A global wave height prediction method based on multi-missing rate view semantic alignment

By constructing teacher and student models and using the STID model for semantic alignment of multiple missing rate views, the problem of missing data in global wave height prediction was solved, achieving stable prediction and efficient inference under different missing rate conditions, and improving the robustness and generalization ability of the model.

CN123069239APending Publication Date: 2026-11-10DALIAN MARITIME UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610881336.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-06-17
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing global wave height prediction methods struggle to achieve a unified model, rapid inference, and stable generalization across global sea areas when faced with missing data. Furthermore, traditional completion models may weaken or amplify local sequence structures, leading to error accumulation and failing to meet operational prediction requirements.

Method used

We employ a semantic alignment method based on multiple missing rate views. By constructing teacher and student models and utilizing the spatial-temporal identity embedding model STID, we perform semantic alignment of multiple missing rate views. Combined with knowledge distillation and multi-view comparative learning, we train the student model to adapt to different missing rate conditions and achieve global wave height prediction.

Benefits of technology

It improves the prediction robustness under conditions of variable missing rates, maintains complete observation semantic information, enhances the joint prediction capability of multiple points in global sea areas, reduces computational complexity and cost, and enhances the model's generalization ability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069239A_ABST
    Figure CN123069239A_ABST
Patent Text Reader

Abstract

The application discloses a global wave height prediction method based on multi-missing rate view semantic alignment, and the method comprises the following steps: acquiring complete sample data, and applying different missing rate masks to complete observation input tensors in the complete sample data to construct incomplete sample data; constructing a teacher model and a student model based on an STID backbone model; training the teacher model by using the complete sample data to obtain a teacher model hidden representation and a teacher model prediction result; constructing a model training total loss function for training the student model based on the teacher model hidden representation and the teacher model prediction result, training the student model by using the incomplete sample data based on the loss function to obtain an optimal prediction model, and realizing global wave height prediction based on multi-missing rate view semantic alignment by using the optimal prediction model. The application solves the problem that the existing method is difficult to meet the unified model, fast reasoning and stable generalization requirements in the global sea area business prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary integration technology of marine engineering, marine meteorological forecasting and artificial intelligence, and in particular to a global wave height prediction method based on semantic alignment of multiple missing rate views. Background Technology

[0002] Significant wave height is a crucial parameter describing sea surface wave states and the marine dynamic environment, with significant applications in areas such as safe ocean navigation, offshore wind power operation and maintenance, marine engineering construction, ship route planning, and marine disaster early warning. Wave evolution across global ocean areas is influenced by multiple factors, including wind fields, pressure fields, air-sea interactions, ocean currents, and regional weather systems, exhibiting pronounced nonlinearity, spatiotemporal coupling, and multi-scale variation characteristics. Therefore, constructing a significant wave height prediction model that can consider both global spatial differences and temporal dynamic evolution characteristics is an important research direction in the field of intelligent ocean forecasting.

[0003] Currently, significant wave height prediction mainly includes numerical models based on physical processes and data-driven machine learning methods. Numerical models are typically based on the spectral energy balance equation, simulating waves through physical processes such as wind input, nonlinear interactions, white wave dissipation, bottom friction, and breaking, thus possessing a strong physical foundation. However, these methods rely on high-quality initial fields, boundary conditions, and parameterization schemes, and suffer from problems such as high computational cost, slow update speed, and local error propagation in large-scale global, high-temporal-resolution, or rapid operational forecasting scenarios. With the development of deep learning technology, multivariate time series prediction models are gradually being applied to wave prediction tasks. These methods can automatically learn nonlinear mapping relationships from historical wind fields, pressure fields, and wave sequences, avoiding explicit solutions to complex physical equations, and possessing high computational efficiency and strong data adaptability. However, in practical applications, global ocean observation and reanalysis data are often affected by factors such as observation equipment failure, remote sensing inversion errors, communication interruptions, data quality control, and data transmission delays, resulting in varying degrees of missing data in historical input sequences. Missing values ​​not only disrupt the periodicity, trend, and local variation characteristics of the original time series, but also cause the missing rate view to change continuously at different time periods, thereby weakening the generalization ability and robustness of traditional deep learning models.

[0004] To address the issue of missing data, existing methods typically employ interpolation, KNN, matrix completion, or deep completion models to recover missing values ​​before inputting the completed results into the prediction model. While these two-stage methods can alleviate some of the missing data problem, some completion methods may smooth or shift the local sequence structure during the missing value recovery process, weakening or amplifying local fluctuations and further introducing error accumulation. Furthermore, when the missing rate view changes over different time periods, traditional completion or prediction models often require separate training for different missing rate views, making it difficult to meet the requirements of a unified model, rapid inference, and stable generalization in operational global maritime forecasting. Summary of the Invention

[0005] This invention provides a global wave height prediction method based on semantic alignment of multiple missing rate views to overcome the above-mentioned technical problems.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: A global wave height prediction method based on semantic alignment of multiple missing rate views, specifically including the following steps: S1: Obtain historical multivariate time series data of pre-defined representative spatial points in global sea areas, and construct input feature vectors by introducing intraday and intraweek time features to obtain a complete observation input tensor; combine the complete observation input tensor and its corresponding significant wave height labels for future times to obtain complete sample data; Random masks with different missing rates are applied to the observation variables in the complete observation input tensor, and intraday and intraweek time features are used as time identity index features to construct incomplete input tensors under different missing rate view conditions. The incomplete input tensors are then combined with the future significant wave height labels in the corresponding complete sample data to obtain incomplete sample data. The observation variables include historical significant wave height, 10 m height U-wind component, 10 m height V-wind component, and sea level pressure. S2: Construct teacher and student models based on the spatial-temporal identity embedding model STID; the STID backbone model includes a fully connected embedding layer, a spatiotemporal embedding splicing layer, a residual multilayer perceptron coding layer, and an output layer; The fully connected embedding layer is used to map the input tensor to a hidden representation vector; The spatiotemporal embedding and splicing layer is used to perform dimensional splicing of the intraday time features, intraweek time features, and spatial embedding features of representative spatial points of the input tensor according to the hidden representation vector, so as to obtain a fused input representation vector that integrates spatiotemporal information. The residual multilayer perceptron coding layer is used to obtain the coded hidden representation vector based on the fused input representation vector; The output layer is used to output future significant wave height prediction data based on the encoded hidden representation vector; S3: Train the teacher model using complete sample data to obtain the teacher model's hidden representation and prediction results; and construct a total model training loss function for training the student model based on the teacher model's hidden representation and prediction results. Train the student model using incomplete sample data based on the total model training loss function to obtain a student model that meets the preset convergence condition and use it as the final prediction model. Use the final prediction model to achieve global wave height prediction based on semantic alignment of multiple missing rate views.

[0007] Furthermore, the formula for obtaining the hidden representation vector described in S2 is:

[0008]

[0009]

[0010]

[0011] In the formula: This represents the hidden representation of the input tensor after it has passed through the embedding layer; Represents the mapping function of the fully connected layer; and This represents the input tensor to be embedded, which corresponds to the input teacher model. When the corresponding input is a student model ; Represents the complete observation input tensor; This indicates an incomplete input tensor. Indicates the first Representative spatial points at time The model input feature vector; Indicates the first Representative spatial points at time Historically significant highs; Indicates intraday time characteristics; Indicates the time characteristics within a week; This represents the U-wind component at a height of 10 m; This represents the V-wind component at a height of 10 m; Indicates air pressure at sea level; Indicates vector transpose; Represents the real space to which the tensor belongs; This represents the number of representative spatial points in the global ocean area; Indicates the length of the historical input; This indicates the number of input features, which includes historical significant wave heights, intraday time features, intraweek time features, 10 m height U-wind component, 10 m height V-wind component, and sea level pressure. Relative to the current time Historical step index for backward tracing; Indicates the first Incomplete input tensors under the condition of missing rate view; Represents a binary mask matrix; This represents element-wise multiplication; Indicates the first The missing rate view corresponding to each input view.

[0012] Furthermore, the fusion input representation vector for obtaining fused spatiotemporal information, as described in S2, is:

[0013] In the formula: This represents the fusion input representation vector that integrates spatiotemporal information; This indicates a tensor concatenation operation; Spatial identity embedding refers to spatial embedding features; This represents the embedding vector corresponding to the intraday time identity feature; This represents the embedding vector of intra-week time identity, which corresponds to the intra-week time feature. Indicates the embedding dimension.

[0014] Furthermore, the formula for obtaining the encoded hidden representation vector described in S2 is as follows:

[0015]

[0016] In the formula: This represents the initial input to the coding layer of the residual multilayer perceptron; Indicates the first Hidden representations output from the coding layer of a multilayer perceptron with layer residuals; This represents the fusion input representation vector that integrates spatiotemporal information; This indicates the number of encoder layers in a residual multilayer perceptron. Indicates the first Nonlinear mapping functions in multilayer perceptrons with layer residuals; Represents the linear rectification activation function; The expression for the output layer is: In the formula: This indicates the output of the predicted future significant wave height; Indicates the output regression layer; This represents the hidden representation vector output by the last layer of the coding layer in a residual multilayer perceptron.

[0017] Furthermore, step S3 specifically includes the following steps: S31: Randomly divide the complete sample data into a complete sample training set and a complete sample validation set according to a preset ratio; and randomly divide the incomplete sample data into an incomplete sample training set and an incomplete sample validation set. S32: With the goal of minimizing the teacher-supervised prediction loss, the teacher model is trained using the complete sample training set to obtain the trained teacher model that outputs the teacher model's hidden representation and the teacher model's prediction results. The teacher-supervised prediction loss is the regression loss between the teacher model's prediction results and their corresponding future significant wave height labels. The teacher model's hidden representation and the teacher model's prediction results are as follows:

[0018] In the formula: The function representation of the teacher model; Indicates the teacher model parameters; This represents the hidden representation output by the last encoder layer of the teacher model; This represents the predicted significant wave height obtained from the teacher model; Represents the complete observation input tensor; The output of the trained teacher model is used as a pseudo-label for the distillation loss during the training of the student model. The student model is trained using an incomplete sample training set to obtain the trained student model corresponding to the hidden representation of the output student model and the prediction result of the student model. The hidden representations and prediction results of the student model are as follows:

[0019] In the formula: The function representation of the student model; Indicates the parameters of the student model; Indicates the first Incomplete input tensors under the condition of missing rate view; The student model is represented in the first... The last hidden representation obtained from the missing rate view input; The student model is represented in the first... The output of the future significant wave height prediction results under the input of the missing rate view; This indicates the number of views showing the missing rate.

[0020] S33: Based on the total loss function of model training, the complete sample validation set and the incomplete sample validation set are respectively input into the trained teacher model and the trained student model. The trained student model is judged according to the preset convergence condition. If it is, the trained student model is confirmed to be the optimal prediction model. The preset convergence condition includes the total loss of model training meeting the preset loss threshold or reaching the maximum number of training rounds. Otherwise, the student model parameters are adaptively updated based on backpropagation, and the student model training process is repeated until the trained student model meets the preset convergence condition to obtain the optimal prediction model; the optimal prediction model is used to realize global wave height prediction based on semantic alignment of multiple missing rate views.

[0021] Furthermore, the total loss function for model training constructed in S33 for:

[0022] In the formula: Indicates the weight of the predicted loss under supervision; Indicates the weight of knowledge distillation loss; Indicates the weights of the multi-view contrast learning loss; This indicates the monitoring and prediction of losses; This represents the total learning loss from multi-view comparison; This indicates that the loss in distillation is hidden. This indicates the predicted distillation loss; The supervised prediction loss for:

[0023] In the formula: Represents the true value of significant wave height at a future time. The student model is represented in the first... Given a missing rate view input, the predicted significant wave height for future time periods; This represents the absolute value operation; The total learning loss of multi-view comparison for:

[0024]

[0025]

[0026]

[0027]

[0028] In the formula: This represents the number of missing data points in the input set containing both positive and negative sample pairs. Low-dimensional contrastive learning representations corresponding to the missing rate views; Represents the cosine similarity function; Representing low-dimensional contrastive learning representations Representation of any two low-dimensional samples in the dataset; Represents the L2 norm; Indicates transpose; This represents the contrast loss between any two low-dimensional sample representations; These represent low-dimensional contrastive learning representations. Different samples in ; Represents an exponential function; Indicates temperature parameter; Indicates the number of samples in the batch; This represents the total number of samples participating in the comparative learning after the two views are stitched together; Indicates the first Missing rate view With the Missing rate view Pairwise contrast loss; Indicated based on the contrast loss formula The contrast loss of the same positive sample in two directions; The student model is represented in the first... Hidden representations under the input condition of missing rate view; The hidden characterization distillation loss is:

[0029] In the formula: Indicates the sign of the average over tensor elements; The predicted distillation loss is: .

[0030] This invention provides a global wave height prediction method based on semantic alignment of multiple missing rate views, with the following beneficial effects: (1) Improving prediction robustness under variable missing rate conditions: This invention constructs multiple missing rate views during the training phase and constrains the student model to learn consistent semantic representations under different missing rate inputs through offline knowledge distillation and multi-view comparative learning. Compared with training the model separately for a single missing rate, this invention can utilize the student model after one training to adapt to preset missing rate and similar missing rate conditions, reduce the complexity of model deployment, and improve robustness in global maritime operational prediction scenarios.

[0031] (2) Maintain complete observation semantic information: This invention first uses complete observations to train the teacher model, and then uses the hidden representations and prediction results of the teacher model to constrain the student model, so that the student model can learn the global periodic information, local fluctuation information and spatial difference information close to the complete observation when dealing with incomplete observations, thereby reducing the damage of missing values ​​to the semantic structure of wind field, sea level pressure and wave height historical sequence.

[0032] (3) Enhancing the joint prediction capability of multiple points in global sea areas: This invention uses pre-set representative spatial points in global sea areas as unified multivariate time series inputs. The spatial identity embedding in the STID backbone model represents the spatial differences between different sea area points, and the temporal identity embedding represents the intraday and intraweek temporal variation characteristics. This structure avoids building a separate model for each spatial point and can achieve a unified output of the significant wave height of representative spatial points in global sea areas at future times.

[0033] (4) Improved training and inference efficiency: This invention uses a lightweight STID model as the backbone model, which has lower computational complexity compared to complex graph neural networks or Transformer structures. At the same time, this invention only uses the student model in the testing phase, without the need for the teacher model to participate in inference, which can reduce the computational cost in practical applications.

[0034] (5) Enhance the generalization ability of the model: This invention constructs positive sample pairs and negative sample pairs for different time period samples by using multi-view comparison learning to construct different time period samples with different missing rates. This enables the model to align multiple missing rate inputs under the same physical state in the representation space, while distinguishing samples under different physical states. This mechanism helps to improve the model's generalization ability for different time periods, different sea conditions and different missing rate combinations. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of the global wave height prediction method based on semantic alignment of multiple missing rate views according to the present invention; Figure 2 This is a schematic diagram of the waveform height prediction network model framework based on semantic alignment of multiple missing rate views in this embodiment; Figure 3 This is a scatter plot of the significant wave height prediction results for the optimal site under a 25% missing rate and a T+1 prediction time in this embodiment. Figure 4 This is a scatter plot of the significant wave height prediction results for the optimal site under a 25% missing rate and a T+3 prediction timeframe in this embodiment. Figure 5 This is a scatter plot of the significant wave height prediction results for the optimal site under a 25% missing rate and a T+6 prediction timeframe in this embodiment. Figure 6 This is a scatter plot of the significant wave height prediction results for the optimal station under a 25% missing rate and a T+12 prediction timeframe in this embodiment. Figure 7 This is a scatter plot of the significant wave height prediction results for the optimal site under a 50% missing rate and a T+1 prediction timeframe in this embodiment. Figure 8 This is a scatter plot of the significant wave height prediction results for the optimal site under a 50% missing rate and a T+3 prediction timeframe in this embodiment. Figure 9 This is a scatter plot of the significant wave height prediction results for the optimal site under a 50% missing rate and a T+6 prediction timeframe in this embodiment. Figure 10 This is a scatter plot of the significant wave height prediction results for the optimal station under a 50% missing rate and a prediction time of T+12 in this embodiment. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] To address the challenges of multi-source coupling of input variables, a large number of spatial points, variable historical sequence missing rates, and insufficient robustness of traditional models in short-term prediction of significant wave heights in global sea areas, representative spatial points in global sea areas can be pre-determined based on uniform sampling using a latitude and longitude grid, sea area coverage, areas of operational interest, and the completeness of historical data. In this embodiment, 99 representative spatial points in global sea areas are selected as the research objects. East-west wind speed at 10 m altitude, north-south wind speed at 10 m altitude, historical significant wave heights, and sea level pressure are used as input variables, and the significant wave height for the next 12 hours is used as the output variable. During the training phase, the teacher model is trained using complete observations, while the student model is trained using incomplete observations constructed with different missing rate masks. During the testing phase, only the trained student model is retained, directly outputting the significant wave height prediction results for the next 12 hours. The multivariate time series data is publicly available ERA5 reanalysis data for global sea areas, including at least historical significant wave heights, the U-wind component at 10 m altitude, the V-wind component at 10 m altitude, and sea level pressure. This embodiment provides a global wave height prediction method based on semantic alignment of multiple missing rate views, using hourly ERA5 reanalysis data from January 1, 2016 to December 31, 2025, such as... Figure 1 As shown, the specific steps include: S1: Obtain historical multivariate time series data of pre-defined representative spatial points in global sea areas, and construct input feature vectors by introducing intraday and intraweek time features to obtain a complete observation input tensor; combine the complete observation input tensor and its corresponding significant wave height labels for future times to obtain complete sample data; Random masks with different missing rates are applied to the observation variables in the complete observation input tensor, and intraday and intraweek time features are used as time identity index features to construct incomplete input tensors under different missing rate view conditions. The incomplete input tensors are then combined with the future significant wave height labels in the corresponding complete sample data to obtain incomplete sample data. The observation variables include historical significant wave height, 10 m height U-wind component, 10 m height V-wind component, and sea level pressure. Specifically, we acquired and organized multivariate time series data from 99 representative spatial points in global sea areas, determined the historical significant wave height, 10 m height U-wind component, 10 m height V-wind component, and sea level pressure for each spatial point at each moment, and constructed the model input feature vector, complete historical observation input tensor, future significant wave height real sequence, and multivariate time series prediction mapping relationship by introducing intraday time feature TOD and intraweek time feature DOW. Among them, the first The input feature vector of a representative spatial point at time t is defined as:

[0039] In the formula: Indicates the first Representative spatial points at time The input feature vector; Indicates the first Representative spatial points at time Historically significant highs; This indicates the time characteristics within a day, corresponding to a 24-hour day; This indicates the time characteristics within a week, corresponding to the 7 days of a week; This represents the U-wind component at a height of 10 m; This represents the V-wind component at a height of 10 m; Indicates air pressure at sea level; Indicates vector transpose; This embodiment constructs a complete observation input tensor based on the historical sequence of the past 12 hours:

[0040] In the formula: Indicates the first Representative spatial points at time The input feature vector; Indicates the current prediction time The corresponding complete historical observation input tensor; Represents the real space to which the tensor belongs; This represents the number of spatial points in global sea areas; Indicates the length of the historical input; The input features are represented in this embodiment, including historical significant wave heights, intraday time features, intraweek time features, 10 m height U-wind component, 10 m height V-wind component, and sea level pressure. This represents the historical step size index relative to the current prediction time. The target output of this embodiment is the significant wave height sequence of representative spatial points over the next 12 hours:

[0041] In the formula: This indicates the true sequence of significant wave heights for the next 12 hours; Indicates the first The representative spatial points will be in the future. Significant wave height at each predicted time; Indicates the predicted length; Indicates the future prediction step size index; Based on this, the multivariate time series prediction mapping relationship is constructed as follows:

[0042] In the formula: This represents the model's output of the significant wave height prediction results for the next 12 hours; Indicates by parameters Predictive models for control; This represents the complete observation input tensor. The core prediction objective of this embodiment is to predict significant wave heights for the next 12 hours using multi-source meteorological and wave information from the past 12 hours. In this embodiment, a random mask is applied to the complete historical observation input tensor to construct incomplete observation inputs under different missing rate conditions. These different missing rate views include views with missing rates of 25%, 50%, 75%, and 90%. The random mask applies to historical significant wave height, the 10 m height U-wind component, the 10 m height V-wind component, and sea level pressure, while intraday and intraweek time indices are not included in the random missing data. The incomplete input tensor is defined as follows:

[0043] In the formula: Indicates the first Incomplete input tensors under various missing rate conditions; random sampling without replacement is performed on the time series location of each representative spatial point, according to the missing rate. Determine the number of missing positions, set the corresponding positions to 0, and set the rest to 1 to obtain a binary mask matrix. Among them, the missing locations corresponding to 25%, 50%, 75%, and 90% missing rates form nested sets in sequence; This represents element-wise multiplication; Indicates the first The missing rate corresponding to the missing rate view; in this embodiment, it is set as follows: The missing rate view is a data matrix, namely 25%, 50%, 75%, and 90%.

[0044] S2: Construct teacher and student models based on the spatial-temporal identity embedding model STID; the STID backbone model includes a fully connected embedding layer, a spatiotemporal embedding splicing layer, a residual multilayer perceptron coding layer, and an output layer; Specifically, such as Figure 2As shown, teacher and student models based on the STID backbone model are constructed using the complete historical observation input tensor (i.e., the complete input tensor) and the incomplete observation input tensor (i.e., the incomplete input tensor). Both the teacher and student models are significant wave height prediction models with STID as the backbone, employing the same STID backbone structure, but their model parameters are independent. The teacher model's input is the complete historical observation input tensor, and its output is the teacher's predicted significant wave height and hidden representations. The student model's input is the incomplete observation input tensor under different missing rates, and its output is the student's predicted significant wave height and hidden representations. The fully connected embedding layer is used to map the input tensor to a hidden representation vector; Specifically, the formula for obtaining the hidden representation vector is:

[0045] In the formula: This represents the hidden representation of the input tensor after it has passed through the embedding layer; Represents the mapping function of the fully connected layer; and This represents the input tensor to be embedded, which corresponds to the input teacher model. When the corresponding input is a student model ; Represents the complete observation input tensor; This indicates an incomplete input tensor; The spatiotemporal embedding and splicing layer is used to perform dimensional splicing of the intraday time features, intraweek time features, and spatial embedding features of representative spatial points of the input tensor according to the hidden representation vector, so as to obtain a fused input representation vector that integrates spatiotemporal information. Specifically, the fusion input representation vector for obtaining fused spatiotemporal information is:

[0046] In the formula: This represents the fusion input representation vector that integrates spatiotemporal information; This indicates a tensor concatenation operation; Spatial identity embedding, or spatial embedding feature, is used to characterize the spatial differences among 99 spatial points in the sea area. The embedding vector representing intraday time identity is the embedding vector corresponding to intraday time features, used to characterize hourly time location. The embedding vector representing the intra-week time identity is the embedding vector corresponding to the intra-week time features, used to characterize the time position within the week cycle; The embedding dimension is represented by the spatial identity embedding, which is replicated in the time dimension. The intraday time identity embedding and the intraweek time identity embedding are replicated in the representative spatial point dimension, so that the spatial identity embedding, intraday time identity embedding, intraweek time identity embedding and the hidden representation vector are aligned in the batch dimension, spatial dimension and time dimension, and spliced ​​in the feature dimension. The residual multilayer perceptron coding layer is used to obtain the coded hidden representation vector based on the fused input representation vector; Specifically, the coding layer of the residual multilayer perceptron consists of L layers of multilayer perceptrons with residual connections, the third... l The layer coding process is represented as follows:

[0047]

[0048] In the formula: This represents the initial input to the coding layer of the residual multilayer perceptron; Indicates the first Hidden representations output from the coding layer of a multilayer perceptron with layer residuals; This represents the fusion input representation vector that integrates spatiotemporal information; This indicates the number of encoder layers in a residual multilayer perceptron. Indicates the first Nonlinear mapping functions in multilayer perceptrons with layer residuals; This represents the linear rectified activation function; the part to the right of the plus sign... This indicates a residual connection.

[0049] The output layer is used to output future significant wave height prediction data based on the encoded hidden representation vector; Specifically, in this embodiment, after L-layer encoding, the regression layer, i.e. the output layer, outputs the prediction results of significant wave heights for the next 12 hours. The expression for the output layer is: In the formula: This indicates the output of the predicted future significant wave height; Indicates the output regression layer; This represents the hidden representation vector output by the last layer of the coding layer in a residual multilayer perceptron.

[0050] S3: Train the teacher model using complete sample data to obtain the teacher model's hidden representation and prediction results; and construct a total model training loss function for training the student model based on the teacher model's hidden representation and prediction results. Train the student model using incomplete sample data based on the total model training loss function to obtain a student model that meets the preset convergence condition and use it as the final prediction model. Use the final prediction model to achieve global wave height prediction based on semantic alignment of multiple missing rate views.

[0051] In this embodiment, a teacher model is trained using complete observation input to obtain the teacher model's hidden representation and prediction results. Simultaneously, a student model is constructed using the same STID structure as the teacher model, and incomplete observations under different missing rate conditions are input into the student model to obtain the corresponding hidden representation and significant wave height prediction results. The semantic alignment in this embodiment refers to: constraining the hidden representations of the student model obtained from the same time sample under different missing rate views to be mutually close in the representation space; constraining them to be close to the teacher model's hidden representation under complete observation input through hidden representation distillation; and constraining the student model's prediction results to be close to the teacher model's prediction results through prediction result distillation. Specifically, this includes the following steps: S31: Divide the complete sample data into a complete sample training set and a complete sample validation set according to a preset ratio; divide the complete sample data into a complete sample training set, a complete sample validation set, and a complete sample test set in chronological order, with division ratios of 70%, 10%, and 20%, respectively; divide the incomplete sample data under different missing rate conditions according to the same time range and sample index to obtain an incomplete sample training set, an incomplete sample validation set, and an incomplete sample test set; and the complete sample test set and the incomplete sample test set do not participate in model training, but are only used in the subsequent model testing stage; S32: With the goal of minimizing the teacher-supervised prediction loss, the teacher model is trained using the complete sample training set to obtain the trained teacher model that outputs the teacher model's hidden representation and the teacher model's prediction results. The teacher-supervised prediction loss is the regression loss between the teacher model's prediction results and their corresponding future significant wave height labels. The teacher model's hidden representation and the teacher model's prediction results are as follows:

[0052] In the formula: The function representation of the teacher model; Indicates the teacher model parameters; This represents the encoded hidden representation vector output by the last encoder layer of the teacher model; This represents the significant wave height prediction results for future moments obtained by the teacher model, i.e., the significant wave height prediction results for the next 12 hours. Represents the complete observation input tensor; The output of the trained teacher model is used as a pseudo-label for the distillation loss during the training of the student model. The student model is trained using an incomplete sample training set to obtain the trained student model corresponding to the hidden representation and prediction result of the output student model. In this embodiment, the parameters of the trained teacher model are frozen during the student model training stage and are only used for the hidden representation and prediction result under the condition of complete observation of the output. The hidden representations and prediction results of the student model are as follows:

[0053] In the formula: The function representation of the student model; Indicates the parameters of the student model; Indicates the first Incomplete input tensors under the condition of missing rate view; The student model is represented in the first... The last hidden representation obtained from the missing rate view input; The student model is represented in the first... The output of the future significant wave height prediction results under the input of the missing rate view; The number of views indicating the missing rate; S33: Based on the total loss function of model training, the complete sample validation set and the incomplete sample validation set are respectively input into the trained teacher model and the trained student model. The trained student model is judged according to the preset convergence condition. If it is, the trained student model is confirmed to be the optimal prediction model. The preset convergence condition includes the total loss of model training meeting the preset loss threshold or reaching the maximum number of training rounds. Otherwise, the student model parameters are adaptively updated based on backpropagation, and the student model training process is repeated until the trained student model meets the preset convergence condition to obtain the optimal prediction model. The optimal prediction model is then used to achieve global wave height prediction based on semantic alignment of multiple missing rate views.

[0054] In a specific embodiment, the total loss function for model training constructed in S33 for:

[0055] In the formula: Indicates the weight of the predicted loss under supervision; Indicates the weight of knowledge distillation loss; Indicates the weights of the multi-view contrast learning loss; This indicates the monitoring and prediction of losses; This represents the total learning loss from multi-view comparison; This indicates that the loss in distillation is hidden. This indicates the predicted distillation loss; The supervised prediction loss for:

[0056] In the formula: Represents the true value of significant wave height at a future time. The student model is represented in the first... Given a missing rate view input, the predicted significant wave height for future time periods; This expression represents the absolute value operation; it corresponds to the mean absolute error constraint in regression prediction. The total learning loss of multi-view comparison for:

[0057]

[0058]

[0059]

[0060]

[0061] In the formula: This represents the number of missing data points in the input set containing both positive and negative sample pairs. Low-dimensional contrastive learning representations corresponding to the missing rate views; Represents the cosine similarity function; Representing low-dimensional contrastive learning representations Representation of any two low-dimensional samples in the dataset; Represents the L2 norm; Indicates transpose; This represents the contrast loss between any two low-dimensional sample representations; These represent low-dimensional contrastive learning representations. Different samples in This means that the sample itself is excluded from the denominator; Represents an exponential function; Indicates temperature parameter; Indicates the number of samples in the batch; This represents the total number of samples participating in the comparative learning after the two views are stitched together; Indicates the first Missing rate view With the Missing rate view Pairwise contrast loss; Indicated based on the contrast loss formula The obtained contrast loss of the same positive sample pair in two directions is defined as positive sample pair, and negative sample pair is defined as different missing rate views of samples at the same time in incomplete sample data. This enables the model to align multiple missing rate inputs under the same physical state in the representation space when learning multi-view contrast. The student model is represented in the first... Hidden representations under the input condition of missing rate view; The hidden characterization distillation loss is:

[0062] In the formula: Indicates the sign of the average over tensor elements; The predicted distillation loss is:

[0063] In this embodiment, offline knowledge distillation constraints are applied to the student model based on the teacher model's hidden representations and prediction results. This allows the student model to learn semantic representations and prediction outputs that approximate complete observations under incomplete input conditions. Hidden representation distillation loss is used to constrain the student model's hidden representations to approach the teacher model's hidden representations; prediction result distillation loss is used to constrain the student model's output under incomplete inputs to approximate the teacher model's prediction results under complete inputs. This embodiment also performs dimensionality reduction on the student model's hidden representations corresponding to different missing rate inputs, constructing low-dimensional contrastive learning representations. A multi-view contrastive learning loss is established based on positive sample pairs between views with different missing rates at the same time and negative sample pairs between samples at different time. This multi-view contrastive learning loss is used to improve the consistency of inputs with different missing rates in the semantic space, enabling the student model to adapt to conditions with variable missing rates. This embodiment uses real future wave height sequences for supervised training of the student model and weights and fuses the supervised prediction loss, hidden representation distillation loss, prediction result distillation loss, and multi-view contrastive learning loss to obtain the total training loss of the student model, i.e., the total training loss function. The optimal parameters of the student model are obtained by minimizing the total loss function during model training. Through joint optimization of the total loss function during model training, the student model can simultaneously learn real wave height supervision information, complete observation semantic knowledge, and robust and consistent representations across views with different missing rates.

[0064] In this embodiment, the optimal parameters of the trained student model are represented as follows:

[0065] In the formula: This represents the optimal parameters of the student model after training; Indicates the total training loss The process of minimizing the parameters; This represents the total loss function during model training. In the model testing phase, this embodiment constructs a test input tensor and a corresponding missing mask based on the actual missing locations. It then inputs the global sea area historical observation tensor under varying missing rates into the trained student model, outputting the significant wave height prediction results for the next 12 hours. The trained student model is then used to output the significant wave height prediction results for the next 12 hours. Figures 3 to 10 As shown, the prediction results are inversely normalized and their performance is evaluated. The predicted significant wave height for the next 12 hours during the testing phase is expressed as follows:

[0066] In the formula: This indicates the predicted significant wave height for the next 12 hours during the testing phase; This represents a historical input tensor with a variable missing rate during the testing phase; This represents the student model after training. This represents the optimal parameters of the student model. At this stage, the teacher model is no longer used, thus reducing inference costs and improving practical deployment efficiency.

[0067] This embodiment evaluates the model prediction results using the coefficient of determination, root mean square error, and site comprehensive score. The formulas for calculating the coefficient of determination and root mean square error are as follows:

[0068] In the formula: Indicates the coefficient of determination; Indicates the root mean square error; Indicates the total number of test samples; Indicates the test sample index; Indicates the first The true significant wave height of each test sample; Indicates the first The predicted significant wave height for each test sample; This represents the mean of the true significant wave heights of the test samples. The larger the value, the stronger the model's ability to explain the real significant changes in wave height; the smaller the RMSE, the lower the model's prediction error, as shown in Table 1. Table 1. Prediction performance of typical stations in global ocean areas at different missing rates and prediction times.

[0069] For the s-th sea area location, which is a representative spatial point, this embodiment calculates its coefficient of determination, root mean square error, and mean true significant wave height based on all test samples of this location, and defines a comprehensive station score:

[0070] In the formula: Indicates the first The overall score of each sea area location; Indicates the first The coefficient of determination at each sea area location; Indicates the first Root mean square error at each sea area location; Indicates the first The mean true significant wave height of each sea area location in the test sample. This comprehensive score considers both the model's ability to fit wave height variation trends and the magnitude of the error relative to the mean wave height level. The larger, The smaller, The higher the value, the better the prediction performance at that location. In the actual evaluation process, this embodiment calculates the missing rate for each missing rate and each prediction timeframe for 99 sea area locations worldwide. and in accordance with Sort the locations from highest to lowest score, and select the five sea areas with the highest overall scores as the sites with the best prediction performance; at the same time, according to The two sea area locations with the lowest comprehensive scores, ranked from lowest to highest, are selected as the stations with the worst prediction performance. This evaluation method in this embodiment, in addition to evaluating the overall error, can further reveal the differences in the model's prediction stability and spatial generalization at different spatial locations. Through the above steps, it can be seen that the method described in this embodiment achieves an integrated fusion of global multi-point meteorological-wave variable input, robust modeling with variable missing rates, and significant wave height prediction for the next 12 hours. This method does not rely on training the model separately for each missing rate; instead, through offline knowledge distillation and multi-view comparative learning, it enables the student model to maintain stable spatiotemporal semantic expression and prediction output under different missing rates, making it suitable for intelligent sea state forecasting tasks in complex global sea environments.

[0071] Compared with existing technologies, the method described in this embodiment achieves the following beneficial technical effects by combining a multi-view representation learning framework with the STID backbone prediction model: (1) Improving prediction robustness under variable missing rate conditions: The method described in this embodiment constructs multiple missing rate views such as 25%, 50%, 75%, and 90% simultaneously during the training phase, and constrains the student model to learn consistent semantic representations under different missing rate inputs through offline knowledge distillation and multi-view comparative learning. Compared with the method of training the model separately for a single missing rate, the method described in this embodiment can utilize the student model after one training to adapt to preset missing rate and similar missing rate conditions, reduce the complexity of model deployment, and improve robustness in global maritime operational prediction scenarios.

[0072] (2) Maintain complete observation semantic information: The method described in this embodiment first uses complete observations to train the teacher model, and then uses the hidden representations and prediction results of the teacher model to constrain the student model, so that the student model can learn global periodic information, local fluctuation information and spatial difference information close to complete observations when dealing with incomplete observations, thereby reducing the damage of missing values ​​to the semantic structure of historical sequences of wind field, air pressure and wave height.

[0073] (3) Enhancing the joint forecasting capability of multiple points in global sea areas: The method described in this embodiment uses 99 spatial points in global sea areas as a unified multivariate time series input. The spatial differences between different sea area points are represented by spatial identity embedding in the STID backbone model, and the intraday and intraweek temporal variation characteristics are represented by temporal identity embedding. This structure avoids building a separate model for each spatial point and can achieve a unified output of significant wave heights for the next 12 hours for all 99 sea area points.

[0074] (4) Improve training and inference efficiency: The method described in this embodiment uses a lightweight STID as the backbone model, which mainly consists of an embedding layer, a multilayer perceptron encoder, and a regression layer. Compared with complex graph neural networks or Transformer structures, it has lower computational complexity. At the same time, the method described in this embodiment only uses the student model in the testing phase, without the need for the teacher model to participate in inference, which can reduce the computational cost in the actual application process.

[0075] (5) Enhancing model generalization ability: The method described in this embodiment constructs positive sample pairs for samples with different missing rates at the same time through multi-view comparative learning, and constructs negative sample pairs for samples at different times. This enables the model to align multiple missing rate inputs under the same physical state in the representation space, while distinguishing samples under different physical states. This mechanism helps improve the model's generalization ability for different time periods, different sea states, and different combinations of missing rates.

[0076] The method described in this embodiment can be widely applied to scenarios such as short-term forecasting of significant wave heights in global sea areas, meteorological support for ocean shipping routes, operation and maintenance scheduling of offshore wind farms, judgment of construction windows for marine engineering projects, risk warning for offshore platforms, and construction of intelligent marine observation systems. By jointly modeling multi-source variables such as 10-meter wind field, historical wave height, and sea-level pressure, this invention can output predictions of significant wave heights for the next 12 hours even when observational data has varying degrees of missing information, providing data support for marine disaster early warning, ship navigation avoidance, and safety management of offshore facilities. Furthermore, the method described in this embodiment uses a unified model to adapt to various missing rate conditions, demonstrating strong engineering deployment value. For large-scale, multi-node, long-term marine forecasting tasks in global sea areas, this method can reduce the cost of repeatedly training models for different missing rates and improve the model's usability under complex data quality conditions. The method described in this embodiment can also be further extended to forecasting other marine meteorological elements such as wind speed, wave period, sea-level pressure anomalies, and ocean current speed, showing promising prospects for widespread application.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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; and these 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 global wave height prediction method based on semantic alignment of multiple missing rate views, characterized in that, Specifically, the following steps are included: S1: Obtain historical multivariate time series data of pre-defined representative spatial points in global sea areas, and construct input feature vectors by introducing intraday and intraweek time features to obtain a complete observation input tensor; combine the complete observation input tensor and its corresponding significant wave height labels for future times to obtain complete sample data; Random masks with different missing rates are applied to the observation variables in the complete observation input tensor, and intraday and intraweek time features are used as time identity index features to construct incomplete input tensors under different missing rate view conditions. The incomplete input tensors are then combined with the future significant wave height labels in the corresponding complete sample data to obtain incomplete sample data. The observation variables include historical significant wave height, 10 m height U-wind component, 10 m height V-wind component, and sea level pressure. S2: Construct teacher and student models based on the spatial-temporal identity embedding model STID; the STID backbone model includes a fully connected embedding layer, a spatiotemporal embedding splicing layer, a residual multilayer perceptron coding layer, and an output layer; The fully connected embedding layer is used to map the input tensor to a hidden representation vector; The spatiotemporal embedding and splicing layer is used to perform dimensional splicing of the intraday time features, intraweek time features, and spatial embedding features of representative spatial points of the input tensor according to the hidden representation vector, so as to obtain a fused input representation vector that integrates spatiotemporal information. The residual multilayer perceptron coding layer is used to obtain the coded hidden representation vector based on the fused input representation vector; The output layer is used to output future significant wave height prediction data based on the encoded hidden representation vector; S3: Train the teacher model using complete sample data to obtain the teacher model's hidden representation and prediction results; and construct a total model training loss function for training the student model based on the teacher model's hidden representation and prediction results. Train the student model using incomplete sample data based on the total model training loss function to obtain a student model that meets the preset convergence condition and use it as the final prediction model. Use the final prediction model to achieve global wave height prediction based on semantic alignment of multiple missing rate views.

2. The global wave height prediction method based on semantic alignment of multiple missing rate views according to claim 1, characterized in that, The formula for obtaining the hidden representation vector described in S2 is: In the formula: This represents the hidden representation of the input tensor after it has passed through the embedding layer; Represents the mapping function of the fully connected layer; and This represents the input tensor to be embedded, which corresponds to the input teacher model. ; When the corresponding input is a student model ; Represents the complete observation input tensor; This indicates an incomplete input tensor; Indicates the first Representative spatial points at time The model input feature vector; Indicates the first Representative spatial points at time Historically significant highs; Indicates intraday time characteristics; Indicates the time characteristics within a week; This represents the U-wind component at a height of 10 m; This represents the V-wind component at a height of 10 m; Indicates air pressure at sea level; Indicates vector transpose; Represents the real space to which the tensor belongs; This represents the number of representative spatial points in the global ocean area; Indicates the length of the historical input; This indicates the number of input features, which includes historical significant wave heights, intraday time features, intraweek time features, 10 m height U-wind component, 10 m height V-wind component, and sea level pressure. Relative to the current time Historical step index for backward tracing; Indicates the first Incomplete input tensors under the condition of missing rate view; Represents a binary mask matrix; This represents element-wise multiplication; Indicates the first The missing rate view corresponding to each input view.

3. The global wave height prediction method based on semantic alignment of multiple missing rate views according to claim 2, characterized in that, The fusion input representation vector for obtaining fused spatiotemporal information, as described in S2, is: In the formula: This represents the fusion input representation vector that integrates spatiotemporal information; This indicates a tensor concatenation operation; Spatial identity embedding refers to spatial embedding features; This represents the embedding vector corresponding to the intraday time identity feature; This represents the embedding vector of intra-week time identity, which corresponds to the intra-week time feature. Indicates the embedding dimension.

4. The global wave height prediction method based on semantic alignment of multiple missing rate views according to claim 3, characterized in that, The formula for obtaining the encoded hidden representation vector described in S2 is: In the formula: This represents the initial input to the coding layer of the residual multilayer perceptron; Indicates the first Hidden representations output from the coding layer of a multilayer perceptron with layer residuals; This represents the fusion input representation vector that integrates spatiotemporal information; This indicates the number of encoder layers in a residual multilayer perceptron. Indicates the first Nonlinear mapping functions in multilayer perceptrons with layer residuals; Represents the linear rectification activation function; The expression for the output layer is: In the formula: This indicates the output of the predicted future significant wave height; Indicates the output regression layer; This represents the hidden representation vector output by the last layer of the coding layer in a residual multilayer perceptron.

5. A global wave height prediction method based on semantic alignment of multiple missing rate views according to claim 4, characterized in that, S3 specifically includes the following steps: S31: Randomly divide the complete sample data into a complete sample training set and a complete sample validation set according to a preset ratio; and randomly divide the incomplete sample data into an incomplete sample training set and an incomplete sample validation set. S32: With the goal of minimizing the teacher-supervised prediction loss, the teacher model is trained using the complete sample training set to obtain the trained teacher model that outputs the teacher model's hidden representation and the teacher model's prediction results. The teacher-supervised prediction loss is the regression loss between the teacher model's prediction results and their corresponding future significant wave height labels. The teacher model's hidden representation and the teacher model's prediction results are as follows: In the formula: The function representation of the teacher model; Indicates the teacher model parameters; This represents the hidden representation output by the last encoder layer of the teacher model; This represents the predicted significant wave height obtained from the teacher model; Represents the complete observation input tensor; The output of the trained teacher model is used as a pseudo-label for the distillation loss during the training of the student model. The student model is trained using an incomplete sample training set to obtain the trained student model corresponding to the hidden representation of the output student model and the prediction result of the student model. The hidden representations and prediction results of the student model are as follows: In the formula: The function representation of the student model; Indicates the parameters of the student model; Indicates the first Incomplete input tensors under the condition of missing rate view; The student model is represented in the first... The last hidden representation obtained from the missing rate view input; The student model is represented in the first... The output of the future significant wave height prediction results under the input of the missing rate view; The number of views indicating the missing rate; S33: Based on the total loss function of model training, the complete sample validation set and the incomplete sample validation set are respectively input into the trained teacher model and the trained student model. The trained student model is judged according to the preset convergence condition. If it is, the trained student model is confirmed to be the optimal prediction model. The preset convergence condition includes the total loss of model training meeting the preset loss threshold or reaching the maximum number of training rounds. Otherwise, the student model parameters are adaptively updated based on backpropagation, and the student model training process is repeated until the trained student model meets the preset convergence condition to obtain the optimal prediction model; the optimal prediction model is used to realize global wave height prediction based on semantic alignment of multiple missing rate views.

6. The global wave height prediction method based on semantic alignment of multiple missing rate views according to claim 5, characterized in that, The total training loss function of the model constructed in S33 for: In the formula: Indicates the weight of the predicted loss under supervision; Indicates the weight of knowledge distillation loss; Indicates the weights of the multi-view contrast learning loss; This indicates the monitoring and prediction of losses; This represents the total learning loss from multi-view comparison; This indicates that the loss in distillation is hidden. This indicates the predicted distillation loss; The supervised prediction loss for: In the formula: Represents the true value of significant wave height at a future time. The student model is represented in the first... Given a missing rate view input, the predicted significant wave height for future time periods; This represents the absolute value operation; The total learning loss of multi-view comparison for: In the formula: This represents the number of missing data points in the input set containing both positive and negative sample pairs. Low-dimensional contrastive learning representations corresponding to the missing rate views; Represents the cosine similarity function; Representing low-dimensional contrastive learning representations Representation of any two low-dimensional samples in the dataset; Represents the L2 norm; Indicates transpose; This represents the contrast loss between any two low-dimensional sample representations; These represent low-dimensional contrastive learning representations. Different samples in ; Represents an exponential function; Indicates temperature parameter; Indicates the number of samples in the batch; This represents the total number of samples participating in the comparative learning after the two views are stitched together; Indicates the first Missing rate view With the Missing rate view Pairwise contrast loss; Indicated based on the contrast loss formula The contrast loss of the same positive sample in two directions; The student model is represented in the first... Hidden representations under the input condition of missing rate view; The hidden characterization distillation loss is: In the formula: Indicates the sign of the average over tensor elements; The predicted distillation loss is: 。