgment of the wave height change; although the model of this invention has local deviations, it can still basically follow the core fluctuation trend of the true value, and compared with LSTM and Informer, it can better preserve the change characteristics of the true value.

[0216] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A wave height prediction system integrating dynamic features and physical gating, characterized in that, include: Data acquisition and preprocessing module, feature selection module, and model prediction module; The data acquisition and preprocessing module collects and preprocesses marine meteorological variable data. The preprocessed marine meteorological variable data is then input into the feature selection module. The feature selection module constructs a Shapley additive interpretation lightweight gradient booster regression model to select the core variable set. The preprocessed marine meteorological variable data and the core variable set are then input into the model prediction module. The model prediction module constructs a physically guided gated feature-aware Transformer prediction model for training. The trained physically guided gated feature-aware Transformer prediction model outputs the significant wave height prediction value. The physically guided gated feature-aware Transformer prediction model consists of two dynamic weight layers, two data embedding layers, two encoder layers, one decoder layer, and a linear prediction layer. The measured values ​​of significant ocean wave heights and the Pearson correlation coefficients of core variables are input into the first dynamic weighting layer to obtain the weighted features. and time stamp matrix ,Will and Input the first data embedding layer and generate the encoder embedding vector using the fusion embedding formula. ,Will Input the encoder and get the encoder output. ; Construct a prediction sequence of target length ,Will After the part to be predicted is padded with 0, it is input into the second data embedding layer, and the decoder embedding vector is generated using the fusion embedding formula. ,Will and Input to decoder, get decoder output ,Will Input linear prediction layer outputs preliminary significant wave height prediction values ; The encoder consists of a gated feature-aware self-attention layer, a LayerNorm layer, a convolutional layer, and a LayerNorm layer in sequence; the decoder consists of a gated feature-aware self-attention layer, a LayerNorm layer, a cross-attention layer, a LayerNorm layer, a convolutional layer, and a LayerNorm layer in sequence.

2. A wave height prediction method integrating dynamic features and physical gating, using the wave height prediction system integrating dynamic features and physical gating as described in claim 1, characterized in that, include: S1. Through the data acquisition and preprocessing module, the monitoring area is determined, marine meteorological variable data are acquired and preprocessed; S2. The feature selection module constructs a Shapley additive interpretation lightweight gradient lift regression model based on the measured values ​​of significant wave heights. The measured values ​​of significant wave heights are used as the target values ​​for prediction, and marine meteorological variables other than the measured values ​​of significant wave heights are used as feature variables. The Shapley additive interpretation lightweight gradient lift regression model is trained, and the importance of the feature variables is analyzed based on the trained model to select the core variable set. S3. The model prediction module constructs a physically guided gated feature-aware Transformer prediction model. It then trains the model using a neural network, employing a hybrid loss function to constrain the model. When the hybrid loss function reaches its minimum value, the trained model is output; otherwise, neural network training continues. The hybrid loss function includes Hubel loss, mean squared error, and mean absolute error. Based on the trained model and preprocessed marine meteorological variable data, the module outputs the optimal predicted wave height.

3. The wave height prediction method integrating dynamic features and physical gating as described in claim 2, characterized in that, S1 includes determining the monitoring area, collecting marine meteorological variable data, and normalizing the marine meteorological variable data: ； In the formula, For normalized marine meteorological variable data, The range of values ​​is , For marine meteorological variable data, This represents the minimum value of the marine meteorological variable data. This represents the maximum value of the marine meteorological variable data.

4. The wave height prediction method integrating dynamic features and physical gating as described in claim 3, characterized in that, S2 includes S2.1, constructing a Shapley additive interpretation lightweight gradient boosting machine regression model, including a data input layer, a core layer, an evaluation layer, an interpretation layer, and a result output layer; the Shapley additive interpretation lightweight gradient boosting machine regression model uses the measured value of significant wave height as the prediction target value, and uses marine meteorological variable data other than the measured value of significant wave height as feature variables to train the Shapley additive interpretation lightweight gradient boosting machine regression model; The data input layer inputs feature variables into the core layer for tree model ensemble training. After training, it outputs the prediction results of significant wave heights and model parameters. The prediction results are input into the evaluation layer to calculate the gain importance score of each feature variable. ; The model parameters are input into the interpretation layer, and the SHAP value of each feature variable is calculated using a tree model interpreter. ;Will and As the analysis results, the input result output layer calculates all samples for each feature variable. absolute value average ,set up Weights and Weights, calculate the first The combined score of each feature variable , Index of the feature variable; S2 includes, S2.2, based on Sort all feature variables in descending order and select the highest-ranking one. These 10 characteristic variables serve as core features, forming a core variable set.

5. The wave height prediction method integrating dynamic features and physical gating as described in claim 4, characterized in that, S3 includes S3.1, inputting the measured values ​​of significant wave heights and the Pearson correlation coefficients of core variables into the first dynamic weighting layer to initialize the feature weight vector. , , As the core variable, For indexing core variables, , For the first Prior weights of the core variables, The value of is the absolute value of the corresponding Pearson correlation coefficient; Selected from marine meteorological variable data 1 feature, construct the input feature matrix : ； In the formula, for The measured value of the significant wave height at a given time, let's assume... An index for selecting features from marine meteorological variable data. , For the first A feature selected from marine meteorological variable data; Calculate the time-step adaptive weight matrix : ； In the formula, For normalization function, It is a 1×1 convolutional layer. It is the ReLU activation function. For batch size, The time step length, Number of core variables Element-by-element multiplication It is the transpose symbol; S3 includes S3.2, calculation : ； In the formula, Use the Sigmoid activation function; Build : ； ； In the formula, For time step index, For the first The time feature vector at each time step For the first The month after time step normalization For the first The day after time step normalization For the first The time step is normalized to an hour.

6. The wave height prediction method integrating dynamic features and physical gating as described in claim 5, characterized in that, S3 includes S3.3, the first data embedding layer generated using a fusion embedding formula. The calculation process is as follows: ； In the formula, For time feature embedding function, For numerical feature embedding functions, For the final embedding vector, , for The dimension; S3 includes, S3.4, and will Input encoder, first of all Input-gated feature-aware self-attention layer performs non-random sampling of feature importance: ； ； In the formula, The key tensor in the self-attention layer is used for gating feature perception. , For key In the variance of the last dimension, For the number of attention heads, For each dimension of attention head, for index, For the first The first batch, the first The first attention head, the first The time step, the first A dimensional key tensor For the first The first batch, the first The first attention head, the first Mean of the key tensor at each time step The sampling coefficients are... Let be the time step length of the key tensor. It is a rounding function. The number of samples; according to Sort the time steps from largest to smallest, and select the first... Sampling is performed at each time step to obtain a set of sampling indices. ,according to from Extract the corresponding subtensor to obtain ; Calculate the initial attention score : , ； In the formula, To query a tensor, for Time step length, Scaling factor It is the transpose symbol; Will Perform hysteresis gating: ； In the formula, and For the weights and biases of the gated linear layer, This is the gate strength coefficient. The attention score after gating; based on By fusing local and global attention through learnable parameters, the training process is adaptively optimized. ； In the formula, The fused attention matrix, For learnable parameters, ; Calculate the context vector : ； In the formula, For value tensors, , for Dimensions ; Output : ； In the formula, This is a LayerNorm layer.

7. The wave height prediction method integrating dynamic features and physical gating as described in claim 6, characterized in that, S3 includes S3.5, setting the known historical sequence length of the decoder. and the number of future time steps predicted by the decoder ,Will The former The significant measured wave height values ​​and corresponding marine meteorological variables at each time step are: , After a time step The significant measured wave height values ​​and corresponding marine meteorological variables for each prediction step are: ,set up ,Will and By piecing them together in sequence ; Will Input the second dynamic weight layer and generate the decoder time stamp matrix according to step S3.

1. Weighted features of decoder The data embedding layer uses the fusion embedding formula to generate the decoder embedding vector. : ； Will The gated feature-aware self-attention layer of the input decoder is used as input, and according to step S3.4, the output of the gated feature-aware self-attention layer of the decoder is obtained. .

8. The wave height prediction method integrating dynamic features and physical gating as described in claim 7, characterized in that, S3 includes, S3.6, and will The input is fed into the cross-attention layer, where the decoder attention score matrix is ​​first calculated. : ； ； In the formula, For the decoder to query the tensor, , For decoder key tensors, , and The learnable parameter matrix generated during model training. The feature dimension of a single attention head; Compute decoder multi-head context representation : ； ； In the formula, For decoder value tensors, , The learnable parameter matrix generated during model training. For each attention head of the decoder, the context vector is... To swap the first and second dimensions, For tensor reshaping; Calculate the output of the cross-attention layer: ； In the formula, Linear layer; S3 includes S3.7, and the output of the compute decoder. : ； S3 includes, S3.8, and will The input projection layer is mapped to the prediction dimension to obtain preliminary significant wave height predictions: ； In the formula, These are preliminary predictions of significant wave heights. , To extract the last Features of each time step This refers to slicing operations, including taking all elements in the first dimension and starting from the last element in the second dimension. Starting from the first element and moving to the last element, take all elements in the third dimension; The number of input feature dimensions is 1.

9. The wave height prediction method integrating dynamic features and physical gating as described in claim 8, characterized in that, S3 includes S3.9, which calculates the hybrid loss function based on the predicted salient wave height of the ocean and the corresponding measured salient wave height. : ； In the formula, For Huber's loss, Mean square error, The mean absolute error, , , For the weighting coefficients, satisfying ; S3 includes S3.10, inputting the preprocessed marine meteorological data into the trained physical-guided gated feature-aware Transformer prediction model to obtain the optimal salient wave height prediction value.

Images