Marine and atmosphere multi-physics field prediction method and system, electronic equipment and medium

Abstract

The invention discloses an ocean and atmosphere multi-physical field prediction method and system based on a multi-scale map neural network, electronic equipment and a medium, and mainly solves the problems of one-sidedness and lack of ocean and atmosphere physical field interaction mechanisms in ocean-atmosphere coupling modeling in the prior art. According to the implementation scheme, ocean and atmosphere historical observation data are obtained and preprocessed; a multi-scale graph neural network model comprising an encoder, multiple grids, a processor and a decoder is constructed, the multiple grids are connected with a local high-resolution grid and a global low-resolution grid through shared nodes and cross-scale edges, and bidirectional information flow is achieved; iteratively optimizing parameters of the multi-scale graph neural network model by using the training set; and the performance of the model is evaluated and optimized through the test set, and finally a high-precision prediction result is generated. The method can synchronously predict the multi-physical field parameters such as the hybrid wave height, the wave period, the temperature and the wind speed, remarkably improves the calculation efficiency and the prediction precision, and can be widely applied to global climate analysis, marine disaster early warning and air-sea interaction research.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and specifically relates to a method system, electronic device and medium for predicting ocean and atmospheric environmental elements based on a multi-scale graph neural network, which can be used for global climate prediction and ocean disaster warning. Background Art

[0002] The multi-physical field coupling of the ocean and atmosphere system, such as the interaction of parameters like temperature, wind, and waves, is the core scientific issue in global climate prediction, extreme weather warning, and ocean disaster prevention and control. There is an interaction relationship between the evolution characteristics of ocean environmental elements and atmospheric variables: after waves are generated under the action of the wind field, their propagation will have a significant feedback on the atmospheric boundary layer structure by changing the sea surface roughness. This physical mechanism indicates that the ocean and atmosphere system is a highly coupled dynamic whole, and its complex non-linear interaction has a decisive impact on the prediction accuracy.

[0003] Traditional numerical prediction models rely on solving fluid mechanics and thermodynamics equations, but there are problems such as grid parameterization errors and low computational efficiency. Although deep learning has made progress in the field of meteorological prediction, its application potential in the ocean-atmosphere coupling system has not been fully developed.

[0004] The patent document with the application number CN202410535681.5 discloses a medium and long-term weather forecasting method and system based on a graph neural network. By constructing a multi-grid model covering the globe, it matches weather states with network nodes and applies the graph neural network to process global multi-weather feature data to improve the accuracy and efficiency of weather forecasting. Although this prediction method can efficiently model atmospheric variables such as wind speed, temperature, and air pressure, it completely ignores the dynamic feedback of ocean environmental elements such as significant wave height, sea surface temperature, and wave propagation direction on the atmospheric system, and fails to fully consider the two-way coupling mechanism at the air-sea interface and its potential impact on prediction accuracy, resulting in significantly insufficient prediction reliability in the ocean-atmosphere coupling scenario. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects in the existing ocean-atmosphere coupling prediction, such as one-sided modeling and the lack of physical field interaction mechanism, and propose a method, system, electronic device and medium for coupling prediction of ocean and atmosphere multi-physical fields based on a multi-scale graph neural network. Through the collaborative dynamic modeling of ocean and atmosphere multi-physical fields, it realizes the collaborative prediction of ocean and atmosphere variables with high precision and high efficiency.

[0006] To achieve the above purpose, the technical solutions adopted by the present invention include the following:

[0007] 1. A method for predicting ocean and atmosphere multi-physical fields based on a multi-scale graph neural network, characterized by comprising:

[0008] (1) Obtain historical observational data of the ocean and the atmosphere, preprocess it, and divide it into a training set and a test set;

[0009] (2) Construct a multi-scale graph neural network model including an encoder, a multi-grid, a processor, and a decoder:

[0010] The multi-grid includes a local-scale grid, a global-scale grid, shared nodes, and cross-scale edges, and the local-scale grid and the global-scale grid are connected through the shared nodes and the cross-scale edges;

[0011] (3) Use the training set to iteratively train the multi-scale graph neural network model to obtain a preliminarily trained model;

[0012] (4) Use the test set to test the preliminarily trained model, optimize the parameters of the preliminarily trained model according to the test results, and retrain the model with optimized parameters to obtain a finally trained multi-scale graph neural network model;

[0013] (5) Obtain current observational data and input it into the finally trained multi-scale graph neural network model to obtain a prediction result.

[0014] Furthermore, for the multi-scale graph neural network model, its transmission relationship is as follows:

[0015] The encoder maps the preprocessed historical observational data of the ocean and the atmosphere into the multi-grid;

[0016] The processor extracts multi-scale features of the local-scale grid and the global-scale grid through graph convolution operations, and uses the shared nodes and the cross-scale edges to achieve two-way information flow between the local-scale grid and the global-scale grid, where the shared nodes are used to store feature fusion information, and the cross-scale edges are used to control the information transfer ratio between the grids;

[0017] The decoder maps the multi-scale feature information output by the processor to a grid with the same resolution as the input data through graph deconvolution operations to generate a prediction result of the multi-physical fields of the ocean and the atmosphere in the future period.

[0018] 2. A multi-physical field prediction system for the ocean and the atmosphere based on a multi-scale graph neural network, characterized by including:

[0019] A data processing module for obtaining historical observational data of the ocean and the atmosphere and preprocessing the data;

[0020] A model construction module for constructing a multi-scale graph neural network model;

[0021] A model training module for iteratively training the multi-scale graph neural network model using a training set;

[0022] A model testing module for evaluating the performance of a trained model using a test set;

[0023] A result prediction module for generating prediction results of multiple ocean - atmosphere physical fields in future time periods.

[0024] 3. An electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that when the processor executes the program, it implements the method for predicting multiple ocean - atmosphere physical fields based on a multi - scale graph neural network according to any one of claims 1 to 6.

[0025] The present invention has the following advantages compared with the prior art:

[0026] First, through the coupling of ocean environmental variables and atmospheric environmental variables, the present invention enables the model to model the bidirectional interaction between multiple ocean - atmosphere physical fields. Compared with the prior art solutions that only predict atmospheric variables and completely ignore the dynamic feedback of ocean environmental variables on the atmospheric system, the present invention can complete the collaborative prediction of ocean environmental variables and atmospheric environmental variables, improving the number of predicted variables and the prediction accuracy.

[0027] Second, by constructing a multi - grid structure based on hexagonal grids and combining an encoder, a processor, and a decoder to complete the construction of a multi - scale graph neural network model, the present invention enables the model to capture local features and global features simultaneously, improving the prediction efficiency and computational performance of the model, reducing the waste of computing resources, and increasing the real - time performance of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flowchart for implementing the method for predicting multiple ocean - atmosphere physical fields provided in Embodiment 1 of the present invention;

[0029] Figure 2 It is a structural diagram of a multi - scale graph neural network provided in Embodiment 1 of the present invention;

[0030] Figure 3 It is a block diagram of a system for predicting multiple ocean - atmosphere physical fields provided in Embodiment 2 of the present invention;

[0031] Figure 4 It is a block diagram of the electronic device provided by the present invention;

[0032] Figure 5 It is a simulation result diagram of predicting atmospheric environmental variables of temperature using the method of the present invention;

[0033] Figure 6 It is a simulation result diagram of predicting ocean environmental variables of mixed wave period using the method of the present invention;

[0034] Figure 7 It is a simulation result diagram for predicting marine environmental variables with mixed wave heights using the method of the present invention. Specific implementation manners

[0035] The present invention will be further described in detail below in conjunction with the accompanying drawings and examples.

[0036] Example 1, a multi-physical field prediction method for the ocean and atmosphere based on a multi-scale graph neural network.

[0037] Refer to Figure 1 , the implementation steps of this example are as follows:

[0038] Step 1, obtain historical observation data of the ocean and atmosphere, preprocess it, and divide it into a training set and a test set.

[0039] 1.1) Data acquisition: Obtain historical observation data of the ocean and atmosphere from the ERA5 public dataset, including marine environmental variables such as mixed wave height, mixed wave direction, mixed wave period, etc. and atmospheric environmental variables such as temperature, wind speed, sea level pressure, etc.;

[0040] 1.2) Missing value filling: Fill the missing values of each environmental variable in the historical observation data with zero values to obtain marine and atmospheric environmental variable data without missing values;

[0041] 1.2) Interpolation processing: Perform bilinear interpolation on the marine environmental variable data with a resolution of 0.5°×0.5° and the atmospheric environmental variable data with a resolution of 0.25°×0.25° without missing values respectively, and unify all environmental variable data to a resolution grid of 1°×1° to obtain marine and atmospheric environmental variable data with unified resolution;

[0042] 1.3) Standardization processing: For the marine and atmospheric environmental variable data with unified resolution, calculate the mean μ and standard deviation σ of each environmental variable respectively, and perform standardization according to the following formula:

[0043]

[0044] Among them, X is the original data, and X norm is the standardized data;

[0045] 1.4) Dataset division: Divide the standardized marine and atmospheric environmental variable data into a training set and a test set in a ratio of 8:2, where the training set is used for model training and the test set is used for model performance evaluation.

[0046] Step 2, construct a multi-scale graph neural network model.

[0047] Refer to Figure 2 , the implementation of this step is as follows:

[0048] 2.1) Establish a multi - grid including local - scale grids, global - scale grids, shared nodes, and cross - scale edges. The local - scale grids and the global - scale grids are connected through shared nodes and cross - scale edges;

[0049] The local - scale grids are high - resolution grids with a resolution of 0.5°×0.5° obtained by using a spherical triangulation algorithm to divide the Earth's surface into hexagonal grids and recursively subdividing them on the basis of the hexagonal grids, which are used to extract features of the local range of the Earth's surface;

[0050] The global - scale grids use a spherical triangulation algorithm to divide the Earth's surface into hexagonal grids, ensuring uniform grid distribution and no directional deviation, and are low - resolution grids with a resolution of 2.5°×2.5° obtained by recursively subdividing them on the basis of the hexagonal grids, which are used to extract large - scale background information from the features of the local - scale grids;

[0051] The shared nodes are points connecting the local - scale grids and the global - scale grids, which are used to store the fusion information of local and global features and achieve information sharing between the two;

[0052] The cross - scale edges are edges connecting the local - scale grids and the global - scale grids, which are used to control the information transfer ratio between the two;

[0053] 2.2) Establish an encoder composed of 1 graph convolutional network layer and 2 multi - layer perceptron layers. First, the multi - layer perceptron converts the pre - processed historical ocean and atmosphere observation data with a resolution of 1°×1° into vector features, and then maps the vector features into the multi - grid through graph convolution operations to generate the initial feature representations of the local - scale grids and the global - scale grids;

[0054] 2.3) Establish a processor composed of 8 graph convolutional network layers. It extracts multi - scale features of the local - scale grids and the global - scale grids through graph convolution operations, and uses shared nodes and cross - scale edges to achieve bidirectional information flow between the local - scale grids and the global - scale grids, transferring small - scale features in the local - scale grids to the global - scale grids to update the large - scale background information, and transferring the large - scale background information in the global - scale grids to the local - scale grids to guide local feature extraction;

[0055] 2.4) Establish a decoder composed of 1 graph convolutional network layer and 2 multi - layer perceptron layers. First, the graph trans - convolution operation maps the multi - scale vector features output by the processor into a grid with the same resolution as the input data, and then the multi - layer perceptron completes the conversion of vector feature information into output data to generate the prediction results of the multi - physical fields of the ocean and the atmosphere in the future time period.

[0056] 2.5) Cascade the above encoder, multi-grid, processor, and decoder in sequence to form a multi-scale graph neural network model.

[0057] Step 3: Iteratively train the multi-scale graph neural network model to obtain a preliminarily trained model.

[0058] 3.1) Model initialization: Initialize the training parameters, set the learning rate to 1e-3, and the number of training epochs to 60.

[0059] 3.2) Forward propagation: Input the training set into the multi-scale graph neural network model and pass it through the encoder, processor, and decoder in sequence to generate predicted values.

[0060] 3.3) Loss calculation: Calculate the mean squared error MSE between the true value and the predicted value:

[0061]

[0062] where, y i is the true value, is the predicted value, and N is the number of samples;

[0063] 3.4) Backward propagation and gradient calculation: Use the backpropagation algorithm to calculate the gradients of the model parameters based on the mean squared error result.

[0064] 3.5) Update model parameters: Use the Adam optimizer to update the model parameters based on the gradients.

[0065] 3.6) Repeat the above steps 3.2)-3.5) until the number of training epochs reaches 60 to obtain a preliminarily trained model.

[0066] Step 4: Optimize the parameters of the preliminarily trained model and retrain it to obtain the final trained model.

[0067] 4.1) Model testing: Input the test set into the preliminarily trained multi-scale graph neural network model and obtain the predicted values of the model through forward propagation.

[0068] 4.2) Performance evaluation: Calculate the root mean squared error RMSE and mean absolute error MAE between the predicted value and the true value, and use them as evaluation values to measure the prediction performance of the model. The calculation formulas are as follows:

[0069]

[0070] where, y i is the true value, is the predicted value, and N is the number of samples;

[0071] 4.3) Parameter Tuning: Tune the learning rate and number of training epochs, which are the model training parameters, according to the evaluation value to obtain a multi-scale graph neural network model with optimized parameters.

[0072] 4.4) Retraining: Input the training set into the multi-scale graph neural network model with optimized parameters and retrain it following the same steps as the preliminary training to obtain the finally trained multi-scale graph neural network model.

[0073] Step 5: Obtain the current observed data and input it into the finally trained multi-scale graph neural network model to get the prediction result.

[0074] 5.1) Data Acquisition: Obtain the current ocean and atmosphere observation data from the ERA5 public dataset, including ocean environmental variables such as combined wave height, combined wave direction, combined wave period, etc., and atmospheric environmental variables such as temperature, wind speed, sea level pressure, etc.

[0075] 5.2) Preprocess the current ocean and atmosphere observation data using the missing value filling, interpolation, and standardization processes in Step 1.

[0076] 5.3) Model Prediction Result: Input the preprocessed current observed data into the finally trained multi-scale graph neural network model to obtain the multi-physical field prediction data of the ocean and atmosphere for the future period.

[0077] Embodiment 2: An ocean and atmosphere multi-physical field prediction system based on a multi-scale graph neural network.

[0078] Refer to Figure 3 , this example includes: a data processing module 1, a model construction module 2, a model training module 3, a model testing module 4, and a result prediction module 5. Its working principle is as follows:

[0079] The data processing module 1 is used to obtain the historical ocean and atmosphere observation data from the public data source ECMWF and preprocess it, including missing value filling, resolution unification, and standardization. Divide the standardized data into a training set and a test set, and send the training set and the test set to the model training module 3 and the model testing module 4 respectively.

[0080] The model construction module 2 is used to construct a multi-scale graph neural network model, which includes an encoder composed of 1 graph convolutional network layer and 2 multi-layer perceptron layers, a multi-grid composed of a local scale grid, a global scale grid, shared nodes, and cross-scale edges, a processor composed of 8 graph convolutional network layers, and a decoder composed of 1 graph convolutional network layer and 2 multi-layer perceptron layers.

[0081] The model training module 3 is used to iteratively train the multi-scale graph neural network model using the training set. It receives the training set sent by the data processing module 1 and the multi-scale graph neural network model sent by the model construction module 2, inputs the training set into the model, generates prediction values through forward propagation, calculates the mean square error MSE between the prediction values and the true values, calculates the gradient of the mean square error with respect to the model parameters through the backpropagation algorithm, and uses the Adam optimizer to update the model parameters until the preset number of training epochs is reached, obtaining a preliminarily trained model.

[0082] The model testing module 4 is used to evaluate the performance of the preliminarily trained model using the test set. It receives the test set sent by the data processing module 1, inputs the test set into the preliminarily trained model, generates prediction results through forward propagation, calculates the root mean square error RMSE and the mean absolute error MAE between the prediction results and the true values as evaluation values, tunes hyperparameters such as the learning rate and the number of training epochs according to the evaluation values, and retrains the model using the tuned parameters to obtain the finally trained model.

[0083] The result prediction module 5 is used to generate the prediction results of the ocean and atmosphere multi-physical fields for future time periods through the finally trained model. It obtains the current observation data, performs the same preprocessing operations on the current observation data as on the training data, inputs the preprocessed data into the finally trained model, and generates the prediction results of the ocean and atmosphere multi-physical fields for future time periods.

[0084] Embodiment 3, the present invention also provides an electronic device.

[0085] Referring to Figure 4 , the electronic device includes: a memory 1, a processor 2, an input / output interface 3, and a bus 4. Among them, the memory 1, the processor 2, and the input / output interface 3 achieve information interaction with each other inside the device through the bus 4.

[0086] The memory 1 can be implemented in the form of a read-only memory ROM, a random access memory RAM, a solid-state drive SSD, a mechanical hard drive HDD, a USB flash drive, etc. The memory 1 can store all the application programs required by the present invention.

[0087] The processor 2 can be implemented using a central processing unit CPU, a microprocessor, or one or more integrated circuits, and is used to execute all the application programs required by the present invention to implement the technical solutions provided by the embodiments of the present invention.

[0088] The input / output interface 3 is used to connect input / output devices to achieve information input and output. The input / output devices can be configured as components in the device or external devices to provide corresponding functions. The input devices can include a keyboard, a mouse, a touch screen, etc., and the output devices can include a display, etc.

[0089] The bus 4 is a communication line used to connect various components, such as the memory 1, the processor 2, and the input / output interface 3, and can perform information transmission and data interaction between the components.

[0090] It should be noted that although the above devices only show the memory 1, the processor 2, the input / output interface 3, and the bus 4, in the specific implementation process, the electronic device can be a distributed computing node system, a server group / server, a desktop computer, a laptop computer, etc. Among them, the server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery network CDN, and big data and artificial intelligence platforms. In other words, the meteorological prediction method based on the weather forecasting system can be executed by software or hardware installed on the terminal device or the server device.

[0091] The present invention also provides a non-transitory computer-readable storage medium, which stores computer instructions for executing the ocean and atmosphere multi-physical field prediction method based on the multi-scale graph neural network in the embodiment. That is, the method steps in the above embodiment are implemented as software or computer code that can be stored in a recording medium such as a CDROM, RAM, floppy disk, hard disk, or magneto-optical disk, or are implemented by computer code originally stored in a remote recording medium or a non-temporary machine-readable medium and to be stored in a local recording medium downloaded through the network, so that the method represented herein can be stored on such a recording medium and processed by software using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware such as an ASIC or FPGA).

[0092] The effects of the present invention can be further illustrated by the following simulation experiments.

[0093] 1. Simulation experiment conditions:

[0094] Hardware conditions:

[0095] a) CPU: Intel(R) Xeon(R) CPU E5-2630 v4 @ 2.20GHz;

[0096] b) GPU: NVIDIA TITAN V 12G

[0097] c) Hard disk: HDD 2TB

[0098] Operating system: Ubuntu 22.04LTS

[0099] Software conditions: Visual Studio Code

[0100] 2. Simulation experiment content and results:

[0101] Simulation 1: Under the above simulation conditions, the present invention is used to predict the temperature of the atmosphere after 6 hours, and the prediction result is visually compared with the true value. The results are as Figure 5 shown, where 5(a) is the true value and 5(b) is the predicted value.

[0102] Simulation 2: Under the above simulation conditions, the present invention is used to predict the mixed wave period of the ocean after 6 hours, and the prediction result is visually compared with the true value. The results are as Figure 6 shown, where 6(a) is the true value and 6(b) is the predicted value;

[0103] Simulation 3: Under the above simulation conditions, the present invention is used to predict the mixed wave height of the ocean after 6 hours, and the prediction result is visually compared with the true value. The results are as Figure 7 shown, where 7(a) is the true value and 7(b) is the predicted value.

[0104] From Figure 5 , Figure 6 , Figure 7 the simulation results, it can be seen that the present invention can accurately predict the multi-physical fields of the ocean and the atmosphere, and can simultaneously complete the prediction of ocean environmental elements and atmospheric environmental elements, increasing the number of variable predictions and improving the utilization rate of computing resources.

[0105] The above description is only several specific examples of the present invention and does not constitute any limitation to the present invention. Obviously, for professionals in the field, after understanding the content and principle of the present invention, various modifications and changes in form and details may be made without departing from the principle and structure of the present invention. However, these corrections and changes based on the idea of the present invention are still within the protection scope of the claims of the present invention.

[0106] It should be noted that the step numbers in the specification and claims of the present invention are only for a clear description of the implementation scheme of the present invention for easy understanding, and their sequence numbers are not limited.

Claims

1. A method for predicting multi-physical fields of ocean and atmosphere based on multi-scale graph neural network, characterized in that, Including: (1) Obtain historical observation data of the ocean and the atmosphere, preprocess it, and divide it into a training set and a test set; (2) Construct a multi-scale graph neural network model including an encoder, a multi-grid, a processor, and a decoder: The multi-grid includes a local-scale grid, a global-scale grid, shared nodes, and cross-scale edges, and the local-scale grid and the global-scale grid are connected through shared nodes and cross-scale edges; (3) Iteratively train the multi-scale graph neural network model using the training set to obtain a preliminarily trained model; (4) Test the preliminarily trained model using the test set, optimize the parameters of the preliminarily trained model according to the test results, and retrain the model with optimized parameters to obtain a finally trained multi-scale graph neural network model; (5) Obtain current observation data and input it into the finally trained multi-scale graph neural network model to obtain a prediction result.

2. The method according to claim 1, characterized in that, In (1), the historical observation data of the ocean and the atmosphere is preprocessed and divided into a training set and a test set, and its implementation includes the following: (1a) Fill the missing values of each environmental variable in the historical observation data by zero-value filling to obtain ocean and atmosphere environmental variable data without missing values; (1b) Interpolate the ocean and atmosphere environmental variable data without missing values, that is, perform bilinear interpolation on the atmospheric environmental variable data with a resolution of 0.25°×0.25° and the ocean environmental variable data with a resolution of 0.5°×0.5°, and unify all environmental variable data onto a resolution grid of 1°×1° to obtain ocean and atmosphere environmental variable data with unified resolution; (1c) Standardize the ocean and atmosphere environmental variable data with unified resolution, that is, calculate the mean and standard deviation of each environmental variable data respectively, and map each environmental variable data to the interval of 0-1 to obtain standardized ocean and atmosphere environmental variable data; (1d) Divide the standardized ocean and atmosphere environmental variable data into a training set and a test set according to a ratio of 8:

2.

3. The method according to claim 1, wherein In (2), for the multi-scale graph neural network model, its transmission relationship is as follows: The encoder maps the preprocessed historical observation data of the ocean and the atmosphere into the multi-grid; The processor extracts multi-scale features of the local-scale grid and the global-scale grid through graph convolution operations, and realizes two-way information flow between the local-scale grid and the global-scale grid using shared nodes and cross-scale edges, where the shared nodes are used to store feature fusion information, and the cross-scale edges are used to control the information transfer ratio between the grids; The decoder maps the multi-scale feature information output by the processor to the same resolution grid as the input data through graph deconvolution operations to generate a prediction result of the multi-physical field of the ocean and the atmosphere in the future period.

4. The method according to claim 1, characterized in that, In (2), for the local-scale grid, global-scale grid, shared nodes, and cross-scale edges in the multi-grid, their structural functions are as follows: The local-scale grid uses a high-resolution grid with a resolution of 0.5°×0.5° to extract features in a local range of the earth's surface; The global-scale grid uses a low-resolution grid with a resolution of 2.5°×2.5° to extract large-scale background information from the features of the local-scale grid; The shared node is a point connecting the local-scale grid and the global-scale grid, and is also used to store the fusion information of local and global features to facilitate information sharing between the two; The cross-scale edge is an edge connecting the local-scale grid and the global-scale grid, and is also used to control the information transfer ratio between the two.

5. The method according to claim 1, wherein The preliminary training of the multi-scale graph neural network model using the training set in (3) is implemented as follows: (3a) Input the training set into the multi-scale graph neural network model, initialize the training parameters: the learning rate is 1e-3, the number of training epochs is 60, and perform forward propagation; (3b) Calculate the mean square error MSE between the true value and the predicted value obtained by forward propagation, and use the backpropagation algorithm to calculate the gradient of the model parameters according to the mean square error result; (3c) Update the model parameters using the Adam optimizer according to the obtained gradient; (3d) Repeat the above steps (3b)-(3c) until the number of training epochs reaches 60 to obtain the preliminarily trained model.

6. The method according to claim 1, characterized in that, (4) The parameter tuning and retraining of the preliminarily trained model using the test set is implemented as follows: (4a) Input the test set into the preliminarily trained multi-scale graph neural network model, and obtain the prediction result of the model through forward propagation; (4b) Calculate the root mean square error RMSE and the mean absolute error MAE between the prediction result and the true value, and use the calculation results as evaluation values; (4c) Tune the learning rate and the number of training epochs parameters according to the evaluation values to obtain the multi-scale graph neural network model with optimized parameters; (4d) Input the training set into the multi-scale graph neural network model with optimized parameters, and retrain it according to the same steps as the preliminary training to obtain the finally trained multi-scale graph neural network model.

7. An ocean and atmosphere multi-physical field prediction system based on a multi-scale graph neural network, characterized in that, including: A data processing module for obtaining historical observation data of the ocean and the atmosphere and preprocessing the data; A model construction module for constructing a multi-scale graph neural network model; A model training module for iteratively training the multi-scale graph neural network model using the training set; A model testing module for evaluating the performance of the trained model using the test set; A result prediction module for generating prediction results of the multi-physical fields of the ocean and the atmosphere in the future time period.

8. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the multi-physical field prediction method of the ocean and the atmosphere based on the multi-scale graph neural network according to any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions for causing the computer to execute the multi-physical field prediction method of the ocean and the atmosphere based on the multi-scale graph neural network according to any one of claims 1 to 6.

Images