Marine temperature prediction method based on Kupman theory and comparative learning

Abstract

The invention discloses an ocean temperature prediction method based on the Kupman theory and comparative learning. The method mainly comprises a data processing module, a dual-core modeling module, a loss optimization module and a result output module. The data processing module comprises a Golden ocean and environment monitoring service data set acquisition unit, a 7: 1: 2 proportion data segmentation and loading unit and a time window-based normalization and anti-normalization unit; the dual-core modeling module comprises an EDMD module and an FE-LSTM module; the loss optimization module comprises a GCL calculation unit fusing multiple factors and a parameter optimization unit adopting an AdamW optimizer; and the result output module combines the prediction results and evaluates and outputs the prediction results through R2, MAE and RMSE. According to the method, the non-stationarity of the ocean data is effectively processed, the prediction accuracy of the sea surface temperature is improved, 1-7-day prediction is supported, reliable data is provided for climate research, ocean management and disaster early warning, and the technical shortages of a traditional method are filled up.

Description

TECHNICAL FIELD

[0001] The present application relates to a method for predicting ocean temperature based on Koopman theory and contrastive learning. BACKGROUND

[0002] At present, ocean temperature prediction is of great significance in many fields such as climate research and ocean resource development. Precise ocean temperature prediction can provide key support for climate model construction, ocean ecological protection, etc. The existing ocean temperature prediction methods, such as the traditional numerical simulation method, have the problems of high computational complexity and insufficient description of complex ocean dynamic processes. The method based on a single machine learning model is also difficult to fully capture the multi-scale and nonlinear characteristics of ocean temperature data. Since the ocean system is a complex dynamic system containing various physical, chemical and biological processes, the ocean temperature is affected by many factors such as ocean currents, monsoons and solar radiation, and has strong nonlinearity and spatiotemporal heterogeneity, so a more efficient and accurate prediction method is needed to improve the accuracy of ocean temperature prediction and provide reliable basis for related field decision-making. Generally speaking, for ocean temperature prediction, the traditional numerical method relies on complex physical equation modeling, which has high computational cost and is easily limited by initial field and parameterization scheme; although a single deep learning model can learn data patterns, it does not sufficiently mine the essential characteristics of the ocean dynamic system. However, these methods are difficult to achieve accurate and efficient prediction when dealing with the multi-component and multi-scale characteristics of ocean temperature data. For accurate prediction of ocean temperature, it is necessary to combine the theory of ocean dynamic system (such as Koopman theory which can describe the linear dynamic characteristics of the system) and advanced machine learning technology (such as contrastive learning which can enhance feature representation) to fully utilize the dynamic characteristics and statistical patterns of ocean temperature data. In addition, the time-invariant component and the time-varying component of ocean temperature have different generation mechanisms and evolution laws, and to achieve accurate prediction of these two components and integrate the results, there are difficulties in method fusion and accuracy balance, therefore it is necessary to construct a prediction framework that integrates multiple theories and technologies to achieve this goal. SUMMARY

[0003] To solve the problems of existing ocean temperature prediction methods in complex ocean dynamic description, multi-scale feature capture, etc., and to improve the accuracy and efficiency of ocean temperature prediction, the present application proposes a method for predicting ocean temperature based on Koopman theory and contrastive learning, which can fully integrate the theory of ocean dynamic system and advanced machine learning technology, accurately predict ocean temperature, provide strong support for climate research, ocean resource development, etc., and has important practical application value.

[0004] The present application can be implemented by the following technical solutions: The application discloses a marine temperature prediction method based on Koopman theory and contrastive learning, and specifically comprises the following steps: firstly, obtaining marine surface temperature (SST) data of a research area, and performing time window division and normalization processing on the data; then, separating the normalized data into time-invariant components and time-varying components through a Fourier filter, performing time series prediction on the time-invariant components by using a feature enhancement network combined with a long short-term memory network (LSTM), and performing time series prediction on the time-varying components by using an extended dynamic mode decomposition (EDMD) module; and finally, merging the two types of prediction results, and performing data denormalization on the merged results to obtain the final marine temperature prediction results.

[0005] Further, when the marine surface temperature data of the research area is obtained, an authoritative marine database covering the target sea area and having a suitable time span is selected to collect marine surface temperature related data containing different latitudes and time periods.

[0006] Further, the time window division is to divide continuous marine temperature data into multiple equal-length time segments according to a set time interval; and the data normalization is to convert the data in each time window to a unified numerical range by calculating the mean and standard deviation, so as to eliminate the dimension effect.

[0007] Further, the Fourier filter performs frequency spectrum analysis on the normalized data, decomposes the data into components of different frequencies, and separates out parts reflecting different characteristics of the marine temperature, such as long-term trend and short-term fluctuation.

[0008] Further, when the feature enhancement network is used to perform time series prediction, the network first extracts and enhances the features of the input data to highlight the key feature information, and then combines a long short-term memory and other time series processing mechanisms to realize the prediction of the marine temperature.

[0009] Further, when the EDMD module is used to perform time series prediction, the dynamic evolution characteristics of the data are modeled based on the Koopman theory, the dynamic mode decomposition model is constructed to capture the evolution law of the marine temperature data, and then the prediction is realized.

[0010] Further, the prediction result merging is to integrate the prediction results of the feature enhancement network and the EDMD model according to corresponding rules; and the data denormalization is to restore the merged prediction results to the numerical range of the original data according to the mean and standard deviation during the previous normalization, so as to obtain the final available marine temperature prediction results.

[0011] Further, the training process of the neural network is as follows: first, initialize the EDMDNet model parameters; then load the training set data and preprocess; then separate the time-invariant component and the time-varying component through Fourier filtering; then perform multi-module parallel calculation and update the model parameters based on the loss function optimization; then determine whether the maximum iteration number or the early stopping condition is reached, if reached, save the optimal model parameters, if not reached, return to the step of loading the training set data and repeat the subsequent process.

[0012] Further, when initializing the EDMDNet model parameters, based on the characteristics of the ocean temperature data in the study area and the EDMDNet network structure, the weights and bias parameters of each layer in the network are randomly initialized, and the initial values of the core parameters such as the dimension of the Kuppers space and the dimension of the LSTM hidden layer are set, laying a foundation for subsequent model training.

[0013] Further, when loading the training set data, ocean surface temperature data covering different time periods and different locations in the study area is selected; preprocessing includes data cleaning to remove outliers and missing values, and format conversion operations to ensure that the data can be effectively read and used by the model.

[0014] Further, Fourier filter is used to perform spectral analysis on the preprocessed data, and the data is decomposed into components of different frequencies, so as to separate the time-invariant component reflecting the long-term stable trend of ocean temperature and the time-varying component reflecting the short-term fluctuation dynamic characteristics.

[0015] Further, the multi-module parallel calculation refers to simultaneously using the feature enhancement network module and the EDMD module to process and calculate the time-invariant component and the time-varying component respectively; the loss optimization and parameter updating are achieved by defining a suitable loss function to calculate the error between the predicted results and the true values, and using an optimization algorithm to update the model parameters in a backward propagation manner to minimize the loss.

[0016] Further, a maximum iteration number threshold is set, and the validation set loss is monitored, if the validation set loss does not decrease for consecutive multiple iteration rounds, the early stopping condition is triggered; when the maximum iteration number is reached or the early stopping condition is met, the current optimal model parameters are saved, otherwise the next round of training iteration is continued.

[0017] Compared with the prior art, the present application has the following advantages: 1) High prediction accuracy: the fusion of the linear feature description ability of Kuppers theory for ocean dynamic system and the feature enhancement advantage of contrastive learning can more accurately capture the multi-scale and nonlinear features of ocean temperature, and improve the prediction accuracy; 2) Adaptability: can effectively deal with the spatio-temporal heterogeneity of ocean temperature data, has good prediction effect on different regions and different time periods of ocean temperature, and has wide application range; 3) Good efficiency: compared with the traditional numerical simulation method, the calculation complexity is reduced while ensuring the prediction accuracy, and the prediction efficiency is improved; 4) Good interpretability: combined with the Copernicus theory, the prediction results can be explained from the perspective of the ocean dynamic system, enhancing the interpretability of the method and facilitating its application in related fields. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is the prediction method flowchart of the present application; Figure 2 is the neural network training flowchart of the present application; Figure 3 is the overall architecture schematic diagram of the algorithm model of the present application; Figure 4 is the EDMD module schematic diagram of the present application; Figure 5 is the FENet module schematic diagram of the present application. DETAILED DESCRIPTION

[0019] The embodiments of the present application will be described below through specific specific embodiments, and those skilled in the art can easily understand other advantages and effects of the present application from the content disclosed in the present specification.

[0020] As shown in FIGS. 1-5, the present application is a marine temperature prediction method based on the Copernicus theory and contrastive learning, which is developed around the logic of “data processing - model training - sub-module prediction - result output”. First, determine the research area, select two sub-regions of the East China Sea Kuroshio basin affected by the East Asian monsoon, Tsushima current and Kuroshio, select 24 different latitude and longitude nodes in each region, and obtain the marine data (including 10 variables such as sea surface temperature and salinity) of Copernicus Marine Environment Monitoring Service (CMEMS) from 1993 to 2021; then, according to the prediction method flowchart shown in FIG. 1, the data is divided into time windows (the window length is set to 96, covering the short-term fluctuations and periodic trends of SST) and normalized (the data is standardized based on the mean and standard deviation of the time window).

[0021] Referring to FIG. 2, the model training link trains the neural network process diagram, initializes the EDMDNet model parameters (determines the core hyperparameters such as the Kupmann space dimension 128, the LSTM hidden layer dimension 64, etc.), loads the training set data divided according to 7:1:2 and completes the preprocessing, separates the time invariable / variable components through the Fourier filter, then performs multi-module parallel calculation (synchronous operation of the FENet module and the EDMD module) and combines the global contrast loss (GCL) to optimize and update the parameters, and continuously iterates until the maximum training round (50 rounds) or the early stop condition (validation set loss without continuous 10 rounds of decline) is reached, and finally the optimal model parameters are saved.

[0022] In the prediction stage, according to the algorithm model overall architecture diagram shown in FIG. 3, the normalized data is input into the trained model: the time invariable component is input into the FENet module shown in FIG. 5, the one-dimensional discrete cosine transform (DCT) is used to obtain the frequency vector of each variable, the tensor matrix is stacked, the full connection layer is used to extract the dependence between variables to enhance the features, and then the LSTM network is used to capture the time sequence dependence; the time variable component is input into the EDMD module shown in FIG. 4, is segmented into multiple small segments, is converted to the Kupmann space through the MLP encoder, the correlation matrix is constructed to fit the Kupmann operator, the operator is used to predict the future Kupmann embedding, and the prediction result is output through the MLP decoder. Finally, the prediction results of the two types of components are combined, are restored to the original data range through the inverse normalization, and the final marine temperature prediction value is obtained, which provides data support for climate science, marine management and disaster prevention and reduction.

[0023] The above only describes the preferred embodiments of the present application and does not limit the present application. Any modification, equivalent replacement and improvement within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting ocean temperature based on Koopman theory and contrastive learning, characterized in that, Specifically, the following steps are included: First, ocean surface temperature (SST) data for the study area were acquired, and the data were divided into time windows and normalized. Next, the normalized data is separated into time-invariant components and time-varying components by using a Fourier filter. The time-invariant components are predicted using a feature enhancement network combined with a long short-term memory network (LSTM), and the time-varying components are predicted using an extended dynamic mode decomposition (EDMD) module. Finally, the two types of prediction results are merged, and the merged results are inversely normalized to obtain the final ocean temperature prediction result.

2. The ocean temperature prediction method based on Koopman theory and contrastive learning according to claim 1, characterized in that, When acquiring ocean surface temperature data for the study area, an authoritative ocean database covering the target sea area and with an appropriate time span is selected to collect ocean surface temperature-related data containing different latitudes and longitudes and time periods.

3. The ocean temperature prediction method based on Koopman theory and contrastive learning according to claim 1, characterized in that, The time window division involves dividing continuous ocean temperature data into multiple time segments of equal length according to a set time interval; data normalization involves converting the data within each time window to a uniform numerical range by calculating the mean and standard deviation, thereby eliminating the influence of dimensions.

4. The ocean temperature prediction method based on Koopman theory and contrastive learning according to claim 1, characterized in that, The Fourier filter performs spectral analysis on the normalized data, decomposing the data into components of different frequencies, thereby separating the parts that reflect different characteristics such as the long-term trend and short-term fluctuations of ocean temperature.

5. The ocean temperature prediction method based on Koopman theory and contrastive learning according to claim 1, characterized in that, When using the feature enhancement network for time series prediction, the network first extracts and enhances the features of the input data to highlight key feature information, and then combines time series processing mechanisms such as long short-term memory to achieve the prediction of ocean temperature.

6. The ocean temperature prediction method based on Koopman theory and contrastive learning according to claim 1, characterized in that, When using the extended dynamic mode decomposition module for time series prediction, the dynamic evolution characteristics of the data are modeled based on Koopman theory. By constructing a dynamic mode decomposition model, the evolution law of ocean temperature data is captured, thereby achieving prediction.

7. The ocean temperature prediction method based on Koopman theory and contrastive learning according to claim 1, characterized in that, The prediction result merging is the process of integrating the prediction results of the feature enhancement network and the EDMD model according to corresponding rules; Data denormalization involves using the mean and standard deviation from the previous normalization to restore the merged prediction results to the numerical range of the original data, thus obtaining the final usable ocean temperature prediction results.

8. The ocean temperature prediction method based on Koopman theory and contrastive learning according to claim 1, characterized in that, The training process of its neural network is as follows: First, initialize the EDMDNet model parameters; then load the training set data and perform preprocessing; then separate the time-invariant components and time-varying components through Fourier filtering; then perform parallel computation of multiple modules and optimize and update the model parameters based on the loss function; then determine whether the maximum number of iterations or the early stopping condition has been reached. If it has been reached, save the optimal model parameters; if it has not been reached, return to the step of loading the training set data and repeat the subsequent process.

9. The ocean temperature prediction method based on Koopman theory and contrastive learning according to claim 8, characterized in that, When initializing the EDMDNet model parameters, based on the characteristics of the ocean temperature data in the study area and the EDMDNet network structure, the weights and bias parameters of each layer in the network are randomly initialized. At the same time, the initial values ​​of the core parameters of the Koopman space dimension and the LSTM hidden layer dimension are set to lay the foundation for subsequent model training.

10. The ocean temperature prediction method based on Koopman theory and contrastive learning according to claim 8, characterized in that, When loading the training set data, ocean surface temperature data covering different time periods and locations in the study area are selected; preprocessing includes data cleaning, removal of outliers and missing values, and format conversion operations to ensure that the data can be effectively read and used by the model.

11. The ocean temperature prediction method based on Koopman theory and contrastive learning according to claim 8, characterized in that, Fourier filters are used to perform spectral analysis on the preprocessed data, decomposing the data into components of different frequencies, thereby separating the time-invariant component reflecting the long-term stable trend of ocean temperature, as well as the time-varying component reflecting the short-term dynamic characteristics of fluctuations.

12. The ocean temperature prediction method based on Koopman theory and contrastive learning according to claim 8, characterized in that, The multi-module parallel computing refers to simultaneously utilizing the feature enhancement network module and the EDMD module to process and calculate the time-invariant and time-varying components, respectively; loss optimization and parameter updating involves defining an appropriate loss function, calculating the error between the predicted result and the true value, and using an optimization algorithm to backpropagate and update the model parameters to minimize the loss.

13. The ocean temperature prediction method based on Koopman theory and contrastive learning according to claim 8, characterized in that, Set a maximum iteration threshold and monitor the change in validation set loss. If the validation set loss no longer decreases after several consecutive iterations, trigger the early stopping condition. When the maximum number of iterations is reached or the early stopping condition is met, save the current optimal model parameters; otherwise, continue to the next round of training iterations.

Images