Sea surface temperature prediction method based on space-time expert mixture and exogenous variables and corresponding electronic equipment
Through the sea surface temperature prediction method based on space-time experts' mixed and exogenous variables, the problem of insufficient generalization ability of existing models on different regions and time scales is solved, and more accurate and uniform sea surface temperature prediction is achieved.
Patent Information
- Application Number
- CN202411889790.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-20
- Publication Date
- 2025-04-11
AI Technical Summary
The existing sea surface temperature prediction model has insufficient generalization ability on different regions and time scales, resulting in uneven distribution of predictive performance.
The sea surface temperature prediction method based on space-time expert mixing and exogenous variables is adopted. Through the space-time expert module, expert routing module and expert mixing module, combined with the cross-variable attention layer and graph neural network, the spatiotemporal characteristics of sea surface temperature data are extracted and weighted summed to accurately predict sea surface temperature.
By selecting the appropriate expert model for training, it can avoid dynamic environmental deviations in different marine areas, improve the prediction accuracy of specific points, and maintain good prediction performance in complex marine environments.
Smart Images
Figure CN119808568A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of ocean surface temperature prediction, and in particular to a method for predicting ocean surface temperature based on spatiotemporal expert mixing and exogenous variables. Background Art
[0002] The ocean occupies about three quarters of the Earth's surface. Sea surface temperature (SST) has an important impact on the global ecological environment and climate. The prediction of SST is of great significance for predicting global environmental and climate change. At present, the methods for predicting sea surface temperature can be roughly divided into two categories, namely physical models and data-driven models. Physical models are mainly based on traditional mathematical calculation methods, in which the various processes of ocean dynamics described are based on simplified equations and parameters. Physical models require a lot of domain knowledge to design formulas, and there are problems such as large uncertainty in the parameterization process and high computational complexity. Data-driven models are based on a large amount of observational data, using traditional machine learning and deep learning models to automatically learn the mapping relationship in the data. At the same time, data-driven models do not require or rarely require knowledge of the ocean and atmosphere and the corresponding equations.
[0003] With the rapid development of deep learning technology, prediction methods based on spatiotemporal features have gradually become a research hotspot. The complexity of spatiotemporal data requires that the model can capture both spatial and temporal dependencies. To this end, many studies have adopted long short-term memory (LSTM) networks and convolutional neural networks (CNNs) to capture the temporal and spatial dimension characteristics of sea surface temperature data, respectively. However, existing spatiotemporal prediction models are usually trained for a unified spatiotemporal model for the entire region, ignoring the characteristics of different regions or time scales, which may lead to insufficient generalization of the model. The expert mixture model provides an effective way to solve this problem. The expert mixture model combines multiple "expert" models and dynamically selects the most suitable expert for processing according to the characteristics of the input data, thereby selecting different model paths for different spatiotemporal data and exogenous variables. This method can flexibly respond to changes in different time scales and spatial characteristics, and improves the ability to model spatiotemporal data. The expert mixture model has been successfully applied in many fields, especially in processing large-scale data, nonlinear relationships and high-dimensional features. In addition, in addition to being affected by the spatiotemporal influence of the space in which it is located, the sea surface temperature is also affected by other variables, which can provide valuable information for sea surface temperature prediction, such as the longitude and latitude wind speed of the sea surface. Summary of the invention
[0004] The object of the present invention is to provide a sea surface temperature prediction method based on spatiotemporal expert mixing and exogenous variables, which can solve the problem of uneven distribution of prediction performance.
[0005] Based on the above purpose, the present invention adopts the following technical solution:
[0006] A sea surface temperature prediction method based on spatiotemporal expert mixing and exogenous variables, characterized by comprising the following steps:
[0007] S1. Data acquisition and preprocessing: Acquire ocean multivariate observation data and preprocess the data; ocean multivariate observation data includes sea surface temperature data and other exogenous variable data;
[0008] S2. Model building: building a sea surface temperature prediction model, including a spatiotemporal expert module, an expert routing module, and an expert hybrid module;
[0009] S3, obtaining a spatiotemporal feature vector: inputting the preprocessed data in step S1 into the spatiotemporal expert module in step S2, extracting the influence of exogenous variables on the sea surface temperature and the spatiotemporal features of the sea surface temperature data itself, and obtaining a spatiotemporal feature vector.
[0010] S4. Obtain path weights: Use the preprocessed sea surface temperature data in step S1 as the input of the expert routing module in step S2, and use the routing function to generate path weights to determine the probability of selecting each spatiotemporal expert module for the current data.
[0011] S5, output prediction: according to the path weight obtained in step S4, the outputs of all expert models obtained in step S3 are weighted and summed to obtain the final output;
[0012] S6, model optimization: According to the prediction output result obtained in step S5, the model is trained by minimizing the mean square error between the prediction result and the true value, and all learnable parameters in the model are optimized and updated to make the model output closer to the true value.
[0013] Preferably, the specific process of data acquisition and preprocessing in step S1 includes:
[0014] The specific process of the preprocessing includes screening the data points according to the selected exogenous variable data and the sea mark temperature data in the study area, and then performing Z-score standardization on each variable data:
[0015]
[0016] In the formula, x represents the data of different variables, Z represents the normalized data, μ and σ represent the mean and standard deviation of different variables respectively;
[0017] After processing the training set, we can get data X with the shape of T×N×C, where T is the total time length, N is the valid points in the study area, and C is the number of variables.
[0018] Preferably, the specific process of building the model in step S2 includes:
[0019] There are multiple spatiotemporal expert modules, each of which is modeled for a different time scale; each spatiotemporal expert module consists of a cross-variable attention layer and a spatiotemporal representation layer, the cross-variable attention layer uses the SST data and other exogenous variables observed in step S1 for sequence-level embedding, and then performs feature interaction through the attention mechanism to extract the impact of exogenous variables on SST; the spatiotemporal representation layer extracts spatial features through a graph neural network and uses a one-dimensional hole convolution and a gated network to extract temporal features; different spatiotemporal expert modules use convolution kernels of different sizes to learn spatiotemporal features at different time scales and capture multi-scale temporal dependencies and pattern changes; the expert routing module is responsible for selecting appropriate spatiotemporal experts, and the expert mixing module performs weighted summation on the outputs of the selected spatiotemporal expert modules through the weights provided by the routing module to obtain the output result, thereby achieving accurate prediction of future sea surface temperature.
[0020] Preferably, the specific process of obtaining the spatiotemporal feature vector in step S3 includes:
[0021] S301, data encoding: Use a multi-layer perceptron to encode the pre-processed data X along the time dimension into a global label G∈R h×N×C ; where h is the hyperparameter of the global tag size and C is the number of variables;
[0022] S302, attention mechanism interaction: Apply the attention mechanism to the global mark between different variables at each point to interact, so as to extract the impact of exogenous variables on the sea surface temperature; and globally mark the sea surface temperature after interaction G SST ∈R h×N Splice to each point of sea surface temperature in the preprocessed data X SST ∈R T×N , get the sea surface temperature data X with global label SST ∈R (T +h)×N ;
[0023] S303, obtain spatial feature vector: apply the globally labeled SST data to the graph neural network to obtain the feature vector H after spatial interaction S ;
[0024] S304, obtaining a spatiotemporal feature vector: the spatial feature vector H obtained in step S303 S A one-dimensional dilated convolution followed by a gated linear unit (GLU) is used to capture the information of the sea surface temperature in the time dimension; finally, the spatiotemporal feature vector H of each point is obtained. ST .
[0025] Preferably, the specific process of obtaining the spatial feature vector in step S303 includes:
[0026] Construct the adjacency matrix E∈R N×N , initialize E by generating two learnable vectors vec1 and vec2:
[0027] E = Softmax(ReLU(vec1×vec2))
[0028] Spatial interactions according to the adjacency matrix:
[0029]
[0030] Where σ is the sigmoid activation function, and W is the learnable matrix;
[0031] H i Aggregation and X i The node features adjacent to the spatial node are finally obtained. S ;
[0032] Preferably, the specific process of obtaining the path weight in step S4 includes:
[0033] Add a noise component to the routing function and multiply the input by a trainable weight matrix W before applying the Softmax function g And add adjustable Gaussian noise:
[0034]
[0035] Where R(·) represents the entire routing function, W r and W noise is a learnable parameter used for weight generation; path weights are selected, the first K path weights are retained and the rest are set to 0.
[0036] Preferably, add a balanced expert loss function L eq :
[0037]
[0038] L eq (X SST )=W eq CV(Equalization(X SST )) 2
[0039] Where W eq is an adjustable scaling factor and CV is the operation that calculates the coefficient of variation.
[0040] Preferably, the specific process of outputting the prediction in step S5 includes:
[0041] The final prediction output is obtained by weighted summing the outputs of all expert models, with the weights provided by the routing module; the expression of the final output is:
[0042]
[0043] In the formula, R(X SST ) i and E i (X) denotes the output of the routing module and the output of the i-th expert model for a given input, respectively.
[0044] Preferably, the final output obtained in step S5 is combined The model is trained with the true value Y by minimizing The total loss function can be written as:
[0045]
[0046] Where θ is all learnable parameters in the model;
[0047] The data obtained in step S1 is used as input and model training is performed, and the trained model can be applied to sea surface temperature prediction.
[0048] An electronic device includes a storage device and a processor. A computer program is stored in the storage device. When the processor executes the computer program, any step in any of the above-mentioned sea surface temperature prediction methods based on spatiotemporal expert mixing and exogenous variables is implemented.
[0049] The beneficial effects of the present invention include:
[0050] The present invention uses a dedicated routing network to select appropriate expert models for training, which can avoid the problem of uneven distribution of prediction performance in different regions caused by ignoring the dynamic environment deviation in different ocean areas in traditional methods, eliminate the deviation between individual areas, and improve the prediction accuracy of specific points.
[0051] The present invention sets up an attention mechanism, which can add exogenous variable information to the sea surface temperature data, can maintain good prediction performance in complex ocean environments, and is capable of handling sea areas with large seasonal and spatial variations and high sensitivity to atmospheric conditions and human activities. It significantly improves the accuracy and reliability of the sea surface temperature prediction model in complex ocean dynamic environments. BRIEF DESCRIPTION OF THE DRAWINGS
[0052] Figure 1 It is a schematic diagram of the method flow of the present invention;
[0053] Figure 2It is a schematic diagram of the overall structure of the model of the present invention.
[0054] Figure 3 Schematic diagram of the cross-variable attention layer of the present invention;
[0055] Figure 4 Schematic diagram of the spatiotemporal expert module of the present invention. DETAILED DESCRIPTION
[0056] Example 1
[0057] The following is a further explanation of the present invention in conjunction with specific embodiments. Figure 1 As shown, this embodiment is a sea surface temperature prediction method based on spatiotemporal expert mixing and exogenous variables, comprising the following steps:
[0058] S1. Data acquisition and preprocessing: Acquire ocean multivariate observation data and preprocess the data; ocean multivariate data includes sea surface temperature data and other exogenous variable data. In practical applications, data from different sources can be selected according to specific needs.
[0059] In this embodiment, the sea surface temperature (SST) data set is selected from the OISST V2.1 (Optimum Interpolation Sea Surface Temperature, OISST) data set, and other exogenous variable data select the meridional sea surface wind speed (VSSW) and latitudinal sea surface wind speed (USSW) data sets provided by CCMP as exogenous variables; the spatial resolution of these data sets is 0.25°, and the temporal resolution is uniformly based on a 1-day scale.
[0060] The Bohai Sea was selected as the case study area in this example due to its diverse and challenging environment. The Bohai Sea is a semi-enclosed body of water in northeastern China with an area of approximately 77,000 square kilometers and an average depth of 18 meters. Its sea temperature is significantly affected by river discharge, especially the flow of the Yellow River, resulting in significant seasonal and spatial variability. The shallow depth and semi-enclosed nature of the area also make it highly sensitive to atmospheric conditions and human activities. Table 1 gives basic statistical information on different variables in the Bohai Sea.
[0061] Table 1 Basic statistical information of Bohai Sea
[0062] variable Mean Standard Deviation Minimum 25％ 50％ 75％ Maximum SST(℃) 13.18 8.34 -1.80 5.09 13.15 21.15 31.30 VSSW(m / s) -0.17 3.81 -18.96 -2.72 -0.31 2.33 18.93 USSW(m / s) 0.28 4.50 -22.45 -2.33 1.12 3.50 17.11
[0063] The specific process of preprocessing ocean multivariate observation data is as follows:
[0064] Points in the wind speed data set are selected based on valid points in the sea surface temperature data in the study area, that is, data from non-ocean areas such as land and islands are excluded; then Z-score standardization is performed on each variable data:
[0065]
[0066] Wherein, x represents data of different variables, Z represents normalized data, μ and σ represent mean values and standard deviations of different variables, respectively; in this embodiment, the data set is divided into a training set and a test set in chronological order; specifically, the first 80% of the time series data is selected as the training set, and the last 20% of the data is selected as the test set for model training and performance verification; after processing the training set, data X of the shape of T×N×C can be obtained, where T is the total time length, N is the valid points in the study area, and C is the number of variables (3 in this embodiment, namely, sea surface temperature, meridional wind speed, and zonal wind speed).
[0067] S2. Model building: Building a sea surface temperature prediction model, such as Figure 2 As shown in , it includes a spatiotemporal expert module, an expert routing module and an expert mixing module; the spatiotemporal expert module includes multiple spatiotemporal expert modules, each of which is modeled for different time scales; each spatiotemporal expert module consists of a cross-variable attention layer and a spatiotemporal representation layer, as shown in Figure 4 As shown in the figure, the cross-variable attention layer uses the SST data and other exogenous variables observed in step S1 for sequence-level embedding, and then performs feature interaction through the attention mechanism to extract the impact of exogenous variables on SST; the spatiotemporal representation layer extracts spatial features through a graph neural network and uses one-dimensional hole convolution and a gating network to extract temporal features; different spatiotemporal expert modules use convolution kernels of different sizes to learn spatiotemporal features at different time scales and capture multi-scale temporal dependencies and pattern changes; the expert routing module is responsible for selecting appropriate spatiotemporal experts, and the expert mixing module performs weighted summation on the outputs of the selected spatiotemporal expert modules through the weights provided by the routing module to obtain the output result, thereby achieving accurate prediction of future sea surface temperatures.
[0068] S3, obtaining the spatiotemporal feature vector: using the preprocessed data to input the spatiotemporal expert module in step S2, extracting the impact of exogenous variables on the sea surface temperature and the spatiotemporal characteristics of the sea surface temperature data itself, and obtaining the spatiotemporal feature vector. The specific process includes:
[0069] S301, data encoding: Use a multi-layer perceptron to encode the pre-processed data X along the time dimension into a global label G∈R h×N×C ; where h is the hyperparameter of the global tag size and C is the number of variables, such as Figure 3 As shown;
[0070] S302, attention mechanism interaction: Apply the attention mechanism to the global mark between different variables at each point to interact, so as to extract the impact of exogenous variables on the sea surface temperature; and globally mark the sea surface temperature after interaction G SST∈R h×N Splice to each point of sea surface temperature in the preprocessed data X SST ∈R T×N , get the sea surface temperature data X with global label SST ∈R (T +h)×N ,like Figure 3 As shown;
[0071] S303, obtaining spatial feature vectors: Figure 4 As shown in the figure, the SST data with global labels is applied to the graph neural network to obtain the feature vector after spatial interaction. The specific process is as follows:
[0072] Construct the adjacency matrix E∈R N×N , initialize E by generating two learnable vectors vec1 and vec2:
[0073] E = Softmax(ReLU(vec1×vec2))
[0074] Spatial interactions according to the adjacency matrix:
[0075]
[0076] Where σ is the sigmoid activation function, W is the learnable matrix; H i Aggregation and X i The node features adjacent to the spatial node are finally obtained. S .
[0077] S304, obtaining the spatiotemporal feature vector: Figure 4 As shown, the spatial feature vector H obtained in step S303 S A one-dimensional dilated convolution followed by a gated linear unit (GLU) is used to capture information on the SST time dimension; finally, the spatiotemporal feature vector H of each point is obtained. ST .
[0078] S4. Obtain path weights: Use the preprocessed sea surface temperature data in step S1 as the input of the expert routing module in step S2, and use the routing function to generate path weights to determine the probability of selecting each spatiotemporal expert module for the current data.
[0079] In this embodiment, adding a noise component to the traditional routing function helps to achieve load balancing; before using the Softmax function, the input is multiplied by a trainable weight matrix W g And add adjustable Gaussian noise, the specific process is as follows:
[0080]
[0081] Where R(·) represents the entire routing function, W r and W noise is a learnable parameter used for weight generation; path weights are selected, the first K path weights are retained and the rest are set to 0.
[0082] To avoid the routing function tending to converge to a state where the same few experts always produce greater weights, this embodiment adds a balanced expert loss function L eq , all experts are encouraged to have equal importance:
[0083]
[0084] L eq (X SST )=W eq CV(Equalization(X SST )) 2
[0085] Where W eq is an adjustable scaling factor and CV is the operation that calculates the coefficient of variation.
[0086] S5, output prediction: According to the path weight obtained in step S4, the output of all expert models obtained in step S3 is weighted and summed to obtain the aggregated spatiotemporal feature vector
[0087]
[0088] In the formula, R(X SST ) i and E i (X) represents the output of the routing module for a given input and the spatiotemporal feature vector H output by the i-th expert model. ST .
[0089] like Figure 2 As shown, the aggregated spatiotemporal feature vector It will be processed through a fully connected layer to generate the final prediction result
[0090]
[0091] Among them, FC represents the fully connected layer. The entire routing distribution and aggregation diagram is as follows Figure 2-4 shown.
[0092] S6, model optimization: In this embodiment, the prediction result obtained by minimizing step S5 is The model is trained by calculating the mean square error between the predicted value and the true value Y. The optimization process aims to update all learnable parameters in the model so that the model output is closer to the true value. In addition, considering the generalization ability of the model, the optimization goal also includes the regularization term L eq , which is used to balance the complexity of the model and avoid overfitting; the formula of the total loss function is as follows:
[0093]
[0094] Where θ is all learnable parameters in the model.
[0095] Example 2
[0096] In this embodiment, the training set data obtained in step S1 of Example 1 is input into the steps of the sea surface temperature prediction method based on spatiotemporal expert mixing and exogenous variables provided in Example 1 for model training; the test set data is then used to verify the model performance, and the sea surface temperature results for the next 30 days, 90 days, 180 days, and 365 days are predicted respectively.
[0097] In this embodiment, FC-LSTM and ConvLSTM time series prediction models are selected as comparison to test the performance of the model proposed in the present invention; root mean square error (RMSE) and mean absolute error (MAE) are selected as evaluation criteria. The specific results are shown in Table 2.
[0098] Table 2 Comparative test results of different models
[0099]
[0100] It can be seen from Table 2 that the model (ours) proposed in the present invention is superior to the comparison model under different prediction lengths. The sea surface temperature prediction method based on spatiotemporal expert mixing and exogenous variables proposed in the present invention can effectively improve the prediction accuracy of sea surface temperature.
[0101] The above description is only a further explanation of the present invention in combination with specific embodiments. All descriptions do not limit the scope of protection of the present invention. Any changes or replacement schemes that can be easily thought of by any technician in this field within the technical scope disclosed by the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.
Claims
1. A sea surface temperature prediction method based on spatiotemporal expert mixing and exogenous variables, characterized in that: The following steps are involved: S1. Data acquisition and preprocessing: acquiring ocean multivariate observation data and preprocessing the data; the ocean multivariate observation data includes sea surface temperature data and other exogenous variable data; S2. Model building: building a sea surface temperature prediction model, including a spatiotemporal expert module, an expert routing module, and an expert hybrid module; S3, obtaining a spatiotemporal feature vector: inputting the preprocessed data in step S1 into the spatiotemporal expert module in step S2, extracting the influence of exogenous variables on the sea surface temperature and the spatiotemporal features of the sea surface temperature data itself, and obtaining a spatiotemporal feature vector. S4. Obtain path weights: Input the preprocessed sea surface temperature data in step S1 into the expert routing module in step S2, and use the routing function to generate path weights to determine the probability of selecting each spatiotemporal expert module for the current data. S5, output prediction: according to the path weight obtained in step S4, the outputs of all spatiotemporal expert modules obtained in step S3 are weighted and summed to obtain the final output; S6, model optimization: According to the prediction output result obtained in step S5, the model is trained by minimizing the mean square error between the prediction result and the true value, and all learnable parameters in the model are optimized and updated to make the model output closer to the true value.
2. The sea surface temperature prediction method based on spatiotemporal expert mixing and exogenous variables according to claim 1 is characterized in that: The specific process of data acquisition and preprocessing in step S1 includes: The data points are screened according to the selected exogenous variable data and the sea mark temperature data in the study area, and then the Z-score standardization is performed on each variable data: In the formula, x represents the data of different variables, Z represents the normalized data, μ and σ represent the mean and standard deviation of different variables respectively; After processing the training set, we can get data X with the shape of T×N×C, where T is the total time length, N is the valid points in the study area, and C is the number of variables.
3. The sea surface temperature prediction method based on spatiotemporal expert mixing and exogenous variables according to claim 2 is characterized in that: The specific process of constructing the model in step S2 includes: There are multiple spatiotemporal expert modules, each of which is modeled for a different time scale; each spatiotemporal expert module includes a cross-variable attention layer and a spatiotemporal representation layer, the cross-variable attention layer uses the sea surface temperature data and other exogenous variables obtained in step S1 to perform sequence-level embedding, and then performs feature interaction through the attention mechanism to extract the impact of other exogenous variables on the sea surface temperature data; the spatiotemporal representation layer extracts spatial features through a graph neural network and uses a one-dimensional hole convolution and a gated network to extract temporal features; different spatiotemporal expert modules use convolution kernels of different sizes to learn spatiotemporal features at different time scales and capture multi-scale time dependencies and pattern changes; the expert routing module is responsible for selecting an appropriate spatiotemporal expert module, and the expert mixing module performs weighted summation on the outputs of the selected spatiotemporal expert modules through the weights provided by the routing module to obtain the output result.
4. The sea surface temperature prediction method based on spatiotemporal expert mixing and exogenous variables according to claim 3 is characterized by: The specific process of obtaining the spatiotemporal feature vector in step S3 includes: S301, data encoding: Use a multi-layer perceptron to encode the pre-processed data X along the time dimension into a global label G∈R h ×N×C ; where h is the hyperparameter of the global tag size and C is the number of variables; S302, attention mechanism interaction: Apply the attention mechanism to the global mark between different variables at each point to interact, so as to extract the impact of exogenous variables on the sea surface temperature; and globally mark the sea surface temperature after interaction G SST ∈R h×N Splice to each point of sea surface temperature in the preprocessed data X SST ∈R T×N , get the sea surface temperature data X with global label SST ∈R (T+h)×N ; S303, obtaining a spatial feature vector: applying the globally labeled sea surface temperature data to a graph neural network to obtain a feature vector after spatial interaction; S304, obtaining a spatiotemporal feature vector: a one-dimensional dilated convolution is performed on the spatial feature vector obtained in step S303, followed by a gated linear unit (GLU) to capture information on the time dimension of the sea surface temperature; and finally the spatiotemporal feature vector of each point is obtained.
5. The sea surface temperature prediction method based on spatiotemporal expert mixing and exogenous variables according to claim 4 is characterized in that: The specific process of obtaining the spatial feature vector in step S303 includes: Construct the adjacency matrix E∈R N×N , initialize E by generating two learnable vectors vec1 and vec2: E = Softmax(ReLU(vec1×vec2)) According to the adjacency matrix, the spatial feature vector H of each point is obtained by spatial interaction. i : Where σ is the sigmoid activation function, W is the learnable matrix, and N is the valid point in the study area; H i Aggregation and X i The node features adjacent to the spatial node are finally obtained. S .
6. The sea surface temperature prediction method based on spatiotemporal expert mixing and exogenous variables according to claim 1 is characterized in that: The specific process of obtaining the path weight in step S4 includes: A noise component is added to the routing function, and the input is multiplied by a trainable weight matrix W before applying the Softmax function. g And add adjustable Gaussian noise: Where R(·) represents the entire routing function, W r and W noise is a learnable parameter used for weight generation; path weights are selected, the first K path weights are retained and the rest are set to 0.
7. The sea surface temperature prediction method based on spatiotemporal expert mixing and exogenous variables according to claim 6 is characterized by: Add balanced expert loss function L eq : L eq (X SST )=W eq ·CV(Equalization(X SST )) 2 Where W eq is an adjustable scaling factor and CV is the operation that calculates the coefficient of variation.
8. The sea surface temperature prediction method based on spatiotemporal expert mixing and exogenous variables according to claim 7 is characterized by: The specific process of outputting the prediction in step S5 includes: The final prediction output is obtained by weighted summing the outputs of all expert models, with the weights provided by the routing module; the final output The expression is: In the formula, R(X SST ) i and E i (X) denotes the output of the routing module and the output of the i-th expert model for a given input, respectively.
9. The sea surface temperature prediction method based on spatiotemporal expert mixing and exogenous variables according to claim 8 is characterized in that: Combined with the final output obtained in step S5 The model is trained with the true value Y by minimizing The total loss function can be written as: Where θ is all learnable parameters in the model; The data obtained in step S1 is used as input and model training is performed, and the trained model can be applied to sea surface temperature prediction.
10. An electronic device comprising a storage device and a processor, wherein a computer program is stored in the storage device, wherein: When the processor executes the computer program, any step in the sea surface temperature prediction method based on spatiotemporal expert mixing and exogenous variables as described in any one of claims 1 to 9 is implemented.