Arctic sea ice short-term forecast method and system combining physical constraints and twin networks

Abstract

The present invention discloses a method and system for short-term Arctic sea ice forecasting that combines physical constraints with a twin network. The method comprises the following steps: acquiring and preprocessing Arctic sea ice data to obtain a training set and a test set; performing modal decomposition on the sea ice density data in the training set based on empirical orthogonal decomposition to obtain time series information and spatial modal information; obtaining a physical constraint dataset based on the time series information and spatial modal information; constructing an empirical orthogonal decomposition loss function; constructing a twin network; constructing and training a multi-factor forecast model that combines physical constraints based on the empirical orthogonal decomposition loss function, the twin network, and the physical constraint dataset; testing the trained multi-factor forecast model that combines physical constraints based on the test set; and completing a short-term forecast of Arctic sea ice density. The present invention can effectively improve the accuracy of SIC forecasts and the interpretability of forecast results, providing forecast data support for Arctic resource development and navigation.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning application for Arctic sea ice forecasting, and specifically relates to a method and system for short-term Arctic sea ice forecasting that combines physical constraints with twin networks. Background Art

[0002] Currently, related work has attempted to apply deep learning to sea ice forecasting. By combining models with different network structures in the computer field with sea ice elements, it can achieve better results than sea ice climate states or traditional statistical forecasting methods. Currently, there are three main types of deep learning models applied to sea ice forecasting and estimation: the first is networks based on time cycles, such as long short-term memory networks (LSTMs), which are good at processing time series information. The second is networks based on convolutional computing, such as convolutional neural networks (CNNs), which are good at processing spatial feature information. The third is networks based on spatiotemporal computing, such as convolutional long short-term memory networks (ConvLSTMs), which can process both spatial and temporal information. Or they are improvements and variants based on the above-mentioned networks.

[0003] However, the LSTM model cannot exploit information across spatial dimensions. In the LSTM model, time series at different spatial grid points are considered separate sample data, and the interactions between adjacent grid points or grid points within adjacent regions cannot be preserved. Therefore, when using the LSTM model for spatial field forecasts of SIC, it is necessary to convert the two-dimensional spatial data into a one-dimensional grid for calculation. While this approach increases the number of training samples, it loses spatial structure, particularly information on the local variations in sea ice.

[0004] The CNN model has inherent limitations. It can only process information in the spatial dimension, not time series. This is because the convolution kernels in the CNN model can only process spatial information within the "receptive field." In the channel dimension, the convolution kernel linearly adds all channels, making it impossible to distinguish temporal information in this dimension. The positive order logic of the time series (from Ti-1 to Ti) cannot be preserved in the channel dimension of the CNN because the CNN treats different channels as different types of factors at the same time. In the channel dimension of the CNN, there is no time chain communication or transmission between the results of different convolution kernels; each convolution kernel performs convolution calculations independently. Therefore, the CNN model is only suitable for spatial field forecasting at the same time. If data from different time periods is stacked in the channel dimension of the CNN model, it will cause problems such as disordered spatiotemporal information and logical errors in the time series.

[0005] Compared to the limitations of LSTM and CNN models, networks based on spatiotemporal computation (such as ConvLSTM) have better applicability for the three-dimensional spatiotemporal forecasting of SIC. Existing technologies have not improved ConvLSTM to build a complete short-term Arctic sea ice forecast model. Summary of the Invention

[0006] This invention aims to address the shortcomings of existing technologies by proposing a short-term Arctic sea ice forecasting method and system that combines physical constraints with a twin network. This method ensures a strict positive order logic for the time series in the data. Furthermore, it incorporates convolutional computations similar to those in CNN models to efficiently extract spatial feature information.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] The short-term Arctic sea ice forecasting method combining physical constraints with twin networks includes the following steps:

[0009] Acquire and preprocess Arctic sea ice data to obtain a training set and a test set; the Arctic sea ice data includes sea ice density data and meteorological and hydrological data;

[0010] Based on empirical orthogonal decomposition, modal decomposition is performed on the sea ice density data in the training set to obtain time series information and spatial modal information;

[0011] Based on the time series information and the spatial modal information, improving the training set to obtain a physical constraint data set;

[0012] Based on the mean square error loss function, the time series information and the spatial modal information are introduced to construct an empirical orthogonal decomposition loss function;

[0013] On the basis of the convolutional long short-term memory network and the combination of the long short-term memory network, a branch for calculating the time series information is added to construct a twin network;

[0014] Based on the empirical orthogonal decomposition loss function, the twin network, and the physical constraint data set, constructing and training a multi-factor prediction model incorporating physical constraints, and testing the trained multi-factor prediction model incorporating physical constraints based on the test set;

[0015] Based on the tested multi-factor forecast model combined with physical constraints, a short-term forecast of Arctic sea ice concentration is completed.

[0016] Preferably, the method for preprocessing the Arctic sea ice data is:

[0017] Using bilinear interpolation calculation, interpolate the spatial grid points of the meteorological and hydrological data onto the grid points of the sea ice density data to unify the spatial format of the Arctic sea ice data;

[0018] Splicing the Arctic sea ice data in time dimension and selecting the longitude and latitude range according to the actual sea ice research area;

[0019] Based on the latitude and longitude range, performing channel dimension splicing on the Arctic sea ice data spliced ​​by the time dimension, and setting missing values ​​and land grid points in the Arctic sea ice data as singular values;

[0020] Normalizing the Arctic sea ice data spliced ​​by the channel dimension and assigning values ​​to the singular values ​​to obtain the processed Arctic sea ice data;

[0021] The processed Arctic sea ice data is divided into a training set and a test set to complete the preprocessing of the Arctic sea ice data.

[0022] Preferably, according to the variance contribution rates of the time series information and the spatial modal information, the number of modes that meets a preset variance contribution rate sum threshold is selected as the base state time series information and the base state spatial modal information.

[0023] Preferably, the method for obtaining the physical constraint data set is:

[0024] Reconstructing the time series information and the spatial modal information back into a spatial field corresponding to the modality;

[0025] The spatial field is supplemented into the channel dimension and combined with the training set to obtain the physical constraint dataset.

[0026] Preferably, the method for constructing the empirical orthogonal decomposition loss function is:

[0027] Based on the mean square error loss function and the actual value and the predicted value of the sea ice concentration data, a grid-by-grid loss value of the sea ice concentration data is calculated;

[0028] Projecting the true value and the predicted value of the sea ice concentration data onto the base state spatial modal information to obtain base state time series information of the true value and the predicted value;

[0029] Calculating a loss value of the base state time series information based on the base state time series information of the true value and the forecast value;

[0030] De-dimensionalizing the loss value of the sea ice concentration data and the loss value of the base state time series information, and then dividing them by their respective variances to obtain the de-dimensionalized sea ice concentration data loss and the base state time series information loss;

[0031] The de-dimensionalized sea ice concentration data loss and the base state time series loss are added according to preset weights to obtain the empirical orthogonal decomposition loss function.

[0032] Preferably, the method for constructing the twin network is:

[0033] Projecting the data in the training set onto the ground state spatial modal information to obtain the ground state time series information of the training set;

[0034] Using the ground state time series information of the training set to predict the ground state time series information of subsequent moments using a long short-term memory network;

[0035] Reconstruct the ground state spatial modal information of the training set and the predicted ground state time series information to obtain the prediction results of the long short-term memory network branch;

[0036] Based on the training set and the convolutional long short-term memory network, obtaining a prediction result of a convolutional long short-term memory network branch;

[0037] The prediction results of the long short-term memory network branch are convolutionally fused with the prediction results of the convolutional long short-term memory network branch to obtain the final prediction result, thereby completing the construction of the twin network.

[0038] Preferably, the method for constructing and training a multi-factor prediction model incorporating physical constraints is:

[0039] Based on preset weights, the empirical orthogonal decomposition loss functions with different weights are obtained, and based on the twin network, multiple groups of multi-factor prediction models with different weights of the empirical orthogonal decomposition loss functions are obtained;

[0040] Based on the physical constraint data set, multiple groups of multi-factor prediction models with different weights of empirical orthogonal decomposition loss functions are trained, and by comparing the multiple groups of training results, a multi-factor prediction model combined with physical constraints with optimal weights is obtained;

[0041] The input end, model structure and model loss function of the multi-factor prediction model combined with physical constraints with optimal weights are improved respectively to obtain the optimal multi-factor prediction model combined with physical constraints; wherein, the model structure is the twin network, and the model loss function is the empirical orthogonal decomposition loss function.

[0042] The present invention also provides an Arctic sea ice short-term forecast system combining physical constraints and twin networks, which is used to implement the Arctic sea ice short-term forecast method, including:

[0043] A data preprocessing module is used to acquire and preprocess Arctic sea ice data to obtain a training set and a test set; the Arctic sea ice data includes sea ice density data and meteorological and hydrological data;

[0044] A modal decomposition module is used to perform modal decomposition on the sea ice density data in the training set based on empirical orthogonal decomposition to obtain time series information and spatial modal information;

[0045] A training set improvement module, configured to improve the training set based on the time series information and the spatial modal information to obtain a physical constraint data set;

[0046] A loss function construction module is used to introduce the time series information and the spatial modal information based on the mean square error loss function to construct an empirical orthogonal decomposition loss function;

[0047] A twin network construction module is used to add a branch for calculating the time series information on the basis of the convolutional long short-term memory network and the combination of the long short-term memory network to construct a twin network;

[0048] A model construction module is used to construct and train a multi-factor prediction model incorporating physical constraints based on the empirical orthogonal decomposition loss function, the twin network, and the physical constraint dataset, and to test the trained multi-factor prediction model incorporating physical constraints based on the test set;

[0049] The forecast module is used to complete the short-term forecast of Arctic sea ice density based on the multi-factor forecast model combined with physical constraints that has passed the test.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] Compared with the pure data-driven deep learning forecasting model, the present invention introduces the EOF physical information of sea ice density, and based on a variety of meteorological and oceanographic elements, constructs a multi-factor forecasting model combined with physical constraints. Based on the empirical orthogonal decomposition loss function, the twin network and the physical constraint data set, a multi-factor forecasting model combined with physical constraints is constructed and trained, and the trained multi-factor forecasting model combined with physical constraints is tested based on the test set; in response to the problem that the current pure data-driven SIC short-term forecast model lacks physical law constraints, the twin network is combined with EOF physical information to propose SEOF_SICNet, which improves the forecast accuracy of SIC and the interpretability of the forecast results, and provides forecast data support for Arctic resource development and waterway navigation. Based on the multi-factor forecasting model combined with physical constraints that has passed the test, the short-term forecast of Arctic sea ice density is completed. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0053] Figure 1 This is a flow chart of a method for short-term Arctic sea ice forecasting that combines physical constraints with a twin network according to an embodiment of the present invention;

[0054] Figure 2 This is a diagram comparing the improved effects of the EOF loss function under different weights according to an embodiment of the present invention;

[0055] Figure 3 This figure compares the improvement effects of different physical constraint combination methods in an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only 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 making creative efforts are within the scope of protection of the present invention.

[0057] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0058] Example 1

[0059] like Figure 1 As shown in Figure 1, the short-term Arctic sea ice forecasting method combining physical constraints with twin networks includes the following steps:

[0060] Arctic sea ice data are acquired and preprocessed to obtain training and test sets. The Arctic sea ice data include sea ice concentration data and meteorological and hydrological data. In this embodiment, the Arctic sea ice data are the sea ice concentration data nsidc-0079, version 3 from the National Snow and Ice Data Center (NSIDC) and the sea surface temperature (SST), skin temperature (SKT), and 2-meter temperature (T2M) data from the fifth generation reanalysis data (ERA5) of the European Meteorological Agency. The various meteorological and hydrological data from ERA5 are interpolated onto the NSIDC sea ice concentration (SIC) grid.

[0061] A further embodiment is that the method for preprocessing Arctic sea ice data is:

[0062] Using bilinear interpolation, the spatial grid points of meteorological and hydrological data were interpolated onto the grid points of sea ice concentration data, thus unifying the spatial format of Arctic sea ice data. The script interface provided by the official NSIDC and ERA5 database websites was used to batch download daily-resolution SIC, SST, SKT, and T2M data for the Arctic Ocean. The spatial grid points of SST, SKT, and T2M were interpolated onto the SIC grid points, thus unifying the spatial format of the various data elements.

[0063] The Arctic sea ice data are spliced ​​in the time dimension, and the longitude and latitude range is selected according to the actual sea ice research area. In this embodiment, the research range is 69°N to 90°N, -180°W to 180°E, and the SIC, SST, SKT and T2M data are spliced ​​in the time dimension in turn, and the longitude and latitude range is selected according to the actual research area.

[0064] Based on the latitude and longitude range, the Arctic sea ice data spliced ​​by the time dimension are spliced ​​in the channel dimension, and the missing values ​​and land grid points in the Arctic sea ice data are set as singular values; the four factors are spliced ​​in the channel dimension, and the missing values ​​and land grid points in the four factors are set to nan.

[0065] The Arctic sea ice data spliced ​​by channel dimension are normalized and the singular values ​​are assigned to obtain the processed Arctic sea ice data; the data are normalized to [0,100], and the normalization formula is:

[0066] x new =(x old -min) / (max-min)*100

[0067] Among them, x old represents the data before normalization, x new Represents normalized data.

[0068] Singular values ​​are processed by setting the nan value of SIC to 0; the nan values ​​of SST, SKT and T2M are set to -100 to distinguish them from normal sea ice density and temperature values.

[0069] The processed Arctic sea ice data is divided into a training set and a test set to complete the preprocessing of the Arctic sea ice data. According to actual needs, historical data of 10 years or more is selected as the training set, and data of 1 year or more is reserved as the test set. The input time step and prediction time step of the model are set, and the training data and test data are obtained respectively by sliding the time window. The processed data is stored in an npz file. In this embodiment, the data from 2009 to 2020 is used as the training set, and the data from 2021 is used as the test set. The input time step of the model is set to 10 days, and the prediction time step is set to 10 days.

[0070] Based on empirical orthogonal decomposition, modal decomposition is performed on the sea ice density data in the training set to obtain time series information and spatial modal information;

[0071] A further embodiment is that

[0072] The EOF (Empirical Orthogonal Decomposition) analysis formula is used to decompose the historical SIC data into PCs and EOFs. The EOF analysis formula is as follows:

[0073] X m×n =EOF m×m ×PC m×n

[0074] Where X is the original data, EOF is the decomposed spatial modal information, PC is the decomposed time series information, m represents the number of grid points in the spatial field, and n represents the length of the observation time series of each grid point.

[0075] According to the variance contribution rate of each mode, the number of modes with a sum of variance contribution rates of 85% or above (preset variance contribution rate sum threshold) is selected as the ground state PCs (ground state time series information) and ground state EOFs (ground state spatial modal information) according to actual needs.

[0076] Based on time series information and spatial modal information, the training set is improved to obtain a physical constraint dataset;

[0077] A further implementation method is that the method for obtaining the physical constraint dataset is:

[0078] Reconstruct the time series information and spatial modal information back into the spatial field corresponding to the mode; specifically, according to actual needs, select three or more main modes related to the physical laws of sea ice, and reconstruct the EOF spatial mode and time mode of each mode back into the spatial field corresponding to the mode.

[0079] The spatial field is supplemented into the channel dimension and combined with the training set to obtain a physically constrained dataset.

[0080] Based on the mean square error loss function, time series information and spatial modal information are introduced to construct the empirical orthogonal decomposition loss function;

[0081] A further implementation method is to construct an empirical orthogonal decomposition loss function as follows:

[0082] Based on the mean square error (MSE) loss function, the grid-wise loss value of sea ice density data is calculated based on the actual and predicted values ​​of sea ice density data.

[0083] The calculation formula of MSE is as follows:

[0084]

[0085] Where x is the true value of SIC, is the predicted value of SIC.

[0086] Project the true and predicted values ​​of sea ice density data onto the base-state spatial modal information to obtain the base-state time series information of the true and predicted values;

[0087] Based on the base state time series information of the true value and the predicted value, the loss value of the base state time series information is calculated;

[0088] Specifically, the predicted and true values ​​of SIC of each batch are projected onto the basis EOFs to obtain the PCs of the predicted and true values. The PCs loss value is calculated as follows:

[0089]

[0090] The loss values ​​of sea ice concentration data and the loss values ​​of base-state time series information are de-dimensionalized and then divided by their respective variances to obtain the de-dimensionalized sea ice concentration data loss and base-state time series information loss.

[0091] De-dimensionalize the loss value of SIC and the loss value of PCs, and remove the variance of each. The calculation formula of variance is as follows:

[0092]

[0093] Where x is the average value. At this time, the SIC loss and PCs loss after dimensioning are:

[0094]

[0095]

[0096] Where train represents the SIC data or PCs data of the entire training set.

[0097] The dedimensionalized sea ice density data loss and the ground-state time series loss are added according to the preset weights to obtain the empirical orthogonal decomposition loss function.

[0098] Combining SIC loss and PCs loss, we first preset a set of weights, and then determine the optimal weights based on experimental results. SIC loss and PCs loss are added together according to the preset weights to construct the EOF loss function. The formula is as follows:

[0099] EOF_Loss=SIC_Loss+ω×PCs_Loss

[0100] Where ω is the weight of the PCs loss value, which is given by prior knowledge and comparative experiments. Four comparative values ​​are preset: 0.1, 0.2, 0.5, and 1.

[0101] Based on the combination of Convolutional Long Short-Term Memory (ConvLSTM) and Long Short-Term Memory (LSTM), a branch for calculating time series information is added to construct a twin network.

[0102] A further implementation method is to construct a twin network by:

[0103] The data in the training set are projected onto the ground state spatial modal information to obtain the ground state time series information of the training set; the time length is 10.

[0104] The ground state time series information of the training set is used to predict the ground state time series information of subsequent moments using a long short-term memory network; the number of hidden layer neurons in the LSTM is 768, and the time length is 10.

[0105] Reconstruct the ground state spatial modal information of the training set and the predicted ground state time series information to obtain the prediction results of the long short-term memory network branch;

[0106] Based on the training set and the convolutional long short-term memory network, the prediction results of the convolutional long short-term memory network branch are obtained;

[0107] The prediction results of the long short-term memory network branch are convolutionally fused with the prediction results of the convolutional long short-term memory network branch to obtain the final prediction results and complete the construction of the twin network.

[0108] Specifically, the ConvLSTM layer consists of three ConvLSTM layers and one Conv2D layer. The number of convolution kernels in the ConvLSTM layer is 512, 512, and 256, respectively; the convolution kernel sizes are 5, 3, and 3, respectively. The convolution kernel size of the Conv2D layer is set to 1; the number of convolution kernels is set to the number of channels of the input data. The activation function during training is set to ReLU, the model optimizer is set to Adam, the learning rate is initially set to 0.001, and dynamic descent is updated using ReduceLROnPlateau.

[0109] S6.5: The results of the ConvLSTM branch are convolved with the results of the LSTM branch through a Conv2D layer to obtain the final prediction result. The convolution kernel size of the Conv2D layer is set to 1; the number of convolution kernels is set to the number of channels of the output data.

[0110] Based on the empirical orthogonal decomposition loss function, the twin network, and the physical constraint dataset, a multi-factor prediction model with physical constraints is constructed and trained. The trained multi-factor prediction model with physical constraints is tested on the test set.

[0111] A further embodiment is to construct and train a multi-factor prediction model incorporating physical constraints by:

[0112] Based on the preset weights, the empirical orthogonal decomposition loss functions with different weights are obtained, and based on the twin network, multiple sets of multi-factor prediction models with different weights of the empirical orthogonal decomposition loss functions are obtained;

[0113] Based on the physical constraint data set, multiple groups of multi-factor prediction models with different weights of empirical orthogonal decomposition loss functions are trained. By comparing the results of multiple groups of training, the multi-factor prediction model with the optimal weight combined with physical constraints is obtained; that is, according to the preset weights (for example, 0.1, 0.2, 0.5 and 1), EOF loss functions with different weights are constructed, marked as EOF_Loss_01, EOF_Loss_02, EOF_Loss_05 and EOF_Loss_10 respectively, multiple groups of models are trained, and the optimal weight model is selected according to evaluation indicators such as MAE (mean absolute error). Figure 2 As shown, taking the 2021 NSIDC Arctic sea ice density data as an example, the improvement effect of the EOF loss function under different weights is compared. In the figure, EOF_Loss_01 is a weight value of 0.1, EOF_Loss_02 is a weight value of 0.2, EOF_Loss_05 is a weight value of 0.5, and EOF_Loss_10 is a weight value of 1.

[0114] The input end, model structure and model loss function of the multi-factor prediction model combined with physical constraints with optimal weights are improved respectively to obtain the optimal multi-factor prediction model combined with physical constraints; among them, the model structure is a twin network, and the model loss function is an empirical orthogonal decomposition loss function.

[0115] The model input, model structure and model loss function are improved respectively. The specific improvement methods include: constructing a new training data set at the input end, adding the spatial modal information of empirical orthogonal decomposition on the basis of the original data; adding a time series prediction branch to the model structure, predicting the time series obtained by empirical orthogonal decomposition based on the long short-term memory network, and performing convolution fusion with the basic branch; improving the loss function at the model output end, adding the dedimensionalized sea ice density data loss and time series loss according to the optimal weights, and constructing a loss function with added physical constraints.

[0116] Different improvement methods and multiple combination methods are compared to improve the model, and the optimal SEOF_SICNet model is selected according to the evaluation indicators.

[0117] Specifically, such as Figure 3 As shown, taking the 2021 NSIDC Arctic sea ice density data as an example, the improvement effects of different combinations of physical constraints are compared (the forecast accuracy, Pearson correlation coefficient and structural similarity are compared). In the figure, ConvLSTM_MSE is the basic model, ConvLSTM_EOF_Loss is the model for improving the loss function, ConvLSTM_EOF_Factor is the model for improving the training data set, and Siam_ConvLSTM_EOF is the model combined with the twin network.

[0118] Based on a tested multi-factor forecast model combined with physical constraints, a short-term forecast of Arctic sea ice concentration is completed.

[0119] Example 2

[0120] The present invention also provides an Arctic sea ice short-term forecasting system that combines physical constraints with a twin network, and a method for implementing the Arctic sea ice short-term forecasting method, including:

[0121] The data preprocessing module is used to obtain and preprocess Arctic sea ice data to obtain training and test sets; Arctic sea ice data includes sea ice density data and meteorological and hydrological data;

[0122] The modal decomposition module is used to perform modal decomposition on the sea ice density data in the training set based on empirical orthogonal decomposition to obtain time series information and spatial modal information;

[0123] The training set improvement module is used to improve the training set based on time series information and spatial modal information to obtain a physical constraint data set;

[0124] The loss function construction module is used to introduce time series information and spatial modal information based on the mean square error loss function to construct the empirical orthogonal decomposition loss function;

[0125] The twin network construction module is used to add branches for calculating time series information based on the convolutional long short-term memory network and the combination of long short-term memory network to build a twin network;

[0126] The model building module is used to build and train a multi-factor prediction model with physical constraints based on the empirical orthogonal decomposition loss function, the twin network, and the physical constraint dataset, and to test the trained multi-factor prediction model with physical constraints based on the test dataset;

[0127] The forecast module is used to complete the short-term forecast of Arctic sea ice density based on a tested multi-factor forecast model combined with physical constraints.

[0128] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A short-term Arctic sea ice forecasting method combining physical constraints with twin networks, characterized by: The following steps are involved: Acquire and preprocess Arctic sea ice data to obtain a training set and a test set; the Arctic sea ice data includes sea ice density data and meteorological and hydrological data; Based on empirical orthogonal decomposition, modal decomposition is performed on the sea ice density data in the training set to obtain time series information and spatial modal information; Based on the time series information and the spatial modal information, improving the training set to obtain a physical constraint data set; Based on the mean square error loss function, the time series information and the spatial modal information are introduced to construct an empirical orthogonal decomposition loss function; On the basis of the convolutional long short-term memory network and the combination of the long short-term memory network, a branch for calculating the time series information is added to construct a twin network; Based on the empirical orthogonal decomposition loss function, the twin network, and the physical constraint data set, constructing and training a multi-factor prediction model incorporating physical constraints, and testing the trained multi-factor prediction model incorporating physical constraints based on the test set; Based on the tested multi-factor forecast model combined with physical constraints, a short-term forecast of Arctic sea ice concentration is completed; According to the variance contribution rates of the time series information and the spatial modal information, selecting a modality number that meets a preset variance contribution rate sum threshold as the base state time series information and the base state spatial modal information; The method for obtaining the physical constraint data set is: Reconstructing the time series information and the spatial modal information back into a spatial field corresponding to the modality; Supplementing the spatial field into the channel dimension and combining it with the training set to obtain the physical constraint dataset; The method for constructing the twin network is: Projecting the data in the training set onto the ground state spatial modal information to obtain the ground state time series information of the training set; Using the ground state time series information of the training set to predict the ground state time series information of subsequent moments using a long short-term memory network; Reconstruct the ground state spatial modal information of the training set and the predicted ground state time series information to obtain the prediction results of the long short-term memory network branch; Based on the training set and the convolutional long short-term memory network, obtaining a prediction result of a convolutional long short-term memory network branch; The prediction results of the long short-term memory network branch are convolutionally fused with the prediction results of the convolutional long short-term memory network branch to obtain the final prediction result, thereby completing the construction of the twin network.

2. The Arctic sea ice short-term forecast method combining physical constraints and twin networks according to claim 1 is characterized in that: The method for preprocessing the Arctic sea ice data is: Using bilinear interpolation calculation, interpolate the spatial grid points of the meteorological and hydrological data onto the grid points of the sea ice density data to unify the spatial format of the Arctic sea ice data; Splicing the Arctic sea ice data in time dimension and selecting the longitude and latitude range according to the actual sea ice research area; Based on the latitude and longitude range, performing channel dimension splicing on the Arctic sea ice data spliced ​​by the time dimension, and setting missing values ​​and land grid points in the Arctic sea ice data as singular values; Normalizing the Arctic sea ice data spliced ​​by the channel dimension and assigning values ​​to the singular values ​​to obtain the processed Arctic sea ice data; The processed Arctic sea ice data is divided into a training set and a test set to complete the preprocessing of the Arctic sea ice data.

3. The Arctic sea ice short-term forecast method combining physical constraints and twin networks according to claim 1 is characterized in that: The method for constructing the empirical orthogonal decomposition loss function is: Based on the mean square error loss function and the actual value and the predicted value of the sea ice concentration data, a grid-by-grid loss value of the sea ice concentration data is calculated; Projecting the true value and the predicted value of the sea ice concentration data onto the base state spatial modal information to obtain base state time series information of the true value and the predicted value; Calculating a loss value of the base state time series information based on the base state time series information of the true value and the forecast value; De-dimensionalizing the loss value of the sea ice concentration data and the loss value of the base state time series information, and then dividing them by their respective variances to obtain the de-dimensionalized sea ice concentration data loss and the base state time series information loss; The de-dimensionalized sea ice density data loss and the ground state time series information loss are added according to preset weights to obtain the empirical orthogonal decomposition loss function.

4. The Arctic sea ice short-term forecast method combining physical constraints and twin networks according to claim 1 is characterized in that: The method for building and training a multi-factor forecasting model that incorporates physical constraints is: Based on preset weights, the empirical orthogonal decomposition loss functions with different weights are obtained, and based on the twin network, multiple groups of multi-factor prediction models with different weights of the empirical orthogonal decomposition loss functions are obtained; Based on the physical constraint data set, multiple groups of multi-factor prediction models with different weights of empirical orthogonal decomposition loss functions are trained, and by comparing the multiple groups of training results, a multi-factor prediction model combined with physical constraints with optimal weights is obtained; The input end, model structure and model loss function of the multi-factor prediction model combined with physical constraints with optimal weights are improved respectively to obtain the optimal multi-factor prediction model combined with physical constraints; wherein, the model structure is the twin network, and the model loss function is the empirical orthogonal decomposition loss function.

5. The Arctic sea ice short-term forecast system combining physical constraints and twin networks is characterized by: A method for implementing the short-term Arctic sea ice forecasting method according to any one of claims 1 to 4, comprising: A data preprocessing module is used to acquire and preprocess Arctic sea ice data to obtain a training set and a test set; the Arctic sea ice data includes sea ice density data and meteorological and hydrological data; A modal decomposition module is used to perform modal decomposition on the sea ice density data in the training set based on empirical orthogonal decomposition to obtain time series information and spatial modal information; A training set improvement module, configured to improve the training set based on the time series information and the spatial modal information to obtain a physical constraint data set; A loss function construction module is used to introduce the time series information and the spatial modal information based on the mean square error loss function to construct an empirical orthogonal decomposition loss function; A twin network construction module is used to add a branch for calculating the time series information on the basis of the convolutional long short-term memory network and the combination of the long short-term memory network to construct a twin network; A model construction module is used to construct and train a multi-factor prediction model incorporating physical constraints based on the empirical orthogonal decomposition loss function, the twin network, and the physical constraint dataset, and to test the trained multi-factor prediction model incorporating physical constraints based on the test set; The forecast module is used to complete the short-term forecast of Arctic sea ice density based on the multi-factor forecast model combined with physical constraints that has passed the test.

Images