A storm surge forecasting method combining empirical orthogonal decomposition and deep learning

Abstract

The present invention discloses a storm surge forecasting method that combines empirical orthogonal decomposition and deep learning, belonging to the field of ocean storm surge forecasting. The method includes the following steps: a. determining the wind field area based on the location of the forecast ocean station, etc.; b. collecting and organizing the 10-meter wind anomaly data matrix from the reanalysis data set, and collecting observed storm surge data from the ocean station; c. performing empirical orthogonal decomposition on the data matrix to obtain the principal mode matrix and the principal component matrix; d. using the principal components in the principal component matrix as forecast factors and the storm surge data as forecast quantities to construct a sample library; and performing model training to obtain a long short-term memory neural network forecasting model; e. collecting the 10-meter wind anomaly data matrix for future forecasts, calculating the principal component values, and then substituting them into the forecast model to obtain future storm surge data. The present invention can improve the accuracy of storm surge forecasts, and the calculated root mean square error is low compared to the actual observed storm surge values.

Description

Technical Field

[0001] The present invention relates to the field of ocean storm surge forecasting, and in particular to a storm surge forecasting method combining empirical orthogonal decomposition and deep learning. Background Art

[0002] A storm surge disaster refers to an abnormal rise in sea levels caused by strong atmospheric disturbances, such as tropical cyclones and extratropical cyclones, which causes the tide level in the affected sea area to far exceed the normal tide level. The severity of a storm surge disaster is closely related to meteorological conditions. Currently, there are three types of storm surge forecasts: the first type is the empirical formula method. This method uses experience to quantify different prediction effects into one or more parameters based on statistical analysis of observational data. This method can combine the maximum storm surge water level and typhoon characteristics such as air pressure and wind speed through empirical formulas. The second type is the fluid dynamics model method, which uses fluid dynamics and atmospheric driving to establish direct or coupled numerical models to predict storm surges. The fluid dynamics model method is currently the most common method. However, its disadvantage is that it requires precise and detailed fluid dynamics equations, terrain data, typhoon models, boundary conditions, and a large amount of computation. The third type is the artificial neural network method. Artificial neural network methods do not require the establishment of complex data models; they only require input data to adjust the network's internal connections. Relying on their powerful self-learning, self-organizing, and self-adaptive capabilities, they can automatically extract and combine mapping rules by learning from system input and output sample pairs, and automatically adjust network structure parameters to adapt to environmental changes. This provides them with powerful approximation and fitting capabilities. Based on this, artificial neural networks can effectively simulate the nonlinear relationship between storm surges and meteorological conditions and have been widely used.

[0003] However, the existing storm surge neural network forecasting models have a relatively simple introduction of meteorological data. One method uses single-point meteorological data observed near the forecast site, and the other uses key parameters of typhoon observations (such as the typhoon's real-time path, maximum wind speed, central air pressure, and the radius of a level 7 wind circle). This method can only predict simple typhoon storm surges.

[0004] The meteorological conditions encountered in actual storm surge forecasts are often complex, influenced by a combination of multiple weather systems. Examples include storm surge disasters caused by the combination of cold air and extratropical cyclones, and storm surge disasters caused by the combination of northward typhoons and surface high pressure. For storm surge disasters forced by complex weather systems, relying solely on simple meteorological data fails to fully account for the impact of these conditions, and the accuracy of neural network simulation forecasts cannot be guaranteed. Summary of the Invention

[0005] To address these technical issues, the present invention proposes a storm surge forecasting method that combines empirical orthogonal decomposition and deep learning. This method, centered around forecasting stations, selects a wide range of meteorological data covering the impacting weather systems, enhancing the model's forecasting capabilities. Furthermore, the method uses empirical orthogonal decomposition to process the meteorological data, preserving the vast majority of the data while avoiding overfitting the neural network model due to excessive forecasting variables.

[0006] The technical solution adopted by the present invention is:

[0007] A storm surge forecasting method combining empirical orthogonal decomposition and deep learning includes the following steps:

[0008] a. Determine the wind farm area based on the location of the forecast ocean station and the characteristics of historical storm surge processes;

[0009] b. Collect and organize 10-meter wind anomaly data matrices from the reanalysis dataset and observe storm surge data from ocean stations based on wind farm regions;

[0010] c. Perform empirical orthogonal decomposition on the 10-meter wind anomaly data matrix to obtain the principal mode matrix and principal component matrix;

[0011] d. Using the principal components in the principal component matrix as prediction factors and the storm surge data as prediction quantities, a sample library is constructed; a long short-term memory neural network is used for model training to obtain a long short-term memory neural network prediction model;

[0012] e. Collect the 10-meter wind anomaly data matrix for future forecasts;

[0013] f. Calculate the principal component values ​​of the 10-meter wind anomaly data matrix for future forecasts;

[0014] g. Substitute the principal component values ​​of the calculated future forecast 10-meter wind anomaly data matrix into the long short-term memory neural network forecast model trained in step d to obtain the predicted future storm surge data.

[0015] The beneficial technical effects of the present invention are as follows:

[0016] Storm surge forecasting relies not only on local meteorological data but also on large-scale data covering the entire weather system. This method, centered on the forecast site, selects meteorological data covering all aspects of the influencing weather system. This allows for a more comprehensive incorporation of meteorological forcing signals into the neural network model, thereby improving the model's forecasting capabilities.

[0017] If a large number of meteorological forcing signals are directly introduced, the model will have too many prediction factors and too many neurons in the neural network model. This will not only increase the training time and convergence difficulty of the long short-term memory neural network model, but also easily cause overfitting of the factor model. As a result, although the model works well in the training phase, the effect drops sharply in the test phase. This method uses the method of empirical orthogonal decomposition to process meteorological data and extract the main modes of the meteorological field. The main modes can explain most of the variance of the meteorological field with only a small number of variables. Replacing the original meteorological field data with the main modes and introducing the neural network model not only reduces the difficulty of model training, but also avoids the problem of overfitting of the model. When performing model forecasting, the future wind field can be directly projected onto the main modes in the model training phase and then substituted into the neural network model. This avoids performing the empirical orthogonal decomposition operation again, which shortens the calculation time and has better forecast accuracy.

[0018] In summary, the present invention can significantly improve the accuracy of storm surge forecasts, and the calculated root mean square error is lower than the actual observation value of storm surge. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Schematic diagram of the flow of the storm surge forecasting method combining empirical orthogonal decomposition and deep learning of the present invention;

[0020] Figure 2 Schematic diagram of the wind farm space range and the location of the Weifang Port Ocean Station in a specific application example of the present invention;

[0021] Figure 3 1 is a spatial distribution diagram of the 1st to 10th main modes in a specific application example of the present invention;

[0022] Figure 4 11 is a spatial distribution diagram of the 11th to 20th main modes in a specific application example of the present invention;

[0023] Figure 5 The histogram of explained variance of the first 20 eigenvalues ​​in the specific application example of the present invention is shown;

[0024] Figure 6 This is a diagram showing the root mean square error results of model training in a specific application example of the present invention;

[0025] Figure 7 This is a graph showing the loss function results of the model training in a specific application example of the present invention;

[0026] Figure 8 The present invention's forecasting method, existing numerical model forecasts, and observed storm surge curves are shown. DETAILED DESCRIPTION

[0027] When developing storm surge neural network prediction models, it is important to incorporate meteorological grid data from a sufficiently large area, centered around the forecast site, to account for the forcing effects of different weather systems on storm surge hazards. However, this presents a new challenge: large-scale meteorological grid data introduces a large number of data variables. Directly incorporating these variables into the neural network model increases the difficulty of training the model and can easily lead to model overfitting. Based on this, this method proposes using empirical orthogonal decomposition to process meteorological data, extracting the primary modes of the meteorological grid data. These modes are then used to replace the full-scale meteorological data, thereby avoiding the overfitting problem of the neural network.

[0028] like Figure 1 As shown in the figure, a storm surge forecasting method combining empirical orthogonal decomposition and deep learning is proposed. The method is mainly divided into two stages: model training stage and model forecasting stage.

[0029] The model training phase includes the following steps:

[0030] a. Determine the wind field area based on the location of the forecast ocean station and the characteristics of historical storm surge processes. The area should be large enough to cover the main weather systems that affect the storm surge at the forecast ocean station.

[0031] b. Collect and organize the 10-meter wind anomaly data matrix from the reanalysis dataset according to the wind field area , collecting observed storm surge data from ocean stations .in: is the grid number of wind field data; is the length of the time series during the model training phase; , Indicates the The grid point Wind speed value at a moment; , Indicates the The storm surge value at that moment.

[0032] c. Yes Perform empirical orthogonal decomposition to obtain the main mode matrix and the principal component matrix .in, , Represents the value of the i-th parameter in the j-th mode; , represents the value of the jth mode at the tth moment.

[0033] Step c specifically includes:

[0034] c1. Calculation The covariance matrix of : . express The transpose of .

[0035] c2, covariance matrix Perform eigendecomposition to obtain the eigenvalue matrix and the corresponding eigenvector matrix ,satisfy .in, is a diagonal matrix, , is the covariance matrix The eigenvalues ​​are arranged from large to small. The size of each eigenvalue reflects the contribution rate corresponding to the mode, that is, the explained variance; Each column in , is the same as The corresponding eigenvector.

[0036] c3, the eigenvector matrix is ​​the main mode matrix, that is .

[0037] c4, principal component matrix for The projection in the main modal matrix is . express The transpose of .

[0038] d. The principal component matrix The first 20 principal components As a predictor, As a forecast quantity, build a sample library The long short-term memory neural network method is used for model training to obtain the neural network prediction model. , Storm surge calculated by the neural network model, It is a long short-term memory neural network model.

[0039] The subsequent model prediction includes the following steps:

[0040] e. Collect the 10-meter wind anomaly data matrix for future forecasts , is the time series length of the model prediction stage.

[0041] f. Calculate the principal components of the wind anomaly data for future forecasts .

[0042] Specifically include:

[0043] f1. Select the main modal matrix The first 20 modes in .

[0044] f2, calculation Projected values ​​in the first 20 main modes , which is the principal component value of the wind anomaly data for future forecasts.

[0045] g. Substitute the long short-term memory neural network model trained in step d ,Right now , and obtain the future storm surge predicted by the neural network model .

[0046] h. Comparison of observed future storm surges and future storm surges predicted by neural network models , calculate the root mean square error. It can also be further compared with the root mean square error calculated for other forecast products to evaluate the forecasting effect of the neural network model of the present invention.

[0047] The present invention will be further described below with reference to specific application examples.

[0048] Take the establishment of the storm surge forecast model for Weifang Port Ocean Station as an example.

[0049] Model training phase:

[0050] a. Based on the location of Weifang Port Ocean Station, the spatial range of the wind farm is selected as 116°~126°E, 33°~42°N ( Figure 2 ).

[0051] b. Download the hourly ERA5 reanalysis data from 2018 to 2022. There are 1149 data sets in total. The first 60% of the data (851 sets) are selected for model training, and the last 40% of the data (568 sets) are selected for model forecasting. There are 1517 grid points in 116°~126°E and 33°~42°N. Considering that wind speed is a vector and includes two variables, the number of grid points is 3034. Select the 10-meter wind field data and perform anomaly processing on the data to obtain the matrix ,in , .

[0052] Collect hourly tide level data from Weifang Port Tide Station from 2018 to 2022, and analyze the tide level data to obtain storm surge data. .

[0053] c. Perform empirical orthogonal decomposition on the first 60% of wind field anomaly data to obtain the eigenvalues , main mode matrix and the principal component matrix . Draw the spatial distribution of the first 20 eigenvalues ​​( Figure 3 、 Figure 4) and histograms ( Figure 5 ), and the cumulative explained variance was calculated to be 90.3%.

[0054] d. Select the first 20 principal components in the principal component matrix as predictors , select the first 60% of the storm surge data of Weifang Port To predict the amount, a sample library is constructed. The model is trained using the long short-term memory neural network method. The model training results are as follows: Figure 6 and Figure 7 As shown, the trained neural network and parameters are stored.

[0055] Model prediction stage:

[0056] e. Select 40% of the wind field anomaly data matrix after sorting (in, ).

[0057] f. The matrix formed by the first 20 main modes , do matrix multiplication , and obtain the projection of 40% wind field anomaly data in the first 20 modes .

[0058] g. Substitute into the neural network forecast model ,Right now , and obtained the storm surge forecast for Weifang Port .

[0059] h. Comparison of storm surge observed at Weifang Port and storm surge predicted by neural network model The root mean square error of the calculation is 13.2 cm. The root mean square error of the storm surge numerical model forecast for Weifang Port during the same period was 16.9 cm, which is 21.9% lower than that of the neural network model established by this method. Figure 8 This is a storm surge curve for an example in the forecast stage.

[0060] Parts not described in the above methods can be achieved by adopting or drawing on existing technologies.

[0061] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A storm surge forecasting method combining empirical orthogonal decomposition and deep learning, characterized by The following steps are involved: a. Determine the wind farm area based on the location of the forecast ocean station and the characteristics of historical storm surge processes; b. Collect and organize 10-meter wind anomaly data matrices from the reanalysis dataset and observe storm surge data from ocean stations based on wind farm regions; c. Perform empirical orthogonal decomposition on the 10-meter wind anomaly data matrix to obtain the principal mode matrix and principal component matrix; d. Using the principal components in the principal component matrix as prediction factors and the storm surge data as prediction quantities, a sample library is constructed; a long short-term memory neural network is used for model training to obtain a long short-term memory neural network prediction model; e. Collect the 10-meter wind anomaly data matrix for future forecasts; f. Calculate the principal component values ​​of the 10-meter wind anomaly data matrix for future forecasts; g. Substituting the principal component values ​​of the calculated future forecast 10-meter wind anomaly data matrix into the long short-term memory neural network forecast model trained in step d to obtain the predicted future storm surge data; In step b: 10-meter wind anomaly data matrix X M×T , where M is the number of grids for wind field data, and T is the length of the time series during the model training phase; represents the wind speed value of the i-th grid point at the t-th moment; Collect observed storm surge data Z from ocean stations T , z T =[z1,…,z t ,…,z T ],z t ,t=1,…,T represents the storm surge value at the tth moment; In step c: for matrix X M×T Perform empirical orthogonal decomposition to obtain the main mode matrix EOF M×M and the principal component matrix PC M×T ;in, Represents the value of the i-th parameter in the j-th mode; represents the value of the jth mode at the tth moment; Step c specifically includes: c1. Calculate the 10-meter wind anomaly data matrix X M×T The covariance matrix of : X M×T R Represents X M×T The transpose of c2, covariance matrix S M×M Perform eigendecomposition to obtain the eigenvalue matrix Λ M×M and the corresponding eigenvector matrix V M×M , satisfying S M×M ×V M×M =V M×M ×Λ M×M ;in, is a diagonal matrix, λ i ,i=1,…,M, is the covariance matrix S M×M The eigenvalues ​​are arranged from large to small, and the size of each eigenvalue reflects the contribution rate corresponding to the mode; Each column in V(i,:), i=1,…,M, is related to λ i The corresponding eigenvector; c3, the eigenvector matrix is ​​the main mode matrix, that is, EOF M×M =V M×M ; c4, principal component matrix PC M×T For X M×T Projection in the main modal matrix, PC M×T =EOF M×T R ×X M×T ;EOF M×T R Indicates EOF M×T The transpose of In step d: the principal component matrix PC M×T The first 20 principal components As a forecast factor, the storm surge data Z T As a forecast quantity, construct a sample library (PC 20×T ,Z T ); The obtained long short-term memory neural network prediction model Z′ T =F(PC 20×T ), Z′ T is the storm surge data calculated by the long short-term memory neural network forecast model, and F is the long short-term memory neural network forecast model; In step e: the future forecast 10-meter wind anomaly data matrix is ​​X M×Q , Q is the length of the time series in the model forecast stage; Step f specifically includes: f1, select the main mode matrix EOF M×M The first 20 modal EOFs in 20×M ; f2, calculate the matrix X M×Q Projected values ​​in the first 20 main modes That is, the principal component value of the 10-meter wind anomaly data matrix for future forecasts.

2. A storm surge forecasting method combining empirical orthogonal decomposition and deep learning according to claim 1, characterized in that: In step a: The scope of the determined wind field area should be able to cover the main weather systems that affect the storm surge forecast at the ocean station.

3. The storm surge forecasting method combining empirical orthogonal decomposition and deep learning according to claim 1, characterized in that: In step g: The principal component values ​​PC of the 10-meter wind anomaly data matrix of the future forecast 20×Q Substitute into the long short-term memory neural network prediction model F, that is, Z′ Q =F(PC 20×Q ), and obtain the future storm surge data Z′ predicted by the long short-term memory neural network forecast model Q .

Images