Physical information and data-driven dual-flow coupled sea surface temperature prediction method and system

Abstract

The present invention belongs to the technical field of sea temperature prediction, and discloses a physical information and data driven dual-flow coupling sea surface temperature prediction method and system, comprising step S1, physical information data driven dual-flow coupling; step S2, physical equation constraint; step S3, dual-flow loss calculation; the present invention takes multiple physical characteristics as input, designs a physical information data driven dual-flow coupling module, separates the mutual influence of physical information and data drive, learns the changing law of ocean surface temperature by itself, and improves the accuracy of sea surface temperature prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of sea temperature prediction, and in particular relates to a physical information and data driven dual flow coupled sea surface temperature prediction method and system. Background Art

[0002] The most advanced sea surface temperature (SST) prediction method currently combines numerical models with data-driven methods, using the physical information contained in the numerical model to constrain the data-driven neural network model. This structure combines the neural network with the partial differential equations of the numerical model, learns the physical operators in the partial differential equations through the neural network, and then completes the prediction of the sea surface temperature through the calculation of the equation, and designs the loss function to implement the parameter update of the neural network. Therefore, the model can make the neural network follow the laws of physics during the training process, thereby improving the prediction accuracy and interpretability of the model. However, this method has the following problems:

[0003] First, this method only uses longitude and latitude coordinates and sea surface temperature as inputs. However, changes in sea surface temperature involve the joint action of multiple environmental factors. Using only coordinates and past sea surface temperatures as inputs causes the model to be unable to learn complete spatiotemporal physical information, thus limiting the model's predictive ability.

[0004] Second, this method combines partial differential equations to design a physical constraint neural network. However, due to the great uncertainty of ocean movement, using only physical equation constraints cannot accurately describe the law of ocean surface temperature changes, which affects the prediction accuracy of this method.

[0005] Third, this method uses the model prediction results at future moments and the true values ​​at future moments to calculate the loss. It does not consider that the true values ​​at past moments can also supervise the model during the iteration of the physical equations, resulting in the loss design scheme lacking some physical information constraints, thereby reducing the prediction ability. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides a physical information and data-driven dual-flow coupled sea surface temperature prediction method and system. When predicting the sea surface temperature, the physical information flow and the data-driven flow that affect the sea surface temperature change are combined. While maintaining physical rationality, the data information is fully utilized to improve the accuracy of the sea surface temperature prediction.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0008] The physical information and data-driven dual-flow coupled sea surface temperature prediction method includes the following steps:

[0009] Step S1, physical information data drives dual-flow coupling:

[0010] The spatial physical set features are extracted through two branches: physical information flow and data driven flow. and ; and input into the two-stream convolutional LSTM interactive time unit to obtain the spatial and temporal features of the input data;

[0011] Step S2, physical equation constraints:

[0012] The hidden information outputted by the last time unit in the data-driven flow is taken as the final output of the data-driven flow, and the hidden information outputted by the last time unit in the physical information flow is taken as the final output of the physical information flow. , which is used to substitute into the ocean mixed layer heat balance equation to realize the combination of physical laws and data; then the automatic differentiation framework is used to realize automatic differentiation and calculate the predicted results of sea surface temperature;

[0013] Step S3, calculation of dual flow loss:

[0014] The loss consists of two parts. The first part, Loss1, is the MSE loss of the predicted and true sea surface temperature at the nth moment output by the data-driven stream in the dual-stream loss module. The second part, Loss2, is the average of the MSE losses of the predicted and true sea surface temperature at each moment of the model in the first n moments in the physical constraint module.

[0015] Further, step S1 is as follows:

[0016] Step S11, design two branches, physical information flow and data driven flow, to extract spatial physical set features respectively:

[0017] In the data-driven flow, the ocean physical characteristics are used as input. First, several ocean physical characteristics at n moments are decomposed according to the time dimension to obtain the data-driven flow input feature X. Then, the feature at each moment is Each ocean physical feature in is input into different feature processing units, spatial physical features are extracted and aggregated, and spatial physical aggregate features are output. ;

[0018] In the physical information flow, firstly, the longitude x, latitude y and sea surface temperature T at n moments are taken as input Y, and the same operation as the data-driven flow is performed on it to obtain the spatial physical set characteristics ;

[0019] Step S12: and Input to the two-stream convolutional LSTM interactive time unit, the convolutional LSTM unit calculates the forget gate, candidate state and output gate through the convolution operation, and uses these gates to update the unit state and hidden state, thereby capturing the spatial and temporal features of the input data at the same time;

[0020] The dual-stream convolutional LSTM interactive time unit includes a data-driven stream LSTM unit and a physical information stream LSTM unit.

[0021] Further, in step S12, and Input into the two-stream convolutional LSTM interactive time unit for calculation. LSTM is a special type of recurrent neural network suitable for processing and predicting time series data. The LSTM unit consists of multiple gating mechanisms, including input gate, forget gate and output gate. The calculation formula is as follows:

[0022] ( );

[0023] ( );

[0024] ( );

[0025] ( );

[0026] ( );

[0027] Formulas (1)-(5) represent the calculation process of data-driven flow and physical information flow, where is the output of the forget gate, is the weight matrix of the forget gate, is the bias, is the hidden state at the previous moment, is the input at the current moment, is the sigmoid activation function, is the output of the input gate, is the weight matrix of the input gate, is the bias, is a candidate memory cell, is the weight matrix of the candidate states, is the bias of the candidate state, is the updated cell state, is the cell state at the previous moment, is the output gate, is the weight of the output gate, is the bias of the output gate, represents the Hadamard product;

[0028] ( );

[0029] ( );

[0030] Using formula (6) and formula (7), the hidden state of the data-driven stream LSTM unit and the hidden state of the physical information stream LSTM unit at each moment are subtracted to form a new hidden state, which is then transmitted to the LSTM unit at the next moment.

[0031] In the formula and Respectively represent The hidden state of the LSTM unit of the data-driven flow and physical information flow at each moment, and Respectively represent The degree of hidden state generation of the data-driven flow and physical information flow LSTM unit at each moment, and Respectively represent The data-driven flow and physical information flow state of the LSTM unit at each moment.

[0032] Furthermore, in step 2, the details are as follows:

[0033] Get the hidden information of the last time unit output in the data-driven flow ,make As the final output of the data-driven flow;

[0034] Obtain the hidden information output of the last time unit in the physical information flow ,make , as the final output of the physical information flow, is used to substitute into the ocean mixed layer heat balance equation to achieve the combination of physical laws and data and ensure that the model output meets the physical constraints; then the automatic differentiation framework is used to calculate Direction and Partial derivatives of direction and To approach and , the final calculation Complete the pair Approximation; where T represents the sea surface temperature, x and y represent the latitude and longitude information of T, and u and v represent the average latitudinal and longitudinal current velocities at the depth of the mixed layer;

[0035] Then the calculated replace Substituting into the mixed layer heat balance equation, the final sea surface temperature prediction of the model is calculated, as shown in formula (8):

[0036] ( );

[0037] in represents the average ocean temperature of the mixed ocean layer, represents the depth of the mixed layer, and represent the specific heat capacity and density of seawater respectively, and is the vertical current velocity and seawater temperature under the mixed layer, R represents the algebraic expression formed by other variables in the equation;

[0038] Last pair Based on the forward integration method, the discretization solution is performed to finally obtain the desired prediction field , as shown in formula (9):

[0039] ( );

[0040] The integral solver outputs , , …, The prediction results of sea surface temperature are output for the physical information flow.

[0041] Furthermore, in step 3, the hidden information output by the last LSTM unit in the data-driven flow is Prediction results as data-driven streams , Loss1 represents the MSE loss of the predicted value and the true value of the sea surface temperature at the nth moment predicted by the data-driven flow, Loss2 represents the average value of the error between the predicted value and the true value of each moment of the model in the first n moments, and Loss3 adds Loss1 and Loss2 as the final loss, realizing the combination of physical information flow and data-driven flow. The calculation of Loss1, Loss2 and Loss3 is shown in formulas (10)-(12);

[0042] ( );

[0043] ( );

[0044] ( );

[0045] Where m is the number of samples, and in formula (10) and are the predicted and actual sea surface temperature at time n predicted by the data-driven flow in the i-th sample, respectively; in formula (11) and They represent the predicted and actual values ​​of the sea surface temperature output by the model at time j in the i-th sample, respectively. .

[0046] Furthermore, the ocean physical characteristics include downward longwave radiation DLR, upward longwave radiation ULR, downward shortwave radiation DSR, upward shortwave radiation USR, latent heat LH, sensible heat SH, sea surface temperature T, average temperature below the ocean mixed layer Sub_SST, momentum flux u component MCF_u, momentum flux v component MCF_v, vertical ocean current velocity OCV below the ocean mixed layer, and mixed layer thickness MLT.

[0047] The present invention also provides a physical information and data driven dual-flow coupled sea surface temperature prediction system, which is used to implement the physical information and data driven dual-flow coupled sea surface temperature prediction method as described above, including a physical information data driven dual-flow coupling module, a physical equation constraint module and a dual-flow loss module; the physical information data driven dual-flow coupling module includes two branches, a physical information flow and a data driven flow, which respectively include a physical information flow LSTM unit and a data driven flow LSTM unit, which are used to obtain the spatial characteristics and temporal characteristics of the input data;

[0048] The physical equation constraint module is used to obtain the hidden information of the last time unit output in the physical information flow as the final output of the physical information flow , substitute it into the ocean mixed layer heat balance equation to realize the combination of physical laws and data; then realize automatic differentiation through the automatic differentiation framework, and calculate the predicted results of sea surface temperature;

[0049] The dual-flow loss module is used to calculate losses.

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

[0051] First, the present invention uses multiple physical features as input, takes into account the joint effects of multiple environmental factors involved in changes in ocean surface temperature, learns comprehensive physical information through a large number of features, and improves prediction capabilities.

[0052] Second, the present invention designs a physical information data-driven dual-stream coupling module, which separates the mutual influence of physical information and data-driven, enables the model to pay more attention to the non-physical information in the data-driven flow, learn the changing laws of ocean surface temperature by itself, and improves the accuracy of the prediction.

[0053] Third, the present invention not only uses the model prediction results at future moments and the true values ​​at future moments to calculate the loss, but also considers using the predicted values ​​at past moments and the true values ​​at past moments in the iterative process of the physical equation to supervise the model, thereby strengthening the physical information constraints of the loss design scheme and improving the prediction ability. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0055] Figure 1 It is a schematic diagram of the system structure of the present invention;

[0056] Figure 2 A training loss variation curve during training for one embodiment of the present invention;

[0057] Figure 3 A comparison diagram of the prediction result and the true value of an embodiment of the present invention;

[0058] Figure 4 is an MSE loss curve of the prediction process of an embodiment of the present invention;

[0059] Figure 5 The RMSE loss curves of the first day predicted by the present invention, U-Net, Conv-LSTM and Swin-Transformer;

[0060] Figure 6 The RMSE loss curves of the third day predicted by the present invention, U-Net, Conv-LSTM and Swin-Transformer;

[0061] Figure 7 The RMSE loss curves of the present invention, U-Net, Conv-LSTM and Swin-Transformer for predicting the 7th day;

[0062] Figure 8 The RMSE loss curves of the present invention, U-Net, Conv-LSTM and Swin-Transformer for predicting the 10th day;

[0063] Fig. 9 The RMSE loss curves of the present invention, U-Net, Conv-LSTM and Swin-Transformer predicted on the 14th day;

[0064] Fig.10The MAE loss curves of the first day predicted by the present invention, U-Net, Conv-LSTM and Swin-Transformer.

[0065] Fig.11 The MAE loss curves of the third day predicted by the present invention, U-Net, Conv-LSTM and Swin-Transformer;

[0066] Fig.12 The MAE loss curves of the present invention, U-Net, Conv-LSTM and Swin-Transformer predicted on the 7th day;

[0067] Fig.13 The MAE loss curves of the present invention, U-Net, Conv-LSTM and Swin-Transformer predicted on the 10th day;

[0068] Fig.14 The MAE loss curves for the 14th day predicted by the present invention, U-Net, Conv-LSTM and Swin-Transformer;

[0069] Fig.15 The present invention and U-Net, Conv-LSTM and Swin-Transformer predict the first day curve;

[0070] Fig.16 The present invention and U-Net, Conv-LSTM and Swin-Transformer predict the third day curve;

[0071] Fig.17 The present invention predicts the 7th day with U-Net, Conv-LSTM and Swin-Transformer curve;

[0072] Fig.18 The present invention and U-Net, Conv-LSTM and Swin-Transformer predict the 10th day curve.

[0073] Fig.19 The present invention and U-Net, Conv-LSTM and Swin-Transformer predict the 14th day curve. DETAILED DESCRIPTION

[0074] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0075] Combination Figure 1 As shown, the present invention designs a physical information data driven dual-flow coupling module, a physical equation constraint module and a dual-flow loss module, and a physical information and data driven dual-flow coupling sea surface temperature prediction method, comprising the following steps:

[0076] Step S1, physical information data drives dual-flow coupling:

[0077] In the physical information data driven dual flow coupling module, the physical information flow and the data driven flow are designed. The spatial physical collection features are extracted through the two branches of the physical information flow and the data driven flow. and ; and input it into the two-stream convolutional LSTM interactive time unit to obtain the spatial and temporal features of the input data.

[0078] The details are as follows:

[0079] Step S11, design two branches, namely, physical information flow and data driven flow, to extract spatial physical set features respectively.

[0080] In the data-driven flow, the ocean physical characteristics are used as input. Several ocean physical characteristics at each moment are decomposed according to the time dimension to obtain the data-driven flow input characteristics , and then the characteristics of each moment Each ocean physical feature in is input into different feature processing units (different convolutional layers) to extract and aggregate spatial physical features and output spatial physical aggregate features. .

[0081] Among them, the ocean physical characteristics include downward longwave radiation DLR, upward longwave radiation ULR, downward shortwave radiation DSR, upward shortwave radiation USR, latent heat LH, sensible heat SH, sea surface temperature T, average temperature below the ocean mixed layer Sub_SST, momentum flux u component MCF_u, momentum flux v component MCF_v, vertical ocean current velocity below the ocean mixed layer OCV, and mixed layer thickness MLT.

[0082] As an example, Downward long wave radiation , upward long wave radiation , downward shortwave radiation , upward shortwave radiation , latent heat , Sensible heat , sea surface temperature at 108 meters , momentum flux u component , momentum flux v component , vertical ocean current velocity at 108 meters , mixed layer thickness ,longitude ,latitude 14 ocean physical features (each feature dimension is .in and is the length and width of the space) as input, where Respectively represent the downward long-wave radiation values ​​at time 0, 1, ..., n-1, represent the upward long-wave radiation at time 0, 1, ..., n-1 respectively, They represent the thickness of the mixing layer at the 0th, 1st, ..., n-1th moment respectively. The meanings of other letters are similar and will not be elaborated here.

[0083] First, All physical characteristics at a moment are decomposed according to the time dimension, and the dimension is Data-driven streaming input features , , and then the features of each moment Input different feature processing units respectively to For example, , , …, Input into different convolutional layers to extract the unique spatial physical features of the physical element , ,… . Then , , …, Perform aggregation operations to obtain spatial physical collection features , the dimension is .

[0084] Similarly, in the physical information flow, first Longitude at that moment ,latitude and sea surface temperature As input , and perform the same operation as the data-driven flow to obtain the spatial physical set characteristics , the dimension is .in Figure 1 In the example, the longitudes at the 0th, ..., n-1th moments are ,latitude and sea surface temperature ,longitude ,latitude and sea surface temperature as input.

[0085] Step S12: and The input is sent to the two-stream convolutional LSTM interactive temporal unit. The convolutional LSTM unit calculates the forget gate, candidate state, and output gate through convolution operation, and uses these gates to update the cell state and hidden state, thereby capturing both the spatial and temporal features of the input data.

[0086] The dual-stream convolutional LSTM interactive time unit includes a data-driven stream LSTM unit and a physical information stream LSTM unit.

[0087] In step S12, and Input into the two-stream convolutional LSTM interactive time unit for calculation. LSTM is a special type of recurrent neural network, which is particularly suitable for processing and predicting time series data. The LSTM unit consists of multiple gating mechanisms, mainly including input gate, forget gate and output gate. The calculation formula is as follows:

[0088] (1);

[0089] (2);

[0090] (3);

[0091] (4);

[0092] (5);

[0093] Formulas (1)-(5) represent the calculation process of data-driven flow and physical information flow, where is the output of the forget gate, is the weight matrix of the forget gate, is the bias, is the hidden state at the previous moment, is the input at the current moment, is the sigmoid activation function, is the output of the input gate, is the weight matrix of the input gate, is the bias, is a candidate memory cell, is the weight matrix of the candidate states, is the bias of the candidate state, is the updated cell state, is the cell state at the previous moment, is the output gate, is the weight of the output gate, is the bias of the output gate,' ' represents the Hadamard product;

[0094] (6);

[0095] (7);

[0096] Using formula (6) and formula (7), the hidden state of the data-driven stream LSTM unit and the hidden state of the physical information stream LSTM unit at each moment are subtracted to form a new hidden state, which is then transmitted to the LSTM unit at the next moment.

[0097] In the formula and Respectively represent The hidden state of the LSTM unit of the data-driven flow and physical information flow at each moment, and Respectively represent The degree of hidden state generation of the data-driven flow and physical information flow LSTM unit at each moment, and Respectively represent The data-driven flow and physical information flow state of the LSTM unit at each moment.

[0098] Step S2, physical equation constraints:

[0099] The hidden information outputted by the last time unit in the data-driven flow is taken as the final output of the data-driven flow, and the hidden information outputted by the last time unit in the physical information flow is taken as the final output of the physical information flow. , which is used to substitute into the ocean mixed layer heat balance equation to realize the combination of physical laws and data; then the automatic differentiation framework is used to realize automatic derivation and calculate the predicted results of sea surface temperature.

[0100] In step 2, the details are as follows:

[0101] Get the hidden information of the last time unit output in the data-driven flow ,make As the final output of the data-driven flow;

[0102] Obtain the hidden information output of the last time unit in the physical information flow ,make , as the final output of the physical information flow, is used to substitute into the ocean mixed layer heat balance equation to achieve the combination of physical laws and data and ensure that the model output meets the physical constraints; then the automatic differentiation framework is used to calculate Direction and Partial derivatives of direction and To approach and , the final calculation Complete the pair of the approximation; is the sea surface temperature, , express The longitude and latitude information, and represents the latitudinal and meridional current velocities averaged over the mixed layer depth;

[0103] Then the calculated replace Substituting into the mixed layer heat balance equation, the final sea surface temperature prediction of the model is calculated, as shown in formula (20):

[0104] (8);

[0105] in represents the average ocean temperature of the mixed ocean layer, represents the depth of the mixed layer, and represent the specific heat capacity and density of seawater respectively, and are the vertical current velocity and seawater temperature below the mixed layer, Represents the algebraic expression formed by other variables in the equation;

[0106] Last pair Based on the forward integration method, the discretization solution is performed to finally obtain the desired prediction field , as shown in formula (9):

[0107] (9);

[0108] The integral solver outputs , , …, The prediction results of sea surface temperature are output for the physical information flow.

[0109] Step S3, calculation of dual flow loss:

[0110] The loss consists of two parts. The first part, Loss1, is the MSE loss of the predicted and true sea surface temperature at the nth moment output by the data-driven stream in the dual-stream loss module. The second part, Loss2, is the average of the MSE losses of the predicted and true sea surface temperature at each moment of the model in the first n moments in the physical constraint module.

[0111] Specifically, in step 3, the hidden information output by the last LSTM unit in the data-driven flow is Prediction results as data-driven streams , Loss1 represents the first The MSE loss of the predicted value and the true value of the sea surface temperature at the moment, Loss2 represents the MSE loss of the predicted value and the true value of the sea surface temperature at the moment, and Loss The average value of the error between the predicted value and the true value of the model at each moment. Loss3 adds Loss1 and Loss2 as the final loss, realizing the combination of physical information flow and data-driven flow. The calculation of Loss1, Loss2 and Loss3 is shown in formulas (10)-(12);

[0112] (10);

[0113] (11);

[0114] (12);

[0115] in is the number of samples, in formula (10) and Respectively Data-driven flow prediction in samples The predicted and actual sea surface temperature at the time; in formula (11) and Respectively represent In the samples The predicted and actual sea surface temperature output by the model at the moment, .

[0116] When training the network model of the present invention, a loss threshold ε is preset. In each training batch, the predicted value of the model is calculated by forward propagation, and the loss loss3 is calculated; the parameters of the model are updated by back propagation; and it is checked whether loss3 is less than ε. If loss3 is less than ε, the training is stopped or other measures are taken (such as adjusting the learning rate, saving the model, etc.). If loss3 is greater than or equal to ε, the training is continued.

[0117] As an example of the present invention, the following introduces training (whether the training is end-to-end, sample description, description of neural network hyperparameters, description of optimization method).

[0118] In order to verify the technical effect, this scheme selected historical ocean physical characteristics data of a coastal area as the input data set. The time range of the data is from January 1, 2010 to December 31, 2019, and the spatial resolution is 0.25°×0.25°. Each data set contains daily observation data such as sea surface temperature, downward longwave radiation, upward longwave radiation, latent heat and sensible heat.

[0119] For each data set, this method uses 80% of the samples as the training set, 10% of the samples as the validation set, and the remaining 10% as the test set. The model uses the root mean square error (RMSE), mean absolute error (MAE) and determination coefficient ( ) are used for evaluation. These model evaluation indicators are commonly used technical means in this field and will not be described in detail. The root mean square error (RMSE) measures the predicted value. With the true value The smaller the value, the higher the prediction accuracy. The specific formula is:

[0120] ;

[0121] The mean absolute error (MAE) is the average of the absolute errors between the predicted value and the true value. The specific formula is:

[0122] ;

[0123] Coefficient of determination ( ) represents the model's ability to explain actual data changes. The specific formula is:

[0124] .

[0125] The model receives data from the past 7 days to predict the sea surface temperature for the next 1, 3, 7, 10, and 14 days. The experiment was conducted on a single NVIDIA Titan RTX GPU, and the network was trained for 300 epochs using the Adam optimizer, with the convolution kernel size set to 3 and the minimum batch size set to 64. During training, the learning rate was adjusted to 1e-4. During training, the training loss (training loss) changes as follows Figure 2 As shown in the figure, it can be seen that after 300 epochs of training, the final Loss of the model remains around 0.3.

[0126] Use the trained model to make predictions and get a comparison between the predicted results and the true values. Figure 3 As shown. It can be seen that the predicted value and the true value remain within a very small gap. The prediction loss of the prediction process, namely the MSE loss, is as follows Figure 4 shown.

[0127] In order to prove the effectiveness of the proposed scheme, this scheme also selected U-Net, Conv-LSTM and Swin-Transformer for comparative experiments with the proposed method. The detailed information of the experimental results is shown in Tables 1 to 3 below, which are about RMSE, MAE and comparison.

[0128] Table 1 Comparison of RMSE between the present invention and the existing methods

[0129]

[0130] Table 2 Comparison of MAE between the present invention and the existing method

[0131]

[0132] Table 3 The present invention and the existing method Comparison

[0133]

[0134] In all forecast time scales, the method of the present invention always shows the lowest RMSE value, especially in the 1-day and 3-day forecasts, which are 0.3432 and 0.6249 respectively, which are significantly better than U-Net (0.3701 and 0.6697), Conv-LSTM (0.3789 and 0.718) and Swin-Transformer (0.368 and 0.6846). This shows that the method of the present invention has higher accuracy in short-term forecasts.

[0135] The predicted RMSE loss curve is as follows Figures 5 to 9 The present invention also performs well in MAE index, with 1-day and 3-day MAE values ​​of 0.2351 and 0.4679 respectively, both lower than other methods. This further proves the effectiveness of the present invention in reducing prediction errors, especially in short-term predictions, showing stronger stability and reliability. The predicted MAE loss curve is shown in Figure 10 to Figure 14 As shown. In terms of indicators, the proposed method also performs best in 1-day and 3-day predictions, which are 0.9846 and 0.9408 respectively, showing a good ability to explain data variation. The values ​​are generally low, especially in the longer-term forecasts, showing a clear downward trend. The curve diagram is as follows Figure 15 to Figure 19 shown.

[0136] In summary, the method of the present invention is superior to U-Net, Conv-LSTM and Swin-Transformer in various evaluation indicators, especially in short-term prediction, showing higher accuracy and stability, which proves its potential and advantages in practical applications.

[0137] As an example, during the model prediction process, a sample is randomly selected and the input parameters of the model are obtained as shown in Table 4 below.

[0138] Table 4 Input parameters of the model

[0139]

[0140] Among them, the first column is the input ocean physical characteristics, the second column is the average value of these physical characteristics, and the third column is the units of these physical characteristics.

[0141] The output sea surface temperature prediction parameters are as follows:

[0142] Table 5 Sea surface temperature prediction and true value parameters

[0143]

[0144] The table shows the output sea surface temperature prediction values ​​and the true values. The loss value of the model in short-term predictions (such as 1 day and 3 days) is relatively low, which shows its effectiveness and reliability in short-term predictions.

[0145] As another embodiment of the present invention, a physical information and data driven dual flow coupled sea surface temperature prediction system is designed to implement the physical information and data driven dual flow coupled sea surface temperature prediction method as described above, such as Figure 1 As shown, it includes a physical information data driven dual-stream coupling module, a physical equation constraint module and a dual-stream loss module; the physical information data driven dual-stream coupling module includes two branches, a physical information flow and a data driven flow, which respectively include a physical information flow LSTM unit and a data driven flow LSTM unit, which are used to obtain the spatial characteristics and temporal characteristics of the input data.

[0146] The physical equation constraint module is used to obtain the hidden information of the last time unit output in the physical information flow as the final output of the physical information flow , substitute it into the ocean mixed layer heat balance equation to realize the combination of physical laws and data; then realize automatic differentiation through the automatic differentiation framework, and calculate the predicted results of sea surface temperature.

[0147] The dual-flow loss module is used to calculate the loss, including the first The MSE loss Loss1 between the predicted value and the true value of the sea surface temperature at the moment, The average value of the error between the predicted value and the true value of the model at each moment is Loss2, and the model loss Loss3 is obtained by adding Loss1 and Loss2.

[0148] The data processing process of each module and the calculation of loss can be found in the previous records and will not be repeated here.

[0149] It should be noted that the model mentioned in the present invention refers to the physical information and data-driven dual-flow coupled sea surface temperature prediction system constructed by the present invention, which is used for sea surface temperature prediction, including a physical information data-driven dual-flow coupling module, a physical equation constraint module and a dual-flow loss module.

[0150] In summary, the present invention (1) constructs the data-driven flow and the physical information flow separately in the physical information data-driven dual-flow coupling module, separates the mutual influence of physical information and data-driven, makes the model pay more attention to the non-physical information in the data-driven flow, and learns the changing law of the sea surface temperature by itself, thereby improving the accuracy of model prediction. (2) In the dual-flow loss module, not only the predicted value and the real value of the future moment in the data-driven flow are calculated, but also the predicted value and the real value of the past moment are calculated in the physical information flow. Finally, the losses of the data-driven flow and the physical information flow are combined to supervise the model, thereby improving the prediction ability of the model.

[0151] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Any changes, modifications, additions or substitutions made by ordinary technicians in this technical field within the essential scope of the present invention should fall within the protection scope of the present invention.

Claims

1. Physical information and data-driven dual-flow coupled sea surface temperature prediction method, characterized in that: The following steps are involved: Step S1: Physical information data drives dual-flow coupling: The spatial physical set feature X' is extracted through the two branches of physical information flow and data driven flow. i and Y′ i ; and input into the two-stream convolutional LSTM interactive time unit to obtain the spatial and temporal features of the input data; Step S2, physical equation constraints: The hidden information output by the last time unit in the data-driven flow is obtained as the final output K of the data-driven flow n ; The hidden information outputted by the last time unit in the physical information flow is taken as the final output u′ of the physical information flow, and u′ is used to substitute into the ocean mixed layer heat balance equation to realize the combination of physical laws and data; then u′ is automatically differentiated through the automatic differentiation framework, and the predicted result of the sea surface temperature is calculated; Step S3, calculation of dual flow loss: The loss consists of two parts. The first part, Loss1, is the MSE loss of the hidden state and the true value of the first n moments of the data-driven flow output; the second part, Loss2, is the average of the MSE losses of the predicted value of the sea surface temperature and the true value at each moment obtained by the physical equation constraints described in step 2 for the first n moments.

2. The physical information and data-driven dual-flow coupled sea surface temperature prediction method according to claim 1 is characterized in that: Step S1 is as follows: Step S11, design two branches, physical information flow and data driven flow, to extract spatial physical set features respectively: In the data-driven flow, the ocean physical characteristics are used as input. First, several ocean physical characteristics at n moments are decomposed according to the time dimension to obtain the data-driven flow input feature X. Then, the feature X at each moment is i Each ocean physical feature in is input into different feature processing units, the spatial physical features are extracted and aggregated, and the spatial physical set feature X′ is output. i ; In the physical information flow, firstly, the longitude x, latitude y and sea surface temperature T at n moments are taken as input Y, and the same operation as the data-driven flow is performed on it to obtain the spatial physical set feature Y′ i ; Step S12: X' i and Y′ i Input to the two-stream convolutional LSTM interactive time unit, the convolutional LSTM unit calculates the forget gate, candidate state and output gate through the convolution operation, and uses these gates to update the unit state and hidden state, thereby capturing the spatial and temporal features of the input data at the same time; The dual-stream convolutional LSTM interactive time unit includes a data-driven stream LSTM unit and a physical information stream LSTM unit.

3. The physical information and data-driven dual-flow coupled sea surface temperature prediction method according to claim 2 is characterized in that: In step S12, X' i and Y′ i Input into the two-stream convolutional LSTM interactive time unit for calculation. LSTM is a recursive neural network used to process and predict time series data. The LSTM unit consists of multiple gating mechanisms, including input gate, forget gate and output gate. The calculation formula is as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ) (1); i t =σ(W i ·[h t-1 ,x t ]+b i ) (2); the T =σ(W o ·[h t-1 ,x t ]+b o ) (5); Formulas (1)-(5) represent the calculation process of data-driven flow and physical information flow, where f t is the output of the forget gate, W f is the weight matrix of the forget gate, b f is the bias, h t-1 is the hidden state at the previous moment, x t is the input at the current moment, σ is the sigmoid activation function, i t is the output of the input gate, W i is the weight matrix of the input gate, b i is the bias, is a candidate memory cell, W c is the weight matrix of the candidate state, b c is the bias of the candidate state, C t is the updated cell state, C t-1 is the unit state at the previous moment, o t is the output gate, W o is the weight of the output gate, b o is the bias of the output gate, * represents the Hadamard product; Using formula (6) and formula (7), the hidden state of the data-driven flow LSTM unit and the hidden state of the physical information flow LSTM unit at each moment are subtracted to form a new hidden state, which is then transmitted to the LSTM unit at the next moment; where and They represent the hidden states of the data-driven flow and physical information flow LSTM units at the tth moment, respectively. and They represent the hidden state generation degree of the LSTM unit of the data-driven flow and physical information flow at the Tth moment, respectively. and They represent the states of the data-driven flow and physical information flow LSTM units at the tth moment respectively.

4. The physical information and data-driven dual-flow coupled sea surface temperature prediction method according to claim 1 is characterized in that: In step 2, the details are as follows: Get the hidden information of the last time unit output in the data-driven flow make As the final output of the data-driven flow; Obtain the hidden information output of the last time unit in the physical information flow make As the final output of the physical information flow, it is used to substitute into the ocean mixed layer heat balance equation to achieve the combination of physical laws and data and ensure that the model output meets the physical constraints; then the partial derivatives of u′ in the x and y directions are calculated through the automatic differentiation framework. and To approach and Final calculation Complete the pair Approximation; where T represents the sea surface temperature, x and y represent the latitude and longitude information of T, and u and v represent the average latitudinal and longitudinal current velocities at the depth of the mixed layer; Then the calculated replace Substituting into the mixed layer heat balance equation, the final sea surface temperature prediction of the model is calculated, as shown in formula (8): Where T m is the average ocean temperature of the ocean mixed layer, h m represents the depth of the mixed layer, C p and ρ represent the specific heat capacity and density of seawater respectively, w e and T d is the vertical current velocity and seawater temperature under the mixed layer, R represents the algebraic expression formed by other variables in the equation; Last pair Based on the forward integration method, the integral is solved and the desired prediction field T′ is finally obtained. m+1 , as shown in formula (9): The integral solution operation outputs T′1, T′2, …, T′ n The prediction results of sea surface temperature are output for the physical information flow.

5. The physical information and data-driven dual-flow coupled sea surface temperature prediction method according to claim 1 is characterized in that: In step 3, the hidden information output by the last LSTM unit in the data-driven flow The prediction result K as a data-driven flow n , Loss1 represents the MSE loss of the hidden state and the true value at the nth moment predicted by the data-driven flow, Loss2 represents the average MSE loss of the predicted value and the true value at each moment obtained by the physical equation constraint described in step 2 in the first n moments, and Loss3 adds Loss1 and Loss2 as the final loss, realizing the combination of physical information flow and data-driven flow. The calculation of Loss1, Loss2 and Loss3 is shown in formulas (10)-(12); Loss3 = Loss1 + Loss2 (12); Where m is the number of samples, K in formula (10) in and T in are the sea level at time n predicted by the data-driven flow in the i-th sample. Surface temperature prediction value and actual value; T ij and T ij They represent the predicted value and actual value of the sea surface temperature output by the model at time j in the ith sample, respectively, j = 0, 1, 2, ..., n.

6. The physical information and data-driven dual-flow coupled sea surface temperature prediction method according to claim 2 is characterized in that: The ocean physical characteristics include downward longwave radiation DLR, upward longwave radiation ULR, downward shortwave radiation DSR, upward shortwave radiation USR, latent heat LH, sensible heat SH, sea surface temperature T, average temperature below the ocean mixed layer Sub_SST, momentum flux u component MCF_u, momentum flux v component MCF_v, vertical ocean current velocity OCV below the ocean mixed layer, and mixed layer thickness MLT.

7. Physical information and data driven dual-flow coupled sea surface temperature prediction system, characterized by: Used to implement the physical information and data-driven dual-flow coupling sea surface temperature prediction method as described in any one of claims 1 to 6, including a physical information data-driven dual-flow coupling module, a physical equation constraint module and a dual-flow loss module; the physical information data-driven dual-flow coupling module includes two branches, a physical information flow and a data-driven flow, which respectively include a physical information flow LSTM unit and a data-driven flow LSTM unit, for obtaining the spatial characteristics and temporal characteristics of the input data; The physical equation constraint module is used to obtain the hidden information outputted by the last time unit in the physical information flow as the final output u′ of the physical information flow, and substitute u′ into the ocean mixed layer heat balance equation to realize the combination of physical laws and data; then, automatic differentiation is realized through the automatic differentiation framework, and the predicted result of the sea surface temperature is calculated; The dual-flow loss module is used to calculate losses.

Images