Intelligent black tide forecasting method and system based on ocean three-dimensional environment information

Abstract

The invention provides a black tide intelligent forecasting method and system based on ocean three-dimensional environment information, and relates to the technical field of artificial intelligence and ocean meteorological service industries, and the method comprises the following steps: obtaining an ocean three-dimensional environment field through multiple ways, and carrying out the preprocessing of data; making an ocean large model training data set; a large black tide forecasting model is constructed, and parameters of an optimal forecasting model are obtained; generating a future marine environment forecasting field and a black tide forecasting result; and pushing through a website and a WeChat official account. According to the method, accurate forecasting of the black tide by combining the ocean three-dimensional environment information and the artificial intelligence model is proposed for the first time, the limitation of a traditional black tide forecasting method is broken through, and higher forecasting accuracy and broadness are achieved. Through the system, more detailed black tide flow and path forecast can be provided, rapid response and real-time early warning of dynamic changes of the black tide can be realized, and the system has extremely high practical application value.

Description

Technical Field

[0001] The present invention relates to a method and system for intelligently forecasting a Kuroshio tide based on three-dimensional ocean environmental information, and relates to the technical fields of artificial intelligence and marine meteorological service industries. Background Art

[0002] The Kuroshio is the strongest western boundary current in the North Pacific and one of the most famous ocean currents in the world. Changes in the Kuroshio have an important impact on the ecology, climate and environment of the surrounding areas, especially changes in its flow and path, which best reflect the main characteristics of the Kuroshio. Therefore, accurate prediction of the Kuroshio flow and path has important scientific significance and practical application value.

[0003] At present, the Kuroshio forecast mainly relies on two methods: numerical prediction model and deep learning model. The numerical prediction model simulates the changes of the Kuroshio by solving complex ocean dynamics equations. However, since the Kuroshio is significantly affected by the nonlinear interaction of multi-scale physical processes, a high-resolution numerical model combined with a large number of ensemble members is required to characterize the uncertainty of the forecast. Although this method has high accuracy, it has a huge computational cost and takes a long time, making it difficult to meet the needs of real-time business forecasting.

[0004] Another method is the Kuroshio forecast based on a deep learning model. Compared with numerical models, deep learning models have the advantages of low computational cost and fast reasoning speed, but they still have many limitations in practical applications. For example, a patent application with application number 202311075315.8 discloses a Kuroshio path intelligent forecasting method based on deep learning, including Kuroshio path data acquisition, Kuroshio path forecasting model construction, Kuroshio path forecasting model training, and Kuroshio path forecasting. A patent application with application number 202411180635.4 discloses a Kuroshio path forecasting system based on multi-source data fusion, in which the data set to be calculated is input into the path forecasting model for prediction to generate a prediction result. Both methods realize the Kuroshio path forecast by performing modal decomposition of the sea surface height, but do not consider the key three-dimensional ocean dynamic process. In addition, the patent application with application number 202311076098.4 discloses a Kuroshio path prediction method based on complex empirical orthogonal functions, which mainly relies on sea surface satellite data, buoys and ship observation data. However, due to the sparseness of deep observation data and the lack of full consideration of the impact of deep flow fields on Kuroshio forecasts, the prediction accuracy of the deep learning model is limited. More importantly, as a complex ocean current system, the changes of the Kuroshio involve the interaction of multiple factors, and most existing methods only focus on the surface Kuroshio path prediction, lacking the ability to predict the Kuroshio current flow and three-dimensional path. Summary of the invention

[0005] The purpose of the present invention is to overcome the deficiencies in the prior art and to provide a method and system for intelligent prediction of the Kuroshio Current based on three-dimensional ocean environmental information, so as to overcome the limitations of current Kuroshio Current flow and path prediction and achieve higher prediction accuracy.

[0006] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0007] In a first aspect, the present invention provides a method for intelligent prediction of Kuroshio based on three-dimensional ocean environment information, comprising:

[0008] Obtain the three-dimensional ocean environment field;

[0009] Preprocess the ocean three-dimensional environment field to obtain the ocean three-dimensional feature vector;

[0010] Input the ocean three-dimensional feature vector into the trained Kuroshio prediction model to obtain the future ocean three-dimensional environmental field;

[0011] The Kuroshio prediction model is built based on the Swin Transformer model as the backbone network. Its input is the three-dimensional feature vector of the ocean and its output is the future three-dimensional environmental field of the ocean.

[0012] Furthermore, the three-dimensional ocean environment field includes sea surface height, temperature, salinity, and latitudinal and longitudinal flow velocities.

[0013] Furthermore, the method for preprocessing the ocean three-dimensional environment field includes:

[0014] First, standardize the ocean three-dimensional environmental field;

[0015] And the missing values ​​of land areas in the standardized data are replaced by 0.

[0016] Furthermore, the training method of the Kuroshio prediction model includes:

[0017] Obtain the ocean 3D environment field through multiple channels and pre-process the data:

[0018] According to the preprocessed ocean 3D environment field, a large ocean model training data set is produced:

[0019] The Kuroshio prediction model is trained according to the training data set, and the parameters of the model are updated to obtain the trained Kuroshio prediction model.

[0020] Furthermore, the three-dimensional ocean environment field is obtained through multiple channels, including:

[0021] Download GLORY (GlobalOcean Physics Reanalysis) and HYCOM high-resolution ocean reanalysis data from multiple operational agency websites: https: / / data.marine.copernicus.eu / product / GLOBAL_MULTIYEAR_PHY_001_030 / download?dataset=cmems_mod_glo_phy_my_0.083deg_P1D-m_202311 and https: / / www.hycom.org / dataserver / gofs-3pt1 / reanalysis;

[0022] Among them, the data time range is from 1993 to 2021, the spatial range is 12°N-36°N, 115°E-140°E, the depth range is from the surface to 1200 meters, and variables include temperature, salinity, longitude and latitude current velocity and sea surface height data.

[0023] Furthermore, based on the preprocessed ocean three-dimensional environment field, a large ocean model training data set is produced, including:

[0024] First, all preprocessed data are divided into training, validation and test sets. The training set is from 1993 to 2017, the validation set is from 2018, and the test set is from 2019 to 2020.

[0025] After that, the initial field and the data in the forecast duration range are paired, and the data is saved as input and label paired training data according to the date according to the forecast duration. The three-dimensional ocean environment field at the forecast time is used as input, and the three-dimensional ocean environment field in the next 10 days is used as the label. After combination, it is saved in npy, hdf5 or nc format.

[0026] Further, the Kuroshio prediction model is trained according to the training data set, and the parameters of the model are updated to obtain the trained Kuroshio prediction model, including:

[0027] When training the Kuroshio prediction model based on the training data set, AdamW is used as the optimizer, MSE is used as the loss function, and the optimization system is built through the Pytorch framework to optimize the model weights. The parameters that obtain the minimum error on the validation set are determined as the optimal weights.

[0028] The training process is iterated 200 times in total. If the loss function does not decrease for 20 consecutive times, the training process is terminated early.

[0029] Furthermore, an autoregressive training strategy is used when training the Kuroshio prediction model; the autoregressive training strategy includes:

[0030] The training process is divided into 6 steps. First, the forecast period is set to 1 day, the model is trained to obtain a set of optimal model parameters, and then the forecast period is gradually extended to 2, 4, 6, 8 and 10 days, and the model is fine-tuned.

[0031] Furthermore, the loss function includes: the error of the predicted environmental field, the error of the flow rate and path position, and the error of the curl calculated from the longitude and latitude flow velocity.

[0032] Furthermore, the loss function is:

[0033]

[0034]

[0035]

[0036]

[0037]

[0038] in, , , , Respectively represent the number of longitude grid points, latitudinal grid points, channels and prediction duration, and are the predicted and observed ocean environment fields, and is the predicted and observed Kuroshio flow, and are the predicted and observed Kuroshio path indices, and are the predicted and observed curl fields, respectively.

[0039] Furthermore, the model consists of an encoding embedding module, a transformer module and an output module.

[0040] The encoding part first encodes the ocean year, month, day and seabed topography information based on sine and cosine functions and then superimposes it with the input ocean environment field.

[0041] Afterwards, the image segmentation part divides the input ocean three-dimensional environment field (such as sea surface height, temperature, salinity, etc.) into patches, each patch is a fixed-size area (for example, a 4×4 area), and then is embedded into a high-dimensional feature space through linear transformation.

[0042] Generate a vector of fixed dimension. Assume the input data is ,in is the number of channels, and The number of grid points in the latitude and longitude directions respectively. It is divided into several patches by block division. ,in is the patch size. After linear embedding, the feature vector of the patch is obtained ,in is the embedding dimension.

[0043] The core part of the transformer module is the windowed multi-head self-attention and window offset mechanism. The multi-head self-attention mechanism divides the image into windows of fixed size, performs local self-attention calculation in each window, and the window offset mechanism can effectively avoid over-reliance on local information. This module receives the data output by the encoding embedding module and transmits it to the transformer module. In the transformer module, the data is processed by the windowed multi-head self-attention and window offset mechanisms.

[0044] The output module includes a multi-layer perceptron, which processes the features from the transformer module and finally outputs the ocean environment field, including the three-dimensional ocean temperature, salinity, current velocity and sea surface height data for the next day.

[0045] Furthermore, the three-dimensional ocean environment field is obtained, including:

[0046] The latest ocean data is downloaded from various websites regularly, and after pre-processing, the data is input into the ocean intelligent forecasting model to forecast the future ocean environment.

[0047] Furthermore, the method also includes: extracting the Kuroshio path and calculating the Kuroshio flow from the three-dimensional ocean environment field predicted by the model; the method of extracting the Kuroshio path and calculating the Kuroshio flow includes:

[0048] The Kuroshio flow is calculated by integrating the flow velocity on each section. The surface Kuroshio path is defined by extracting the 0.7 m sea surface height contour, and the deep Kuroshio path is defined by the grid points with the largest flow velocity.

[0049] Furthermore, the method also includes:

[0050] The model prediction results as well as the black current flow and path are pushed through the website and WeChat public account.

[0051] Furthermore, the model prediction results and the black current flow and path are pushed through the website and WeChat public account, including:

[0052] After obtaining the forecast results, the forecast results are automatically saved locally with a specific file name, and the program is set to automatically read the pictures in the current path and upload all the pictures to the website and WeChat public account. In addition, an early warning signal can be sent according to the black current flow and path status.

[0053] In a second aspect, the present invention provides a Kuroshio intelligent forecasting system based on three-dimensional ocean environmental information, including a processor and a storage medium;

[0054] The storage medium is used to store instructions;

[0055] The processor is used to operate according to the instructions to execute the steps of the method according to the first aspect.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] 1. This method is the first to use the three-dimensional ocean environment information to predict the Kuroshio flow, and the loss function takes into account the environmental field, Kuroshio and curl prediction errors at the same time. It can provide the ocean environmental field related to the Kuroshio, and can also consider the changes in the Kuroshio flow and three-dimensional path from an overall perspective, and has achieved good prediction results.

[0058] 2. This method constructs a Kuroshio forecasting system that integrates forecasting and early warning, which can perform multi-GPU inference locally and transmit the forecast results to the website and WeChat public account in real time.

[0059] 3. The present invention proposes an intelligent Kuroshio forecasting method based on three-dimensional ocean environmental information. The system makes full use of three-dimensional ocean environmental information, combined with a deep learning model, can comprehensively consider the ocean dynamic process, and provide high-precision Kuroshio flow and path forecasts. At the same time, the system supports real-time push of forecast results through websites and WeChat public accounts, etc., realizing the intelligence and business of Kuroshio forecasting, and providing strong technical support for scientific research and practical applications in related fields. The present invention has important application value and practical significance for marine economy and disaster prevention and mitigation. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is a flow chart of the Kuroshio intelligent forecasting method based on the three-dimensional ocean environment information of the present invention;

[0061] Figure 2 It is a model architecture diagram of the Kuroshio intelligent forecasting method based on the three-dimensional ocean environment information of the present invention. DETAILED DESCRIPTION

[0062] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0063] Embodiment 1:

[0064] like Figure 1 As shown, this embodiment 1 provides a Kuroshio intelligent forecasting method based on three-dimensional ocean environmental information, including:

[0065] S001: Obtain the three-dimensional ocean environment field through multiple channels and pre-process the data: Download the three-dimensional ocean reanalysis data from multiple business agency websites, including ocean maps and temperature, salinity, flow field and sea surface height data at different depths of each point on the ocean map. Then standardize the data and fill in missing values.

[0066] S002: Create a large ocean model training data set: According to different forecast durations, organize the data into pairs (x, y), where x represents the input data and y represents the label data of different forecast durations.

[0067] S003: Construct a large Kuroshio prediction model and obtain the parameters of the optimal forecast model: Optimize the Kuroshio intelligent forecast model constructed through the training data set and obtain the optimal model parameters.

[0068] S004: Generate future ocean environment field and Kuroshio forecast results; download the ocean data of the day as model input, perform three-dimensional ocean environment field forecast, and obtain Kuroshio flow and path forecast.

[0069] S005: Push through the website and WeChat public account: Save the forecast results in the form of pictures and then push them through various methods.

[0070] Embodiment 2:

[0071] Based on the same inventive concept as in Example 1, this embodiment further discloses a Kuroshio intelligent forecasting method based on three-dimensional ocean environmental information, including:

[0072] S001: Obtain the three-dimensional ocean environment field through multiple channels and pre-process the data: Download the three-dimensional ocean reanalysis data from multiple business agency websites, including temperature, salinity, flow field and sea surface height data at different depths. Then standardize the data and fill in missing values.

[0073] In step S001, multi-channel acquisition of the ocean three-dimensional environmental field refers to downloading GLORY (Global Ocean Physics Reanalysis) and HYCOM (Hybrid Coordinate Ocean Model) high-resolution ocean reanalysis data. The download URLs are https: / / data.marine.copernicus.eu / product / GLOBAL_MULTIYEAR_PHY_001_030 / download?dataset=cmems_mod_glo_phy_my_0.083deg_P1D-m_202311 and https: / / www.hycom.org / dataserver / gofs-3pt1 / reanalysis. The data time range is from 1993 to 2021, the spatial range is 12°N-36°N, 115°E-140°E, the depth range is from the surface to 1200 meters, and the variables include temperature, salinity, latitude and longitude current velocity, and sea surface height data. The spatial resolution can be adaptive and is not limited to a specific resolution.

[0074] In step S001, the data is preprocessed by first performing a standardization operation on the data, such as the calculation formula is:

[0075]

[0076] In formula (1) Indicates the data that needs to be standardized. and They are obtained by calculating the mean and standard deviation of each grid point at different depths for different variables. After that, the nan (missing data) values ​​in the land area in the standardized data are replaced by 0.

[0077] S002: Create a large ocean model training data set: According to different forecast durations, organize the data into pairs (x, y), where x represents the input data and y represents the label data of different forecast durations.

[0078] In step S002, a large ocean model training data set is produced. First, all standardized data are divided into training, validation and test sets. Among them, the training set is 1993-2017, the validation set is 2018, and the test set is 2019-2020. After that, the data in the initial field and the forecast duration range are paired, and the data is saved as input and label paired training data according to the date according to the forecast duration. Specifically, the three-dimensional ocean environment field at the forecast time is used as input, and the three-dimensional ocean environment field in the next 10 days is used as a label, and the combination is saved in npy, hdf5 or nc format.

[0079] S003: Construct a large Kuroshio prediction model and obtain the parameters of the optimal forecast model: Optimize the Kuroshio intelligent forecast model constructed through the training data set and obtain the optimal model parameters.

[0080] In step S003, the specific steps of constructing the Kuroshio prediction model are as follows:

[0081] like Figure 2 As shown in the figure, the Kuroshio prediction model uses the Swin Transformer model as the backbone network. The model consists of an encoding embedding module (that is, Figure 2 The encoding part first encodes the ocean year, month, day and seabed topography information based on sine and cosine functions and then superimposes it with the input ocean environment field to obtain the superimposed ocean environment field. After that, the image segmentation part divides the superimposed ocean environment field into patches, each patch is a fixed-size area (for example, a 4×4 area), and then embeds it into a high-dimensional feature space through linear transformation to generate a fixed-dimensional vector. Assume that the input data is ,in is the number of channels, and The number of grid points in the latitude and longitude directions respectively. It is divided into several patches by block division. ,in is the patch size. After linear embedding, the feature vector of the patch is obtained ,in is the embedding dimension.

[0082] The core parts of the transformer module are windowed multi-head self-attention and multi-layer perceptron. The multi-head self-attention mechanism divides the image into fixed-size windows, performs local self-attention calculations in each window, and the multi-layer perceptron maps the array to the feature space. In addition, each part is normalized to help the model better learn data features.

[0083] The output module includes a multi-layer perceptron, which processes the features from the transformer module and finally outputs the ocean environment field, including the three-dimensional ocean temperature, salinity, current velocity and sea surface height data for the next day.

[0084] In this embodiment, the output of the trained Kuroshio prediction model is the ocean three-dimensional environment field for the next day. If the ocean three-dimensional environment field for the Nth day needs to be predicted, the model prediction result is re-input into the trained Kuroshio prediction model and repeated N times. During the training process, the training process is divided into multiple steps. First, the forecast time is set to 1 day, the model is trained, and a set of optimal model parameters are obtained. Then, the forecast time is gradually extended to 2, 4, 6, 8 and 10 days, and the model is fine-tuned to obtain a trained Kuroshio prediction model.

[0085] After the ocean three-dimensional environment field (such as sea surface height, temperature, salinity and current velocity) is input into the model, the ocean year, month, day and seabed topography information are encoded based on sine and cosine functions and fused with the input data. Then, patch division is performed, and each patch is a fixed-size area (for example, a 4×4 area). Each patch is embedded into the high-dimensional feature space through linear transformation to generate a fixed-dimensional vector. Assume that the input data is ,in is the number of channels, and The number of grid points in the latitude and longitude directions respectively. It is divided into several patches by block division. ,in is the patch size. After linear embedding, the feature vector of the patch is obtained ,in is the embedding dimension.

[0086] During the model training process, the optimizer uses AdamW and the loss function uses MSE. Then, the optimization system is built through the Pytorch framework to optimize the model weights, and the parameters that obtain the minimum error on the validation set are determined as the optimal weights. The loss function consists of three errors: 1) the error of the predicted environmental field; 2) the error of the flow and path position; 3) the curl error calculated from the longitude and latitude flow velocity, the formula is ,in, is the zonal velocity, is the meridional velocity. The training process was iterated 200 times in total. To prevent overfitting, the training process was terminated early if the loss function did not decrease for 20 consecutive times. The Kuroshio prediction model adopts an autoregressive training strategy. Specifically, we divided the training process into 6 steps. First, the forecast duration was set to 1 day to obtain a set of optimal model parameters. On this basis, the forecast duration was gradually extended to 2, 4, 6, 8 and 10 days, and the model was fine-tuned.

[0087] S004: Generate future ocean environment field and Kuroshio forecast results; download the ocean data of the day as model input, perform three-dimensional ocean environment field forecast, and obtain Kuroshio flow and path forecast.

[0088] In step S004, generating the future ocean environment field and Kuroshio forecast results means downloading the latest ocean data from a website, preprocessing the data, and inputting the ocean intelligent forecast model to forecast the future ocean environment field. Among them, Kuroshio forecast refers to extracting the Kuroshio path and calculating the Kuroshio flow from the three-dimensional ocean environment field predicted by the model. Specifically, the Kuroshio flow is determined by integrating the flow velocity on each section, the surface Kuroshio path can be defined by extracting the 0.7m contour line of the sea surface height, and the deep Kuroshio path can be defined by the grid point with the largest flow velocity.

[0089] S005: Push through the website and WeChat public account: Save the forecast results in the form of pictures and then push them through various methods.

[0090] In step S005, the specific steps of pushing through the website and WeChat public account. After obtaining the forecast result of Example 6, the system first automatically saves the forecast result to local storage in a specific file name format. Subsequently, the program automatically reads the relevant image files in the current directory according to the file path, including forecast images, Kuroshio path change images, flow change images, etc., to ensure that all forecast image files can be efficiently and accurately uploaded to the specified platform.

[0091] In order to ensure that the forecast results can be widely disseminated and monitored in real time, the system has an embedded automated script that regularly checks the updated files in the file path according to the preset time interval and automatically uploads the updated content to the public website platform and WeChat public account specified by the user. By connecting to the network platform API, the program can complete the upload process without manual intervention.

[0092] In addition, in order to improve the visibility of forecast information and the user's understanding, the system automatically generates visual charts, which include not only the temporal and spatial changes of the Kuroshio path and flow, but also forecast data charts and error analysis charts of the three-dimensional ocean environment. In order to facilitate further analysis by users, the system will also generate detailed reports including data analysis results, forecast accuracy, model evaluation reports, etc.

[0093] To enhance the system's intelligent early warning function, the system monitors the changes in the black current flow and path in real time, and determines whether there are abnormal fluctuations or extreme changes through set thresholds. When the black current flow predicted by the model reaches the set danger threshold, or the path changes dramatically, the system will immediately issue an early warning signal. These early warning information will be pushed to relevant users via email, text messages, and WeChat public accounts to ensure that relevant departments can take necessary emergency measures at the first time.

[0094] Embodiment three:

[0095] This embodiment provides a Kuroshio intelligent forecasting system based on three-dimensional ocean environmental information, including a processor and a storage medium;

[0096] The storage medium is used to store instructions;

[0097] The processor is used to operate according to the instructions to execute the steps of the method according to the first embodiment.

[0098] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0099] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0100] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0101] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0102] The above are only preferred embodiments of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A Kuroshio intelligent forecasting method based on three-dimensional ocean environmental information, characterized in that: include: Obtain the three-dimensional ocean environment field and its time to be predicted; Preprocessing the ocean three-dimensional environmental field to obtain a preprocessed ocean three-dimensional environmental field; The preprocessed ocean three-dimensional environmental field and its time are input into the trained Kuroshio forecast model to obtain the future ocean three-dimensional environmental field; The Kuroshio forecast model is built based on the Swin Transformer model as the backbone network. Its input is the three-dimensional ocean environment field and its time, and its output is the future three-dimensional ocean environment field.

2. The Kuroshio intelligent forecasting method based on ocean three-dimensional environmental information according to claim 1 is characterized in that: The three-dimensional ocean environment field includes sea surface height, temperature, salinity, and latitudinal and longitudinal current velocities.

3. The Kuroshio intelligent forecasting method based on ocean three-dimensional environmental information according to claim 1 is characterized in that: The methods for preprocessing the ocean three-dimensional environment field include: First, standardize the ocean three-dimensional environmental field; And the missing values ​​of land areas in the standardized data are replaced by 0.

4. The Kuroshio intelligent forecasting method based on ocean three-dimensional environmental information according to claim 1 is characterized in that: The training method of the Kuroshio forecast large model includes: Obtain the ocean three-dimensional environment field and its time in multiple ways and perform preprocessing: According to the preprocessed ocean 3D environment field, a large ocean model training data set is produced: The Kuroshio forecast model is trained according to the training data set, and the parameters of the model are updated to obtain the trained Kuroshio forecast model.

5. The Kuroshio intelligent forecasting method based on ocean three-dimensional environmental information according to claim 4 is characterized in that: Multiple ways to obtain the ocean 3D environment, including: Download GLORY and HYCOM ocean reanalysis data from multiple operational agency websites; Among them, the time range of the data is from 1993 to 2021, the spatial range is 12°N-36°N, 115°E-140°E, and the depth range is from the surface to 1200 meters.

6. The method for intelligent Kuroshio forecasting based on three-dimensional ocean environment information according to claim 5 is characterized in that: According to the preprocessed ocean 3D environment field, a large ocean model training data set is produced, including: The data in the initial field and the forecast time range are paired, with the three-dimensional ocean environment field at the forecast time as input and the three-dimensional ocean environment field 10 days in the future as the label. After combination, they are saved as a training data set with paired input and label, and divided into training set, test set and validation set.

7. The method for intelligent prediction of Kuroshio based on three-dimensional ocean environment information according to claim 6, characterized in that: The Kuroshio forecast model is trained according to the training data set, and the parameters of the model are updated to obtain the trained Kuroshio forecast model, including: The Kuroshio prediction model was trained based on the training set and the test set. The optimizer used AdamW and the loss function used MSE. The optimization system was built through the Pytorch framework to optimize the model weights. The parameters that obtained the minimum error on the validation set were determined as the optimal weights. The training process is iterated 200 times in total. If the loss function does not decrease for 20 consecutive times, the training process is terminated early. The autoregressive training strategy is used when training the Kuroshio forecast model; the autoregressive training strategy includes: The training process is divided into multiple steps. First, the forecast period is set to 1 day, and the model is trained to obtain a set of optimal model parameters. Then, the forecast period is gradually extended to 2, 4, 6, 8 and 10 days, and the model is fine-tuned to obtain a trained Kuroshio forecast model. The loss function includes the error of the predicted environmental field, the error of the flow rate and the path position, and the curl error calculated from the longitude and latitude flow velocity; The loss function is expressed as: ； ； ； ； ； Among them, MSE1 represents the forecast environmental field error, MSE2 represents the flow error, MSE3 represents the path position error, and MSE4 represents the curl error calculated from the longitude and latitude flow velocity. , , , Respectively represent the number of longitude grid points, latitudinal grid points, channels and forecast duration, and are the predicted and observed ocean environment fields, and is the predicted and observed Kuroshio flow, and are the predicted and observed Kuroshio path indices, and are the predicted and observed curl fields, respectively.

8. The method for intelligent prediction of Kuroshio based on three-dimensional ocean environment information according to claim 1, characterized in that: The Kuroshio forecast model includes: Coding and embedding module: used to encode the moment of the ocean three-dimensional environment field to be predicted based on sine and cosine functions and superimpose it with the input ocean three-dimensional environment field, and divide the superimposed ocean three-dimensional environment field into regions of fixed size, and then embed it into the high-dimensional feature space through linear transformation to generate a vector of fixed dimension; Transformer module: used to divide the vector into fixed-size windows through the multi-head attention mechanism, perform local self-attention calculation in each window, and map the array to the feature space through the multi-layer perceptron to obtain the features; Output module: used to process the features from the transformer module through a multi-layer perceptron, and finally output the three-dimensional ocean environment field at the future moment, including the future three-dimensional ocean temperature, salinity, flow rate and sea surface height data.

9. The method for intelligent prediction of Kuroshio based on three-dimensional ocean environment information according to claim 1, characterized in that: The method further comprises: extracting the Kuroshio path and calculating the Kuroshio flow from the three-dimensional ocean environment field predicted by the Kuroshio prediction large model; The method of extracting the Kuroshio path and calculating the Kuroshio flow from the three-dimensional ocean environment field predicted by the Kuroshio forecast model includes: The Kuroshio flow is calculated by integrating the flow velocity on each section. The surface Kuroshio path is defined by extracting the 0.7 m sea surface height contour, and the deep Kuroshio path is defined by the grid points with the largest flow velocity.

10. An intelligent Kuroshio forecasting system based on three-dimensional ocean environmental information, characterized in that: including processor and storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1-9.

Images