Ocean data prediction method and system based on VIT module

By employing an artificial intelligence approach based on the VIT module, the problems of accuracy and real-time performance in ocean surface height and eddy forecasting have been solved, enabling more accurate and faster ocean data prediction and adapting to the needs of changing marine environments.

CN118673335BActive Publication Date: 2024-11-15青岛国实科技集团有限公司
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411161617.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-23
Publication Date
2024-11-15
Estimated Expiration
2044-08-23

AI Technical Summary

Technical Problem

Existing methods for predicting ocean surface height and eddies are insufficient in terms of accuracy, real-time performance, and adaptability, and cannot meet the decision-making needs of marine-related industries.

Method used

By employing an artificial intelligence approach based on the VIT module, a sea surface height prediction model is constructed by acquiring sea surface height data. Combined with a vortex prediction model, this enables accurate prediction of future ocean data.

Benefits of technology

It improves the accuracy and real-time performance of long-term trend forecasts, enables rapid output of forecast results, identifies changes in the marine environment, provides more refined forecast results and better spatial resolution, and reduces the demand for computing resources.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118673335B_ABST
    Figure CN118673335B_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of ocean forecasting, and specifically discloses an ocean data forecasting method and system based on a VIT module, comprising: a data acquisition step: acquiring existing sea surface height data within a first set time; a future sea surface height data acquisition step: obtaining future sea surface height data within a second set time in the future through a sea surface height prediction model based on the VIT module based on the existing sea surface height data; a vortex data acquisition step: obtaining vortex data corresponding to the future sea surface height data through a vortex prediction model based on the future sea surface height data; wherein the future sea surface height data includes a sea surface height value, and the vortex data includes a vortex position, a vortex radius and a vortex rotation direction. The invention improves the speed and accuracy of sea surface height forecasting, and reduces the forecasting cost at the same time.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of ocean forecasting technology, and in particular to an ocean data prediction method and system based on a VIT module. Background Art

[0002] The current mainstream ocean surface height forecast and eddy forecast are based on numerical models. However, in practice, this model has great limitations on the accuracy, real-time and adaptability of the forecast, and cannot provide stronger support for decision-making and applications in ocean-related industries and fields.

[0003] Therefore, in order to solve the above defects, it is urgent to develop an ocean data prediction method and system based on the VIT module using artificial intelligence to overcome the above defects. Summary of the invention

[0004] In view of the above problems, the present invention provides an ocean data prediction method based on a VIT module, which includes:

[0005] Data acquisition step: acquiring existing sea surface height data within a first set time;

[0006] The step of obtaining future sea surface height data is as follows: based on the existing sea surface height data, obtaining future sea surface height data within a second set time in the future through a sea surface height prediction model based on a VIT module;

[0007] Vortex data obtaining step: obtaining vortex data corresponding to the future sea surface height data through a vortex prediction model based on the future sea surface height data;

[0008] The future sea surface height data includes the sea surface height value, and the vortex data includes the vortex position, vortex radius and vortex rotation direction.

[0009] In the above-mentioned ocean data prediction method, the step of obtaining future sea surface height data comprises:

[0010] Steps for obtaining the time series data set: obtain the sea surface height remote sensing data from the operational satellite, and process the data to form a time series data set;

[0011] Model building and obtaining step: after building an initial sea surface height prediction model based on the VIT module, the initial sea surface height prediction model is trained and tested through the time series data set to obtain the final sea surface height prediction model;

[0012] Future sea surface height data prediction step: obtaining the future sea surface height data of the i+1th day through the sea surface height prediction model according to the existing sea surface height data of the i-th day and the n days before the i-th day;

[0013] Among them, i is the current date, and n is a positive integer.

[0014] In the above-mentioned ocean data prediction method, the step of obtaining the time series data set includes:

[0015] Obtaining daily average sea surface height remote sensing data;

[0016] After the sea surface height remote sensing data are sorted according to a set order, normalization processing is performed on the sea surface height remote sensing data;

[0017] Dividing the normalized sea surface height remote sensing data into training data and true values ​​corresponding to the training data;

[0018] The time series data set is constructed using the training data and the true value.

[0019] In the above-mentioned ocean data prediction method, the model building and obtaining step comprises:

[0020] Building a block embedding layer, which divides the input data into n×n small blocks, maps each small block into a one-dimensional embedding vector, and encodes the position information into the embedding vector;

[0021] Constructing an encoder layer, the encoder layer comprising a plurality of first VIT module units and block aggregation units, the first VIT module units and the block aggregation units being alternately arranged, and reducing the length and width of each small block and increasing the number of feature channels through the encoder layer until a first target is reached and then stopping;

[0022] Constructing a decoder layer, the decoder layer comprising a plurality of second VIT module units and block expansion units, the second VIT module units and the block expansion units being alternately arranged, and increasing the length and width of each small block and reducing the number of feature channels through the decoder layer until a second target is reached and then stopping;

[0023] Construct an MLP linear layer and obtain the prediction result by predicting each small block;

[0024] The initial sea surface height prediction model is trained and tested by using the time series data set to obtain the final sea surface height prediction model.

[0025] In the above-mentioned ocean data prediction method, each of the first VIT module unit and the second VIT module unit includes at least one VIT module.

[0026] In the above-mentioned ocean data prediction method, the vortex data acquisition step comprises:

[0027] Filtering step: filtering the future sea surface height data;

[0028] Vortex data prediction step: obtaining vortex data corresponding to the future sea surface height data through a vortex prediction model based on the future sea surface height data after filtering.

[0029] The present invention also provides an ocean data prediction system based on a VIT module, which includes:

[0030] A data acquisition unit, which acquires existing sea surface height data within a first set time;

[0031] A future sea surface height data obtaining unit, which obtains future sea surface height data within a second set time in the future based on the existing sea surface height data through a sea surface height prediction model based on a VIT module;

[0032] a vortex data obtaining unit, which obtains vortex data corresponding to the future sea surface height data through a vortex prediction model based on the future sea surface height data;

[0033] The future sea surface height data includes the sea surface height value, and the vortex data includes the vortex position, vortex radius and vortex rotation direction.

[0034] In the above-mentioned ocean data prediction system, the future sea surface height data obtaining unit comprises:

[0035] The time series data set acquisition module acquires the sea surface height remote sensing data from the operational satellite and processes the data to form a time series data set;

[0036] A model building module, after building an initial sea surface height prediction model based on the VIT module, trains and tests the initial sea surface height prediction model through the time series data set to obtain the final sea surface height prediction model;

[0037] A future sea surface height data prediction module, which obtains the future sea surface height data of the i+1th day through the sea surface height prediction model according to the existing sea surface height data of the i-th day and the n days before the i-th day;

[0038] Among them, i is the current date, and n is a positive integer.

[0039] In the above-mentioned ocean data prediction system, the model building and obtaining module includes:

[0040] Building a block embedding layer, which divides the input data into n×n small blocks, maps each small block into a one-dimensional embedding vector, and encodes the position information into the embedding vector;

[0041] Constructing an encoder layer, the encoder layer comprising a plurality of first VIT module units and block aggregation units, the first VIT module units and the block aggregation units being alternately arranged, and reducing the length and width of each small block and increasing the number of feature channels through the encoder layer until a first target is reached and then stopping;

[0042] Constructing a decoder layer, the decoder layer comprising a plurality of second VIT module units and block expansion units, the second VIT module units and the block expansion units being alternately arranged, and increasing the length and width of each small block and reducing the number of feature channels through the decoder layer until a second target is reached and then stopping;

[0043] Construct an MLP linear layer and obtain the prediction result by predicting each small block;

[0044] The initial sea surface height prediction model is trained and tested by using the time series data set to obtain the final sea surface height prediction model.

[0045] In the above-mentioned ocean data prediction system, the vortex data acquisition unit comprises:

[0046] A filtering module, for filtering the future sea surface height data;

[0047] The vortex data prediction module obtains vortex data corresponding to the future sea surface height data through a vortex prediction model based on the future sea surface height data after filtering.

[0048] Compared with the traditional numerical model ocean surface height forecast and vortex forecast, the present invention has the following advantages:

[0049] 1. Accurate long-term trend prediction: Artificial intelligence technology helps to make accurate long-term trend predictions and provide more accurate predictions of marine climate change and the evolution of marine ecosystems.

[0050] 2. Guarantee of real-time forecasting: Compared with traditional numerical models, which re-run complex model calculations for each forecast, AI forecasting can output forecast results more quickly. Since the parameters of the model have been trained, only new observation data needs to be input for inference, which greatly reduces the calculation time of the forecast. This real-time performance enables AI forecasting to capture changes and evolutions in the marine environment in a timely manner and quickly provide accurate forecast information.

[0051] 3. Artificial intelligence ocean forecast products can identify potential laws and patterns by learning and analyzing a large amount of historical observation data, thereby providing more accurate forecast results. Deep learning algorithms have excellent performance in ocean data processing and pattern recognition, and can automatically extract features, discover hidden associations and nonlinear relationships, thereby improving the accuracy and reliability of forecasts.

[0052] 4. AI ocean forecast products also have better spatial resolution and customizability. Traditional numerical models often use larger grid spacing due to limitations in computing efficiency and computing resources, resulting in limited spatial resolution of forecast results. AI-based methods can flexibly process data of different resolutions and can customize forecasts according to user needs, providing more refined forecast results.

[0053] Other features and advantages of the present invention will be described in the following description, and partly become obvious from the description, or be understood by implementing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0055] Figure 1 is a flow chart of the ocean data prediction method of the present invention;

[0056] Figure 2 for Figure 1 The step-by-step flow chart of step S2;

[0057] Figure 3 for Figure 1 A step-by-step flow chart of step S3;

[0058] Figure 4 It is a structural schematic diagram of the ocean data prediction system of the present invention;

[0059] Figure 5 It is a structural schematic diagram of the sea surface height prediction model of the present invention. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] The exemplary embodiments and descriptions of the present invention are used to explain the present invention, but are not intended to limit the present invention. In addition, elements / components with the same or similar reference numerals used in the drawings and embodiments are used to represent the same or similar parts.

[0062] The terms “first”, “second”, “S1”, “S2”, etc. used in this document do not specifically refer to an order or sequence, nor are they used to limit the present invention. They are only used to distinguish elements or operations described with the same technical terms.

[0063] The directional terms used in this document, such as up, down, left, right, front or back, etc., are only used to refer to the directions of the drawings. Therefore, the directional terms used are used to illustrate and not to limit the present invention.

[0064] The words “include,” “including,” “have,” “contain,” etc. used in this document are open-ended terms, meaning including but not limited to.

[0065] As used herein, "and / or" includes any or all combinations of the items described.

[0066] As used herein, “plurality” includes “two” and “more than two”; as used herein, “plurality groups” includes “two groups” and “more than two groups”.

[0067] Certain terms used to describe the present application are discussed below or elsewhere in this specification to provide additional guidance to those skilled in the art regarding the description of the present application.

[0068] See also Figure 1 , Figure 1 FIG. 1 is a flow chart of the ocean data prediction method of the present invention. Figure 1 As shown, a method for predicting ocean data based on a VIT module of the present invention includes:

[0069] Data acquisition step S1: acquiring existing sea surface height data within a first set time;

[0070] Future sea surface height data acquisition step S2: based on the existing sea surface height data, obtaining the future sea surface height data within a second set time in the future through a sea surface height prediction model based on the VIT module;

[0071] Vortex data acquisition step S3: based on the future sea surface height data, obtaining vortex data corresponding to the future sea surface height data through a vortex prediction model;

[0072] The future sea surface height data includes the sea surface height value, and the vortex data includes the vortex position, vortex radius and vortex rotation direction.

[0073] It should be noted that the first set time of the present invention can be a certain day or a certain period of time. Specifically, the certain day can be the current day or a certain day before the current day, and the certain period of time can be several consecutive or discontinuous days including the current day and the days before the current day, or several consecutive or discontinuous days before the current day that do not include the current day. In this embodiment, the first set time including the current day and several consecutive days before the current day is taken as an example for explanation, but the present invention is not limited to this.

[0074] It should be noted that the second set time of the present invention can be a certain day or a certain period of time after the current day. Specifically, the certain day can be any day after the current day, and the certain period of time can be several consecutive or discontinuous days after the current day. In this embodiment, the second set time includes the day after the current day as an example for explanation, but the present invention is not limited to this.

[0075] Please refer to Figure 2 and Figure 5 , Figure 2 for Figure 1 The step-by-step flow chart of step S2; Figure 5 Schematic diagram of the structure of the sea surface height prediction model of the present invention. Figure 2 and Figure 5 Please combine Figure 1 The future sea surface height data acquisition step S2 comprises:

[0076] Time series data set acquisition step S21: acquiring sea surface height remote sensing data from an operational satellite, and processing the data to form a time series data set;

[0077] Model building and obtaining step S22: after building an initial sea surface height prediction model based on the VIT module, the initial sea surface height prediction model is trained and tested through the time series data set to obtain the final sea surface height prediction model;

[0078] Future sea surface height data prediction step S23: obtaining the future sea surface height data of the i+1th day through the sea surface height prediction model according to the existing sea surface height data of the i-th day and the n days before the i-th day;

[0079] Among them, i is the current date, and n is a positive integer.

[0080] The step S21 of obtaining the time series data set includes:

[0081] Obtaining daily average sea surface height remote sensing data;

[0082] After the sea surface height remote sensing data are sorted according to a set order, normalization processing is performed on the sea surface height remote sensing data;

[0083] Dividing the normalized sea surface height remote sensing data into training data and true values ​​corresponding to the training data;

[0084] The time series data set is constructed using the training data and the true value.

[0085] Specifically, we first need to obtain the operational satellite sea surface height remote sensing data and pre-process the data. This application uses the daily average sea surface height remote sensing data provided by Aviso. After sorting the obtained satellite data by date, the data is normalized. The formula is as follows:

[0086] ;

[0087] According to the formula, the range of all data values ​​can be compressed to between 0.1 and 1. This has the following advantages over 0-1 normalization:

[0088] Avoid values ​​that are too small: In the ocean prediction process, land areas are invalid areas and will be masked to distinguish them. This will cause the original 0-1 normalization to produce many zero values, which may cause problems in numerical calculations, such as numerical stability and accuracy. Compressing the value to 0.1-1 can remove these zero values.

[0089] Improved numerical stability: In some computations, such as backpropagation in deep learning, very small values ​​can lead to the vanishing gradient problem; squeezing to 0.1-1 can alleviate this problem.

[0090] Then the data is divided into training data and true value. The division method is, assuming that the current date is i, the data of the 9 days before i is concatenated with i to form a continuous ten-day sea surface height data; the data of i+1 day is taken as the true value. These eleven data constitute a data tuple required for model training. The ten-day continuous sea surface height data is the training data, and the i+1 day is the true value required to predict the training data. All the obtained data can be made into a time series data set for model training according to the above rules.

[0091] The model building and obtaining step S22 includes:

[0092] Building a block embedding layer, which divides the input data into n×n small blocks, maps each small block into a one-dimensional embedding vector, and encodes the position information into the embedding vector;

[0093] Constructing an encoder layer, the encoder layer comprising a plurality of first VIT module units and block aggregation units, the first VIT module units and the block aggregation units being alternately arranged, and reducing the length and width of each small block and increasing the number of feature channels through the encoder layer until a first target is reached and then stopping;

[0094] Constructing a decoder layer, the decoder layer comprising a plurality of second VIT module units and block expansion units, the second VIT module units and the block expansion units being alternately arranged, and increasing the length and width of each small block and reducing the number of feature channels through the decoder layer until a second target is reached and then stopping;

[0095] Construct an MLP linear layer and obtain the prediction result by predicting each small block;

[0096] The initial sea surface height prediction model is trained and tested by using the time series data set to obtain the final sea surface height prediction model.

[0097] Specifically, the data that needs to be processed for ocean prediction is spatiotemporal data, which are interrelated in both time and space dimensions. Therefore, this application uses the VIT module as the basic model and makes improvements on this basis. The specific network structure is as follows: Figure 5As shown. The workflow of the VIT structure is to first pre-process the input data, divide it into n×n small blocks, and map each small block into a one-dimensional embedding vector through a block embedding layer, and encode the position information into the embedding vector. After generating the embedding vector, you can enter the VIT module. First, the embedding vector will pass through the normalization module. Because this application is aimed at the data of sea surface height and considers spatial characteristics, batch normalization is selected as the normalization module. Then the normalized embedding vector is input into the attention module to analyze the spatiotemporal correlation of each small block and adjust the numerical value. In order to ensure the integrity of the features and the gradient does not disappear during the back propagation process, a residual module is added after the attention module. Repeat the normalization once, then change the vector dimension through MLP, and finally make a residual to complete a VIT process. Combine and stack multiple VIT modules to finally realize the function of sea surface height prediction. The VIT module can be used as an encoder and decoder, so you only need to encode the original data and then decode it. Through the constraints of the loss function, the prediction capability can be realized.

[0098] In the encoding stage, two layers of VIT modules are first passed through, and the feature dimension and resolution remain unchanged. Then, a downsampling operation is performed through the block aggregation layer, and the length and width of each block will become half of the original, and then the number of feature channels of each block will become twice the original. Repeat this encoding operation to make the length and width of the block become one-eighth of the original, and the number of feature channels of the block becomes 8 times the original. In the decoding stage, the operation process is exactly the opposite of the encoding. First, two layers of VIT modules are passed through, and the feature dimension and resolution remain unchanged. Then, an upsampling operation is performed through the block expansion layer, and the length and width of each block will become twice the original, and then the number of feature channels of each block will become half of the original. Repeat this decoding operation until each block is restored to the dimension at the time of input, and finally a MLP linear layer is used to predict the value of the sea surface height of each grid point. The loss function of the network is MSELoss, and the formula is as follows:

[0099] ;

[0100] Where N is the total amount of data, x i is the predicted value, y i is the true value.

[0101] It should be noted that each of the first VIT module unit and the second VIT module unit includes at least one VIT module.

[0102] The future sea surface height data prediction step S23 specifically includes:

[0103] After the network training is completed and the network parameters are fixed, it can be put into commercial operation. According to the settings of the data set and network structure, this network can use 10 days of sea surface height data to predict the sea surface height value for the next day. Based on this, the sea surface height forecast for the next seven days can be realized. The specific process is as follows:

[0104] Assume that the current date is i. First, the result of day i+1 is predicted based on the data from 9 consecutive days before i to day i. Then the data of day i+9 before i is removed and the data of day i+1 is added. In this way, the result of day i+2 can be predicted based on the data from 8 consecutive days before i to day i+1. Similarly, the prediction results of day i+1 to day i+7 can be obtained. Since the prediction of the results of the following days requires the previous prediction results, as the number of days predicted increases, the error is gradually accumulated, and the accuracy of the prediction results is gradually reduced.

[0105]

[0106] Subsequent tests have shown that the network can complete 7-day sea surface height predictions in seconds on a single-CPU device, and its calculation speed is faster than traditional numerical model calculations.

[0107] The two most concerned indicators in the field of ocean forecasting are timeliness and accuracy. Artificial intelligence methods have defeated traditional numerical models in these two indicators. Numerical models need to predict and solve a large number of partial differential equations. The computational complexity increases exponentially with the increase in the resolution of the forecast area. The calculation process is time-consuming and power-consuming. In most cases, supercomputers are required to complete the calculation. After completing model training, artificial intelligence methods only need to perform forward propagation in the model reasoning process, mainly matrix operations, which can greatly reduce the consumption of computing resources. This project does not require the use of supercomputers. The CPU of a home computer can support model operations, and the timeliness of calculations is greatly improved. At the same time, after testing, the forecast accuracy of artificial intelligence methods has exceeded that of traditional numerical models. Therefore, artificial intelligence forecasting is an economical and affordable forecasting method.

[0108] The technology in this application has achieved significant innovative breakthroughs in the fields of image processing and pattern recognition. Its core advantage lies in the use of the industry-leading Vision Transformer (VIT) module, which fully exploits the huge potential of this module as an encoder and decoder. Compared with the traditional convolutional neural network (CNN), the VIT module significantly expands the receptive field of the model through the self-attention mechanism, enabling it to capture a wider range of contextual information.

[0109] Specifically, the VIT module divides the input image into multiple small blocks, and then uses the multi-head self-attention mechanism in the Transformer architecture to process these small blocks in parallel. This processing method not only improves computational efficiency, but also enables each small block to be associated with all other small blocks in the image, thereby achieving a deep understanding of global dependencies.

[0110] In addition, the VIT module performs well in analyzing the correlation between blocks in the image. It can identify and enhance those signals that may be far apart in space but have correlation. This long-distance correlation is crucial for understanding complex image structures and patterns. For example, in weather forecasting, distant cloud patterns may have a direct connection with the weather conditions in the current area, and the VIT module can capture this long-distance influence.

[0111] Another significant advantage of the VIT module is its ability to aggregate spatiotemporal information. When processing sequence data or video streams, the VIT module can not only process information at a single time point, but also take into account changes in the time series, thus providing the model with richer dynamic features. This comprehensive use of spatiotemporal information enables the VIT module to perform excellent performance in prediction tasks.

[0112] Please refer to Figure 3 , Figure 3 for Figure 1 The step-by-step flow chart of step S3 in FIG. Figure 3 Please combine Figure 1 , the vortex data acquisition step S3 comprises:

[0113] Filtering step S31: filtering the future sea surface height data;

[0114] Vortex data prediction step S32: obtaining vortex data corresponding to the future sea surface height data through a vortex prediction model based on the future sea surface height data after filtering.

[0115]

[0116] Specifically, the future sea surface height data is shown in the above table. The rows in the table are dimensions, the columns are longitude, and NaN represents land. After completing the prediction of the sea surface height, the vortex can be predicted based on the sea surface height data. After adjusting its filter parameters, the existing vortex prediction model py-eddy-tracker can be used to identify and predict the basic characteristics of the vortex such as position, radius, and rotation direction based on the sea surface height data.

[0117] For example, when a predicted value is 0.911111, after it is judged that it exceeds the threshold, a weighted calculation is performed based on the eight upper, lower, left and right values ​​of 0.773018, 0.763941, 0.763082, 0.773018, 0.758167, 0.773449, 0.768415, and 0.767276 to obtain a replacement value, and 0.911111 is replaced by the replacement value. For another example, when a predicted value is 0.601111, after it is judged that it is lower than the threshold, a weighted calculation is also performed to obtain a replacement value, and 0.601111 is replaced by the replacement value, thereby ensuring the accuracy of the recognition and prediction results.

[0118] Please refer to Figure 4 , Figure 4 FIG. 1 is a schematic diagram of the structure of the ocean data prediction system of the present invention. Figure 4 As shown, an ocean data prediction system based on a VIT module of the present invention, wherein any one of the above-mentioned ocean data prediction methods is applied, and the ocean data prediction system comprises:

[0119] The data acquisition unit 11 acquires the existing sea surface height data within a first set time;

[0120] A future sea surface height data obtaining unit 12 is configured to obtain future sea surface height data within a second set time in the future based on the existing sea surface height data through a sea surface height prediction model based on a VIT module;

[0121] A vortex data obtaining unit 13, which obtains vortex data corresponding to the future sea surface height data through a vortex prediction model based on the future sea surface height data;

[0122] The future sea surface height data includes the sea surface height value, and the vortex data includes the vortex position, vortex radius and vortex rotation direction.

[0123] Wherein, the future sea surface height data obtaining unit 12 comprises:

[0124] The time series data set acquisition module 121 acquires the sea surface height remote sensing data from the operational satellite and processes the data to form a time series data set;

[0125] A model building module 122, after building an initial sea surface height prediction model based on the VIT module, trains and tests the initial sea surface height prediction model through the time series data set to obtain the final sea surface height prediction model;

[0126] A future sea surface height data prediction module 123 is configured to obtain the future sea surface height data of the i+1th day through the sea surface height prediction model according to the existing sea surface height data of the ith day and the n days before the ith day;

[0127] Among them, i is the current date, and n is a positive integer.

[0128] Wherein, the model building acquisition module includes:

[0129] Building a block embedding layer, which divides the input data into n×n small blocks, maps each small block into a one-dimensional embedding vector, and encodes the position information into the embedding vector;

[0130] Constructing an encoder layer, the encoder layer comprising a plurality of first VIT module units and block aggregation units, the first VIT module units and the block aggregation units being alternately arranged, and reducing the length and width of each small block and increasing the number of feature channels through the encoder layer until a first target is reached and then stopping;

[0131] Constructing a decoder layer, the decoder layer comprising a plurality of second VIT module units and block expansion units, the second VIT module units and the block expansion units being alternately arranged, and increasing the length and width of each small block and reducing the number of feature channels through the decoder layer until a second target is reached and then stopping;

[0132] Construct an MLP linear layer and obtain the prediction result by predicting each small block;

[0133] The initial sea surface height prediction model is trained and tested by using the time series data set to obtain the final sea surface height prediction model.

[0134] Wherein, the vortex data obtaining unit 13 comprises:

[0135] A filtering module 131 performs filtering processing on the future sea surface height data;

[0136] The vortex data prediction module 132 obtains vortex data corresponding to the future sea surface height data through a vortex prediction model based on the future sea surface height data after filtering.

[0137] In summary, the present invention is based on the VIT module and a reasonably designed neural network, which realizes the functions of sea surface height prediction and vortex identification for the next seven days, which is of great significance to marine disaster warning and prevention and navigation and fishery support. At the same time, the present invention improves the speed and accuracy of sea surface height forecasting while reducing the forecasting cost.

[0138] Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent substitutions for some of the technical features therein; and these modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting ocean data based on the VIT module, characterized in that: include: Data acquisition step: acquiring existing sea surface height data within a first set time; The step of obtaining future sea surface height data is as follows: based on the existing sea surface height data, obtaining future sea surface height data within a second set time in the future through a sea surface height prediction model based on a VIT module; Vortex data obtaining step: obtaining vortex data corresponding to the future sea surface height data through a vortex prediction model based on the future sea surface height data; The future sea surface height data includes the sea surface height value, and the vortex data includes the vortex position, vortex radius and vortex rotation direction; Wherein, the step of obtaining future sea surface height data comprises: Steps for obtaining the time series data set: obtain the sea surface height remote sensing data from the operational satellite, and process the data to form a time series data set; Model building and obtaining step: after building an initial sea surface height prediction model based on the VIT module, the initial sea surface height prediction model is trained and tested through the time series data set to obtain the final sea surface height prediction model; Future sea surface height data prediction step: obtaining the future sea surface height data of the i+1th day through the sea surface height prediction model according to the existing sea surface height data of the i-th day and the n days before the i-th day; Where i is the current date and n is a positive integer; The model building and obtaining step includes: Building a block embedding layer, which divides the input data into n×n small blocks, maps each small block into a one-dimensional embedding vector, and encodes the position information into the embedding vector; Constructing an encoder layer, the encoder layer comprising a plurality of first VIT module units and a block aggregation unit, wherein the first VIT module units and the block aggregation units are alternately arranged, and the encoder layer is used to reduce the length and width of each small block and increase the number of feature channels until a first target is reached and then stop, wherein the block aggregation unit is used to perform a downsampling operation, the length and width of each small block will become half of the original, and then the number of feature channels of each small block will become twice the original; Constructing a decoder layer, the decoder layer comprising a plurality of second VIT module units and a block expansion unit, the second VIT module units and the block expansion unit being alternately arranged, and the decoder layer increasing the length and width of each small block and reducing the number of feature channels until a second target is reached and then stopping, wherein the block expansion unit performs an upsampling operation, the length and width of each small block will become twice the original, and then the number of feature channels of each small block will become half of the original; Construct an MLP linear layer and obtain the prediction result by predicting each small block; The initial sea surface height prediction model is trained and tested by using the time series data set to obtain the final sea surface height prediction model.

2. The ocean data prediction method according to claim 1, characterized in that: The step of obtaining the time series data set comprises: Obtaining daily average sea surface height remote sensing data; After sorting the sea surface height remote sensing data according to a set order, normalizing the sea surface height remote sensing data; The normalized sea surface height remote sensing data is divided into training data and true values ​​corresponding to the training data; The time series data set is constructed using the training data and the true value.

3. The ocean data prediction method according to claim 2, characterized in that: Each of the first VIT module unit and the second VIT module unit includes at least one VIT module.

4. The ocean data prediction method according to claim 1, characterized in that: The vortex data acquisition step comprises: Filtering step: filtering the future sea surface height data; Vortex data prediction step: based on the future sea surface height data after filtering, the vortex prediction model is used to obtain the vortex data corresponding to the future sea surface height data.

5. An ocean data prediction system based on the VIT module, characterized in that: include: A data acquisition unit, which acquires existing sea surface height data within a first set time; A future sea surface height data obtaining unit, which obtains future sea surface height data within a second set time in the future based on the existing sea surface height data through a sea surface height prediction model based on a VIT module; a vortex data obtaining unit, which obtains vortex data corresponding to the future sea surface height data through a vortex prediction model based on the future sea surface height data; The future sea surface height data includes the sea surface height value, and the vortex data includes the vortex position, vortex radius and vortex rotation direction; Wherein, the future sea surface height data obtaining unit comprises: The time series data set acquisition module acquires the sea surface height remote sensing data from the operational satellite and processes the data to form a time series data set; A model building module, after building an initial sea surface height prediction model based on the VIT module, trains and tests the initial sea surface height prediction model through the time series data set to obtain the final sea surface height prediction model; A future sea surface height data prediction module, which obtains the future sea surface height data of the i+1th day through the sea surface height prediction model according to the existing sea surface height data of the i-th day and the n days before the i-th day; Where i is the current date and n is a positive integer; Wherein, the model building acquisition module includes: Building a block embedding layer, which divides the input data into n×n small blocks, maps each small block into a one-dimensional embedding vector, and encodes the position information into the embedding vector; Constructing an encoder layer, the encoder layer comprising a plurality of first VIT module units and a block aggregation unit, wherein the first VIT module units and the block aggregation units are alternately arranged, and the encoder layer is used to reduce the length and width of each small block and increase the number of feature channels until a first target is reached and then stop, wherein the block aggregation unit is used to perform a downsampling operation, the length and width of each small block will become half of the original, and then the number of feature channels of each small block will become twice the original; Constructing a decoder layer, the decoder layer comprising a plurality of second VIT module units and a block expansion unit, the second VIT module units and the block expansion unit being alternately arranged, and the decoder layer increasing the length and width of each small block and reducing the number of feature channels until a second target is reached and then stopping, wherein the block expansion unit performs an upsampling operation, the length and width of each small block will become twice the original, and then the number of feature channels of each small block will become half of the original; Construct an MLP linear layer and obtain the prediction result by predicting each small block; The initial sea surface height prediction model is trained and tested by using the time series data set to obtain the final sea surface height prediction model.

6. The ocean data prediction system according to claim 5, characterized in that: The vortex data acquisition unit comprises: A filtering module, for filtering the future sea surface height data; The vortex data prediction module obtains vortex data corresponding to the future sea surface height data through a vortex prediction model based on the future sea surface height data after filtering.