A three-dimensional hydrodynamic environment prediction system considering the vertical structural characteristics of the ocean

Abstract

This invention discloses a three-dimensional hydrodynamic environment forecasting system considering the vertical structural characteristics of the ocean, belonging to the interdisciplinary field of deep learning and ocean forecasting. It includes: a data acquisition device for acquiring three-dimensional hydrodynamic environment data of the ocean area to be measured; a data storage and management device for persistently storing, segmenting, statistically analyzing, and managing the three-dimensional hydrodynamic environment data; a data loading device for setting loading parameters according to the forecasting task and processing the three-dimensional hydrodynamic environment data into high-dimensional matrix data according to the loading parameters; a model training device for training a deep learning model based on the high-dimensional matrix data to obtain a forecasting model; a forecasting device for generating three-dimensional hydrodynamic environment forecasting results based on the forecasting model and real-time acquired data; and a visualization device connected to the forecasting device for graphically displaying the forecasting results. This invention significantly improves the timeliness and practicality of three-dimensional hydrodynamic environment forecasting for ocean areas.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of deep learning and ocean forecasting, and in particular relates to a three-dimensional hydrodynamic environment forecasting system that takes into account the vertical structural characteristics of the ocean. Background Technology

[0002] Currently, three-dimensional marine hydrodynamic environment forecasting primarily relies on marine numerical models, which are methods for simulating marine physical processes based on physical equations and numerical analysis. However, the ocean is an extremely complex chaotic system, and human understanding of it is insufficient to construct a high-accuracy marine numerical model. With the accumulation of marine data and the development of deep learning, deep learning methods are gradually being applied to three-dimensional marine hydrodynamic environment forecasting.

[0003] Deep learning is an empirical model based on mathematical statistics and optimization theory. While ocean forecasting models built using this method have lower physical interpretability compared to numerical ocean models, they often require fewer computational resources, offer faster forecasting speeds, and generally achieve higher accuracy. However, existing deep learning-based three-dimensional ocean hydrodynamic environment forecasting models are limited, often relying on single forecasting elements and failing to consider the coupling relationships between different variables and water layers. This leads to information waste and limits the model's forecasting performance. Summary of the Invention

[0004] To address the problem of information waste caused by the lack of consideration for the coupling relationships between different variables and water layers in existing deep learning-based three-dimensional marine hydrodynamic environment forecasting models, this invention proposes an intelligent three-dimensional marine hydrodynamic environment forecasting system that considers the vertical structural characteristics of the ocean. This forecasting system can effectively integrate and construct the coupling relationships between different variables and water layers, thereby improving the accuracy of three-dimensional marine hydrodynamic environment forecasting.

[0005] To address the aforementioned technical problems, this invention provides a three-dimensional hydrodynamic environment prediction system that considers the vertical structural characteristics of the ocean, comprising:

[0006] Data acquisition device, used to acquire three-dimensional hydrodynamic environment data of the marine area to be measured;

[0007] A data storage and management device, connected to the data acquisition device, is used for persistent storage, segmentation, statistics and management of the three-dimensional hydrodynamic environment data;

[0008] A data loading device, connected to the data storage and management device, is used to set loading parameters according to the forecast task and process the three-dimensional hydrodynamic environment data into high-dimensional matrix data according to the loading parameters.

[0009] A model training device, connected to the data loading device, is used to train a deep learning model based on the high-dimensional matrix data to obtain a prediction model.

[0010] The forecasting device is connected to the model training device and is used to generate three-dimensional marine hydrodynamic environment forecasting results based on the forecasting model and real-time acquired data.

[0011] A visualization device, connected to the forecasting device, is used to graphically display the forecast results.

[0012] Preferably, the data acquisition device includes:

[0013] The data download module is used to automatically download three-dimensional variable data of the ocean area to be measured according to preset parameters; wherein, the preset parameters include a variable list, latitude and longitude range, depth range, time range, and data format.

[0014] Preferably, the data storage and management device includes:

[0015] The database module is used to receive and store downloaded 3D variable data;

[0016] The segmentation module is used to segment the three-dimensional variable data into two-dimensional spatial field data at a single time step;

[0017] The statistics module is used to calculate the mean and standard deviation of the two-dimensional spatial field data;

[0018] The management module is used to perform integrity checks, slice access, querying, and deletion operations on the two-dimensional spatial field data.

[0019] Preferably, the data loading device includes:

[0020] The parameter setting module is used to set the forecast lead, characteristic variables, label variables, and the corresponding water layer list;

[0021] The dimension stacking module is used to stack two-dimensional spatial field data into high-dimensional matrix data according to the loading parameters;

[0022] The statistical information loading module is used to load the mean and standard deviation of the high-dimensional matrix data from the database module.

[0023] The normalization module is used to perform Z-score normalization on the high-dimensional matrix data based on the mean and standard deviation;

[0024] The dataset partitioning module is used to divide the normalized high-dimensional matrix data into training and test sets.

[0025] Preferably, the model training device includes:

[0026] The model building module is used to build an initial prediction model based on the SimVP deep learning model architecture; wherein the initial prediction model includes an encoder, a transformer, and a decoder;

[0027] The training execution module is used to train the initial prediction model using the training set to obtain the prediction model; wherein, the MSE loss function and Adam optimizer are used for optimization during the training process.

[0028] Preferably, the encoder includes Ns hidden layer convolutional kernels with a number of hid_S ConvNormReLU blocks, used for downsampling and feature extraction of the input high-dimensional matrix data;

[0029] The converter includes an Inception parallel convolution module with Nt hidden layer convolution kernels and a number of hid_T, which is used to learn the nonlinear coupling law between different water layers and different variables.

[0030] The decoder includes Ns unConvNormReLU blocks with a hidden layer convolution kernel size of hid_S, used to upsample the feature map to reconstruct the target frame sequence.

[0031] Preferably, the system further includes a performance testing device connected to the model training device, used to perform performance testing on the forecast model based on a test set, wherein RMSE and PCC metrics are used to evaluate the predictive performance of the forecast model.

[0032] Preferably, the performance testing device includes:

[0033] The RMSE calculation module is used to calculate the root mean square error between the predicted result and the actual value.

[0034] The PCC calculation module is used to calculate the Pearson correlation coefficient between the predicted results and the actual values.

[0035] Preferably, the forecasting device includes:

[0036] The real-time data acquisition module is used to acquire the latest three-dimensional hydrodynamic environment data of the marine area to be measured in real time.

[0037] The real-time forecasting module is used to input the latest three-dimensional hydrodynamic environment data of the marine area to be measured into the forecasting model and generate marine three-dimensional hydrodynamic environment forecasting results in real time.

[0038] Compared with the prior art, the present invention has the following advantages and technical effects:

[0039] This invention automatically downloads and acquires three-dimensional hydrodynamic environmental data of marine areas through a data acquisition device, which significantly improves data acquisition efficiency, reduces human intervention, and ensures the integrity and timeliness of the data.

[0040] This invention uses a data storage and management device to persistently store, segment, statistically analyze, and manage the acquired three-dimensional hydrodynamic environment data, achieving efficient data storage and rapid retrieval, and significantly improving data management efficiency and data quality.

[0041] This invention uses a data loading device to set loading parameters according to the forecast task and efficiently processes the data into high-dimensional matrix data, thereby achieving rapid data loading and memory optimization and significantly improving data processing efficiency.

[0042] This invention trains a deep learning model based on high-dimensional matrix data using a model training device, effectively capturing the nonlinear coupling patterns between different water layers and different variables, and significantly improving the model's prediction accuracy and generalization ability.

[0043] This invention significantly improves the model's ability to model complex marine environments by using the SimVP deep learning model architecture to construct a forecast model, and effectively reduces the model's computational complexity and training time.

[0044] This invention achieves effective feature extraction from input data, learning of nonlinear coupling patterns, and reconstruction of target frame sequences through the synergistic action of encoders, converters, and decoders, significantly improving the prediction accuracy of the model.

[0045] This invention employs the MSE loss function and Adam optimizer for model training, and introduces a dynamic learning rate decay strategy, which significantly accelerates the model convergence speed and effectively avoids oscillations and overfitting problems during the training process.

[0046] This invention uses a performance testing device to evaluate the performance of forecast models based on RMSE and PCC indices, thereby achieving a quantitative assessment of the model's forecast accuracy and significantly improving the interpretability and reliability of the model's performance.

[0047] This invention achieves real-time, high-precision forecasting of the three-dimensional hydrodynamic environment of marine areas through a real-time data acquisition module and a real-time forecasting module, significantly improving the timeliness and practicality of the forecast.

[0048] This invention uses a visualization device to intuitively display forecast results to users in a graphical manner, which significantly improves the readability of forecast results and user experience, making it easier for users to quickly understand and apply forecast information.

[0049] In addition to seawater temperature, this invention includes zonal seawater temperature, meridional seawater temperature, and seawater salinity as forecast variables for three-dimensional marine hydrodynamic environment forecasting, thus increasing the abundance of forecast variables. Therefore, the three-dimensional marine hydrodynamic environment forecasting system constructed by this invention can provide data support for my country's marine development. Attached Figure Description

[0050] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0051] Figure 1 This is a schematic diagram of the system structure according to an embodiment of the present invention;

[0052] Figure 2 This is a schematic diagram illustrating the construction of vertical structural features according to an embodiment of the present invention. Detailed Implementation

[0053] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0054] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0055] like Figure 1 As shown, this embodiment provides a three-dimensional hydrodynamic environment prediction system that considers the vertical structural characteristics of the ocean, including:

[0056] Data acquisition device, used to acquire three-dimensional hydrodynamic environment data of the marine area to be measured;

[0057] A data storage and management device, connected to a data acquisition device, is used for persistent storage, segmentation, statistics, and management of three-dimensional hydrodynamic environment data;

[0058] The data loading device, connected to the data storage and management device, is used to set loading parameters according to the forecast task and process the three-dimensional hydrodynamic environment data into high-dimensional matrix data according to the loading parameters.

[0059] The model training device, connected to the data loading device, is used to train a deep learning model based on high-dimensional matrix data to obtain a prediction model.

[0060] The forecasting device, connected to the model training device, is used to generate three-dimensional marine hydrodynamic environment forecasting results based on the forecasting model and real-time acquired data.

[0061] A visualization device, connected to the forecasting device, is used to graphically display the forecast results.

[0062] Furthermore, the data acquisition device includes:

[0063] The data download module is used to automatically download three-dimensional variable data of the ocean area to be measured according to preset parameters; the preset parameters include a variable list, latitude and longitude range, depth range, time range, and data format.

[0064] Furthermore, the data download module in this embodiment is an automated data download script designed for the GLORYS12V1 reanalysis dataset. This script is written based on the official data download API (Application Programming Interface) provided by Copernicus Marine Service and can automatically download target data to the local database according to preset parameters, including the list of three-dimensional variables to be downloaded, latitude and longitude range, depth range, time range, and data format.

[0065] Furthermore, the data storage and management device includes:

[0066] The database module is used to receive and store downloaded 3D variable data;

[0067] The segmentation module is used to segment three-dimensional variable data into two-dimensional spatial field data at a single time step;

[0068] The statistics module is used to calculate the mean and standard deviation of two-dimensional spatial field data;

[0069] The management module is used to perform integrity checks, slice access, querying, and deletion operations on two-dimensional spatial field data.

[0070] Furthermore, the data storage and management device in this embodiment includes a software module for persistent storage, segmentation, statistics, and management of downloaded data. Specifically, persistent storage refers to the persistent preservation of data using a disk or solid-state drive; segmentation refers to dividing high-dimensional data (three-dimensional spatial field time series) into two-dimensional data (single-time-step two-dimensional spatial field); statistics refers to calculating the mean and standard deviation of the two-dimensional physical field time series of all variable data in the database for each water layer; and management refers to functions for quickly verifying data integrity, quickly accessing complete or sliced ​​data, quickly querying data, and quickly deleting data.

[0071] Furthermore, the data loading device includes:

[0072] The parameter setting module is used to set the forecast lead, characteristic variables, label variables, and the corresponding water layer list;

[0073] The dimension stacking module is used to stack two-dimensional spatial field data into high-dimensional matrix data according to the loading parameters;

[0074] The statistics loading module is used to load the mean and standard deviation of high-dimensional matrix data from the database module.

[0075] The normalization module is used to perform Z-score normalization on high-dimensional matrix data based on the mean and standard deviation.

[0076] The dataset partitioning module is used to divide the normalized high-dimensional matrix data into training and test sets.

[0077] Furthermore, the data loading device in this embodiment is a lazy data loader used to quickly load training and test samples corresponding to any forecast task. It can quickly load data into memory on demand, thereby achieving fast loading and optimizing memory management. The data source for loading by this device is two-dimensional data after database segmentation. Its specific loading process can be decomposed into: setting loading parameters, stacking feature and label dimensions (giving learnable vertical structural feature characteristics), loading statistical information corresponding to features, feature normalization, and splitting the dataset.

[0078] Specifically, the parameter setting module sets the loading parameters, which include setting the forecast lead, characteristic variables, label variables, and the list of water layers corresponding to the characteristic and label variables.

[0079] like Figure 2 As shown, the dimension stacking module stacks two-dimensional spatial field data into high-dimensional matrix data according to the loading parameters. Specifically, it stacks a two-dimensional matrix with a dimension order of (Height, Width) into a high-dimensional matrix with a dimension order of (Batch, Depth, Variable, Height, Width) based on the set loading parameters. The dimensions of this high-dimensional matrix represent the sample size, number of water layers, number of variables, latitude grid points, and longitude grid points, respectively, thereby constructing the required learnable vertical structural features or label samples.

[0080] The statistical information loading module loads the corresponding statistical information, specifically the mean and standard deviation of each feature variable for each water layer from the database.

[0081] The normalization module performs Z-score normalization on high-dimensional matrix data based on the mean and standard deviation. Specifically, it performs Z-score normalization on the feature variables based on the mean and standard deviation. The calculation formula is as follows:

[0082] ;

[0083] in, For normalized data, Features before normalization express The average value, express The standard deviation.

[0084] The dataset partitioning module divides the dataset into training and test sets along the batch dimension according to a given ratio, specifically by dividing the normalized dataset (composed of features and labels) into training and test sets, and randomly shuffling the sample order in the training set. Based on this, lazy data loaders are constructed for the training and test sets respectively.

[0085] Furthermore, the model training device includes:

[0086] The model building module is used to build an initial prediction model based on the SimVP deep learning model architecture; the initial prediction model includes an encoder, a transformer, and a decoder.

[0087] The training execution module is used to train the initial prediction model using the training set to obtain the prediction model; during the training process, the MSE loss function and the Adam optimizer are used for optimization.

[0088] Furthermore, the encoder includes Ns hidden layer convolutional kernels with a number of hid_S ConvNormReLU blocks, used for downsampling and feature extraction of the input high-dimensional matrix data;

[0089] The converter includes an Inception parallel convolution module with Nt hidden layers and hid_T kernels, used to learn the nonlinear coupling laws between different water layers and different variables;

[0090] The decoder includes Ns unConvNormReLU blocks with hid_S hidden layer convolutional kernels, used to upsample the feature maps to reconstruct the target frame sequence.

[0091] Furthermore, the model training device in this embodiment specifically trains and constructs a pre-trained model that considers the vertical structural features of the ocean based on the training set according to a given deep learning model architecture and hyperparameters. The hyperparameters of the device include model architecture parameters, loss function, optimizer type, learning rate, and training termination condition.

[0092] The deep learning model architecture used in this embodiment is the SimVP (Simpler Yet Better VideoPrediction) model architecture. This architecture is a neural network architecture for frame sequence (video) prediction. It receives a high-dimensional matrix (a batch of frame sequences) and outputs prediction results (a batch of frame sequences) through its learned high-dimensional mapping space. The batch of frame sequences corresponds to a high-dimensional matrix with dimensions in the order (Batch, Depth, Variable, Height, Width). Based on the construction of this matrix, the SimVP model architecture can learn the nonlinear coupling laws between different water layers and different variable physical fields, thereby constructing a pre-trained model that considers the vertical structural characteristics of the ocean.

[0093] The hyperparameters specifically include:

[0094] The loss function used is the MSE (Mean Square Error) loss function, and its mathematical formula is shown below:

[0095] ;

[0096] Where MSE is the loss, N is the sample size, and y i Let y be the true value of the i-th sample. i ’ This is the predicted value for the i-th sample;

[0097] The optimizer type used is the Adam (Adaptive Moment Estimation) optimizer;

[0098] The learning rate is dynamic, and it decays with the number of training epochs. The decay formula is as follows:

[0099] ;

[0100] Where l is the learning rate, l0 is the initial learning rate, α is the decay coefficient, and n is the number of training rounds;

[0101] The training termination condition is determined by a given upper limit on the number of training epochs or the loss update status on the training set. If the difference between the previous training loss and the current training loss is less than a given threshold before the given upper limit on the number of training epochs, training stops; otherwise, it stops when the upper limit on the number of training epochs is reached.

[0102] Furthermore, the system also includes a performance testing device connected to the model training device, used to test the performance of the forecast model based on a test set, wherein the RMSE and PCC indices are used to evaluate the predictive performance of the forecast model.

[0103] Furthermore, the performance testing device in this embodiment is used to test the model training device. This module uses RMSE (Root Mean Square Error) and PCC (Pearson Correlation Coefficient) to measure the error magnitude and linear correlation between the predicted results and the true values, respectively. The mathematical formulas are as follows:

[0104] ;

[0105]

[0106] Where RMSE is the root mean square error, N is the sample size, and y i Let y be the true value of the i-th sample. i ’ Let be the predicted value for the i-th sample; PCC is the correlation coefficient, y is the true value, y' is the correction result, i is the sample number, and N is the sample size. Let y be the average value. Let y' be the average value.

[0107] Furthermore, the performance testing apparatus includes:

[0108] The RMSE calculation module is used to calculate the root mean square error between the predicted result and the actual value.

[0109] The PCC calculation module is used to calculate the Pearson correlation coefficient between the predicted results and the actual values.

[0110] Furthermore, the forecasting device includes:

[0111] The real-time data acquisition module is used to acquire the latest three-dimensional hydrodynamic environment data of the marine area to be measured in real time.

[0112] The real-time forecasting module is used to input the latest three-dimensional hydrodynamic environment data of the marine area to be measured into the forecasting model and generate real-time three-dimensional hydrodynamic environment forecasting results.

[0113] As a supplementary embodiment, the specific parameters involved in the aforementioned specific steps are as follows in this embodiment:

[0114] The required list of variables to download is [seawater temperature, seawater salinity, zonal seawater current velocity, meridional seawater current velocity]; longitude range is 100~118°E; latitude range is 8~26°N; depth range is 0~2000m; time range is January 1, 1993 to December 31, 2020; and data format is NetCDF (Network Common Data Form).

[0115] The forecast lead time was set to 1, 2, and 3 days, respectively. The feature and label variables were seawater temperature, seawater salinity, zonal seawater current velocity, and meridional seawater current velocity. The list of water layers corresponding to the feature and label variables was [0, 2, 3, 4, 5, 6, 8, 10, 11, 13, 16, 18, 22, 25, 29, 34, 40, 47, 56, 66, 78, 92, 110, 131, 156, 186, 222, 266, 318, 380, 454, 541, 644, 763, 902, 1062, 1245, 1452, 1684] (in meters).

[0116] The given proportion for splitting the dataset is set to the first 95% of the data in the training set and the last 5% in the test set.

[0117] Ns, hid_S, Nt, and hid_T are set to 4, 16, 8, and 256, respectively; for the dynamic learning rate, l0 and α are set to 0.1 and 0.95, respectively; the maximum number of training rounds is set to 100, and the threshold for the difference between the previous training loss and the current training loss is set to 0.0001.

[0118] To verify the superiority of the intelligent prediction system for three-dimensional marine hydrodynamic environment that considers the vertical structural characteristics of the ocean proposed in this embodiment, two sets of experiments were conducted, one considering the vertical structural characteristics of the ocean and the other not. The experimental results of this embodiment are as follows:

[0119] Depend on and It can be seen that, in the test set, considering the vertical structure characteristics of the ocean, the RMSE corresponding to the zonal seawater current velocity forecast 1, 2 and 3 days in advance are 0.063, 0.081 and 0.078 m / s, respectively, which are better than the 0.260, 0.282 and 0.293 m / s in the corresponding unconsidered case; the PCC are 0.854, 0.797 and 0.778, respectively, which are better than the 0.771, 0.740 and 0.705 in the corresponding unconsidered case. Similarly, considering the characteristics of the ocean's vertical structure, the RMSEs for the meridional seawater current velocity forecasts 1, 2, and 3 days in advance are 0.058, 0.080, and 0.072 m / s, respectively, which are better than the 0.267, 0.274, and 0.304 m / s for the unconsidered cases; the PCCs are 0.845, 0.799, and 0.780, respectively, which are better than the 0.760, 0.761, and 0.706 for the unconsidered cases. Likewise, considering the characteristics of the ocean's vertical structure, the RMSEs for the seawater temperature forecasts 1, 2, and 3 days in advance are 0.191, 0.259, and 0.291℃, respectively, which are better than the 0.275, 0.294, and 0.3℃ for the unconsidered cases. 22℃ is optimal; the PCC values ​​are 0.920, 0.938, and 0.925, respectively, which are better than the 0.767, 0.754, and 0.709 in the corresponding unconsidered cases. Similarly, considering the vertical structure characteristics of the ocean, the RMSE values ​​for seawater salinity forecasts 1, 2, and 3 days in advance are 0.094, 0.140, and 0.132 PSU, respectively, which are better than the 0.283, 0.296, and 0.321 PSU in the corresponding unconsidered cases; the PCC values ​​are 0.744, 0.758, and 0.730, respectively, which are better than the 0.750, 0.739, and 0.689 in the corresponding unconsidered cases. The latter forecast is slightly higher than the former for the one-day advance forecast, and the former is better for the remaining advance forecast values. Overall, the forecast performance of the intelligent forecasting system for the three-dimensional hydrodynamic environment of the ocean considering the vertical structure characteristics of the ocean is significantly higher than that without considering the vertical structure characteristics, which fully verifies the technical advantages of the present invention.

[0120] Table 1

[0121]

[0122] Table 2

[0123]

[0124] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A three-dimensional hydrodynamic environmental forecasting system that accounts for vertical structure features of the ocean, characterized by, The system comprises: a data acquisition device for automatically downloading three-dimensional variable data of a to-be-tested marine area according to preset parameters; wherein the preset parameters include a variable list, a latitude and longitude range, a depth range, a time range, and a data format; a data storage and management device connected with the data acquisition device, for persistently storing, segmenting, counting, and managing the three-dimensional variable data; a data loading device connected with the data storage and management device, for setting loading parameters according to a prediction task, and processing the three-dimensional variable data into high-dimensional matrix data according to the loading parameters; a model training device connected with the data loading device, for training a deep learning model based on the high-dimensional matrix data to obtain a prediction model; a prediction device connected with the model training device, for generating a marine three-dimensional hydrodynamic environment prediction result based on the prediction model and real-time acquired data; a visualization device connected with the prediction device, for graphically displaying the prediction result. The data loading device comprises: a parameter setting module for setting a prediction lead time, characteristic variables, label variables, and corresponding water layer lists; a dimension stacking module for stacking two-dimensional spatial field data into high-dimensional matrix data according to the loading parameters; specifically, stacking a two-dimensional matrix with a dimension order of (Height, Width) into a high-dimensional matrix with a dimension order of (Batch, Depth, Variable, Height, Width) based on the set loading parameters, wherein the dimensions of the high-dimensional matrix represent sample quantity, water layer quantity, variable quantity, latitude grid, and longitude grid, respectively. The model training device comprises: a model construction module for constructing an initial prediction model based on a SimVP deep learning model architecture; wherein the initial prediction model comprises an encoder, a transformer, and a decoder; a training execution module for training the initial prediction model using a training set to obtain the prediction model; wherein an MSE loss function and an Adam optimizer are used for optimization during the training process; the encoder comprises Ns hidden layers of ConvNormReLU blocks with a convolution kernel number of hid_S, for downsampling and feature extraction of the input high-dimensional matrix data; the transformer comprises Nt hidden layers of Inception parallel convolution modules with a convolution kernel number of hid_T, for learning nonlinear coupling rules between different water layers and different variables; the decoder comprises Ns hidden layers of unConvNormReLU blocks with a convolution kernel number of hid_S, for upsampling feature maps to reconstruct a target frame sequence.

2. The system of claim 1, wherein the data storage and management device comprises: a database module for receiving and storing the downloaded three-dimensional variable data; a segmentation module for segmenting the three-dimensional variable data into two-dimensional spatial field data of a single time step; a statistical module for calculating the mean and standard deviation of the two-dimensional spatial field data. The management module is configured to perform integrity check, slice access, query, and deletion operation on the two-dimensional spatial field data. 3.The system of claim 1, wherein, The data loading device further comprises: The statistical information loading module is configured to load the mean and standard deviation corresponding to the high-dimensional matrix data from the database module; The normalization module is configured to perform Z-score normalization on the high-dimensional matrix data based on the mean and standard deviation; The data set division module is configured to divide the normalized high-dimensional matrix data into a training set and a test set. 4.The system of claim 1, wherein, The system further comprises a performance testing device connected with the model training device, configured to perform performance testing on the prediction model based on the test set, wherein the prediction performance of the prediction model is evaluated by using RMSE and PCC indexes. 5.The system of claim 4, wherein, The performance testing device comprises: The RMSE calculation module is configured to calculate the root mean square error between the prediction result and the true value; The PCC calculation module is configured to calculate the Pearson correlation coefficient between the prediction result and the true value. 6.The system of claim 1, wherein, The prediction device comprises: The real-time data acquisition module is configured to acquire the latest three-dimensional hydrodynamic environment data of the to-be-tested marine area in real time; The real-time prediction module is configured to input the latest three-dimensional hydrodynamic environment data of the to-be-tested marine area into the prediction model to generate a real-time marine three-dimensional hydrodynamic environment prediction result.

Images