Typhoon wave height prediction method based on deep learning moe-transformer model

Abstract

The application discloses a typhoon wave height prediction method based on a deep learning MOE-Transformer model, and specific steps are as follows: the method steps are as follows: collecting historical typhoon data of a research area, combining a central pressure difference formula to construct a virtual typhoon data set; through a Holland typhoon empirical model and a wind wave growth relationship, processing the historical and virtual typhoon data set to obtain a wind field, a pressure field and an effective wave height field; taking an ERA5 data set as a benchmark, correcting the above field data by using an error model to form a typhoon space-time fusion database containing meteorological data, the effective wave height field and typhoon data; training and testing the MOE-Transformer model to obtain a typhoon wave effective wave height prediction model for typhoon wave height prediction. The method improves the multi-task adaptability and generalization ability, and improves the precision and timeliness of typhoon wave height prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of typhoon wave prediction, in particular to a typhoon wave height prediction method based on a deep learning MOE-Transformer model, and more particularly to predicting the effective wave height of a typhoon in the generation, development, maturity and decay time phases of a typhoon under different intensity typhoon typical paths. BACKGROUND

[0002] Typhoon waves are a hot issue in ocean dynamics research in recent years. During a typhoon, extreme waves occur, with wave heights reaching tens of meters, which has an important impact on the construction and safe operation of offshore and marine engineering. Accurate prediction of typhoon waves has important scientific significance and engineering application value for offshore and marine engineering.

[0003] Currently, typhoon wave prediction methods are mainly divided into empirical prediction methods, numerical prediction methods and intelligent prediction methods based on machine learning. Empirical methods are only applicable to a certain region and lack widespread applicability. Numerical simulation methods require a large amount of computing resources. Intelligent prediction based on machine learning takes into account the model's generality and fast computing speed, providing a new direction for typhoon wave prediction. Traditional RNN prediction has limited memory time;

[0004] The MOE-Transformer model used in the present application can view the information of the entire time series simultaneously due to the expert mixing mechanism and self-attention mechanism, and can give the relationship between wind speed and effective wave height under different typhoon intensities. SUMMARY

[0005] To address the problem of insufficient data samples in typhoon wave deep learning model prediction, the present application uses virtual typhoons to compensate for the lack of samples, calculates the effective wave height distribution in the study area using the wind wave growth relationship, and predicts the effective wave height based on the deep learning model MOE-Transformer to improve the accuracy of typhoon wave prediction.

[0006] To achieve the above purpose, the present application realizes the following technical solutions:

[0007] A typhoon wave height prediction method based on a deep learning MOE-Transformer model, comprising the following steps:

[0008] S1. Collect historical typhoon data in the study area as a historical typhoon dataset, and construct a virtual typhoon dataset for the study area using Markov chains, kernel density functions and central pressure difference formulas;

[0009] S2. Process the historical typhoon and virtual typhoon datasets using the Holland typhoon empirical model and the wind wave growth relationship to obtain the wind field, pressure field and effective wave height field of both;

[0010] S3, select the ERA5 dataset as the reference dataset, correct the wind field, pressure field and significant wave height field in the historical typhoon and virtual typhoon dataset using the error model, and the corrected dataset constitutes a typhoon spatio-temporal fusion database;

[0011] The database includes meteorological data of the study area: wind field, pressure field; significant wave height value of the study area: significant wave height field; typhoon data: longitude and latitude of the typhoon center, near-center maximum wind speed and central pressure;

[0012] S4, the typhoon spatio-temporal fusion database is divided into a training set, a validation set and a test set;

[0013] S5, select the typhoon data, meteorological data of the study area and significant wave height value of the study area in the t-12 to t period as the input sequence, use the MOE-Transformer model to construct the mapping relationship between the typhoon data, meteorological data of the study area and significant wave height value of the study area, and predict the significant wave height value of the study area at t+1, t+3, t+6 and t+12 respectively., as the target output result of the model;

[0014] S6, train and validate the MOE-Transformer model based on the data in the training set and the validation set, obtain the typhoon wave significant wave height prediction model; evaluate and test the typhoon wave significant wave height prediction model based on the data in the test set, and use the typhoon wave significant wave height prediction model to predict the typhoon wave height.

[0015] Further, in S1, the virtual typhoon dataset of the study area is constructed by Markov chain, kernel density function and central pressure difference formula, and the specific method comprises:

[0016] Based on historical typhoon data, the study area is divided into grids, the frequency of historical typhoon generation in the study area and the generation and termination probability at the starting point of each historical typhoon are counted by kernel density function, and the starting point distribution of virtual typhoon is obtained by combining the probability and starting point distribution; extract the near-center maximum wind speed, typhoon moving speed and typhoon moving direction during the historical typhoon as the initial information in the Markov chain; sample the three sampling parameters at a fixed time interval to obtain the change value of the three sampling parameters until the virtual typhoon terminates;

[0017] According to the historical typhoon data, the mapping relationship between the difference △P of the environmental pressure and the central pressure of the historical typhoon, the near-center maximum wind speed, the typhoon center latitude and the typhoon moving speed is established, and the central pressure of the virtual typhoon is valued, and finally the virtual typhoon dataset is obtained.

[0018] Further, the termination condition of the virtual typhoon includes that the maximum wind speed near the center of the virtual typhoon is less than 9 m / s, or the sea surface temperature at the location of the virtual typhoon is less than 10℃.

[0019] Further, in S2, the wind field, the pressure field and the significant wave height field of the historical typhoon and the virtual typhoon are obtained by processing the data sets of the historical typhoon and the virtual typhoon by the Holland typhoon empirical model and the wind wave growth relationship, and the specific method includes:

[0020] The pressure field model constructed by the Holland typhoon empirical model is:

[0021] ;

[0022] In the formula, P c represents the sea surface pressure at the center of the cyclone; P n represents the peripheral pressure; R max represents the maximum wind speed radius; r is the distance from the calculation point to the center of the typhoon; and B represents the Holland pressure profile parameter.

[0023] The wind field model constructed by the Holland typhoon empirical model is:

[0024] ;

[0025] In the formula, V s is the gradient wind speed at a distance r from the center of the tropical cyclone, ρ a represents the air density; and f represents the Coriolis force.

[0026] The calculation of the significant wave height field is performed by the wind wave growth relationship formula:

[0027] ;

[0028] In the formula, U 10 is the wind speed at 10 m height, F is the equivalent wind area, H s is the significant wave height, and g is the acceleration of gravity.

[0029] Further, in S3, the ERA5 data set is selected as the reference data set, and the wind field, the pressure field and the significant wave height field in the historical typhoon and the virtual typhoon data set are corrected by using the error model, and the corrected data set constitutes a typhoon spatiotemporal fusion database, and the specific method includes:

[0030] The ERA5 reanalysis data is used as the reference data set, and the wind field, the pressure field and the significant wave height field data of the data set during the typhoon passage are extracted by spatiotemporal matching;

[0031] Then a three-dimensional error field historical typhoon data set containing wind field deviation, pressure field deviation and wave height field deviation is constructed by combining the historical typhoon wind field, pressure field and significant wave height field with the reference data set, the data is normalized by Z-score standardization, and then input into the error model for training, wherein the error model is a U-Net error correction model;

[0032] The wind field, pressure field and significant wave height field in the historical typhoon and virtual typhoon data set are input into the trained U-Net error correction model, and finally the corrected historical typhoon and virtual typhoon data set are output.

[0033] Further, during the inputting into the error model for training process, when the root mean square error (RMSE) of the error model training on the validation set is continuously reduced by less than 1e-3 for 5 epochs, and the effective wave height RMSE < 0.08, the near-center maximum wind speed RMSE < 0.05, and the central pressure RMSE < 0.01, the model is considered to be trained.

[0034] Further, in S4, the typhoon spatio-temporal fusion database is divided into a training set, a validation set and a test set according to a ratio of 7:2:1.

[0035] Further, in S6, the MOE-Transformer model is trained and verified based on the training set and the validation set to obtain a typhoon wave effective wave height prediction model based on the MOE-Transformer model, and the specific method comprises:

[0036] S61, the MOE-Transformer model includes an input layer and position encoding, an encoder, a decoder and an output layer, and the hyperparameters are randomly initialized within a preset range;

[0037] S62, the weight parameters of the MOE-Transformer model are initialized, including: input layer position encoding parameters, transformation matrices of query vectors Query, key vectors Key and value vectors Value of each layer of the encoder, and weight parameters of the gate expert network;

[0038] S63, the input sequence is forward propagated through the MOE-Transformer model:

[0039] Adding position coding to the input layer encodes position information into the input vector; encoder processing: perform multi-head self-attention calculation on the query vector, key vector and value vector, divide into n heads, respectively calculate and splice after dimension reduction; perform Masked multi-head self-attention calculation, select expert combination through the gating expert network, standardize and grade typhoons according to the near center maximum wind speed, and based on the MOE mechanism, match typhoons of different grades to the corresponding typhoon wave expert model in the expert combination for simulation and error correction; through the predefined wind speed-expert model mapping rule, the decoupling and optimal scheduling between expert models are realized; then, residual connection and layer normalization output layer processing are performed; the prediction result of the output of the MOE-Transformer model is compared with the target result, the root mean square error RMSE is selected as the loss function, and the loss value of the model is calculated;

[0040] S64, compare the prediction result of the output of the MOE-Transformer model with the target result, select the root mean square error RMSE as the loss function, and calculate the loss value of the model;

[0041] S65, according to the loss value, calculate the gradient through the back propagation algorithm, and update the weight parameters of the MOE-Transformer model to reduce the loss value;

[0042] S66, train the MOE-Transformer model based on the data in the training set and the validation set, and obtain a typhoon wave height prediction model based on the MOE-Transformer;

[0043] S67, after the training is completed, calculate the root mean square error RMSE between the MOE-Transformer prediction value and the true value based on the test set, and evaluate and test the typhoon wave height prediction model.

[0044] Further, the standardization grading includes: tropical depression, tropical storm, strong tropical storm, typhoon, strong typhoon and super strong typhoon.

[0045] A typhoon wave height prediction system based on a deep learning MOE-Transformer model, the system comprises:

[0046] One or more processors;

[0047] Memory for storing one or more programs;

[0048] When the one or more programs are executed by the one or more processors, the one or more processors implement the typhoon wave height prediction method based on the deep learning MOE-Transformer model

[0049] Compared with the prior art, the present application has the advantages of:

[0050] The present application can generate a large number of virtual typhoons in a research area, and solves the problem of insufficient samples in typhoon wave prediction.

[0051] Compared with other time series prediction models, the MOE-Transformer model can effectively capture the dependency relationship in a long time sequence through an expert mixing mechanism and a self-attention mechanism. Different subspace features are extracted through a multi-head attention mechanism. This mechanism can simultaneously focus on different time scales or different patterns, and the introduction of the MOE mechanism further enhances the modeling ability of the model for typhoons of different intensities. Therefore, in complex time series data, information fusion and feature extraction can be performed from multiple perspectives. The MOE-Transformer overcomes the problem of gradient vanishing or gradient explosion that is prone to occur in long sequences. During training, it can be parallelized, and all time steps can be processed simultaneously, improving computational efficiency.

[0052] The method of the present application is suitable for typhoon wave prediction in different geographical regions. Typhoon waves have different growth patterns in different regions (differences in water depth, terrain, etc.), and traditional single models are difficult to adapt to all regions. This model improves the generalization ability and makes it suitable for typhoon wave prediction in multiple regions. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 is a prediction method flowchart of an embodiment of the present application;

[0054] Figure 2 is a water depth and terrain map of a research area according to an embodiment of the present application;

[0055] Figure 3 is a MOE-Transformer deep learning model structure diagram according to an embodiment of the present application;

[0056] Figure 4 is a comparison diagram of different prediction time prediction values and test set significant wave heights according to an embodiment of the present application. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the embodiments of the present application will be described clearly and completely below in combination with the drawings of the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the described embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art belong to the scope of protection of the present application.

[0058] Embodiment:

[0059] As Figure 1As shown, it is an embodiment of the application, which provides a typhoon wave height prediction method based on a deep learning MOE-Transformer model, comprising the following steps:

[0060] Firstly, based on the historical typhoon path data set of China Typhoon Network from 1949 to 2023, the kernel density estimation method is used to generate a virtual typhoon data set with a time span of 1900-2400. Subsequently, the wind field, pressure field and significant wave height field during the historical typhoon and virtual typhoon are calculated by using Holland empirical formula and wind wave growth relationship respectively;

[0061] At the same time, the ERA5 reanalysis data set is used to obtain the reference wind field, pressure field and significant wave height field data.

[0062] In order to improve the calculation accuracy, an error model, i.e. a U-Net deep learning network, is used to correct the error of the wind wave growth relationship calculation result. First, the U-Net deep learning network is trained, verified and tested by using the historical typhoon data. The learned error distribution rule is transferred to the virtual typhoon data set, and the central pressure and maximum near-center wind speed of the typhoon are corrected. The corrected wind field, pressure field and significant wave height field data and the typhoon characteristic parameters (longitude and latitude of the typhoon center, maximum near-center wind speed and central pressure) together constitute a sample database.

[0063] Finally, a MOE-Transformer hybrid expert model is used to establish the nonlinear mapping relationship between the typhoon characteristics and the environmental field. Through the multi-head self-attention mechanism, the spatio-temporal characteristics are captured, and the high-precision prediction of the effective wave height in the research area within 1-12h is realized.

[0064] The whole process realizes the technical process from historical data reconstruction, physical model calculation to deep learning correction and prediction. Specifically:

[0065] The typhoon information from 1949 to 2023 is obtained from China Typhoon Network (https: / / www.typhoon.org.cn / ), and the grid is divided in the western Pacific Ocean. The specific research area is 15°N-42°N, 105°E-150°E, which is divided into 5°×5°. First, the starting point parameter characteristic probability and annual occurrence frequency are counted, and then the moving speed, moving direction and near-center maximum wind speed probability of the typhoon in the sub-region are counted according to the kernel density function until the typhoon terminates. Finally, a virtual typhoon data set affecting the western Pacific Ocean is generated. The data set includes the longitude and latitude of the typhoon center, the near-center maximum wind speed and the central pressure.

[0066] The virtual typhoon path generation method using kernel density estimation method is as follows:

[0067] Based on the historical typhoon database, the spatial frequency distribution characteristics of typhoon generation in the target sea area are statistically analyzed by using the kernel density function to establish the probability density field of typhoon generation location. Combined with the statistical characteristics of the life cycle of historical typhoons, a joint probability model of typhoon generation-termination is constructed. The initial state parameters of the virtual typhoon are generated from the above probability model. The Markov chain method is used to select the typhoon intensity characteristic quantity (near-center maximum wind speed), movement characteristic quantity (moving speed and moving direction) as the key state variable. A state transition model based on fixed time step is established:

[0068] (1)

[0069] P represents the transition probability matrix, represents the state of the random process at time t, K is the probability distribution obtained by kernel density estimation (KDE), represents the near-center maximum wind speed, represents the typhoon moving speed, and θ represents the typhoon moving direction. The Markov chain has no memory and is suitable for modeling random processes such as typhoon paths.

[0070] Based on the least squares method to fit multivariate nonlinear problems, the near-center maximum wind speed, latitude and typhoon moving speed of historical typhoon data are extracted to fit the central pressure of virtual typhoon:

[0071] (2)

[0072] where, is the difference between the environmental pressure and the central pressure of the historical typhoon, is the near-center maximum wind speed, lat is the latitude of the typhoon center, is the typhoon moving speed, and a-e are constants.

[0073] Based on this, all the information of the virtual typhoon has been obtained. The above typhoon data is input into the Holland typhoon empirical model to construct the wind field and pressure field in the study area using the Holland empirical typhoon model.

[0074] The formula of the pressure model is:

[0075] (2)

[0076] where P is the pressure value at the calculation point, P c is the central pressure of the typhoon, P n is the peripheral pressure, R max is the maximum wind speed radius, r is the distance from the calculation point to the typhoon center. B is the Holland pressure profile parameter.

[0077] The formula of the wind field model is:

[0078] (3)

[0079] In the formula, V s is the gradient wind speed at a distance r from the center of the tropical cyclone, f is the Coriolis force parameter, and p a is the air density.

[0080] The moving wind field formula is:

[0081] (4)

[0082] In the formula, is the moving speed of the typhoon;

[0083] The Holland formula with right bias is:

[0084] (5)

[0085] In the formula, C1 and C2 are correction coefficients; λ is the angle between the line connecting the calculation point and the center of the typhoon and the positive east direction; and β is the angle between the gradient wind and the sea surface wind.

[0086] For tropical cyclones in the Northern Hemisphere, the waves generated in the strong wind area on the right side of the typhoon center will propagate in a direction approximately parallel to the propagation direction of the typhoon. On the contrary, on the left side of the typhoon, the waves will propagate in a direction approximately opposite to the propagation direction of the typhoon. Since the wave group speed increases with the increase of the near-center maximum wind speed, the relative value of the moving speed of the typhoon and the near-center maximum wind speed plays a crucial role in determining the maximum significant wave height in the affected area of the typhoon.

[0087] Based on the above discussion, the significant wave height in the study area under typhoon conditions is calculated using the wind wave growth relationship:

[0088] (6)

[0089] In the formula, U 10 is the wind speed at 10 m height, F is the equivalent wind area, H s is the significant wave height, g is the acceleration of gravity, and F is the length of the wind area.

[0090] Figure 2 The study area of the present application is the study area, and six typhoon grades affecting the study area are selected: super typhoon, strong typhoon, typhoon, strong tropical storm, tropical storm and tropical depression file. A U-Net error network correction model is defined, the standard deviations of the wind field, pressure field and significant wave height field of the historical typhoon data set and the ERA5 data set are calculated, and the historical typhoon and virtual typhoon data sets are corrected through the standard deviations, and the typhoon spatiotemporal fusion database is obtained after combination.

[0091] The specific correction process is:

[0092] The three-dimensional error field historical typhoon data set containing wind field deviation, pressure field deviation and wave height field deviation is constructed by combining the wind field, pressure field and significant wave height field of historical typhoon with the reference data set, and then the data is normalized by Z-score standardization and input into the U-Net error correction model for training; when the root mean square error (RMSE) on the validation set is less than 1e-3 and the effective wave height RMSE < 0.08, the near-center maximum wind speed RMSE < 0.05, and the central pressure RMSE < 0.01 in the last 5 epochs, to ensure the convergence of the model;

[0093] The wind field, pressure field and significant wave height field in the historical typhoon and virtual typhoon data set are input into the trained U-Net error correction model, and the corrected historical typhoon and virtual typhoon data set are finally output.

[0094] The U-Net error correction model has an encoder-decoder structure:

[0095] The encoder part adopts a downsampling structure, including a 3×3 convolution layer, a batch normalization layer and an activation function. The downsampling adopts a max pooling operation, and the number of feature channels is expanded to 64-512, effectively extracting multi-scale error features.

[0096] The decoder part realizes feature map upsampling through transpose convolution, and each level contains a 3×3 deconvolution layer and a feature fusion module.

[0097] The spatio-temporal fusion database is specifically:

[0098] The significant wave height field of the historical typhoon and virtual typhoon data set corrected by the U-Net is spatio-temporally aligned with the wind field and pressure field, and the longitude and latitude of the center of the historical typhoon and virtual typhoon, the near-center maximum wind speed and the central pressure are integrated to construct a spatio-temporal fusion database with a spatial grid accuracy of 0.125°×0.125°;

[0099] Figure 3 MOE-Transformer deep learning model structure diagram, MOE means Model of Experts;

[0100] Data set division and standardization processing: the typhoon spatio-temporal fusion database is divided into training set, validation set and test set in the ratio of 7:2:1 by using stratified sampling method. The Z-score standardization processing is performed on various types of feature parameters:

[0101] X_norm = (X - μ) / σ

[0102] In the formula, μ is the mean of the training set, and σ is the standard deviation.

[0103] Input-output sequence construction: the typhoon characteristic field (including typhoon parameters, wind field, pressure field and significant wave height field) of the period from t-12h to t is selected as the input sequence, and the significant wave height field at t+1h, t+3h, t+6h and t+12h is taken as the prediction target, respectively, to construct the space-time sequence sample.

[0104] MOE-Transformer model initialization:

[0105] (1) Input embedding layer: the typhoon center position is converted into spherical coordinate encoding, and various field data are mapped to a d_model-dimensional feature space through linear projection; (2) space-time position encoding: two-dimensional sinusoidal encoding is adopted; model training:

[0106] (1) Forward propagation: the input sequence is extracted through n self-attention mechanisms, and the 6-level typhoon is matched to the corresponding typhoon wave expert model for feature selection;

[0107] (2) Loss calculation: the loss function is used for evaluation;

[0108] (3) Parameter update: the Adam optimizer is used for training, and the initial learning rate is 5*10 -4 ; the optimal model configuration of the model width d model , the number of layers N, the number of attention heads h and the training round number Epochs is selected;

[0109] Model evaluation test:

[0110] The test set data is evaluated, and the spatial root mean square error RMSE of the predicted significant wave height field and the corrected significant wave height field is calculated.

[0111] Table 1 RMSE of significant wave height field prediction and correction results under different prediction lengths

[0112] Forecast horizon 1h 3h 6h 12h RMSE 0.06 0.08 0.09 0.12

[0113] By Figure 4 Taking the sea area of Hainan as an example, the prediction time efficiency of 1h, 3h, 6h and 12h is compared with the calculation results of the significant wave height in the data set, and the root mean square error is 0.06, 0.08, 0.09 and 0.12 respectively, and the calculation results of the method are good.

[0114] In addition, the application also provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize any one of the typhoon wave height prediction methods based on the deep learning MOE-Transformer model.

[0115] The application further provides a typhoon wave height prediction system based on a deep learning MOE-Transformer model, the system comprising:

[0116] one or more processors;

[0117] a memory for storing one or more programs;

[0118] When the one or more programs are executed by the one or more processors, the one or more processors implement the typhoon wave height prediction method based on the deep learning MOE-Transformer model.

[0119] Those skilled in the art will understand that embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer usable program code.

[0120] Of course, the above is only a specific application example of the present application, and the present application has other implementation manners. Any technical solution formed by equivalent replacement or equivalent transformation falls within the protection scope of the present application.

Claims

1. A method for predicting typhoon wave height based on a deep learning MOE-Transformer model, characterized in that, Includes the following steps: S1. Collect historical typhoon data for the study area as a historical typhoon dataset. Construct a virtual typhoon dataset for the study area using Markov chains, kernel density functions, and central pressure difference formulas. Specific methods include: Based on historical typhoon data, the study area is divided into grids. The frequency of historical typhoon formation and the probability of formation and termination at the starting point of each historical typhoon are statistically analyzed using kernel density function. The starting point distribution of virtual typhoons is obtained by combining the probability and the starting point distribution. Three sampling parameters are extracted during historical typhoons: the maximum wind speed near the center, the typhoon's movement speed, and the typhoon's movement direction, as the initial information in the Markov chain. The three sampling parameters are sampled at fixed time intervals to obtain the change values ​​of the three sampling parameters until the virtual typhoon terminates. Based on historical typhoon data, a mapping relationship is established between the difference ΔP between ambient air pressure and historical typhoon center air pressure, the maximum wind speed near the center, the latitude of the typhoon center, and the typhoon's movement speed. The center air pressure of the virtual typhoon is then assigned a value, and finally, a virtual typhoon dataset is obtained. S2. The historical typhoon and virtual typhoon datasets were processed using the Holland typhoon empirical model and the wind and wave growth relationship to obtain the wind field, pressure field and significant wave height field of the two datasets, respectively. S3. The ERA5 dataset was selected as the benchmark dataset. The wind field, pressure field and significant wave height field in the historical typhoon and virtual typhoon datasets were corrected using the error model. The corrected datasets constituted the typhoon spatiotemporal fusion database. The database includes meteorological data for the study area: wind field and pressure field; significant wave height values ​​for the study area: significant wave height field; typhoon data: longitude and latitude of the typhoon center, maximum wind speed near the center, and central pressure. S4. Divide the typhoon spatiotemporal fusion database into a training set, a validation set, and a test set; S5. Select typhoon data, meteorological data of the study area and significant wave height of the study area from t-12 to t as input sequences. Use the MOE-Transformer model to construct the mapping relationship between typhoon data, meteorological data of the study area and significant wave height of the study area. Predict the significant wave height of the study area at t+1, t+3, t+6 and t+12 respectively, and use them as the target output of the model. S6. The MOE-Transformer model is trained and validated based on the data in the training set and validation set to obtain the typhoon wave significant wave height prediction model; the typhoon wave significant wave height prediction model is evaluated and tested based on the data in the test set, and typhoon wave height is predicted using the typhoon wave significant wave height prediction model.

2. The typhoon wave height prediction method according to claim 1, characterized in that, The termination conditions for the virtual typhoon include: the maximum wind speed near the center of the virtual typhoon is less than 9 m / s, or the sea surface temperature at the location of the virtual typhoon is less than 10°C.

3. The typhoon wave height prediction method according to claim 1, characterized in that, In S2, historical typhoon and virtual typhoon datasets are processed using the Holland typhoon empirical model and wind-wave growth relationships to obtain their wind fields, pressure fields, and significant wave height fields, respectively. Specific methods include: The pressure field model constructed using the Holland typhoon empirical model is as follows: ； In the formula, P c P represents the sea surface pressure at the center of the cyclone. n Indicates the external air pressure; R max The radius represents the maximum wind speed; r is the distance from the calculation point to the typhoon center; B represents the Holland pressure profile parameters. The wind field model constructed using Holland's empirical typhoon model is as follows: ； In the formula, V s ρ is the gradient wind speed at a distance r from the center of the tropical cyclone. a The value represents air density; f represents the Coriolis force. The effective wave height field is calculated using the wind-wave growth relationship formula: ； In the formula, U 10 The wind speed is at a height of 10m, F is the equivalent wind zone, and H is the wind speed at a height of 10m. s The effective wave height is given by g, where g is the gravitational acceleration.

4. The typhoon wave height prediction method according to claim 1, characterized in that, In S3, the ERA5 dataset is selected as the baseline dataset. An error model is used to correct the wind field, pressure field, and significant wave height field in historical and virtual typhoon datasets. The corrected datasets constitute a spatiotemporal fusion database of typhoons. Specific methods include: ERA5 reanalysis data was used as the baseline dataset, and wind field, pressure field and significant wave height data during the typhoon's passage were extracted from the dataset through spatiotemporal matching. Then, by combining the wind field, pressure field, and significant wave height field of historical typhoons with the benchmark dataset, a three-dimensional error field historical typhoon dataset containing wind field deviation, pressure field deviation, and wave height field deviation is constructed. Z-score standardization is used for data normalization, and then the data is input into the error model for training. The error model is the U-Net error correction model. The wind field, pressure field, and significant wave height field from the historical typhoon and virtual typhoon datasets are input into the trained U-Net error correction model, and the final output is the corrected historical typhoon and virtual typhoon datasets.

5. The typhoon wave height prediction method according to claim 4, characterized in that, During the training process of the input to the error model, when the root mean square error (RMSE) on the validation set decreases by less than 1e-3 within 5 consecutive epochs and the effective wave height RMSE < 0.08, the maximum wind speed near the center RMSE < 0.05, and the central pressure RMSE < 0.01, the model is considered to have been trained.

6. The typhoon wave height prediction method according to claim 1, characterized in that, In S4, the typhoon spatiotemporal fusion database is divided into a training set, a validation set, and a test set in a ratio of 7:2:

1.

7. The typhoon wave height prediction method according to claim 1, characterized in that, In S6, the MOE-Transformer model is trained and validated based on the data from the training and validation sets to obtain a typhoon wave significant wave height prediction model based on the MOE-Transformer model. The specific method includes: S61. The MOE-Transformer model includes an input layer, a position encoder, an encoder, a decoder, and an output layer, and the hyperparameters are randomly initialized within a preset range. S62. Initialize the weight parameters of the MOE-Transformer model, including: input layer position encoding parameters, transformation matrices of query vectors, key vectors, and value vectors of each layer of the encoder, and weight parameters of the gated expert network. S63. Propagate the input sequence forward using the MOE-Transformer model: Location encoding is added to the input layer to encode location information into the input vector. The encoder processes the query vector, key vector, and value vector, performing multi-head self-attention computation (MOE) to divide them into n heads, calculating them separately, and then concatenating them for dimensionality reduction. Masked MOE computation is then performed, and an expert portfolio is selected through a gated expert network. Typhoons are standardized and classified according to their maximum near-center wind speed. Based on the MOE mechanism, typhoons of different levels are matched to the corresponding typhoon wave expert models in the expert portfolio for simulation and error correction. Decoupling and optimal scheduling between expert models are achieved through predefined wind speed-expert model mapping rules. Then, residual connections and layer normalization are performed on the output layer. Prediction results are generated through linear transformation and a softmax layer. S64. Compare the predicted results of the MOE-Transformer model with the target results, select the root mean square error (RMSE) as the loss function, and calculate the loss value of the model. S65. Based on the loss value, calculate the gradient using the backpropagation algorithm and update the weight parameters of the MOE-Transformer model to reduce the loss value. S66. The MOE-Transformer model is trained based on the data in the training set and validation set to obtain a typhoon wave height prediction model based on MOE-Transformer. S67. After training, the root mean square error (RMSE) between the MOE-Transformer predicted value and the true value is calculated based on the test set, and the typhoon wave height prediction model is evaluated and tested.

8. The typhoon wave height prediction method according to claim 7, characterized in that, The standardized classification includes: tropical depression, tropical storm, severe tropical storm, typhoon, severe typhoon, and super typhoon.

9. A typhoon wave height prediction system based on a deep learning MOE-Transformer model, characterized in that, The system includes: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the typhoon wave height prediction method based on the deep learning MOE-Transformer model as described in any one of claims 1-8.

Images