Sea wave significant wave height intelligent prediction and error correction method and system based on 2D-Geoform

Abstract

The invention discloses an intelligent sea wave significant wave height prediction and error correction method and system based on 2D-Geoform, and relates to the technical field of ocean exploration, and the method comprises the following steps: collecting and preprocessing original data, and constructing a training data set; constructing a 2D-Geoform deep learning model, and performing model training by taking the training data set as input and the prediction error as output until the loss function is converged to obtain an error diagnosis model; and using the error diagnosis model to identify the estimated deviation of the sea wave significant wave height prediction, and subtracting the estimated deviation from the originally predicted sea wave significant wave height to obtain the corrected sea wave significant wave height. According to the method, the accuracy of the numerical wave model and the analysis capability of a mathematical-statistical method can be combined, the accuracy of SWH prediction is improved, and inherent errors in numerical prediction are corrected.

Description

Technical Field

[0001] The present invention relates to the field of ocean detection technology, and more specifically to a method and system for intelligently predicting and correcting the effective wave height of ocean waves based on 2D-Geoformer. Background Art

[0002] The significant wave height (SWH) constitutes an important aspect of ocean surface dynamics, summarizing the average peak wave height of the first third of large waves. This indicator is of great significance and is fundamental to the safety and operational planning of numerous maritime, research and recreational activities. Accurate SWH forecasts are particularly important in the Northwest Pacific region, due to the busy maritime economic activities in the region and the presence of tropical cyclones introduces considerable variability, making reliable predictions a complex and challenging task. Accurate predictions of SWH are not only essential to safeguard maritime navigation, but also help advance scientific knowledge and enhance the enjoyment of marine recreational activities. However, achieving precise forecasts is inherently difficult due to the dynamic interactions between the various factors that affect wave generation, growth and dissipation. These factors include a range of dynamical, physical and environmental conditions, highlighting the significant scientific efforts involved in improving the accuracy of SWH predictions.

[0003] In the field of SWH forecasting, there are two main methods that are widely recognized. The first is based on the principle of physical dynamics, using wave numerical models to accurately simulate the generation, movement and dissipation of ocean waves by calculating and analyzing the basic wave equations. These models combine a series of physical processes such as wave generation, spectral distribution, propagation, nonlinear wave interaction, dissipation mechanism, and refraction and diffraction effects. Famous numerical models include Wave Watch III (WW3), the Nearshore Wave Simulation (SWAN) system and the MASUM model. Various studies have confirmed that these models have shown excellent effectiveness in predicting SWH. At present, operational wave forecasting systems mainly rely on these advanced numerical wave models to provide accurate SWH forecasts.

[0004] Another approach to forecasting SWH is the mathematical-statistical approach, which can be roughly divided into point forecasting and spatiotemporal forecasting. Point forecasting delves into the spectral features, context, and temporal dependencies in SWH series to achieve continuous time series forecasting. This category includes various methods such as wavelet analysis, particle swarm optimization (PSO), extreme learning machine (ELM) methods, Bayesian hyperparameter optimization, elastic network methods, singular value decomposition (SVD), and empirical mode decomposition (EMD). The extended version of point forecasting not only analyzes the temporal evolution of SWH, but also integrates influencing factors such as wind speed, wind direction, duration, wind zone, sea level pressure, and air temperature. Techniques in this area include long short-term memory (LSTM) networks, hierarchical machine learning models, artificial neural networks (ANNs), multiple additive regression trees (MARTs), hybrid models of wavelets and neural networks (WNNs), and pruned radial basis function (GAP-RBF) networks.

[0005] Spatiotemporal forecasting goes a step further, taking into account the temporal evolution and spatial correlation of SWH and the spatiotemporal dynamics of influencing meteorological factors. This approach expands the forecast scope to the entire region, such as convolutional LSTM networks and multivariate 3-layer LSTM-based methods, which provide more comprehensive and regional forecasts for SWH. These sophisticated techniques are designed to capture the complexity of wave dynamics in time and space to improve the accuracy of maritime forecasts.

[0006] Advances in mathematical-statistical methods have enabled substantial progress in SWH forecasting. However, they have not yet achieved forecast accuracy comparable to that of numerical wave models. This gap is attributed to the reliance of traditional methods on long-term observational data, often overlooking the development of numerical wave modeling that can significantly improve SWH forecasting capabilities. Although numerical models offer improved accuracy, their performance depends on the quality of input data, model resolution, and accurate representation of physical processes, and any inaccuracies in these factors can severely affect forecast results.

[0007] Despite these shortcomings, mathematical-statistical methods still excel in revealing complex correlations in data sets, and numerical models are adept at capturing the physical dynamics of SWHs. Therefore, how to combine the accuracy of numerical wave models with the analytical power of mathematical-statistical methods to improve the accuracy of SWH predictions and correct the inherent errors in numerical predictions is a technical problem that technicians in this field need to solve urgently. Summary of the invention

[0008] In view of this, the present invention provides a method and system for intelligent prediction and error correction of ocean wave significant wave height based on 2D-Geoformer, which solves the problems existing in the background technology.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] A method for intelligent prediction and error correction of ocean wave significant wave height based on 2D-Geoformer comprises the following steps:

[0011] Collect raw data and preprocess them to build a training data set;

[0012] Construct a 2D-Geoformer deep learning model, and train the model using the training data set as input and the prediction error as output until the loss function converges to obtain an error diagnosis model;

[0013] The error diagnosis model is used to identify the estimated deviation of the wave significant wave height prediction, and the estimated deviation is subtracted from the original predicted wave significant wave height to obtain the corrected wave significant wave height.

[0014] Optionally, the original data is derived from the predicted value of the significant wave height of the ocean wave of an ocean prediction system, wherein the ocean prediction system is designed based on a surface wave-tide-circulation coupled ocean model.

[0015] Optionally, construct a training data set, which specifically includes the following steps:

[0016] Perform data labeling, data cleaning, normalization and feature processing on the original data to obtain the original data set;

[0017] The coordinates of each original data are obtained, and based on the coordinates, the sea level pressure matrix and the significant wave height matrix corresponding to the coordinates of the original data are extracted.

[0018] Optionally, the 2D-Geoformer deep learning model is built based on the encoder-decoder framework, including: data preprocessing module, encoder, decoder, and output layer;

[0019] The data preprocessing module is located at the base of the encoder and decoder stack, which is used to receive the input of the predicted value of the significant wave height and perform patch embedding;

[0020] The encoder is used to convert the input data into feature maps through the spatiotemporal attention mechanism to establish complex spatiotemporal connections between multivariate data points;

[0021] The decoder analyzes the output of the data preprocessing module and the encoder to generate the final output field;

[0022] The output layer is used to output the decoder's output fields at the same spatial resolution as the input data.

[0023] Optionally, the specific operations in the data preprocessing module are as follows:

[0024] Receive the original predicted significant wave heights, and divide each set of input data into N non-overlapping patches of fixed size;

[0025] The decomposed non-overlapping patches are converted into patch embeddings of a specific dimension through a learnable linear layer;

[0026] The spatial position embedding and temporal embedding are integrated into the patch embedding to maintain the spatiotemporal context of each patch.

[0027] Optionally, the encoder consists of four identical coding blocks, each coding block including two different sub-layers;

[0028] The first sublayer is a multi-head spatiotemporal attention layer, which is used to evaluate dependencies across spatiotemporal dimensions through a scaled dot product attention score function to generate a feature matrix;

[0029] The second sub-layer is a fully connected feed-forward network, which is used to facilitate the flow and integration of information throughout the encoding process by adding residual connections around each sub-layer.

[0030] Optionally, the decoder is composed of four identical decoding blocks, each decoding block includes three different sub-layers, and residual connections are added around each sub-layer;

[0031] The first and third sublayers both include multi-head spatiotemporal attention layers and feedforward networks;

[0032] The second sub-layer is used to perform spatiotemporal attention analysis, integrating the encoder’s feature matrix with the results of the previous sub-layer to produce the final output field.

[0033] Optionally, the output layer consists of a feed-forward network that maps and reshapes the decoder refined representation to its original resolution.

[0034] Optionally, the method further includes:

[0035] Taking the reanalysis dataset ERA5 as the observation benchmark, the trained error diagnosis model is corrected to obtain a corrected error diagnosis model that meets the preset requirements.

[0036] The present invention also provides an intelligent prediction and error correction system for significant wave height of ocean waves based on 2D-Geoformer, comprising:

[0037] The data set building module is used to collect raw data and perform preprocessing to build a training data set;

[0038] Model building module, used to build 2D-Geoformer deep learning model;

[0039] The model training module is used to train the model with the training data set as input and the prediction error as output until the loss function converges to obtain the error diagnosis model;

[0040] The correction module is used to identify the estimated deviation of the prediction of the wave significant wave height by using the error diagnosis model, and subtract the estimated deviation from the original predicted wave significant wave height to obtain the corrected wave significant wave height.

[0041] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a method and system for intelligent prediction and error correction of significant wave height based on 2D-Geoformer, which has the following features:

[0042] Beneficial effects:

[0043] Combining the accuracy of numerical wave models with the analytical capabilities of mathematical-statistical methods, a 2D-Geoformer deep learning model is designed to explain the relationship between the original wave height prediction and the related errors and perform error correction, improve the accuracy of SWH prediction, correct the inherent errors in numerical prediction, and provide important support for maritime navigation, coastal management and broader marine economic activities. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0045] Figure 1 A flow chart of the method for intelligent prediction and error correction of ocean wave significant wave height based on 2D-Geoformer provided by the present invention;

[0046] Figure 2 A structural diagram of the 2D-Geoformer deep learning model provided by the present invention;

[0047] Figure 3 This is a structural diagram of the ocean wave significant wave height intelligent prediction and error correction system based on 2D-Geoformer provided by the present invention. DETAILED DESCRIPTION

[0048] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] In order to improve the accuracy of the prediction of the significant wave height of the ocean wave, the embodiment of the present invention discloses a method for intelligent prediction and error correction of the significant wave height of the ocean wave based on 2D-Geoformer, such as Figure 1 As shown, the following steps are included:

[0050] Collect raw data and preprocess them to build a training data set;

[0051] Construct a 2D-Geoformer deep learning model, and train the model using the training data set as input and the prediction error as output until the loss function converges to obtain an error diagnosis model;

[0052] The error diagnosis model is used to identify the estimated deviation of the wave significant wave height prediction, and the estimated deviation is subtracted from the original predicted wave significant wave height to obtain the corrected wave significant wave height.

[0053] Next, Figure 1 The process shown is explained in detail for a better understanding.

[0054] 1. Data preparation

[0055] In this embodiment, the original data comes from the predicted value of the significant wave height of the ocean prediction system, and the ocean prediction system is designed based on the surface wave-tide-circulation coupled ocean model developed by the First Institute of Oceanography (FIO-COM) of the Ministry of Resources of China.

[0056] Specifically, in order to optimize computing efficiency and resource management, the data collected this time are concentrated in the northwest Pacific region, which presents diverse ocean and climate conditions and is an ideal research object. The surface wave-tide-circulation coupled ocean model runs at high temporal and spatial resolution, with data updated every three hours and a spatial resolution of 0.1°×0.1°, which is easy to manage and reduce the dimension of the data set, helping to enhance the understanding of the ocean conditions along the Maritime Silk Road.

[0057] Furthermore, constructing a training data set specifically includes the following steps:

[0058] Perform data labeling, data cleaning, normalization and feature processing on the original data to obtain the original data set;

[0059] The coordinates of each original data are obtained, and based on the coordinates, the sea level pressure matrix and the significant wave height matrix corresponding to the coordinates of the original data are extracted.

[0060] In this embodiment, the raw data may include wind direction, average wind speed, peak wind speed, dominant wave period, average wave period, wave direction, sea level pressure, air temperature, sea surface temperature, significant wave height, etc.

[0061] After collecting the original data, abnormal wave values ​​can be eliminated, and the missing values ​​of the effective wave height can be filled using filling methods including average value filling and interpolation filling, and zero values ​​can be filled for different grid data on land.

[0062] The normalization formula is:

[0063]

[0064] Where, X * is the standardized data, X is the original data, X is the mean of the original data, and δ is the standard deviation of the original data.

[0065] 2. Model construction

[0066] 2D-Geoformer is an improvement on 3D-Geoformer, which was developed for analyzing three-dimensional multivariate distributions and their spatiotemporal interactions, while 2D-Geoformer focuses on horizontal spatiotemporal attention aspects related to the prediction of wave significant wave height and its prediction error. This improvement takes into account the key relationship between wave significant wave height and the corresponding error.

[0067] In this embodiment, the 2D-Geoformer deep learning model is built based on the encoder-decoder framework, such as Figure 2 As shown, it includes: a data preprocessing module, an encoder, a decoder and an output layer;

[0068] The data preprocessing module is located at the base of the encoder and decoder stack, which is used to receive the input of the predicted value of the significant wave height and perform patch embedding;

[0069] The encoder is used to convert the input data into feature maps through a spatiotemporal attention mechanism to establish complex spatiotemporal connections between multivariate data points; the encoder setting can simplify the calculation of dependencies between arbitrary positions and reduce it to a constant number of operations;

[0070] The decoder analyzes the output of the data preprocessing module and the encoder to generate the final output field;

[0071] The output layer is used to output the decoder's output fields at the same spatial resolution as the input data.

[0072] Furthermore, the specific operations in the data preprocessing module are as follows:

[0073] Receive the original predicted significant wave height (dimension is 25*132*154), each set of input data is divided into N fixed-size non-overlapping patches; where each patch size is 4*4, N=(N lat / h0)×(N lon / w0);

[0074] Through a learnable linear layer, the decomposed non-overlapping patches are converted into patch embeddings of a specific dimension. Among them, the selected embedding dimension is d = 256;

[0075] Since the 2D-Geoformer model lacks the order and convolution operations that traditionally inject spatiotemporal position information, the spatial position is embedded and time embedding is integrated into the patch embedding to maintain the spatiotemporal context of each patch, forming and introduced into the subsequent encoding process.

[0076] Furthermore, the encoder consists of four identical coding blocks, each of which includes two different sub-layers;

[0077] The first sublayer is a multi-head spatiotemporal attention layer, which is used to evaluate dependencies across spatiotemporal dimensions through a scaled dot product attention score function to generate a feature matrix;

[0078] The second sub-layer is a fully connected feed-forward network, which is used to facilitate the flow and integration of information throughout the encoding process by adding residual connections around each sub-layer.

[0079] Specifically, when the encoder receives After that, the multi-head spatiotemporal attention layer of the first encoding block begins to analyze, evaluate the dependencies across spatiotemporal dimensions, and generate a feature matrix; the matrix is ​​linearly transformed as the input of the subsequent encoding block; after four encoding block feature extraction iterations, Efficiently compressed into a feature map matrix It is then passed to the decoder for further retrieval analysis.

[0080] In this embodiment, the multi-head spatiotemporal attention layer includes: a multi-channel connection module and a spatial attention weighted module, wherein the spatial attention weighted module includes multiple dot product attention sub-modules.

[0081] Furthermore, the decoder consists of four identical decoding blocks, each of which includes three different sub-layers, and residual connections are added around each sub-layer;

[0082] The first and third sublayers both include multi-head spatiotemporal attention layers and feedforward networks;

[0083] The second sublayer is used for spatiotemporal attention analysis, which transforms the encoder feature matrix It is integrated with the result of the previous sub-layer to produce the final output field.

[0084] Specifically, the decoder receives two inputs, one is the output of the data preprocessing module at the bottom of the decoder stack, and the other is the feature matrix generated by the encoder. These input data are processed by 4 decoding blocks and the output retrieval field is It retains the same spatial resolution as the initial predicted wave significant height; where T out Indicates the time series length of the search field, which is consistent with the length of the input data.

[0085] Furthermore, the output layer consists of a feed-forward network to map and reshape the original resolution of the decoder refined representation.

[0086] 3. Model training

[0087] In the process of model training, the root mean square error is used as the loss function to quantify the deviation between the retrieval error and the actual error. In addition, the mean absolute error, mean absolute percentage error and goodness of fit can also be used as evaluation indicators to verify the accuracy of the model. The specific formula is:

[0088]

[0089] Where: represents the prediction result, y represents the effective wave height in the data set, represents the average value of the effective wave height in the data set, N represents the number of observations, RMSE represents the root mean square error, MAE represents the mean absolute error, MAPE represents the absolute percentage error, R 2 Indicates goodness of fit.

[0090] For optimization, the Adam algorithm is used with an initial learning rate of 1×10 -4 After each training phase, the accuracy of the model was evaluated using the validation set, and the model parameters that produced the smallest root mean square error were retained.

[0091] Furthermore, the method described in this embodiment also includes: taking the reanalysis data set ERA5 as an observation benchmark, correcting the trained error diagnosis model to obtain a corrected error diagnosis model that meets preset requirements.

[0092] Among them, ERA5 was developed by the European Centre for Medium-Range Weather Forecasts. As a global reanalysis dataset, it combines advanced methods to aggregate different meteorological data. ERA5 data are acquired on a 0.25°×0.25° grid and are updated every three hours. The ERA5 dataset is very accurate in describing the characteristics of the actual significant wave height of ocean waves, and thus becomes the key data for wave analysis. This embodiment uses this reanalysis dataset to correct the prediction error in the SWH numerical prediction.

[0093] Specifically, the sea level pressure and significant wave height can be obtained from the ERA5 dataset, and the sea level pressure matrix and significant wave height matrix can be established.

[0094] 4. Error correction

[0095] The error diagnosis model is used to identify the estimated deviation of the wave significant wave height prediction, and the estimated deviation is subtracted from the original predicted wave significant wave height to obtain the corrected wave significant wave height. This method can improve the accuracy of wave significant wave height prediction and provide great support for maritime navigation, coastal management and a wider range of marine economic activities.

[0096] and Figure 1 Corresponding to the method described above, the embodiment of the present invention also provides an intelligent prediction and error correction system for significant wave height of ocean waves based on 2D-Geoformer, which is used for Figure 1 The specific implementation of the method in the embodiment of the present invention is that the intelligent prediction and error correction system of the significant wave height of the sea wave based on 2D-Geoformer can be applied to computer terminals or various mobile devices, such as Figure 3 As shown, specifically including:

[0097] The data set building module is used to collect raw data and perform preprocessing to build a training data set;

[0098] Model building module, used to build 2D-Geoformer deep learning model;

[0099] The model training module is used to train the model with the training data set as input and the prediction error as output until the loss function converges to obtain the error diagnosis model;

[0100] The correction module is used to identify the estimated deviation of the prediction of the wave significant wave height by using the error diagnosis model, and subtract the estimated deviation from the original predicted wave significant wave height to obtain the corrected wave significant wave height.

[0101] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0102] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for intelligent prediction and error correction of ocean wave significant wave height based on 2D-Geoformer, characterized in that: The following steps are involved: Collect raw data and preprocess them to build a training data set; Construct a 2D-Geoformer deep learning model, and train the model using the training data set as input and the prediction error as output until the loss function converges to obtain an error diagnosis model; The error diagnosis model is used to identify the estimated deviation of the wave significant wave height prediction, and the estimated deviation is subtracted from the original predicted wave significant wave height to obtain the corrected wave significant wave height.

2. The method for intelligent prediction and error correction of ocean wave significant wave height based on 2D-Geoformer according to claim 1 is characterized in that: The original data comes from the predicted value of the significant wave height of the ocean wave from the ocean prediction system, which is designed based on the surface wave-tide-circulation coupled ocean model.

3. The method for intelligent prediction and error correction of ocean wave significant wave height based on 2D-Geoformer according to claim 1 is characterized in that: Constructing a training dataset includes the following steps: Perform data labeling, data cleaning, normalization and feature processing on the original data to obtain the original data set; The coordinates of each original data are obtained, and based on the coordinates, the sea level pressure matrix and the significant wave height matrix corresponding to the coordinates of the original data are extracted.

4. The method for intelligent prediction and error correction of ocean wave significant wave height based on 2D-Geoformer according to claim 1, characterized in that: The 2D-Geoformer deep learning model is built based on the encoder-decoder framework, including: data preprocessing module, encoder, decoder, and output layer; The data preprocessing module is located at the base of the encoder and decoder stack, which is used to receive the input of the predicted value of the significant wave height and perform patch embedding; The encoder is used to convert the input data into feature maps through the spatiotemporal attention mechanism to establish complex spatiotemporal connections between multivariate data points; The decoder analyzes the output of the data preprocessing module and the encoder to generate the final output field; The output layer is used to output the decoder's output fields at the same spatial resolution as the input data.

5. The method for intelligent prediction and error correction of ocean wave significant wave height based on 2D-Geoformer according to claim 4 is characterized in that: The specific operations in the data preprocessing module are as follows: Receive the original predicted significant wave heights, and divide each set of input data into N non-overlapping patches of fixed size; The decomposed non-overlapping patches are converted into patch embeddings of a specific dimension through a learnable linear layer; The spatial position embedding and temporal embedding are integrated into the patch embedding to maintain the spatiotemporal context of each patch.

6. The method for intelligent prediction and error correction of ocean wave significant wave height based on 2D-Geoformer according to claim 4, characterized in that: The encoder consists of four identical coding blocks, each of which includes two different sub-layers; The first sublayer is a multi-head spatiotemporal attention layer, which is used to evaluate dependencies across spatiotemporal dimensions through a scaled dot product attention score function to generate a feature matrix; The second sub-layer is a fully connected feed-forward network, which is used to facilitate the flow and integration of information throughout the encoding process by adding residual connections around each sub-layer.

7. The method for intelligent prediction and error correction of ocean wave significant wave height based on 2D-Geoformer according to claim 6, characterized in that: The decoder consists of four identical decoding blocks, each of which includes three different sub-layers, and residual connections are added around each sub-layer; The first and third sublayers both include multi-head spatiotemporal attention layers and feedforward networks; The second sub-layer is used to perform spatiotemporal attention analysis, integrating the encoder’s feature matrix with the results of the previous sub-layer to produce the final output field.

8. The method for intelligent prediction and error correction of ocean wave significant wave height based on 2D-Geoformer according to claim 4, characterized in that: The output layer consists of a feed-forward network that maps and reshapes the original resolution of the decoder’s refined representation.

9. The method for intelligent prediction and error correction of ocean wave significant wave height based on 2D-Geoformer according to claim 1, characterized in that: The method further comprises: Taking the reanalysis dataset ERA5 as the observation benchmark, the trained error diagnosis model is corrected to obtain a corrected error diagnosis model that meets the preset requirements.

10. A system for realizing the intelligent prediction and error correction method of ocean wave significant wave height based on 2D-Geoformer as claimed in any one of claims 1 to 9, characterized in that: include: The data set building module is used to collect raw data and perform preprocessing to build a training data set; Model building module, used to build 2D-Geoformer deep learning model; The model training module is used to train the model with the training data set as input and the prediction error as output until the loss function converges to obtain the error diagnosis model; The correction module is used to identify the estimated deviation of the prediction of the wave significant wave height by using the error diagnosis model, and subtract the estimated deviation from the original predicted wave significant wave height to obtain the corrected wave significant wave height.

Images