Sea wave probability prediction method and system

Abstract

The invention belongs to the cross technical field of artificial intelligence and marine meteorological prediction, and discloses a sea wave probability prediction method and system, and the method comprises the steps: obtaining historical wind field data and sea wave spectrum data of a target sea area, carrying out the preprocessing and organization of the data, and constructing a training data set; constructing a hybrid expert probability model, wherein the model comprises a gating network and a plurality of expert networks; the training data set is used for training the hybrid expert probability model, the trained model receives input wind field data, and hybrid probability distribution is output through the synergistic effect of the gating network and the expert network; and sampling is carried out from the mixed probability distribution to obtain a predicted sea wave spectrum set, and sea wave risk probability prediction is realized. Complete distribution information can be obtained through one-time forward calculation, a numerical mode set is not needed, the computing power and energy consumption expenditure are remarkably reduced, and high-frequency updating and quasi-real-time business application can be conveniently achieved in resource-limited environments such as shipborne, buoys and offshore stations.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and marine meteorological forecasting, and in particular relates to a method and system for predicting ocean wave probability. Background Technology

[0002] With the continued application of numerical models and artificial intelligence in wave forecasting, the mainstream approaches of existing technologies include: conducting deterministic forecasts at regional or global scales using third-generation wave models based on spectral energy balance equations, combining buoy and satellite observation data for correction and assimilation, and using deep learning models for rapid regression prediction of key elements such as significant wave height, characteristic period, and peak frequency. While these methods have achieved some accuracy improvements in short-term numerical indicators, they still largely fall under the deterministic forecasting path. The forecast results are presented as single values, failing to reflect the range of uncertainty and risk levels, and are therefore insufficient to support scenarios requiring risk quantification, such as maritime operations and shipping disaster prevention.

[0003] In the field of probabilistic forecasting, a few existing studies have attempted to borrow the approach of atmospheric ensemble forecasting. This involves creating an ensemble of atmospheric models by perturbing initial conditions or driving wind fields, and then using this ensemble as input to drive wave models to generate an ensemble wave spectrum. While this method can provide some distribution information, it suffers from two shortcomings: first, wave processes are less sensitive to initial perturbations than atmospheric circulation, resulting in limited differences among ensemble members and making it difficult to form a sufficiently dispersed probability distribution; second, this method requires multiple runs of numerical models, leading to high computational costs and resource consumption, making it difficult to generalize for high-resolution or long-term series forecasting tasks. Meanwhile, current deep learning research is entirely focused on deterministic regression, and a technical route for directly modeling the probability distribution of wave spectra has not yet been established. Therefore, deep learning-based probabilistic wave forecasting methods remain a blank.

[0004] Based on the above analysis, the problems and shortcomings of the existing technology are as follows: (1) The forecast results of traditional wave forecasting methods are presented as a single value, which cannot reflect the range of uncertainty and risk level, and is difficult to support scenarios that require risk quantification, such as offshore operations and shipping disaster prevention.

[0005] (2) In the existing wave forecasting methods based on atmospheric ensembles, the wave process is not as sensitive to initial disturbances as atmospheric circulation, and the differences between ensemble members are limited, making it difficult to form a sufficiently dispersed probability distribution; the computational cost is high and the resource consumption is large, making it difficult to promote in high-resolution or long-term series forecasting tasks.

[0006] (3) Existing deep learning research is all focused on deterministic regression, and a technical route for directly modeling the probability distribution of wave spectrum has not yet been formed. There is still no wave probability prediction method based on deep learning. Summary of the Invention

[0007] To overcome the problems existing in related technologies, the present invention discloses an embodiment of a method and system for predicting ocean wave probability, particularly a method and system for predicting ocean wave probability based on deep learning, the technical solution of which is as follows: This invention is implemented as follows: a method for predicting ocean wave probability includes the following steps: S1. Acquire historical wind field data of the target sea area and wave spectrum data output by wave numerical models based on different physical parameterization schemes. Preprocess and organize the data to build a training dataset. S2, Construct a hybrid expert probabilistic model, which includes a gating network and multiple expert networks; wherein each expert network is configured to learn a mapping relationship of a physical parameterization scheme and output a wave spectrum probability distribution conditioned on historical wind field data; the gating network is configured to calculate the weights of each expert network based on the input historical wind field data; S3 uses a training dataset to train a hybrid expert probability model. The trained model receives input wind field data and outputs a hybrid probability distribution through the synergy of a gating network and an expert network. Sampling is performed from the hybrid probability distribution to obtain a set of predicted wave spectra, thus achieving wave risk probability prediction.

[0008] In step S1, obtaining historical wind field data for the target sea area includes: Obtaining wind field data: 10m height wind speed component in ERA5 reanalysis data The time resolution is 1 hour, and the spatial resolution is... Covering the target sea area and time period; Obtain model output data: Run the source term parameterization scheme of the WW3 wave model within the same sea area and time period. , No. Two-dimensional directional spectrum corresponding to the parameterization scheme for: ; In the formula, For frequency, This indicates the direction of wave propagation.

[0009] In step S1, the data preprocessing includes: Wind speed components are uniformly interpolated to the same grid, and the time series is standardized by dividing the average anemometer of the training set period. The spectral data is compressed using a logarithmic transformation, and the discrete resolution is maintained in the frequency and direction dimensions in accordance with the mode to ensure that the input and output have stable numerical scales.

[0010] In step S1, organizing the data includes: Time window design: The input wind field is formed by splicing together sequences from the past 6-12 hours to create the input vector. This is used to capture the time lag effect of wind on waves, and its expression is: ; In the formula, The duration of the input data. This is the initial time; for Time's up Moment , for Time's up Moment ; Input-output alignment: by wind speed component The obtained input vector As input labels, the output spectra are determined by each parameterization scheme. As output labels, they constitute the training sample pairs. ; Dataset setup: The WW3 mode was driven by ERA5 wind field for long-term operation to obtain spectral data covering typical sea states over many consecutive years; the dataset was divided in chronological order, with the first 80% used as the training set and the last 20% as the test set; 10% of the training set was further divided as the validation set for hyperparameter selection and early model stopping.

[0011] In step S2, the expert network for each parameterization scheme A multilayer perceptron structure is used, with the input being the wind field feature vector. The signal passes through two fully connected hidden layers, each containing 128 neurons and activated by ReLU, and finally outputs the conditional probability distribution of the predicted spectrum. The output layer consists of two parts, one of which is the mean of the predicted spectrum. One part is given directly using linear activation; the other part is the uncertainty covariance of the prediction spectrum. Softplus activation is used to ensure non-negativity, and the covariance matrix is ​​approximated as a diagonal form, thus yielding the conditional distribution. for: ; In the formula, The mean of the graph. For covariance, The mean of the graph is covariance is It follows a normal distribution.

[0012] Furthermore, gating networks are used for adaptive weight allocation among multiple expert networks to control the importance of different experts; gating networks The input is the wind field feature vector. The neurons pass through two fully connected hidden layers, the first containing 64 neurons and the second containing 32 neurons, both using ReLU activation and converted to normalized weights. The expression is: ; Where, For expert network numbering, For the first A gated network, For the first A gated network; The scheme achieves the best average performance by prioritizing the learning of this expert through a bias-guided model for parameterized schemes. Set a fixed positive bias ,plan To guide the gating network to prioritize weight allocation ; During training, entropy regularization constraints are introduced. To prevent all weight from completely collapsing to ; ; Where, This is a weighted index.

[0013] In step S2, the overall probability is obtained by combining the results of the gating network and the expert network. The distribution, expressed as: ; During inference, spectral samples are obtained from this mixed distribution to achieve probability prediction.

[0014] In step S3, training the hybrid expert probability model using the training dataset includes: A mixed distribution is used to apply the negative log-likelihood of the training samples, along with load balancing regularization, to fuse the multi-expert probability outputs into a unified distribution for training. Regularization suppresses weight collapse, and the loss function is... The expression is: ; Where, Weights for entropy regularization constraints; Optimization of the hybrid expert probabilistic model using the Adam optimizer, with an initial learning rate of... Batch size 64, training for 150 epochs.

[0015] In step S3, during the prediction phase, the wind field sequence is input. The model outputs a mixed distribution, yielding multiple sets of wave spectrum data; the distribution and confidence intervals of significant wave heights and peak periods are calculated to achieve risk probability prediction. Significant wave height The calculation formula is: ; In the formula, For the wave spectrum, For frequency; Spectral peak period The calculation formula is: ; In the formula, In order to be in The maximum value in the direction.

[0016] Another object of the present invention is to provide a wave probability prediction system for regulating the wave probability prediction method, the system comprising: The dataset construction module is used to acquire historical wind field data of the target sea area and wave spectrum data output by wave numerical models based on different physical parameterization schemes, preprocess and organize the data, and construct the training dataset. A hybrid expert probabilistic model construction module is used to construct a hybrid expert probabilistic model, which includes a gating network and multiple expert networks. Each expert network is configured to learn a mapping relationship of a physical parameterization scheme and output a wave spectrum probability distribution conditioned on historical wind field data. The gating network is configured to calculate the weights of each expert network based on the input historical wind field data. The wave risk probability prediction module is used to train a hybrid expert probability model using a training dataset. The trained model receives input wind field data and outputs a hybrid probability distribution through the synergy of a gating network and an expert network. By sampling from the hybrid probability distribution, the predicted wave spectrum set is obtained, thus realizing the wave risk probability prediction.

[0017] Combining all the above technical solutions, the beneficial effects of this invention are as follows: First, this invention breaks through the existing path of "deterministic regression or ensemble-driven" approaches, proposing a hybrid expert wave spectrum probability prediction method based on a multi-parameterization scheme. Expert networks are constructed for each of the commonly used source term parameterization schemes in WW3, directly learning the conditional distribution. A fixed positive bias is introduced into the gating network, incorporating the empirical advantages of ST6 as priors in the weight allocation, while entropy regularization suppresses weight collapse, maintaining a suitable diversity among multiple experts. The hybrid expert probability model of this invention outputs the mean and covariance of the spectral distribution, which can directly generate risk products such as quantile bands and overthreshold probabilities, achieving the goal of "accuracy comparable to ST6 while providing uncertainty measurement." Compared to technical solutions relying on atmospheric ensembles, this invention obtains the probability distribution with a single forward computation, significantly reducing computing power and latency, and facilitating engineering deployment.

[0018] Secondly, this invention takes the wind field characteristics within a time window as input, and outputs spectral conditional distributions from expert networks of corresponding parameterized schemes. A gating network applies a fixed positive bias to a specified scheme (ST6) and outputs normalized weights. Entropy regularization is introduced during training to prevent complete weight collapse. The negative log-likelihood of the mixed distribution is used as the main loss and jointly optimized with entropy regularization. In the prediction stage, no atmospheric or wave ensemble integration is required; the probability distribution of the spectrum can be directly output, and significant wave quantiles, over-threshold probabilities, and other risk indicators can be derived. The aforementioned gating mechanism with bias, the multi-expert wave spectrum probability modeling and training / inference process, and the system implementation for generating wave risk products without ensemble conditions constitute the core innovation and scope of protection of this invention.

[0019] Third, this invention enables probability prediction of ocean waves. The model directly outputs the probability distribution of the wave spectrum, providing three types of information: mean, interval, and risk. Compared to deterministic results that only provide single values, this can be used to quantify operational indicators such as "probability of exceeding ××m" and "95% confidence upper bound," significantly reducing the risk of underestimation or false alarms. In the embodiment, the observation sequence is effectively covered by the 5%-95% confidence interval, demonstrating its ability to characterize the uncertainty of future sea states and its decision-making usability. This invention integrates multiple source term parameterization schemes. Although ST6 is more accurate in an average sense, a single scheme may exhibit systematic biases in situations such as limited elongation wind areas, rapid wind direction changes, multi-peak spectra of swells and wind waves, and enhanced refraction and breaking in shallow water / nearshore areas. By gating and fusing various schemes (and applying a priori bias to ST6), the model maintains the accuracy advantage of ST6 under normal sea states and can adaptively introduce information from other schemes to correct biases under complex and extreme sea states, thereby improving robustness and generalization ability and reducing dependence on specific sea areas, seasons, and weather patterns. This invention has high computational efficiency. Traditional wave probability forecasts often rely on atmospheric ensembles, requiring multiple numerical integrations to obtain the ensemble spectrum. The computational cost and latency increase linearly with the size of the ensemble. This solution can obtain complete distribution information with a single forward calculation, without the need for numerical model ensembles. This significantly reduces computing power and energy consumption, making it easier to achieve high-frequency updates and near-real-time operational applications in resource-constrained environments such as shipboard, buoy, and nearshore sites.

[0020] Fourth, most existing methods focus on deterministic regression or ensemble forecasting driven by ensemble wind fields, which are difficult to simultaneously consider risk representation, robustness, and computational efficiency. However, the wave spectrum probabilistic modeling method based on hybrid experts proposed in this invention can directly output the probability distribution of wave spectrum and its derived statistics. While inheriting the average accuracy advantage of ST6, it enhances the applicability under complex and extreme sea conditions by gating and fusing multiple parameterization schemes. Furthermore, it replaces multiple numerical integrations with a single forward calculation, significantly reducing computational overhead and achieving a quantitative representation of uncertainty and risk. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure; Figure 1 This is a flowchart of the wave probability prediction method provided in an embodiment of the present invention; Figure 2 This is a comparison chart of deterministic and probabilistic forecasts provided in an embodiment of the present invention; Figure 3 This is a risk characterization diagram of extreme large wave events based on probability forecasting provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0023] The innovation of this invention lies in its departure from existing methods that rely on deterministic regression or ensemble-driven approaches. Instead, it proposes a hybrid expert method for predicting the probability of ocean waves based on a multi-parameterized scheme. Expert networks are constructed for each of the commonly used source term parameterization schemes in WW3, directly learning the conditional distribution. A fixed positive bias is introduced into the gating network, incorporating the empirical advantages of ST6 as prior knowledge into the weight allocation. Simultaneously, entropy regularization is used to suppress weight collapse, maintaining a suitable diversity among multiple experts. The hybrid expert probability model outputs the mean and covariance of the spectral distribution, directly generating risk products such as quantile bands and out-of-threshold probabilities, achieving the goal of "accuracy comparable to ST6 while providing uncertainty measurement." Compared to methods relying on atmospheric ensembles, this method obtains the probability distribution with a single forward computation, significantly reducing computational power and latency, and facilitating engineering deployment.

[0024] This invention takes wind field characteristics within a time window as input, and outputs spectral conditional distributions from expert networks corresponding to the parameterized schemes. A gating network applies a fixed positive bias to a specified scheme (ST6) and outputs normalized weights. Entropy regularization is introduced during training to prevent complete weight collapse. The negative log-likelihood of the mixed distribution is used as the main loss and jointly optimized with entropy regularization. In the prediction stage, no atmospheric or wave ensemble integration is required; the probability distribution of the spectrum can be directly output, and significant wave quantiles, over-threshold probabilities, and other risk indicators can be derived. The aforementioned gating with bias mechanism, multi-expert wave spectrum probability modeling and training / inference process, and system implementation for generating wave risk products without ensemble conditions constitute the core innovation and scope of protection of this invention.

[0025] Example 1, as Figure 1 As shown, the wave probability prediction method provided in this embodiment of the invention includes the following steps: S1. Acquire historical wind field data of the target sea area and wave spectrum data output by wave numerical models based on different physical parameterization schemes. Preprocess and organize the data to build a training dataset. To achieve probabilistic prediction of wave spectra, this invention constructs a training dataset based on numerical pattern output.

[0026] (1) Data collection Wind field data: 10m height wind speed component in ERA5 reanalysis data The time resolution is 1 hour, and the spatial resolution is... Covering the target sea area and time period; Model output data: Common source term parameterization schemes for running the WW3 wave model within the same sea area and time period. , No. Two-dimensional directional spectrum corresponding to the parameterization scheme for: ; In the formula, For frequency, This indicates the direction of wave propagation.

[0027] Data preprocessing: Wind speed components are first uniformly interpolated to the same grid, and then the time series is divided by the long-term mean and standardized; the spectral data is compressed by logarithmic transformation and the discrete resolution is kept consistent with the model in the frequency and direction dimensions to ensure that the input and output have stable numerical scales.

[0028] (2) Data organization Time window design: To capture the time lag effect of wind field on waves, the input wind field is formed by splicing together sequences from the past 6-12 hours to create the input vector. This is used to capture the time lag effect of wind on waves, and its expression is: ; In the formula, The duration of the input data. This is the initial time; for Time's up Moment , for Time's up Moment ; Input-output alignment: by wind speed component The obtained input vector As input labels, the output spectra are determined by each parameterization scheme. As output labels, they constitute the training sample pairs. ; Dataset Setup: The WW3 model was driven by the ERA5 wind field and run for an extended period to acquire spectral data covering typical sea states over several consecutive years, encompassing various scenarios such as weak winds, wind and wave development, and mixed sea states. The dataset was divided chronologically, with the first 80% used as the training set and the last 20% as the test set to prevent future information leakage. An additional 10% was allocated from the training set as a validation set for hyperparameter selection and early model stopping.

[0029] S2, Construct a hybrid expert probabilistic model, which includes a gating network and multiple expert networks; wherein each expert network is configured to learn a mapping relationship of a physical parameterization scheme and output a wave spectrum probability distribution conditioned on historical wind field data; the gating network is configured to calculate the weights of each expert network based on the input historical wind field data; (1) Expert network; Expert network for each parameterization scheme A multilayer perceptron structure is used, with the input being the wind field feature vector. The signal passes through two fully connected hidden layers, each containing 128 neurons activated by ReLU, and finally outputs the conditional probability distribution of the predicted spectrum. The output layer consists of two parts, one of which is the mean of the predicted spectrum. One part is given directly using linear activation; the other part is the uncertainty covariance of the prediction spectrum. Softplus activation is used to ensure non-negativity, and the covariance matrix is ​​approximated as a diagonal form. This yields the conditional distribution. for: ; In the formula, The mean of the graph. For covariance, The mean of the graph is covariance is It follows a normal distribution.

[0030] (2) Gated network; The core function of gating networks is to adaptively allocate weights among multiple expert networks to control the relative importance of different experts. The input is the wind field feature vector. The neurons pass through two fully connected hidden layers, the first containing 64 neurons and the second containing 32 neurons, both activated using ReLU. Finally, the neurons are converted to normalized weights. : ; In the formula, For expert network numbering, For the first A gated network, For the first A gated network; The scheme achieves the best average performance by prioritizing the learning of this expert through a bias-guided model for parameterized schemes. Set a fixed positive bias ,plan To guide the gating network to prioritize weight allocation ; During training, entropy regularization constraints are introduced. To prevent all weight from completely collapsing to ; ; In the formula, This is a weighted index.

[0031] (3) Mixed output; Combining gating and expert results, the overall probability distribution is obtained: ; During inference, spectral samples are obtained from this mixed distribution to achieve probability prediction.

[0032] S3 uses a training dataset to train a hybrid expert probability model. The trained model receives input wind field data and outputs a hybrid probability distribution through the synergy of a gating network and an expert network. Sampling is performed from the hybrid probability distribution to obtain a set of predicted wave spectra, thus achieving wave risk probability prediction.

[0033] (1) Training process; Loss function: The negative log-likelihood of the training samples using a mixture distribution, combined with load balancing regularization, fuses the probability outputs of multiple experts into a single overall distribution for training. Simultaneously, regularization suppresses weight collapse, improving the accuracy and stability of probability predictions; the expression is: ; In the formula, For loss function, Weights for entropy regularization constraints; Optimization of the hybrid expert probabilistic model using the Adam optimizer, with an initial learning rate of... Batch size 64, training for 150 epochs.

[0034] (2) Probability output; In the prediction phase, the wind field sequence is input. The model outputs a mixed distribution, yielding multiple sets of wave spectrum data. This allows for further calculation of the distribution and confidence intervals of statistical quantities such as significant wave height and peak period, enabling risk probability prediction.

[0035] Significant wave height The calculation formula is: ; In the formula, For the wave spectrum, For frequency; Spectral peak period The calculation formula is: ; In the formula, In order to be in The maximum value in the direction.

[0036] Example 2: The wave probability prediction system provided in this embodiment of the invention includes: The dataset construction module is used to acquire historical wind field data of the target sea area and wave spectrum data output by wave numerical models based on different physical parameterization schemes, preprocess and organize the data, and construct the training dataset. A hybrid expert probabilistic model construction module is used to construct a hybrid expert probabilistic model, which includes a gating network and multiple expert networks. Each expert network is configured to learn a mapping relationship of a physical parameterization scheme and output a wave spectrum probability distribution conditioned on historical wind field data. The gating network is configured to calculate the weights of each expert network based on the input historical wind field data. The wave risk probability prediction module is used to train a hybrid expert probability model using a training dataset. The trained model receives input wind field data and outputs a hybrid probability distribution through the synergy of a gating network and an expert network. By sampling from the hybrid probability distribution, the predicted wave spectrum set is obtained, thus realizing the wave risk probability prediction.

[0037] To further demonstrate the positive effects of the above embodiments, the present invention conducts the following experiments based on the above technical solutions.

[0038] like Figure 2 As shown, during a certain wave development process, wave height forecasts were performed using the ST6 deterministic model and the hybrid expert probabilistic model proposed in this invention, and the results were compared with buoy observations. The results show that the deterministic forecast only outputs a single wave height curve, which is insufficient to provide risk warnings when observed values ​​deviate from this curve. In contrast, the probabilistic model of this invention provides a mean prediction along with an uncertainty range of 5%-95%. Most of the observed results fall within this range, and the mean curve fits the observations well, significantly improving the risk representation ability of the forecast results and enabling a better characterization of future sea state uncertainties.

[0039] like Figure 3As shown, during an extreme high-wave event, wave height forecasts were performed using the ST6 deterministic model and the hybrid expert probabilistic model proposed in this invention, and the results were compared with buoy observations. It can be seen that the ST6 deterministic forecast significantly underestimated the peak wave height, failing to capture the observed extreme values; while the probabilistic model of this invention not only more closely approximates the observed values ​​on the mean curve, but also effectively covers the observed maximum wave height within the 5%-95% confidence interval. These results demonstrate that this invention can provide a reasonable upper bound on risk under extreme scenarios, avoiding underestimation of the probability of disastrous waves, thereby improving the robustness and practical value of the forecast.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting ocean wave probability, characterized in that, The prediction method includes the following steps: S1. Acquire historical wind field data of the target sea area and wave spectrum data output by wave numerical models based on different physical parameterization schemes. Preprocess and organize the data to build a training dataset. S2, Construct a hybrid expert probabilistic model, which includes a gating network and multiple expert networks; wherein each expert network is configured to learn a mapping relationship of a physical parameterization scheme and output a wave spectrum probability distribution conditioned on historical wind field data; the gating network is configured to calculate the weights of each expert network based on the input historical wind field data; S3 uses a training dataset to train a hybrid expert probability model. The trained model receives input wind field data and outputs a hybrid probability distribution through the synergy of a gating network and an expert network. Sampling is performed from the hybrid probability distribution to obtain a set of predicted wave spectra, thus achieving wave risk probability prediction.

2. The wave probability prediction method according to claim 1, characterized in that, In step S1, obtaining historical wind field data for the target sea area includes: Obtaining wind field data: 10m height wind speed component in ERA5 reanalysis data The time resolution is 1 hour, and the spatial resolution is... Covering the target sea area and time period; Obtain model output data: Run the source term parameterization scheme of the WW3 wave model within the same sea area and time period. , No. Two-dimensional directional spectrum corresponding to the parameterization scheme for: ； In the formula, For frequency, This indicates the direction of wave propagation.

3. The wave probability prediction method according to claim 1, characterized in that, In step S1, the data preprocessing includes: Wind speed components are uniformly interpolated to the same grid, and the time series is standardized by dividing the average anemometer of the training set period. The spectral data is compressed using a logarithmic transformation, and the discrete resolution is maintained in the frequency and direction dimensions in accordance with the mode to ensure that the input and output have stable numerical scales.

4. The wave probability prediction method according to claim 1, characterized in that, In step S1, organizing the data includes: Time window design: The input wind field is formed by splicing together sequences from the past 6-12 hours to create the input vector. This is used to capture the time lag effect of wind on waves, and its expression is: ； Where, The duration of the input data. This is the initial time; for Time's up Moment , for Time's up Moment ; Input-output alignment: by wind speed component The obtained input vector As input labels, the output spectra are determined by each parameterization scheme. As output labels, they constitute the training sample pairs. ; Dataset setup: The WW3 mode was driven by ERA5 wind field for long-term operation to obtain spectral data covering typical sea states over many consecutive years; the dataset was divided in chronological order, with the first 80% used as the training set and the last 20% as the test set; 10% of the training set was further divided as the validation set for hyperparameter selection and early model stopping.

5. The wave probability prediction method according to claim 1, characterized in that, In step S2, the expert network for each parameterization scheme A multilayer perceptron structure is used, with the input being the wind field feature vector. The signal passes through two fully connected hidden layers, each containing 128 neurons and activated by ReLU, and finally outputs the conditional probability distribution of the predicted spectrum. The output layer consists of two parts, one of which is the mean of the predicted spectrum. One part is given directly using linear activation; the other part is the uncertainty covariance of the prediction spectrum. Softplus activation is used to ensure non-negativity, and the covariance matrix is ​​approximated as a diagonal form, thus yielding the conditional distribution. for: ； In the formula, The mean of the graph. For covariance, The mean of the graph is covariance is It follows a normal distribution.

6. The wave probability prediction method according to claim 5, characterized in that, Gated networks are used to adaptively allocate weights among multiple expert networks, controlling the importance of different experts; gated networks The input is the wind field feature vector. The neurons pass through two fully connected hidden layers, the first containing 64 neurons and the second containing 32 neurons, both using ReLU activation and converted to normalized weights. The expression is: ； In the formula, For expert network numbering, For the first A gated network, For the first A gated network; The scheme achieves the best average performance by prioritizing the learning of this expert through a bias-guided model for parameterized schemes. Set a fixed positive bias ,plan To guide the gating network to prioritize weight allocation ; During training, entropy regularization constraints are introduced. To prevent all weight from completely collapsing to ; ； In the formula, This is a weighted index.

7. The wave probability prediction method according to claim 6, characterized in that, In step S2, the overall probability is obtained by combining the results of the gating network and the expert network. The distribution, expressed as: ； During inference, spectral samples are obtained from this mixed distribution to achieve probability prediction.

8. The wave probability prediction method according to claim 7, characterized in that, In step S3, training the hybrid expert probability model using the training dataset includes: A mixed distribution is used to apply the negative log-likelihood of the training samples, along with load balancing regularization, to fuse the multi-expert probability outputs into a unified distribution for training. Regularization suppresses weight collapse, and the loss function is... , the expression is: ； In the formula, Weights for entropy regularization constraints; Optimization of the hybrid expert probabilistic model using the Adam optimizer, with an initial learning rate of... Batch size 64, training for 150 epochs.

9. The wave probability prediction method according to claim 1, characterized in that, In step S3, during the prediction phase, the wind field sequence is input. The model outputs a mixed distribution, resulting in multiple sets of wave spectrum data; Calculate the distribution and confidence interval of significant wave heights and peak periods to achieve risk probability prediction; Significant wave height The calculation formula is: ； In the formula, For the wave spectrum, For frequency; Spectral peak period The calculation formula is: ； In the formula, In order to be in The maximum value in the direction.

10. A wave probability prediction system, characterized in that, This system is used to regulate the wave probability prediction method according to any one of claims 1-9, and the system includes: The dataset construction module is used to acquire historical wind field data of the target sea area and wave spectrum data output by wave numerical models based on different physical parameterization schemes, preprocess and organize the data, and construct the training dataset. A hybrid expert probabilistic model construction module is used to construct a hybrid expert probabilistic model, which includes a gating network and multiple expert networks. Each expert network is configured to learn a mapping relationship of a physical parameterization scheme and output a wave spectrum probability distribution conditioned on historical wind field data. The gating network is configured to calculate the weights of each expert network based on the input historical wind field data. The wave risk probability prediction module is used to train a hybrid expert probability model using a training dataset. The trained model receives input wind field data and outputs a hybrid probability distribution through the synergy of a gating network and an expert network. By sampling from the hybrid probability distribution, the predicted wave spectrum set is obtained, thus realizing the wave risk probability prediction.

Images