Sea wave significant wave height prediction model training and sea wave significant wave height prediction method

Abstract

The invention relates to the technical field of ocean engineering, and discloses a method for training a sea wave significant wave height prediction model and predicting the sea wave significant wave height, and the method for training the sea wave significant wave height prediction model comprises the following steps: according to the correlation between historical meteorological characteristic time series data and historical sea wave height time series data, calculating the sea wave significant wave height prediction model; selecting a target meteorological feature from the plurality of meteorological features; based on the decomposition number and penalty factor of the variational mode decomposition, optimizing to obtain a target decomposition number and a target penalty factor, and performing variational mode decomposition on the historical sea wave height time sequence data to obtain an intrinsic mode sub-sequence; the sea wave significant wave height prediction model is trained according to the target historical meteorological characteristic time series data and the intrinsic mode subsequences, the correlation between the meteorological data and the historical sea wave height data is analyzed, and the variational mode decomposition parameters are optimized, so that the input accuracy of the sea wave significant wave height prediction model is improved, and the prediction accuracy of the sea wave significant wave height prediction model is improved. Therefore, the precision of the sea wave significant wave height prediction model is improved.

Description

Technical Field

[0001] This invention relates to the field of marine engineering technology, specifically to the training of a wave significant wave height prediction model and a method for predicting wave significant wave height. Background Technology

[0002] Significant wave height (SWH), as a core parameter describing the energy and propagation characteristics of ocean waves, has significant application value in fields such as marine engineering, maritime transportation, disaster early warning, and renewable energy development. Therefore, predicting SWH is of great importance in the field of marine science, and the accuracy of SWH prediction is directly related to the safety of offshore operations, port design optimization, and the improvement of wave energy power generation efficiency.

[0003] In related technologies, the method for predicting the significant wave height of ocean waves involves inputting multiple meteorological factors and raw wave height data into a deep learning model to obtain the predicted significant wave height. However, this method suffers from the fact that the raw wave height data is affected by periodic tides, random wind field disturbances, and other factors, exhibiting non-stationary characteristics. Furthermore, among the multiple meteorological factors, there may be uncorrelated or low-correlated meteorological factors, resulting in low prediction accuracy of the significant wave height and failing to reflect reality. Summary of the Invention

[0004] In view of this, the present invention provides a training method for a wave significant wave height prediction model and a method for predicting wave significant wave height, in order to solve the problem of low prediction accuracy of wave significant wave height caused by wave significant wave height prediction methods in related technologies.

[0005] In a first aspect, the present invention provides a training method for a significant wave height prediction model, comprising: selecting a target meteorological feature from multiple meteorological features based on the correlation between historical meteorological feature time series data and historical wave height time series data; the target meteorological feature being a strongly correlated meteorological feature whose correlation satisfies a preset condition; optimizing the number of decompositions and the penalty factor of variational mode decomposition based on sample wave height time series data to obtain a target number of decompositions and a target penalty factor; performing variational mode decomposition on historical wave height time series data based on the target number of decompositions and the target penalty factor to obtain an intrinsic mode subsequence of the target number of decompositions; and inputting the target historical meteorological feature time series data and the intrinsic mode subsequence corresponding to the target meteorological feature into the constructed significant wave height prediction model for training.

[0006] This invention utilizes the correlation between historical meteorological feature time-series data and historical ocean wave height time-series data for multiple meteorological features to select target meteorological features with strong correlations among multiple meteorological features. This allows the ocean wave significant wave height prediction model to focus on key influencing factors, reduce interference from irrelevant or weakly correlated meteorological features, and improve the input quality of the ocean wave significant wave height prediction model. This invention optimizes the number of decompositions and the penalty factor in variational mode decomposition (VMD) based on sample ocean wave height time-series data to obtain target decomposition numbers and target penalty factors. By optimizing these target decomposition numbers and penalty factors, variational mode decomposition is performed on historical ocean wave height time-series data. This makes the decomposition of historical ocean wave height time-series data more suitable for the data characteristics. Since sample ocean wave height time-series data is affected by periodic tides and random wind field disturbances, exhibiting non-stationary characteristics, this invention adapts to the non-stationary characteristics of the sample ocean wave height time-series data. The decomposed intrinsic mode subsequences more accurately reflect the wave wave components of different frequencies and characteristics, improving the ability of the ocean wave significant wave height prediction model to capture the patterns of each intrinsic mode subsequence during training, and enhancing the ocean wave significant wave height prediction model's ability to characterize complex ocean wave changes. This invention inputs the target historical meteorological feature time-series data corresponding to the target meteorological features and the intrinsic mode subsequences into the constructed ocean wave significant wave height prediction model for training. Compared with related technologies, this invention enables the wave significant wave height prediction model to fully learn the changing mechanism of wave significant wave height under meteorological driving and its own fluctuation mode, from feature selection and data decomposition to multi-source data input training. The trained wave significant wave height prediction model has better adaptability to different sea conditions in wave significant wave height prediction, thus improving the accuracy of the wave significant wave height prediction model.

[0007] In one optional implementation, a target meteorological feature is selected from multiple meteorological features based on the correlation between historical meteorological feature time-series data and historical wave height time-series data. This includes: dividing the historical meteorological feature time-series data and historical wave height time-series data of each meteorological feature into intervals to obtain target tables of different resolutions; determining a joint probability distribution function and a marginal probability distribution function based on the number of data points in each target table; the joint probability distribution function characterizes the probability of simultaneous occurrence of meteorological feature variables and historical wave height variables; the marginal probability distribution function characterizes the meteorological probability distribution of meteorological feature variables and the wave height probability distribution of historical wave height variables; determining the mutual information of each target table based on the joint probability distribution function and the marginal probability distribution function; selecting the maximum mutual information from multiple mutual information values ​​and comparing the maximum mutual information with a preset mutual information threshold; the maximum mutual information characterizes the correlation between historical meteorological feature time-series data and historical wave height time-series data; if the maximum mutual information is greater than or equal to the preset mutual information threshold, the corresponding meteorological feature is selected as the target meteorological feature; if the maximum mutual information is less than the preset mutual information threshold, the corresponding meteorological feature is discarded.

[0008] In one optional implementation, the number of decompositions and the penalty factor of variational mode decomposition are optimized based on the sample ocean wave height time series data to obtain the target number of decompositions and the target penalty factor. This includes: obtaining a first value range for the number of decompositions and a second value range for the penalty factor; mapping the obtained chaotic sequence to the first and second value ranges to obtain an initial sparrow population; using each sparrow individual in the initial sparrow population to perform variational mode decomposition on the sample ocean wave height time series data; evaluating the decomposition results using a preset fitness function; obtaining the optimal sparrow individual based on the evaluation results; and updating and iterating the sparrow population based on the optimal sparrow individual until a first preset stopping condition is reached to obtain the target number of decompositions and the target penalty factor.

[0009] This invention obtains a first range of values ​​for the number of decompositions and a second range of values ​​for the penalty factor. The obtained chaotic sequence is mapped to these first and second ranges to obtain an initialized sparrow population. This allows the sparrow population to be more evenly distributed throughout the search space, increasing the chance of finding the global optimum and reducing the risk of the algorithm getting trapped in local optima, thereby accelerating convergence. This invention utilizes each individual sparrow in the initialized sparrow population to perform variational mode decomposition on the sample ocean wave height time-series data. A preset fitness function is used to evaluate the decomposition results, and the optimal sparrow individual is obtained based on the evaluation results. The sparrow population is then updated iteratively based on the optimal sparrow individual until a first preset stopping condition is reached, obtaining the target number of decompositions and the target penalty factor, thus improving the flexibility and efficiency of parameter optimization.

[0010] In one optional implementation, updating and iterating the sparrow population based on the optimal sparrow individual includes: updating and iterating the position of the discoverer individual in the sparrow population based on the optimal sparrow individual by combining sine and cosine perturbations and the golden ratio; updating and iterating the position of the follower individual in the sparrow population based on the position of the discoverer individual; and updating and iterating the position of the vigilant individual in the sparrow population based on the reverse logic of the optimal sparrow individual, so as to update and iterate the sparrow population.

[0011] This invention introduces sine and cosine perturbations and the golden ratio when updating the position of individual discoverers. It uses a flexible search mechanism to balance global exploration and local development, thereby improving the performance of the optimization algorithm.

[0012] In one optional implementation, variational mode decomposition is performed on historical wave height time series data based on the target decomposition quantity and the target penalty factor to obtain the intrinsic mode subsequence of the target decomposition quantity. This includes: constructing a variational problem based on the target decomposition quantity and historical wave height time series data; transforming the variational problem into an unconstrained variational problem using the target penalty factor and the Lagrange multiplication operator; performing a frequency domain transformation on the unconstrained variational problem; and iteratively solving the frequency domain problem obtained from the frequency domain transformation until a second preset stopping condition is met to obtain the intrinsic mode subsequence of the target decomposition quantity.

[0013] In one alternative implementation, the training method for the significant wave height prediction model includes: the significant wave height prediction model is a model constructed from a deep learning network based on a self-attention mechanism and a bidirectional long short-term memory network.

[0014] The wave significant height prediction model of this invention is a model constructed based on a deep learning network with a self-attention mechanism and a bidirectional long short-term memory network. Therefore, the wave significant height prediction model can capture the global characteristic trend of wave height and extract local temporal features, thereby improving the wave height prediction accuracy of the wave significant height prediction model over a longer time span.

[0015] Secondly, the present invention provides a method for predicting the significant wave height of ocean waves, comprising: acquiring time-series data of forecast meteorological features corresponding to the target meteorological features at the time to be predicted; inputting the time-series data of forecast meteorological features into the significant wave height prediction model of the first aspect or any corresponding embodiment described above, to obtain multiple intrinsic mode prediction sub-sequences; the target meteorological features are meteorological features with strong correlations that satisfy preset conditions selected from the multiple meteorological features based on the correlation between the historical meteorological feature time-series data of the multiple meteorological features and the historical wave height time-series data respectively; and superimposing the multiple intrinsic mode prediction sub-sequences to obtain the significant wave height prediction data at the time to be predicted.

[0016] The present invention inputs the time series data of the forecast meteorological features into the sea wave significant wave height prediction model of the first aspect or any corresponding embodiment mentioned above, and obtains multiple intrinsic mode prediction subsequences. Since the target meteorological feature is a strongly correlated meteorological feature selected from multiple meteorological features that meets the preset conditions, it is convenient for the sea wave significant wave height prediction model to focus on key influencing factors and reduce the interference of irrelevant or weakly correlated meteorological features, thereby improving the accuracy of the sea wave significant wave height prediction data at the time to be predicted.

[0017] Thirdly, the present invention provides a training device for a significant wave height prediction model, comprising: a meteorological feature selection module, used to select a target meteorological feature from multiple meteorological features based on the correlation between historical meteorological feature time series data and historical wave height time series data; the target meteorological feature is a strongly correlated meteorological feature whose correlation satisfies a preset condition; a parameter optimization module, used to optimize the number of decompositions and the penalty factor of variational mode decomposition based on sample wave height time series data to obtain a target number of decompositions and a target penalty factor; a variational mode decomposition module, used to perform variational mode decomposition on historical wave height time series data based on the target number of decompositions and the target penalty factor to obtain an intrinsic mode subsequence of the target number of decompositions; and a model training module, used to input the target historical meteorological feature time series data and the intrinsic mode subsequence corresponding to the target meteorological feature into the constructed significant wave height prediction model for training.

[0018] Fourthly, the present invention provides a device for predicting the significant wave height of ocean waves, comprising: a model prediction module, configured to acquire time-series data of forecast meteorological features corresponding to the target meteorological features at the time to be predicted, input the time-series data of forecast meteorological features into the significant wave height prediction model of the first aspect or any corresponding embodiment, and obtain multiple intrinsic mode prediction sub-sequences; the target meteorological features are meteorological features with strong correlations selected from multiple meteorological features based on the correlation between the historical meteorological feature time-series data of multiple meteorological features and the historical wave height time-series data, respectively, and the correlation satisfies preset conditions; and a sequence overlay module, configured to overlay the multiple intrinsic mode prediction sub-sequences to obtain the significant wave height prediction data at the time to be predicted.

[0019] Fifthly, the present invention provides a computer device including a processor, a memory, and a processor that are interconnected. The memory stores computer instructions. The processor and the memory execute the computer instructions to perform the training method for the wave significant wave height prediction model of the first aspect or any corresponding embodiment thereof, or the wave significant wave height prediction method of the second aspect.

[0020] In a sixth aspect, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the training method of the wave significant wave height prediction model of the first aspect or any corresponding embodiment thereof, or the wave significant wave height prediction method of the second aspect.

[0021] In a seventh aspect, the present invention provides a computer program product, including computer instructions for causing a computer to execute the training method for the wave significant wave height prediction model of the first aspect or any corresponding embodiment thereof, or the wave significant wave height prediction method of the second aspect. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the specific embodiments or related technologies of the present invention, the drawings used in the description of the specific embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the training method of the effective wave height prediction model for ocean waves according to an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the Transformer model structure according to an embodiment of the present invention.

[0025] Figure 3 This is a structural diagram of the Transformer model-multi-head attention mechanism according to an embodiment of the present invention.

[0026] Figure 4 This is a schematic diagram of the BiLSTM network structure according to an embodiment of the present invention.

[0027] Figure 5 This is a flowchart illustrating another training method for the effective wave height prediction model of ocean waves according to an embodiment of the present invention.

[0028] Figure 6 This is a schematic diagram of the optimization process for the number of target decompositions and the target penalty factor according to an embodiment of the present invention.

[0029] Figure 7 This is a flowchart illustrating another training method for the ocean wave significant wave height prediction model according to an embodiment of the present invention.

[0030] Figure 8 This is a flowchart illustrating a method for predicting the effective wave height of ocean waves according to an embodiment of the present invention.

[0031] Figure 9 This is a schematic diagram illustrating the results of improved SSA-optimized VMD according to an embodiment of the present invention.

[0032] Figure 10 This is a schematic diagram of the effective wave height decomposition of each modal component and residual according to an embodiment of the present invention.

[0033] Figure 11 This is a schematic diagram comparing the prediction results of buoy A according to an embodiment of the present invention.

[0034] Figure 12 This is a schematic diagram comparing the prediction results of the B ocean buoy according to an embodiment of the present invention.

[0035] Figure 13 This is a structural block diagram of a training device for a wave significant height prediction model according to an embodiment of the present invention.

[0036] Figure 14 This is a structural block diagram of a device for predicting the effective wave height of ocean waves according to an embodiment of the present invention.

[0037] Figure 15 This is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0038] 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.

[0039] Significant wave height (SWH), as a core parameter describing wave energy and propagation characteristics, has significant application value in marine engineering, maritime transportation, disaster early warning, and renewable energy development. Its prediction accuracy directly affects the safety of offshore operations, port design optimization, and the improvement of wave energy power generation efficiency. However, the generation and evolution of ocean waves are influenced by the coupled effects of multiple factors such as wind speed, water depth, and topography, exhibiting significant nonlinearity, non-stationarity, and spatiotemporal dependence, making SWH prediction a challenging problem in marine science.

[0040] In recent years, with the improvement of data acquisition technology and computing power, SWH prediction methods have gradually shifted from traditional empirical models to comprehensive models that combine data-driven approaches with physical mechanisms, covering a variety of methods such as traditional time series analysis, physical numerical simulation, machine learning, and deep learning.

[0041] Currently, methods for predicting significant wave height (SWH) include traditional time series models, physical numerical models, machine learning methods, and deep learning methods. Traditional time series models extrapolate based on the statistical regularities of historical data, such as the Autoregressive Moving Average (ARMA) and Autoregressive Integrated Moving Average (ARIMA) models, achieving short-term predictions by capturing the linear correlation of SWH sequences. However, traditional time series models assume the data is stationary and ignore the interaction of multiple factors, making them difficult to adapt to complex sea conditions. Physical numerical models simulate wave propagation processes based on hydrodynamic equations and are suitable for reconstructing wave fields in large-scale sea areas. However, physical numerical models rely on high-resolution meteorological driving data, resulting in high computational costs.

[0042] To overcome the limitations of traditional models, machine learning methods such as Support Vector Machine (SVM) and Random Forest (RF) have been introduced into SWH prediction. SVM, by fusing wind speed and wave period features, significantly improves prediction accuracy. However, machine learning models are susceptible to noise interference when processing long-term time-series data and struggle to capture multi-scale fluctuation characteristics. Furthermore, shallow network models (such as multilayer feedforward neural networks based on error backpropagation) have limited generalization capabilities, leading to significantly increased prediction errors when facing extreme weather events.

[0043] Deep learning offers a new approach to SWH prediction by automatically extracting high-dimensional features from data. Long Short-Term Memory (LSTM) networks and their variants (such as gated recurrent units) have demonstrated outstanding performance in SWH prediction by capturing temporal dependencies through gating mechanisms.

[0044] Currently, the Transformer architecture (a deep learning network based on self-attention) is gradually being applied to wave prediction due to the advantages of self-attention in global dependency modeling. Although deep learning models have made significant progress, ocean wave time series data are affected by periodic tides, random wind field disturbances, and other factors, exhibiting non-stationary characteristics. Related studies mostly use normalization or sliding window smoothing to process the data, but lack decomposition and modeling of the inherent non-stationarity of the signal. For example, while the Deterministic-Stochastic Decomposition (DSD) algorithm separates deterministic and stochastic components, it does not fully integrate signal processing techniques such as wavelet transform or empirical mode decomposition, making it difficult to effectively eliminate the interference of high-frequency noise on model training. Factors affecting the effective wave height include variables such as temperature, air temperature, wind speed, and air pressure, each with varying correlations. Too many uncorrelated or low-correlated input variables will increase the model's prediction accuracy and cause interference, thus reducing its prediction accuracy. The local receptive field makes it difficult to extract the spatiotemporal coupling relationships of multi-source features (such as wind speed, water temperature, and topography) from massive historical data, resulting in insufficient generalization ability under complex sea conditions.

[0045] This invention provides a training method for a wave significant height prediction model. By analyzing the correlation between meteorological data and historical wave significant height data, the variational mode decomposition parameters are optimized to improve the accuracy of the input to the wave significant height prediction model, thereby improving the accuracy of the wave significant height prediction model.

[0046] According to an embodiment of the present invention, a training method embodiment for a prediction model of significant wave height of ocean waves is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0047] This embodiment provides a training method for a predictive model of significant wave height, which can be used with computer equipment. Figure 1 This is a flowchart of a training method for a wave significant wave height prediction model according to an embodiment of the present invention, such as... Figure 1 As shown, the process includes the following steps:

[0048] Step S101: Based on the correlation between historical meteorological feature time series data of multiple meteorological features and historical sea wave height time series data, select a target meteorological feature from multiple meteorological features; the target meteorological feature is a meteorological feature with strong correlation that meets the preset conditions.

[0049] Among them, meteorological features are meteorological elements related to the marine environment and weather conditions that can affect the formation and changes of sea wave height. For example, meteorological features include sea surface temperature, sea surface air temperature, wind direction, wind speed, and air pressure. Historical meteorological feature time series data are specific historical meteorological time series data corresponding to the meteorological features, arranged in chronological order. Historical sea wave height time series data are historical sea wave height observation data sequences arranged in chronological order. In this embodiment of the invention, outliers in the historical meteorological feature time series data are removed, and cubic spline interpolation is used to fill in the removed and missing values. The historical meteorological feature time series data used subsequently are the processed historical meteorological feature time series data. Continuous wavelet transform (CWT) technology is used to remove noise from the historical sea wave height time series data to improve data quality. The historical sea wave height time series data used subsequently are the denoised historical sea wave height time series data.

[0050] In some optional implementations, target meteorological features are selected from multiple meteorological features based on the correlation between historical meteorological feature time-series data and historical wave height time-series data. This includes: dividing the historical meteorological feature time-series data and historical wave height time-series data of each meteorological feature into intervals to obtain target tables of different resolutions; determining the joint probability distribution function and marginal probability distribution function based on the number of data points in each target table; the joint probability distribution function is used to characterize the probability of meteorological feature variables and historical wave height variables occurring simultaneously; the marginal probability distribution function is used to characterize the meteorological probability distribution of meteorological feature variables and the wave height probability distribution of historical wave height variables; determining the mutual information of each target table based on the joint probability distribution function and the marginal probability distribution function; selecting the maximum mutual information from multiple mutual information values ​​and comparing the maximum mutual information with a preset mutual information threshold; the maximum mutual information is used to characterize the correlation between historical meteorological feature time-series data and historical wave height time-series data; if the maximum mutual information is greater than or equal to the preset mutual information threshold, the corresponding meteorological feature is selected as the target meteorological feature; if the maximum mutual information is less than the preset mutual information threshold, the corresponding meteorological feature is discarded.

[0051] Specifically, the historical meteorological feature time series data and historical sea wave height time series data of each meteorological feature are divided into intervals to obtain target tables of different resolutions. This includes: dividing the historical meteorological feature time series data of each meteorological feature into intervals according to different specifications to obtain multiple first intervals; dividing the historical sea wave height time series data into intervals according to different specifications to obtain multiple second intervals; and obtaining target tables of different resolutions from multiple first intervals and multiple second intervals.

[0052] Exemplarily, if the historical meteorological characteristic time series data of seawater temperature is X and the historical wave height time series data is Y, X is divided into m intervals according to the specification of 5°C, and Y is divided into n intervals according to the specification of 1 m, then a target grid of m * n is obtained.

[0053] In some optional embodiments, according to the number of data points in the target table for each resolution, determining the joint probability distribution function and the marginal probability distribution function includes: obtaining the joint probability distribution function according to the quotient of the number of data in a single grid space and the total number of data points in the target table, obtaining the meteorological marginal probability distribution function according to the quotient of the number of data points in a single interval of the historical meteorological characteristic time series data and the total number of data points in the historical meteorological characteristic time series data, obtaining the wave height marginal probability distribution function according to the quotient of the number of data points in a single interval of the historical wave height time series data and the total number of data points in the historical wave height time series data, and forming the marginal probability distribution function from the meteorological marginal probability distribution function and the wave height marginal probability distribution function.

[0054] In some optional embodiments, the formula for determining the mutual information of the target table for each resolution according to the joint probability distribution function and the marginal probability distribution function is:

[0055]

[0056] where I(X, Y) is the mutual information between the historical meteorological characteristic time series data X and the historical wave height time series data Y, p(x, y) is the joint probability distribution function between the historical meteorological characteristic time series data point x and the historical wave height time series data point y, p(x) is the marginal probability distribution function of the historical meteorological characteristic time series data point x, and p(y) is the marginal probability distribution function of the historical wave height time series data point y.

[0057] In some optional embodiments, after determining the mutual information of the target table for each resolution according to the joint probability distribution function and the marginal probability distribution function, the mutual information is normalized, and the mutual information for subsequent maximum mutual information selection is the normalized mutual information.

[0058] Exemplarily, the formula for normalizing the mutual information is:

[0059]

[0060] where M(X, Y) is the normalized mutual information value, max is to take the maximum value, mn < B is the grid division constraint, B is the 0.6th power of the sample size, I(X, Y) is the mutual information between the historical meteorological characteristic time series data X and the historical wave height time series data Y, min is to take the minimum value, m is the number of intervals of the historical meteorological characteristic time series data, and n is the number of intervals of the historical wave height time series data.

[0061] In this embodiment of the invention, target tables of different resolutions are constructed, and the normalized mutual information value corresponding to each target table is determined. The maximum value among the normalized mutual information values ​​corresponding to each target table is selected as the maximum mutual information. For example, the formula for determining the maximum mutual information is:

[0062] MIC(X,Y)=max(M(X,Y))

[0063] Wherein, MIC(X,Y) is the maximum mutual information, M(X,Y) is the normalized mutual information value, and max is the maximum value. In this embodiment of the invention, the closer MIC(X,Y) is to 1, the stronger the correlation between the two variables, and the closer MIC(X,Y) is to 0, the weaker the correlation between the two variables.

[0064] In some optional implementations, the preset mutual information threshold can be set according to the actual situation. For example, the preset mutual information threshold can be 0.2.

[0065] Step S102: Based on the sample ocean wave height time series data, optimize the number of decompositions and the penalty factor of variational mode decomposition to obtain the target number of decompositions and the target penalty factor.

[0066] Among them, the sample wave height time series data is selected from the historical wave height time series data for decomposition quantity and penalty factor optimization.

[0067] In some optional implementations, the sample wave height time series data is input into a preset improved sparrow search algorithm to optimize the number of decompositions and the penalty factor, thereby obtaining the target number of decompositions and the target penalty factor.

[0068] Among them, the preset improved sparrow search algorithm is an algorithm obtained by introducing Singer chaotic sequence and golden sine and cosine strategy into the sparrow search algorithm.

[0069] In some optional implementations, the Sparrow Search Algorithm (SSA), as a novel intelligent optimization algorithm, analyzes the foraging and warning behaviors of sparrows to perform local and global searches in the objective function space. It can effectively approximate the population optimum, offering advantages such as fast convergence and strong global search capability. Traditional SSA algorithms often randomly generate population individuals during initialization, leading to reduced population diversity in later iterations and a tendency to get trapped in local optima. Therefore, this invention introduces the Singer chaotic sequence and the golden sine / cosine strategy into the Sparrow Search Algorithm. The Singer mapping is a parameterized nonlinear chaotic mapping that allows the sparrow population to be more evenly distributed throughout the search space, increasing the chance of finding the global optimum and reducing the risk of the algorithm getting trapped in local optima, thereby accelerating the convergence speed. The golden sine / cosine population update strategy combines the characteristics of the golden ratio and the sine / cosine function, aiming to balance global exploration and local exploitation through a flexible search mechanism, thereby improving the performance of the optimization algorithm.

[0070] Step S103: Based on the target decomposition quantity and the target penalty factor, perform variational mode decomposition on the historical wave height time series data to obtain the intrinsic mode subsequence of the target decomposition quantity.

[0071] Specifically, the variational mode decomposition algorithm is set according to the target decomposition number and the target penalty factor. Historical wave height time series data is input into the set variational mode decomposition algorithm to perform variational mode decomposition on the historical wave height time series data, and obtain the intrinsic mode subsequence of the target decomposition number.

[0072] In this embodiment of the invention, different meteorological characteristics and each effective wave height subsequence often have different dimensions and units. To eliminate the influence of dimensions, facilitate training, and better evaluate experimental results, the preprocessed data is further normalized. The Max-Min method is directly used for normalization, as shown in the following formula:

[0073]

[0074] Where, x ′ The data is normalized, where x is the actual data. min For the maximum value, x min It is a local minimum.

[0075] Step S104: Input the target historical meteorological feature time series data and intrinsic mode subsequence corresponding to the target meteorological feature into the constructed sea wave significant wave height prediction model for training.

[0076] In some optional implementations, the wave significant height prediction model is a model constructed using a deep learning network based on a self-attention mechanism and a bidirectional long short-term memory network. The deep learning network based on the self-attention mechanism is used to extract global features of the target's historical meteorological feature time series data and intrinsic mode subsequences. The model constructed using the bidirectional long short-term memory network is used to construct a mapping relationship between meteorological feature time series data and wave significant height based on the extracted target's historical meteorological feature time series data and intrinsic mode subsequences global features, for the prediction of wave significant height.

[0077] Among them, the deep learning network based on the self-attention mechanism can be the Transformer model. The input of the Transformer model is the time series data of the target historical meteorological features and the intrinsic mode subsequence corresponding to the target meteorological features. The Transformer model is used to extract features from the time series data of the target historical meteorological features to obtain global features. The Bi-directional Long Short-Term Memory Network (Bi-LSTM) is used to predict the significant wave height of the ocean waves based on the global features output by the Transformer model.

[0078] In this embodiment of the invention, the Transformer model is a sequence-to-sequence model based on an attention mechanism, consisting of an encoder and a decoder. The encoder encodes the input target historical meteorological feature time-series data into a high-dimensional feature vector, enabling the decoder to extract more accurately based on the contextual dependencies of the sequence when generating the output sequence. The encoder consists of multiple identical layers, each containing two sub-layers: a multi-head self-attention mechanism and a feedforward neural network, and incorporating residual connections and layer normalization. The decoder structure is similar to the encoder, including masked multi-head self-attention, encoder-decoder attention, feedforward neural networks, residual connections, and layer normalization. The masked multi-head self-attention mechanism only references wave height data within the historical time step and does not consider future information, thus ensuring that the Transformer model maintains its autoregressive characteristics. The encoder-decoder attention mechanism matches the wave height data at the current time step with the key features generated by the encoder, dynamically focusing on wave height information from similar time steps in historical data. This significantly improves the Transformer model's ability to handle long sequences and capture global dependencies. The feedforward neural network further performs a nonlinear transformation on the output of the multi-head self-attention mechanism, enhancing the Transformer model's ability to handle nonlinear relationships. The decoder employs residual connections and layer normalization techniques. Through residual connections, input features are directly added to the output of the neural network layers, enabling the Transformer model to more effectively utilize information from the input features, thereby accelerating the training process, mitigating overfitting, and improving model training efficiency.

[0079] For example, such as Figure 2 The diagram shows the structure of the Transformer model, which includes an encoder and a decoder. In the encoder, the temporal data of ocean wave feature variables (i.e., the historical meteorological feature temporal data of the target) is positionally encoded through input embedding, then processed through a multi-head attention mechanism into residual connections and layer normalization, and then again through a feedforward network into residual connections and layer normalization. The processed result is directly fed into the multi-head attention mechanism of the decoder. In the decoder, the temporal data of significant wave height (i.e., the intrinsic mode subsequence) is positionally encoded through output embedding, then processed through a masked multi-head attention mechanism into residual connections and layer normalization, then again through a multi-head attention mechanism into residual connections and layer normalization, and finally through a feedforward network into residual connections and layer normalization. Finally, linearization and normalized exponential function processing are performed.

[0080] The self-attention mechanism, as the core of the encoder, weights and aggregates each position in the historical meteorological feature time-series data of the input target, assigning higher weights to factors affecting the significant wave height, enabling the encoder to focus on any position in the input intrinsic mode subsequence. The attention mechanism consists of a query matrix (Q), a key matrix (E), and a value matrix (V), used to calculate a weighted sum of values. Q is used to find relevant information about the significant wave height in the intrinsic mode subsequence, such as periodicity, seasonality, and trend. E uses certain features in the intrinsic mode subsequence as key vectors, and Q and E together calculate the attention vector, which is then weighted and normalized using the Softmax function before being output. The attention used by the Transformer is called "scaled dot product attention," which calculates the correlation between a query vector and a set of key-value pairs, thereby generating a weighted value vector.

[0081] For example, the formula for calculating the self-attention mechanism is as follows:

[0082]

[0083] Q = X f W Q

[0084] E=X f W E

[0085] V = X f W V

[0086] Where Attention(Q, E, V) is the self-attention mechanism function for the query matrix Q, key matrix E, and value matrix V, softmax is the activation function, T is the transpose, d is the matrix dimension, and X is the value of the matrix. f Let Q be the input matrix, E be the query matrix, V be the key matrix, and W be the value matrix. Q To query the weight matrix of the matrix, W E W is the weight matrix of the key matrix. V The weight matrix is ​​the value matrix.

[0087] Multi-head self-attention (MFA) is an extension and improvement of single-head self-attention. It divides the model into multiple self-attention heads, forming independent subspaces, allowing the model to simultaneously focus on different aspects of information. These self-attention heads process different parts of the input data, thereby capturing richer contextual information. Subsequently, this information from different aspects is concatenated and linearly transformed to generate the final attention value. The specific calculation process is as follows:

[0088] Multihead(Q,E,V)=Concat(h1,h2,h3)*W

[0089]

[0090] Where Multihead(Q, E, V) represents parallel processing of multiple attention heads, Concat connects the outputs of each attention head, h1 is the computation result of the first attention head, h k The result of the calculation for the k-th attention head is given, where W is the output weight matrix of the multi-head attention, and X is the input matrix. The weight matrix is ​​the query matrix of the k-th attention head. Let the weight matrix be the key matrix of the k-th attention head. Let be the weight matrix of the value matrix of the k-th attention head.

[0091] In this embodiment of the invention, the self-attention mechanism used is a multi-head self-attention mechanism, which is an extension and improvement of the self-attention mechanism, such as... Figure 3 The diagram shows the structure of the Transformer model's multi-head attention mechanism. The query matrix Q, key matrix K, and value matrix V are linearly transformed and then input into the scaling dot product attention mechanism. There are n scaling dot product attention modules. The outputs of the scaling dot product attention mechanism are concatenated and linearly transformed to obtain the multi-head attention. Specifically, the calculation steps of the scaling dot product attention mechanism are as follows: the query matrix Q and key matrix K are multiplied (MatMul), their similarity is calculated, the multiplication result is scaled, the scaled result is masked, the masked result is normalized (SoftMax), and finally, the normalized result is multiplied again. The value matrix V is then directly multiplied. The multi-head self-attention mechanism divides the Transformer model into multiple self-attention heads, forming multiple independent subspaces, allowing the Transformer model to simultaneously focus on different aspects of information. These self-attention heads process different parts of the input data, thereby capturing richer contextual information. This information from different aspects is then concatenated and linearly transformed to generate the final attention value.

[0092] In some alternative implementations, Bi-LSTM combines the architecture of a forward LSTM (Long Short-Term Memory Network) network and a backward LSTM network. LSTM networks are a variant of Recurrent Neural Networks (RNNs). By introducing unique three gating structures and memory units, Bi-LSTM effectively solves the short-term memory defect caused by gradient vanishing in RNNs.

[0093] For example, the formula for LSTM is:

[0094] f t =σ(W f *[h t-1 x t ]+b f )

[0095] i t =σ(W i *[h t-1 x t ]+b i )

[0096]

[0097] o t =σ(W o *[h t-1 x t ]+b o )

[0098]

[0099] Among them, f t Let W be the forgetting gate of the LSTM at time t, σ() be the activation function of the LSTM, and W be the value of the forgetting gate. f Output x to the hidden layer t The weight matrix to the forget gate, h t-1 Let b be the hidden state at time t-1. f For the bias term of the forget gate, i t Let W be the input gate of the LSTM at time t. i Output x to the hidden layer t The weight matrix to the input gate, b i For the bias term of the input gate, W represents the value of the temporary memory cell in the LSTM at time t. c Output x to the hidden layer t The weight matrix to the temporary memory unit, b c For the bias term of temporary memory units, C t Let C be the memory cell value of the LSTM at time t. t-1 The memory cell value of the LSTM at time t-1. For element-wise multiplication, To add element by element, o t W is the output gate of the LSTM at time t. o Output x to the hidden layer t The weight matrix to the output gate, b o h is the bias term for the output gate.t Let t be the hidden state of the feedforward LSTM at time t, and tanh() be the hyperbolic tangent activation function.

[0100] In this embodiment of the invention, the BiLSTM model combines forward LSTM and backward LSTM, which can simultaneously extract the forward and backward historical data features of the data. This overcomes the limitation of LSTM, which can only predict the output of the next moment based on the time series information of the historical moment. It can also better explore the intrinsic relationship between the current effective wave height data and historical and future data, thereby improving the utilization rate of data and the prediction accuracy of the model.

[0101] For example, such as Figure 4 The diagram shown is a schematic of the BiLSTM network structure. Figure 4 The left side shows an LSTM network, including a forget gate, input gate, and output gate. The input X... t The hidden state h from the previous moment t-1 And the memory unit value C from the previous moment. t-1 The value passes through the forget gate, input gate, and output gate in sequence, and the current hidden state h is output. t and C at the current moment t The forget gate obtains its forget gate value through the activation function σ, the input gate obtains its input gate value through the activation function σ and the hyperbolic tangent activation function tanh, and the output gate obtains its output gate value through the activation function σ and the hyperbolic tangent activation function tanh. Figure 4 The rightmost part is a BiLSTM network, which is a combination of forward LSTM and backward LSTM. The n inputs pass through the n LSTMs in the forward layer and the n LSTMs in the backward layer, respectively, and output n y values.

[0102] The training method for the sea wave significant wave height prediction model provided in this embodiment selects a target meteorological feature with strong correlation among multiple meteorological features by comparing the correlation between historical meteorological feature time series data and historical sea wave height time series data. This allows the sea wave significant wave height prediction model to focus on key influencing factors, reduce interference from irrelevant or weakly correlated meteorological features, and improve the input quality of the sea wave significant wave height prediction model. This invention optimizes the number of decompositions and the penalty factor in variational mode decomposition based on sample ocean wave height time-series data to obtain target decomposition numbers and target penalty factors. By optimizing these target decomposition numbers and penalty factors, variational mode decomposition is performed on historical ocean wave height time-series data. This makes the decomposition of historical ocean wave height time-series data more suitable for the data characteristics. Since sample ocean wave height time-series data is affected by periodic tides, random wind field disturbances, etc., it exhibits non-stationary characteristics. This invention adapts to the non-stationary characteristics of sample ocean wave height time-series data, and the decomposed intrinsic mode subsequences more accurately reflect the ocean wave fluctuation components of different frequencies and characteristics. This improves the ability of the ocean wave significant wave height prediction model to capture the patterns of each intrinsic mode subsequence during training, and enhances the ocean wave significant wave height prediction model's ability to characterize complex ocean wave changes. In this invention embodiment, the target historical meteorological feature time-series data corresponding to the target meteorological feature and the intrinsic mode subsequences are input into the constructed ocean wave significant wave height prediction model for training. Compared with related technologies, this invention enables the wave significant wave height prediction model to fully learn the changing mechanism of wave significant wave height under meteorological driving and its own fluctuation mode, from feature selection and data decomposition to multi-source data input training. The trained wave significant wave height prediction model has better adaptability to different sea conditions in wave significant wave height prediction, thus improving the accuracy of the wave significant wave height prediction model.

[0103] This embodiment provides a training method for a predictive model of significant wave height, which can be used with computer equipment. Figure 5 This is a flowchart of another training method for the ocean wave significant wave height prediction model according to an embodiment of the present invention, such as... Figure 5 As shown, the process includes the following steps:

[0104] Step S501: Based on the correlation between historical meteorological feature time-series data and historical sea wave height time-series data for multiple meteorological features, a target meteorological feature is selected from the multiple meteorological features; the target meteorological feature is a strongly correlated meteorological feature whose correlation meets preset conditions. For details, please refer to [link to relevant documentation]. Figure 1 Step S101 of the illustrated embodiment will not be described again here.

[0105] Step S502: Based on the sample ocean wave height time series data, optimize the number of decompositions and the penalty factor of variational mode decomposition to obtain the target number of decompositions and the target penalty factor.

[0106] Specifically, step S502 includes:

[0107] Step S5021: Obtain the first range of values ​​for the number of decompositions and the second range of values ​​for the penalty factor, and map the obtained chaotic sequence to the first and second ranges to obtain the initialized sparrow population.

[0108] Step S5022: Using each sparrow individual in the initialized sparrow population, perform variational mode decomposition on the sample ocean wave height time series data, evaluate the decomposition results using a preset fitness function, and obtain the optimal sparrow individual based on the evaluation results.

[0109] Step S5023: Update and iterate the sparrow population based on the best sparrow individual until the first preset stopping condition is met, and obtain the target decomposition quantity and target penalty factor.

[0110] The process of updating and iterating the sparrow population based on the optimal individual sparrow includes: updating and iterating the position of the discoverer individuals in the sparrow population based on the optimal individual sparrow, using sine and cosine perturbations and the golden ratio; updating and iterating the position of the follower individuals in the sparrow population based on the position of the discoverer individuals; and updating and iterating the position of the vigilant individuals in the sparrow population based on the reverse logic of the optimal individual sparrow, so as to update and iterate the sparrow population.

[0111] In some optional implementations, in the sparrow search algorithm, the sparrow flock is divided into three roles based on foraging strategies: discoverers, followers, and watchdogs. Discoverers are responsible for exploring the general area of ​​food, followers search based on clues from the discoverers, and watchdogs ensure the safety of the foraging area. This division of labor and cooperation allows the sparrow flock to forage efficiently in complex environments while ensuring safety. The position update method for discoverers is as follows:

[0112]

[0113] in, Let be the position of the i-th sparrow in the j-th dimension of the t-th generation, exp be the exponential function, ρ be the exploration step size control parameter, ρ∈(0,1], iter max Let be the maximum number of iterations, Q be a normally distributed random number, L be the first preset matrix, R2 be the warning value (R2∈[0,1]), and ST be the safety threshold (ST∈[0.5,1]).

[0114] Followers will continue to follow the discoverer in search of food. Follower replenishment occurs as follows:

[0115]

[0116] in, Let be the position of the i-th sparrow in the j-th dimension of the t-th generation, Q be a normally distributed random number, and exp be an exponential function. Let n be the worst position in the t-th generation, and n be the population size. Let A be the optimal discoverer position in generation t, and let L be the second preset matrix and L be the first preset matrix.

[0117] The vigilant individuals are randomly distributed throughout the population, comprising 10% to 20% of the population. Their positions are updated as follows:

[0118]

[0119] in, Let i be the position of the i-th sparrow in the j-th dimension of the t-th generation. Let be the global optimal position in generation t, and let β be a random number following a normal distribution with a variance of 1 and a mean of 0. The optimal discoverer position in generation t. f is the worst position in the t-th generation. i f is the fitness value for the current position. g f is the optimal fitness value in this iteration. w This is the worst fitness value in this iteration, where K is a random number in the range [-1, 1] and ε is a preset constant.

[0120] In this embodiment of the invention, the traditional SSA algorithm often randomly generates individuals in the population during initialization. However, the population diversity decreases in the later stages of iteration, making it easy to get trapped in local optima. This embodiment of the invention adds the Singer chaotic sequence and the golden sine and cosine strategy to the traditional SSA algorithm.

[0121] The formula for the Singer chaotic mapping is:

[0122]

[0123] Where, x n+1 Let x be the value of the chaotic variable in the next iteration. n The value of the chaotic variable in the current iteration is μ, which is the control parameter and μ∈[0,1].

[0124] In some alternative implementations, the update formula for the golden sine / cosine update population strategy is typically expressed as:

[0125]

[0126] in, Let i be the value of the i-th sparrow after updating in the j-th dimension. Let λ be the optimal value for the i-th sparrow in the j-th dimension, λ be the golden ratio, v be the frequency factor, and θ be the random phase factor. Let be the current value of the i-th sparrow in the j-th dimension, and let cos be the cosine function.

[0127] In some alternative implementations, the update formula for the golden sine / cosine update population strategy can also be expressed as:

[0128]

[0129] in, Let i be the value of the i-th sparrow after updating in the j-th dimension. Let λ be the optimal value for the i-th sparrow in the j-th dimension, where λ is the golden ratio (usually λ = 0.68) used to adjust the step size to ensure the update is close to the optimal solution, ω is the frequency factor, and θ is the random phase factor. Let be the current value of the i-th sparrow in the j-th dimension, and sin be the sine function.

[0130] In some optional implementations, the preset fitness function can be sample entropy (SampEn). Sample entropy is a parameter value that measures the complexity of time series data. It has two major advantages: it is independent of data length and has better consistency. Therefore, the minimum sample entropy of each subsequence is selected as the fitness function of the improved SSA algorithm to optimize the number of target decompositions and the target penalty factor.

[0131] For example, such as Figure 6 The diagram illustrates the optimization process for the number of target decompositions and the target penalty factor. Inputting valid wave height data (i.e., time-series sample wave height data), the upper and lower limits of the VMD (Variational Mode Decomposition) decomposition quantity K and the penalty factor α are set. The sparrow population is initialized using the Singer nonlinear chaotic mapping strategy. The fitness function is the minimum sample entropy (envelope entropy). The optimal individual is selected, and the position of the discoverer and follower is updated using a golden sine / cosine improvement strategy. 10% to 20% of individuals are randomly selected for updating the position of the watchdog, updating the individual target value, updating the optimal individual position and the optimal function value, and iterating until the iteration coefficients reach the set values. The optimal values ​​of VMD parameters K and α are then output.

[0132] Step S503: Based on the target decomposition quantity and the target penalty factor, perform variational mode decomposition on the historical wave height time series data to obtain the intrinsic mode subsequence of the target decomposition quantity.

[0133] Specifically, step S503 includes:

[0134] Step S5031: Construct a variational problem based on the number of target decompositions and historical wave height time series data.

[0135] Step S5032: Using the objective penalty factor and the Lagrange multiplication operator, the variational problem is transformed into an unconstrained variational problem.

[0136] Step S5033: Perform frequency domain transformation on the unconstrained variational problem, and iteratively solve the frequency domain problem obtained by the frequency domain transformation until the second preset stopping condition is met, so as to obtain the eigenmode subsequence of the target decomposition number.

[0137] In this embodiment of the invention, during the decomposition process of VMD, the number of decompositions K determines the number of modal components, and the penalty factor α determines the signal reconstruction accuracy. However, the traditional method of selecting parameters based on human experience has drawbacks such as subjectivity, inefficiency, and lack of scalability. If the number of decompositions K is set too large, over-decomposition will occur; if the number of decompositions K is too small, some signals will be lost. If the penalty factor α is set too large, the bandwidth of each intrinsic modal subsequence will be smaller; conversely, the bandwidth of the intrinsic modal subsequence signal will be larger, ultimately affecting the decomposition accuracy. Therefore, an improved sparrow search algorithm is used to optimize the number of decompositions K and the penalty factor α.

[0138] In this embodiment of the invention, VMD is an adaptive signal decomposition method with strong non-recursiveness, which can effectively suppress mode aliasing during the signal decomposition process. VMD can decompose historical significant wave height time series data into intrinsic mode subsequences with different randomness and volatility. These intrinsic mode subsequences reflect the regularity of wave height time series data in different frequency bands, reduce the non-stationarity of significant wave height time series data, and enhance the prediction performance of subsequent wave significant wave height prediction models.

[0139] Step S504: Input the target historical meteorological feature time series data and intrinsic mode subsequence corresponding to the target meteorological feature into the constructed sea wave significant wave height prediction model for training. For details, please refer to [link to relevant documentation]. Figure 1 Step S101 of the illustrated embodiment will not be described again here.

[0140] In this embodiment of the invention, a first range of values ​​for the number of decompositions and a second range of values ​​for the penalty factor are obtained. The obtained chaotic sequence is mapped to the first and second ranges to obtain an initialized sparrow population. This allows the sparrow population to be more evenly distributed throughout the search space, increasing the chance of finding the global optimum and reducing the risk of the algorithm getting stuck in local optima, thereby accelerating the convergence speed. This invention utilizes each individual sparrow in the initialized sparrow population to perform variational mode decomposition on the sample ocean wave height time series data. A preset fitness function is used to evaluate the decomposition results, and the optimal sparrow individual is obtained based on the evaluation results. The sparrow population is then updated iteratively based on the optimal sparrow individual until a first preset stopping condition is reached, obtaining the target number of decompositions and the target penalty factor, thus improving the flexibility and efficiency of parameter optimization.

[0141] This embodiment provides a training method for a predictive model of significant wave height, which can be used with computer equipment. Figure 7 This is a flowchart of another training method for the significant wave height model of ocean waves according to an embodiment of the present invention, such as... Figure 7 As shown, the process includes the following steps:

[0142] Data processing and feature selection: Outlier removal and cubic spline filling for missing values ​​are performed on the characteristic time series data (historical meteorological characteristic time series data) and the significant wave height data (historical sea wave height time series data). Based on the characteristic time series data, mutual information feature selection is performed using MIC to obtain strongly correlated features (target meteorological features). CWT denoising is then performed on the significant wave height data.

[0143] The time series decomposition method is improved by using the golden sine and cosine strategy and the Singer chaotic mapping to improve the sparrow search algorithm. The improved sparrow search algorithm is used to search for the optimal number of parameters K and the penalty factor α of VMD. The denoised effective wave height data is input into the optimized VMD for decomposition to obtain multiple intrinsic mode functions (IMFs). The multiple IMFs and strongly correlated feature data are normalized and used to train the Transformer-BiLSTM prediction model.

[0144] Data prediction involves inputting time-series meteorological characteristic data into the Transformer-BiLSTM prediction model for prediction, obtaining multiple intrinsic mode function prediction values, and then superimposing these multiple intrinsic mode function prediction values ​​to obtain the final prediction value.

[0145] This embodiment provides a training method for a predictive model of significant wave height, which can be used with computer equipment. Figure 8 This is a flowchart of a method for predicting the significant wave height of ocean waves according to an embodiment of the present invention, such as... Figure 8 As shown, the process includes the following steps:

[0146] Step S801: Obtain the time series data of the forecast meteorological features corresponding to the target meteorological features at the time to be predicted, and input the time series data of the forecast meteorological features into... Figures 1 to 6 The effective wave height prediction model of ocean waves in the model yields multiple intrinsic mode prediction subsequences; the target meteorological features are meteorological features with strong correlations that meet preset conditions, selected from multiple meteorological features based on the correlation between the historical meteorological feature time series data of multiple meteorological features and the historical ocean wave height time series data.

[0147] Step S802: Superimpose multiple intrinsic mode prediction subsequences to obtain the predicted wave height data for the time to be predicted.

[0148] In this embodiment of the invention, to verify the feasibility of the significant wave height prediction model, measured wave data from Sea A and Sea B were selected. 80% of the data was used as the training set, 10% as the validation set, and 10% as the test set. Significant wave height is influenced by characteristic variables such as gust wind speed, average wind speed, average wave period, main wave period, sea level pressure, wind direction, and air temperature. Selecting strongly correlated input variables and reducing the redundancy of input variables in the prediction model is key to improving the accuracy of significant wave height prediction. The Micro-Information Coefficient (MIC) method was used to select strongly correlated characteristic variables and determine the input parameters that have the greatest influence on wave height prediction. The maximum mutual information coefficients between each characteristic variable and the significant wave height are shown in Table 1.

[0149] Table 1: Schematic table of maximum mutual information coefficients between each characteristic variable and the significant wave height.

[0150]

[0151] If we use 0.2 as the threshold for strong correlation characteristic variables, Table 1 shows that the mean wave period, gust speed, and mean wind speed are highly positively correlated with the significant wave height, while sea level pressure, main wave period, air temperature, and wind direction are weakly correlated with the significant wave height. Therefore, we select the mean wave period, gust speed, and mean wind speed as the input variables for the sea wave significant wave height prediction model.

[0152] In some optional implementations, to better evaluate the predictive performance of the wave significant wave height prediction model, the mean square error E is selected. MSE Root mean square error E RMSE Mean absolute error E MAE Mean absolute percentage error E MAPE The coefficient of determination (R²) is used to evaluate the prediction results of the significant wave height prediction model. The expressions for each evaluation index are as follows:

[0153]

[0154] Among them, E MSE Mean square error, E RMSE For root mean square error, E MAE For mean absolute error, E MAPE H is the mean absolute percentage error, R² is the coefficient of determination, and H i The actual value of the effective wave height. For the effective wave height prediction value, is the average effective wave height, and n is the total number of effective wave height data.

[0155] In this embodiment of the invention, the root mean square error E RMSE Mean absolute error E MAE Mean absolute percentage error E MAPEA smaller R value indicates higher prediction accuracy of the significant wave height prediction model. 2 The value of 1 is used to evaluate the fitting effect of the wave significant height prediction model. The closer the value is to 1, the higher the fitting effect of the wave significant height prediction model.

[0156] In some optional implementations, an improved SSA-VMD adaptive signal decomposition model is used to decompose the effective wave height time series data to reduce its randomness and volatility. The improved SSA algorithm is used to optimize the number of decompositions K and the penalty factor α with the minimum sample entropy as the fitness function. The parameter settings are shown in Table 2.

[0157] Table 2: Schematic diagram of decomposition quantity K and penalty factor α.

[0158] K α Maximum number of iterations Population number [2,10] [10,2500] 10 25

[0159] For example, such as Figure 9 The diagram shows the results of the improved SSA optimization of VMD. It can be seen that the process tends to stabilize when the number of iterations is 2. At this time, the optimal number of decompositions K = 9 and the optimal penalty factor α = 306 are obtained.

[0160] like Figure 10 As shown, this is a schematic diagram of the modal components and residuals of the effective wave height decomposition. The horizontal axis represents the sampling point number, the vertical axis represents the signal amplitude of different intrinsic mode functions (IMFs), and res represents the residual.

[0161] In some optional implementations, to verify the effectiveness of the wave significant height prediction model, three models, LSTM, BiLSTM and Transformer-BiLSTM, were selected for experimental comparison. The model parameter settings are shown in Table 3.

[0162] Table 3: Schematic diagram of model parameter settings.

[0163]

[0164] In some alternative implementations, based on LSTM, BiLSTM, Transformer-BiLSTM, and the wave significant wave height prediction model (MIC-improved SSA optimized VMD-Transformer-BiLSTM prediction model) of the present invention, the wave height values ​​for the next 24 hours (24 data points) are predicted in multiple steps using 240 consecutive time series data points (historical 10 days) as input. Figure 11 This is a diagram showing the comparison of prediction results from ocean buoy A. Figure 11 This presentation shows a comparison of continuous rolling forecasts from ocean buoy A over a period of 20 days, from 00:50 on November 1, 2018 to 23:50 on November 20, 2018. Figure 12This is a diagram comparing the prediction results of Ocean Buoy B. Figure 12 This compares the continuous rolling forecast results of Ocean Buoy B for a total of 20 days, from 00:50 on September 27, 2023 to 23:50 on October 16, 2023. Figure 11 and Figure 12 As shown, the wave prediction curves of the wave significant wave height prediction model are in significantly better agreement with the actual observations than the comparison model. The evaluation indicators for model prediction accuracy are shown in Tables 4 and 5.

[0165] Table 4: Comparison of evaluation indicators of the effective wave height prediction accuracy of each model for ocean buoy A.

[0166]

[0167] Table 5: Comparison of evaluation indicators of the effective wave height prediction accuracy of each model for ocean buoy B.

[0168]

[0169] In the wave height prediction of ocean buoys A and B, the significant wave height prediction model showed a comprehensive and significant improvement in prediction evaluation indicators compared to the Transformer-BiLSTM model, with E... MAE E MAPE E MSE E RMSE The decreases were 45.9% / 35.9%, 3.5% / 2.1%, 72.5% / 62%, and 23.1% / 8.1%, respectively; R 2 The indicators improved by 8.3% / 8.1%, fully verifying the superiority and stability of the wave significant wave height prediction model in wave significant wave height prediction.

[0170] This embodiment also provides a training device for a significant wave height prediction model, which is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0171] This embodiment provides a training device for a prediction model of significant wave height, such as... Figure 13 As shown, it includes:

[0172] The meteorological feature selection module 1301 is used to select a target meteorological feature from multiple meteorological features based on the correlation between the historical meteorological feature time series data of multiple meteorological features and the historical sea wave height time series data; the target meteorological feature is a meteorological feature with strong correlation that meets the preset conditions.

[0173] The parameter optimization module 1302 is used to optimize the number of decompositions and the penalty factor of variational mode decomposition based on the sample ocean wave height time series data, so as to obtain the target number of decompositions and the target penalty factor.

[0174] The variational mode decomposition module 1303 is used to perform variational mode decomposition on historical ocean wave height time series data based on the target decomposition number and the target penalty factor to obtain the intrinsic mode subsequence of the target decomposition number.

[0175] The model training module 1304 is used to input the target historical meteorological feature time series data and intrinsic mode subsequence corresponding to the target meteorological feature into the constructed sea wave significant wave height prediction model for training.

[0176] In some optional implementations, the meteorological feature selection module 1301 includes:

[0177] The target table determination unit is used to divide the historical meteorological feature time series data and historical sea wave height time series data of each meteorological feature into intervals to obtain target tables of different resolutions.

[0178] The probability distribution determination unit is used to determine the joint probability distribution function and the marginal probability distribution function based on the number of data points in the target table at each resolution. The joint probability distribution function is used to characterize the probability of meteorological characteristic variables and historical wave height variables occurring simultaneously. The marginal probability distribution function is used to characterize the meteorological probability distribution of meteorological characteristic variables and the wave height probability distribution of historical wave height variables.

[0179] The mutual information determination unit is used to determine the mutual information of the target table at each resolution based on the joint probability distribution function and the marginal probability distribution function.

[0180] The maximum mutual information determination module is used to select the maximum mutual information from multiple mutual information and compare the maximum mutual information with a preset mutual information threshold; the maximum mutual information is used to characterize the correlation between historical meteorological feature time series data and historical sea wave height time series data.

[0181] The selection unit is used to select the corresponding meteorological feature as the target meteorological feature based on the maximum mutual information being greater than or equal to a preset mutual information threshold, and to discard the corresponding meteorological feature based on the maximum mutual information being less than the preset mutual information threshold.

[0182] In some alternative implementations, the parameter optimization module 1302 includes:

[0183] The population initialization unit is used to obtain a first range of values ​​for the number of decompositions and a second range of values ​​for the penalty factor, and to map the obtained chaotic sequence to the first and second ranges to obtain an initialized sparrow population.

[0184] The evaluation unit is used to perform variational mode decomposition on the sample ocean wave height time series data using each individual sparrow in the initialized sparrow population, evaluate the decomposition results using a preset fitness function, and obtain the optimal individual sparrow based on the evaluation results.

[0185] The update iteration unit is used to update and iterate the sparrow population based on the best sparrow individual until the first preset stopping condition is met, and to obtain the target decomposition quantity and the target penalty factor.

[0186] Specifically, the update iteration unit includes:

[0187] The discoverer update sub-unit is used to update the position of the discoverer individuals in the sparrow population based on the optimal sparrow individual by combining sine and cosine perturbations and the golden ratio.

[0188] The follower update subunit is used to update the positions of follower individuals in the sparrow population based on the position of the discoverer individual.

[0189] The vigilant update subunit is used to update and iterate the position of vigilant individuals in the sparrow population based on the reverse logic that depends on the best individual sparrow, so as to update and iterate the sparrow population.

[0190] In some alternative implementations, the variational mode decomposition module 1303 includes:

[0191] The variational problem construction unit is used to construct variational problems based on the number of target decompositions and historical wave height time series data.

[0192] The problem transformation unit is used to transform variational problems into unconstrained variational problems using the objective penalty factor and the Lagrange multiplication operator.

[0193] The iterative solution unit is used to perform frequency domain transformation on the unconstrained variational problem, and iteratively solve the frequency domain problem obtained by the frequency domain transformation until the second preset stopping condition is met, so as to obtain the eigenmode subsequence of the target decomposition number.

[0194] This embodiment also provides a device for predicting the significant wave height of ocean waves. This device is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0195] This embodiment provides a device for predicting the significant wave height of ocean waves, such as... Figure 14 As shown, it includes:

[0196] The model prediction module 1401 is used to acquire the time series data of the forecast meteorological features corresponding to the target meteorological features at the time to be predicted, and input the time series data of the forecast meteorological features. Figures 1 to 7 The effective wave height prediction model for ocean waves yields multiple intrinsic mode prediction subsequences. The target meteorological feature is a meteorological feature with strong correlation that meets preset conditions, selected from multiple meteorological features based on the correlation between the historical meteorological feature time series data of multiple meteorological features and the historical ocean wave height time series data.

[0197] The sequence overlay module 1402 is used to overlay multiple intrinsic mode prediction subsequences to obtain the effective wave height prediction data of the sea waves at the time to be predicted.

[0198] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0199] In this embodiment, the training device and the prediction device for the significant wave height prediction model are presented in the form of functional units. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0200] This invention also provides a computer device having the above-described features. Figure 13 The training device for the wave significant wave height prediction model shown is as follows: Figure 14 The device shown is for predicting the effective wave height of ocean waves.

[0201] Please see Figure 15 , Figure 15 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of the present invention, such as... Figure 15 As shown, the computer device includes one or more processors 1510, memory 1520, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components are interconnected via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 15Take the 1510 processor as an example.

[0202] Processor 1510 may be a central processing unit, a network processor, or a combination thereof. Processor 1510 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0203] The memory 1520 stores instructions executable by at least one processor 1510 to cause the at least one processor 1510 to perform the method shown in the above embodiments.

[0204] Memory 1520 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, memory 1520 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, memory 1520 may optionally include memory remotely located relative to processor 1510, and this remote memory may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0205] The memory 1520 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 1520 may also include a combination of the above types of memory.

[0206] The computer device also includes a communication interface 1530 for communicating with other devices or communication networks.

[0207] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0208] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0209] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A training method for a predictive model of significant wave height, characterized in that, The method includes: Based on the correlation between historical meteorological feature time series data of multiple meteorological features and historical sea wave height time series data, a target meteorological feature is selected from the multiple meteorological features; the target meteorological feature is a meteorological feature with a strong correlation that satisfies a preset condition. Based on the sample ocean wave height time series data, the number of decompositions and the penalty factor of variational mode decomposition are optimized to obtain the target number of decompositions and the target penalty factor. Based on the target decomposition quantity and the target penalty factor, variational mode decomposition is performed on the historical wave height time series data to obtain the intrinsic mode subsequence of the target decomposition quantity. The target historical meteorological feature time series data corresponding to the target meteorological feature and the intrinsic mode subsequence are input into the constructed sea wave significant wave height prediction model for training.

2. The method according to claim 1, characterized in that, The step of selecting a target meteorological feature from among multiple meteorological features based on the correlation between historical meteorological feature time-series data and historical sea wave height time-series data includes: The historical meteorological feature time series data and the historical sea wave height time series data of each meteorological feature are divided into intervals to obtain target tables of different resolutions; Based on the number of data points in the target table at each resolution, a joint probability distribution function and a marginal probability distribution function are determined; the joint probability distribution function is used to characterize the probability of meteorological characteristic variables and historical wave height variables occurring simultaneously; the marginal probability distribution function is used to characterize the meteorological probability distribution of the meteorological characteristic variables and the wave height probability distribution of the historical wave height variables. Based on the joint probability distribution function and the marginal probability distribution function, the mutual information of the target table at each resolution is determined; The maximum mutual information is selected from multiple mutual information values, and the maximum mutual information is compared with a preset mutual information threshold; the maximum mutual information is used to characterize the correlation between the historical meteorological feature time series data and the historical sea wave height time series data. If the maximum mutual information is greater than or equal to the preset mutual information threshold, the corresponding meteorological feature is selected as the target meteorological feature; if the maximum mutual information is less than the preset mutual information threshold, the corresponding meteorological feature is discarded.

3. The method according to claim 1 or 2, characterized in that, The optimization of the number of decompositions and the penalty factor of variational mode decomposition based on sample ocean wave height time series data yields the target number of decompositions and the target penalty factor, including: Obtain a first range of values ​​for the number of decompositions and a second range of values ​​for the penalty factor, and map the obtained chaotic sequence to the first range and the second range to obtain an initialized sparrow population; Using each individual sparrow in the initial sparrow population, variational mode decomposition is performed on the sample ocean wave height time series data. The decomposition results are evaluated using a preset fitness function, and the optimal individual sparrow is obtained based on the evaluation results. Based on the optimal sparrow individual, the sparrow population is updated and iterated until the first preset stopping condition is reached, thereby obtaining the target decomposition quantity and the target penalty factor.

4. The method according to claim 3, characterized in that, The updating and iteration of the sparrow population based on the optimal sparrow individual includes: Combining sine and cosine perturbations and the golden ratio, the position of the discoverer individual in the sparrow population is updated and iterated based on the optimal sparrow individual; the position of the follower individual in the sparrow population is updated and iterated based on the position of the discoverer individual; and the position of the vigilant individual in the sparrow population is updated and iterated based on the reverse logic of the optimal sparrow individual, so as to update and iterate the sparrow population.

5. The method according to claim 1 or 2, characterized in that, The step of performing variational mode decomposition on the historical wave height time series data based on the target decomposition number and the target penalty factor to obtain the intrinsic mode subsequence of the target decomposition number includes: Based on the target decomposition quantity and the historical wave height time series data, a variational problem is constructed; Using the target penalty factor and the Lagrange multiplication operator, the variational problem is transformed into an unconstrained variational problem; The unconstrained variational problem is transformed into the frequency domain, and the frequency domain problem obtained by the frequency domain transformation is solved iteratively until the second preset stopping condition is met, so as to obtain the intrinsic mode subsequence of the target decomposition number.

6. The method according to claim 1 or 2, characterized in that, The wave significant height prediction model is a model constructed based on a deep learning network with a self-attention mechanism and a bidirectional long short-term memory network.

7. A method for predicting the significant wave height of ocean waves, characterized in that, The method includes: Obtain the time series data of the forecast meteorological features corresponding to the target meteorological features at the time to be predicted, and input the time series data of the forecast meteorological features into the sea wave significant wave height prediction model according to any one of claims 1 to 6 to obtain multiple intrinsic mode prediction subsequences; the target meteorological features are meteorological features with strong correlations that satisfy preset conditions, selected from the multiple meteorological features based on the correlation between the historical meteorological feature time series data of multiple meteorological features and the historical sea wave height time series data respectively. The predicted subsequences of multiple intrinsic modes are superimposed to obtain the predicted wave height data of the sea waves at the time to be predicted.

8. A training device for a predictive model of significant wave height, characterized in that, The device includes: The meteorological feature selection module is used to select a target meteorological feature from the multiple meteorological features based on the correlation between historical meteorological feature time series data and historical sea wave height time series data; the target meteorological feature is a meteorological feature with a strong correlation that satisfies a preset condition. The parameter optimization module is used to optimize the number of decompositions and the penalty factor of variational mode decomposition based on the sample ocean wave height time series data, so as to obtain the target number of decompositions and the target penalty factor. The variational mode decomposition module is used to perform variational mode decomposition on the historical wave height time series data based on the target decomposition number and the target penalty factor to obtain the intrinsic mode subsequence of the target decomposition number. The model training module is used to input the target historical meteorological feature time series data corresponding to the target meteorological feature and the intrinsic mode subsequence into the constructed sea wave significant wave height prediction model for training.

9. A device for predicting the significant wave height of ocean waves, characterized in that, The device includes: The model prediction module is used to acquire the time series data of the forecast meteorological features corresponding to the target meteorological features at the time to be predicted, and input the time series data of the forecast meteorological features into the sea wave significant wave height prediction model according to any one of claims 1 to 6 to obtain multiple intrinsic mode prediction subsequences; the target meteorological features are meteorological features with strong correlations that satisfy preset conditions, selected from the multiple meteorological features based on the correlation between the historical meteorological feature time series data of multiple meteorological features and the historical sea wave height time series data respectively. The sequence overlay module is used to overlay multiple intrinsic mode prediction sub-sequences to obtain the effective wave height prediction data of the sea waves at the time to be predicted.

10. A computer device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the training method for the wave significant height prediction model according to any one of claims 1 to 6 or the wave significant height prediction method according to claim 7.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the training method of the wave significant wave height prediction model according to any one of claims 1 to 6 or the wave significant wave height prediction method according to claim 7.

12. A computer program product, characterized in that, Includes computer instructions for causing a computer to execute the training method of the wave significant wave height prediction model according to any one of claims 1 to 6 or the wave significant wave height prediction method according to claim 7.

Images