Sea surface temperature prediction method and device based on physical constraint

Abstract

The invention relates to a physical constraint-based sea surface temperature prediction method and device, and the method comprises the steps: introducing a seawater density difference as a physical constraint term to construct a loss function, and forcing a prediction result to accord with an ocean dynamics law in a model training process; the method solves the problem of prediction result distortion caused by neglecting physical constraints in a traditional data driving method, and has the advantage of improving the physical consistency and accuracy of the sea surface temperature prediction result.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the technical field of temperature prediction, in particular to a sea surface temperature prediction method and device based on physical constraints. BACKGROUND

[0002] As a basic variable in the sea-air coupling system, sea surface temperature (SST) plays a key role in regulating the marine ecosystem. Accurate sea surface temperature prediction is crucial for understanding climate variability, supporting ecological assessment, and providing a basis for human-related decisions in coastal and marine areas.

[0003] Currently, the main data-driven method for predicting the three-dimensional temperature field of the ocean. Data-driven methods solve the problem of sea surface temperature prediction from a data perspective. Due to the powerful learning ability in the field of deep learning and the development of numerical calculation models, practical and accurate data-driven methods are more reliable than other methods. Although traditional data-driven prediction models can capture the spatio-temporal characteristics of sea surface temperature, such models often ignore the underlying physical laws and violate the principles of ocean dynamics, resulting in a decrease in prediction accuracy. SUMMARY

[0004] The following is a summary of the subject matter described in detail herein. This summary is not intended to limit the scope of the claims.

[0005] The main purpose of the embodiments of the present disclosure is to propose a sea surface temperature prediction method and device based on physical constraints, which can make the prediction results meet the requirements of data accuracy and physical rationality, and improve the prediction accuracy of sea surface temperature.

[0006] In a first aspect, the present application provides a sea surface temperature prediction method based on physical constraints, comprising: acquiring a historical temperature-salinity spatio-temporal sequence of an ocean area collected in a historical period; the historical temperature-salinity spatio-temporal sequence is composed of a historical sea surface temperature spatio-temporal sequence and a corresponding historical salinity spatio-temporal sequence; building a sea surface temperature prediction model, inputting the historical temperature-salinity spatio-temporal sequence into the sea surface temperature prediction model, extracting corresponding spatio-temporal features from the historical temperature-salinity spatio-temporal sequence according to the sea surface temperature prediction model, and predicting a sea surface temperature prediction value and a salinity prediction value at a first time based on the spatio-temporal features; the first time is located after the historical period; determining a constraint term, and constructing a loss function of the sea surface temperature prediction model according to the constraint term, the sea surface temperature prediction value and the salinity prediction value, and the true value of the sea surface temperature and the true value of the sea surface salinity of the ocean area at the first time; the constraint term is the sea water density difference generated by the sea surface temperature and salinity predicted by the sea surface temperature prediction model; The sea surface temperature prediction model is optimized by backpropagation based on the loss function to obtain the trained sea surface temperature prediction model. In response to a temperature prediction command for a target ocean area, the spatiotemporal sequence of temperature and salinity collected in the target ocean area is input into the trained sea surface temperature prediction model to obtain the sea surface temperature of the target ocean area at a future time, as output by the sea surface temperature prediction model.

[0007] The sea surface temperature prediction method based on physical constraints provided in this embodiment has at least the following beneficial effects: Compared with existing technologies, traditional data-driven methods rely solely on historical data statistical patterns and may ignore the impact of salinity changes on density stratification. This method introduces seawater density difference as a physical constraint to construct a loss function, which forces the prediction results to conform to the laws of ocean dynamics during model training. This solves the problem of distortion in prediction results caused by ignoring physical constraints in traditional data-driven methods, and has the advantage of improving the physical consistency and accuracy of sea surface temperature prediction results.

[0008] In a second aspect, this application provides a sea surface temperature prediction device based on physical constraints, the device comprising: The data acquisition module is used to acquire historical temperature and salinity spatiotemporal sequences of marine areas collected in historical time periods; the historical temperature and salinity spatiotemporal sequences consist of historical sea surface temperature spatiotemporal sequences and corresponding historical salinity spatiotemporal sequences; The model training module is used to build a sea surface temperature prediction model. The historical temperature and salinity spatiotemporal sequence is input into the sea surface temperature prediction model to extract corresponding spatiotemporal features from the historical temperature and salinity spatiotemporal sequence according to the sea surface temperature prediction model, and based on the spatiotemporal features, predict the sea surface temperature and salinity values ​​at a first moment; the first moment is after the historical time period. Determine the constraint terms, and construct the loss function of the sea surface temperature prediction model based on the constraint terms, the predicted sea surface temperature and salinity values, and the actual sea surface temperature and salinity values ​​of the ocean area at the first moment; the constraint terms are the seawater density difference generated by the sea surface temperature and salinity predicted by the sea surface temperature prediction model. The sea surface temperature prediction model is optimized by backpropagation based on the loss function to obtain the trained sea surface temperature prediction model. The model prediction module is used to respond to the temperature prediction command of the target ocean area by inputting the spatiotemporal sequence of temperature and salinity collected in the target ocean area into the trained sea surface temperature prediction model to obtain the sea surface temperature of the target ocean area at a future time output by the sea surface temperature prediction model.

[0009] A third aspect of this application provides an electronic device including at least one controller and a memory for communicatively connecting to the controller; the memory stores instructions executable by the at least one controller to cause the at least one controller to perform a physical constraint-based sea surface temperature prediction method as described above.

[0010] A fourth aspect of this application provides a computer-readable storage medium storing computer-executable instructions for causing a computer to perform a physical constraint-based sea surface temperature prediction method as described above.

[0011] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating a physical constraint-based sea surface temperature prediction method provided in an embodiment of this application. Figure 2 This is a schematic diagram of the prediction process of the sea surface temperature prediction model provided in the embodiments of this application; Figure 3 This is a schematic diagram comparing the RMSE and ACC indices of the ConvLSTM model provided in the embodiments of this application; Figure 4 This is a schematic diagram comparing the RMSE and ACC indices of the TCN-ConvLSTM model provided in the embodiments of this application; Figure 5 This is a schematic diagram comparing the RMSE and ACC indices of the PG-ConvLSTM model provided in the embodiments of this application; Figure 6 This is a schematic diagram comparing the RMSE and ACC indices of the PG / TCN-ConvLSTM model provided in the embodiments of this application; Figure 7 This is a schematic diagram of the structure of a sea surface temperature prediction device based on physical constraints provided in an embodiment of this application; Figure 8 This is a schematic diagram of the structure of an electronic device provided in one embodiment of this application. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0015] In the description of this application, the use of terms such as "first," "second," etc., is for the purpose of distinguishing technical features only and should not be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated or the order of the technical features indicated.

[0016] In the description of this application, it should be understood that the orientation descriptions, such as up, down, etc., are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed or function in a specific orientation, and therefore should not be construed as a limitation of this application.

[0017] like Figure 1 One embodiment of this application provides a method for predicting sea surface temperature based on physical constraints, the method comprising: Step S110: Obtain the historical temperature and salinity spatiotemporal sequence of the ocean area collected in historical time periods; the historical temperature and salinity spatiotemporal sequence consists of the historical sea surface temperature spatiotemporal sequence and the corresponding historical salinity spatiotemporal sequence. The historical temperature and salinity spatiotemporal sequence refers to a data set containing the spatiotemporal distribution of historical sea surface temperature and salinity. For example, it can be constructed using satellite remote sensing data and ocean buoy observation data. The temporal resolution can be daily or weekly, and the spatial resolution can be 0.25°×0.25° grid data.

[0018] Step S120: A sea surface temperature prediction model is built by inputting historical temperature and salinity spatiotemporal sequences into the model. The model extracts corresponding spatiotemporal features from these sequences and, based on these features, predicts the sea surface temperature and salinity values ​​at the first moment in time; this first moment is after the historical time period. The sea surface temperature prediction model is a deep learning architecture capable of extracting spatiotemporal features and performing multivariate predictions. Specifically, it can be implemented using a convolutional long short-term memory network, capturing spatial features through convolutional layers and temporal dependencies through recurrent structures. The model simultaneously predicts sea surface temperature and salinity, providing a foundation for physical constraint calculations.

[0019] Step S130: Determine the constraint terms, and construct the loss function of the sea surface temperature prediction model based on the constraint terms, the predicted sea surface temperature and salinity values, and the actual sea surface temperature and salinity values ​​of the ocean area at the first moment. The constraint terms are the seawater density differences caused by the sea surface temperature and salinity predicted by the sea surface temperature prediction model. The constraint terms transform physical laws into computable optimization objectives.

[0020] Step S140: Perform backpropagation optimization on the sea surface temperature prediction model according to the loss function to obtain the trained sea surface temperature prediction model; during the training phase, calculate the seawater density difference corresponding to the predicted temperature and salinity, add it as a regularization term to the loss function, and adjust the model parameters through backpropagation so that the prediction results satisfy the density conservation constraint while reducing data errors.

[0021] Step S150: In response to the temperature prediction command for the target ocean area, the spatiotemporal sequence of temperature and salinity collected in the target ocean area is input into the trained sea surface temperature prediction model to obtain the sea surface temperature of the target ocean area at future time output by the sea surface temperature prediction model.

[0022] In the prediction phase, the target area data is input into the trained model, and the output is a temperature prediction value that conforms to physical laws.

[0023] Compared to existing technologies, traditional data-driven methods rely solely on historical statistical patterns, potentially neglecting the impact of salinity variations on density stratification. This method, however, introduces seawater density difference constraints, integrating ocean thermodynamic principles into the model learning process. This avoids predictions that violate hydrostatic equilibrium, effectively resolving the disconnect between data-driven models and physical laws, ensuring that predictions simultaneously meet the requirements of data accuracy and physical plausibility. For long-term prediction tasks, the introduction of physical constraints can suppress error accumulation and enhance the model's generalization ability to complex ocean processes.

[0024] In some embodiments, the sea surface temperature prediction model is constructed based on a convolutional long short-term memory network; Step S120 involves inputting the historical temperature-salinity spatiotemporal sequence into the sea surface temperature prediction model. This allows the model to extract corresponding spatiotemporal features from the historical temperature-salinity spatiotemporal sequence and, based on these features, predict the sea surface temperature and salinity values ​​for the first moment. This includes: Step S1210: Input the historical temperature-salinity spatiotemporal sequence into the convolutional long short-term memory network to extract the corresponding spatiotemporal features from the historical temperature-salinity spatiotemporal sequence based on the convolutional long short-term memory network, and predict the sea surface temperature and salinity at the first moment based on the spatiotemporal features.

[0025] Convolutional Long Short-Term Memory (ConvLSTM) is a deep learning model that combines convolutional neural networks and long short-term memory network structures. The network extracts spatial features through convolutional operations and captures temporal dependencies through gating mechanisms, thereby achieving joint modeling of the spatiotemporal sequence of the marine environment.

[0026] Specifically, historical temperature-salinity spatiotemporal sequences are input into a convolutional long short-term memory network (LSTM) with time steps as the dimension. At each time step, the network performs convolution operations on the temperature and salinity fields, extracting spatial feature maps. Information flow is then controlled through forget gates, input gates, and output gates, passing the hidden state from the current time step to the next. Through iterative processing of the sequence data, the network ultimately outputs a hidden state vector that fuses spatiotemporal features, and generates predicted temperature and salinity values ​​for future time steps based on this vector.

[0027] Compared to existing technologies, traditional methods typically employ independent convolutional neural networks to process spatial features or use recurrent neural networks alone to model time series, failing to effectively capture the nonlinear interactions of marine environmental parameters across the spatiotemporal dimensions. Convolutional Long Short-Term Memory (LSTM) networks, however, reduce model complexity through kernel parameter sharing while maintaining long-term series modeling capabilities through gating mechanisms, enabling them to more accurately reflect the coordinated changes in temperature and salinity during marine dynamic processes.

[0028] In some embodiments, the sea surface temperature prediction model is constructed based on a convolutional long short-term memory network and a temporal convolutional network; Step S120 inputs the historical temperature-salinity spatiotemporal sequence into the sea surface temperature prediction model, so as to extract the corresponding spatiotemporal features from the historical temperature-salinity spatiotemporal sequence according to the sea surface temperature prediction model, and predicts the sea surface temperature and salinity values ​​at the first moment based on the spatiotemporal features, including: Step S1221: Input the historical temperature-salinity spatiotemporal sequence into the convolutional long short-term memory network to extract the corresponding spatiotemporal features from the historical temperature-salinity spatiotemporal sequence according to the convolutional long short-term memory network; Step S1222: Input the spatiotemporal features into the temporal convolutional network to predict the sea surface temperature and salinity at the first moment based on the temporal convolutional network.

[0029] Among them, convolutional long short-term memory networks are used to capture spatial correlations and temporal dependencies from input spatiotemporal sequences. Temporal convolutional networks (TCNs) are temporal modeling modules built on causal convolutions, specifically implemented using a dilated convolution stacked structure, for efficient modeling of long sequence temporal dependencies.

[0030] Specifically, historical temperature and salinity spatiotemporal sequences are input into a convolutional long short-term memory network. The network extracts gridded features in the spatial dimension using three-dimensional convolutional kernels, and simultaneously establishes memory units in the temporal dimension through a gating mechanism, forming spatiotemporal features that include spatial distribution patterns and temporal evolution laws. Subsequently, the temporal convolutional network performs dilated convolution operations on the extracted spatiotemporal features, increasing the receptive field to capture long-term temporal dependencies, and finally outputs predicted values ​​for temperature and salinity at future times.

[0031] Compared to existing technologies, traditional methods typically use recurrent neural networks alone to process spatiotemporal sequences, making it difficult to simultaneously handle spatial feature extraction and long-term temporal modeling. Our proposed solution, however, employs a phased processing mechanism. A convolutional long short-term memory network is dedicated to spatial feature capture, while a temporal convolutional network focuses on temporal dimension modeling, effectively overcoming the limitations of a single network structure in learning complex spatiotemporal patterns. Through this technical solution, we achieve hierarchical modeling of the spatiotemporal evolution of ocean temperature and salinity fields. The introduction of physical constraints ensures that the prediction results conform to ocean dynamics principles, improving prediction accuracy while enhancing the physical interpretability of the model output.

[0032] In some embodiments, the loss function includes: ; ; in, For loss function, Let be the mean square error function. It is the first Predicted values ​​of the training samples. It corresponds to the actual value. The training samples consist of a set of historical sea surface temperature spatiotemporal data and corresponding historical salinity spatiotemporal data from the historical sea surface temperature spatiotemporal sequence and the corresponding historical salinity spatiotemporal sequence. For constraint terms The weight hyperparameters, The mapping function for averaging. This is the ReLU function.

[0033] The mean squared error function is a mathematical function that measures the prediction error by calculating the average of the squared differences between the predicted and the true values. Specifically, it can be implemented by dividing the sum of squared errors of a batch of samples by the number of samples, and is used to drive the model to learn the statistical regularities of the data itself. The direction of optimization, which can be determined through cross-validation or grid search, is used to balance data errors and physical constraints. The ReLU function is an activation function that non-linearly truncates the input value. Specifically, it can be implemented by retaining the positive part and setting the negative part to zero, and is used to trigger constraints only when the density increases abnormally due to the prediction of sea surface temperature and salinity.

[0034] Compared to existing technologies, traditional data-driven methods only use mean squared error as the loss function, which may lead to prediction results that violate the continuity of seawater density. Our method, however, introduces density difference constraints based on physical equations. This approach effectively integrates the data-driven model with physical laws, maintaining the spatiotemporal feature extraction capabilities of deep learning models while guiding the model to generate prediction results that conform to ocean dynamics principles. This solves the problem of accumulated prediction bias caused by neglecting physical laws in traditional methods.

[0035] In some embodiments, ; in, To use the first set of training samples After inputting the sea surface temperature prediction model, the model outputs the predicted sea surface temperature and salinity values ​​for the first set of training samples. For the second set of training samples After inputting into the sea surface temperature prediction model, the sea surface temperature prediction and salinity prediction values ​​of the second set of training samples output by the sea surface temperature prediction model are: For based on To obtain the density of seawater, For based on The obtained seawater density.

[0036] Furthermore, the seawater density was calculated according to the seawater state equation of the TEOS-10 standard.

[0037] The EOS-10 standard refers to the thermodynamic seawater equation jointly published by the International Commission on International Oceanography and the International Union of Geodesy and Geophysics. Specifically, it uses temperature, salinity, and pressure as input parameters to calculate seawater density through polynomial fitting or table lookup, ensuring that the density calculation conforms to internationally recognized physical standards. The seawater equation of state is a mathematical expression based on thermodynamic principles that describes the relationship between seawater density and temperature, salinity, and pressure. It can be calculated using the Gibbs function defined in the TEOS-10 standard or its simplified form to achieve high-precision density calculation.

[0038] In some embodiments, prior to obtaining the historical temperature, salinity, and spatiotemporal sequence of the ocean area collected in historical time periods in step S100, the method further includes: Step S101 involves labeling, handling outliers and missing values, normalizing, and reshaping the historical sea surface temperature spatiotemporal series and the corresponding historical salinity spatiotemporal series.

[0039] Tagging refers to adding spatiotemporal dimension identification information to the data. Specifically, it can be achieved by combining latitude and longitude coordinates with timestamps, enabling the model to identify the correlation between different regions and time periods.

[0040] Outlier and missing value handling refers to detecting and correcting outliers or missing values ​​in the data. This can be achieved by using median filtering combined with linear interpolation to eliminate noise interference caused by sensor errors or transmission interruptions.

[0041] Normalization refers to mapping data with different dimensions to a unified numerical range. Specifically, it can be achieved by using maximum and minimum value scaling or Z-score standardization to avoid training bias caused by differences in the scale of the dependent variables in the model.

[0042] Data sequence reshaping refers to adjusting the dimensional arrangement of data structures. This can be achieved by using sliding window truncation or matrix rearrangement to convert the original data into a tensor format that meets the model input requirements.

[0043] This application improves the effectiveness of input data from three dimensions—data quality, distribution consistency, and structural adaptability—through a systematic preprocessing process, thereby enhancing the model's ability to capture dynamic changes in the ocean.

[0044] This application also provides an embodiment of a method for predicting sea surface temperature based on physical constraints, which includes the following parts: (1) First, the sea surface temperature prediction problem is modeled as a spatiotemporal series multi-step regression problem: that is, given historical data... Historical sea surface temperature at any moment Predicting future salinity based on historical levels Sea surface temperature at any given time The model's input is spatiotemporal sequence data, represented by the following dimensions: ,in, The number of training samples, To input the time step, For spatial dimensions, These are the characteristic numbers for temperature and salinity. The output of the sea surface temperature prediction model is the future... Step's sea surface temperature series: The physical constraints of the sea surface temperature prediction model are embedded through the seawater density equation, which links sea surface temperature and salinity through the equation. The prediction results must satisfy the physical law that seawater density increases with depth, i.e. , This refers to the depth of the seawater.

[0045] (2) The models used in this method include ConvLSTM networks and TCN networks, which are described in detail below: ConvLSTM replaces fully connected layers with convolutional layers, capturing spatial correlations through convolutional operations and temporal dependencies through LSTM units, enabling it to process spatiotemporal sequence data such as meteorological fields. ConvLSTM consists of an input layer, a feature extraction layer, and an output layer.

[0046] Input layer: Used to receive historical temperature-salinity spatiotemporal sequence data, with dimensions of [missing information]. .

[0047] Feature extraction layer: Consists of a 3-layer ConvLSTM2D network (each layer with 10 3×3 convolutional kernels, ReLU activation, and sequence return), extracting spatiotemporal features through convolutional gating units (input gate, forget gate, and output gate). The input gate, forget gate, output gate, and cell state update equations of the ConvLSTM unit are as follows: ; ; ; ; ; in, for Input feature map at time step 1 This is the hidden state from the previous moment. This represents the cell state at the previous moment. These are the input gate, forget gate, and output gate, respectively. , representing convolution weights, For convolution operations, Represents element-wise multiplication. This is the Sigmoid activation function.

[0048] Output layer: 3D convolutional layer (Conv3D) combined with the Sigmoid activation function maps spatiotemporal features to the prediction dimension. ; in, It is the input feature map. It is a convolution kernel. It's a bias. It is the sigmoid activation function.

[0049] TCN is a convolution-based temporal model composed of multiple TCN blocks. Each block contains: a Conv2D layer with dilated convolutions, a batch normalization layer, and a ReLU activation layer. TCN captures long sequence dependencies through dilated convolutions and residual connections. Dilated convolutions increase the receptive field by inserting gaps into the convolutional kernel, as shown in the formula: ; in, For the output feature map, the first One location, The kernel size is [size]. As the expansion factor, For convolution kernel weights, This represents the input feature map.

[0050] (3) The model's loss function; To ensure that the sea surface temperature prediction model conforms to known physical laws, the density changes resulting from the predicted sea surface temperature and salinity are primarily constrained by seawater density. This is mainly calculated based on the seawater state equation of the TEOS-10 standard. ; in: It is a reference density, and it is a temperature. The function, This is a coefficient related to sea surface temperature. Among them, Used to describe the linear effect of salinity on seawater density. It is used to describe the nonlinear effect of salinity on seawater density. It is a constant coefficient that describes the effect of the square of salinity on seawater density. Salinity. The formulas for calculating it are as follows: ; ; ; ; Next, the density difference of seawater under different inputs is calculated: ; in, These represent the sea surface temperature and salinity predicted by the model, respectively, through forced... The non-negativity of the prediction ensures that the prediction conforms to thermodynamic stability.

[0051] The mean squared error (MSE) is used in the model to measure this difference: ; in, It is the actual value. These are the model's predicted values. This refers to the number of samples. By minimizing this loss function, the model can continuously adjust its parameters to make the predictions closer to the true values. However, in the model, while ensuring that the training error is minimized, physical constraints must also be met. Therefore, a combined loss function is used as the model's loss function: ; in, It is the final loss function. It is the mean squared error loss. These are the weight hyperparameters of the physical constraint terms. It's due to the difference in seawater density. ,use The function is designed to ensure that the loss function is only affected when the density difference is positive.

[0052] like Figures 3 to 6 The following provides a set of experimental examples; The model training and testing in this study were completed collaboratively using an online parallel GPU cluster and a local development environment. On the hardware side, an online parallel GPU cluster was used, comprising 8 distributed computing units, with each node configured with an NVIDIA A100 GPU. The local development environment used PyCharm as the integrated development environment (IDE) for easy model debugging and code management. The software environment employed a deep learning framework based on PyTorch.

[0053] Training parameter configuration: batch size is set to 32, initial learning rate is set to 0.01 (using cosine annealing for decay), total training epochs are 100, and the Adam optimizer is used. =0.9, =0.999), and a Dropout layer (random drop probability set to 0.2) is introduced to suppress overfitting.

[0054] Experimental data; The sea surface temperature and salinity data used in this embodiment are from the EN4.2.1 dataset released by the Hadley Centre of the UK Met Office. The data covers the period from January 1980 to December 2023, spatially covering the global ocean (90°N to 90°S, 0°E to 360°E). The raw data includes temperature and salinity observations at different depths. The raw data underwent the following preprocessing: Tagging extraction and storage: Temperature and salinity data were extracted from the raw data in the form of tags and concatenated by month. In order to reduce the data memory usage, data from 6 representative depth layers (0, 5, 11, 17, 22, 26) were selected from 42 depth layers for tagging and storage.

[0055] Outlier and missing value handling: Filter outliers and missing values ​​and record their locations, storing them as NAN values. Before building the model for prediction, restore the NAN values ​​to 0.

[0056] Normalization: The temperature and salinity data are scaled to the [0,1] interval to eliminate the dimensional influence between different features.

[0057] ; Data sequence reshaping: Adjusting the data dimensions to a form acceptable to the model (number of samples, input time steps, spatial dimensions, number of feature channels).

[0058] By using the fusion architecture of ConvLSTM and TCN, this model can capture the large-scale spatial distribution characteristics of sea surface temperature (such as the morphological matching between the equatorial high temperature zone and the polar low temperature zone). Moreover, the density difference regularization mechanism introduced by the model effectively avoids the systematic bias of the prediction results and ensures physical consistency.

[0059] Model error analysis: The impact of the learning rate: A high initial learning rate (0.01) accelerates early convergence (loss decreases by 80% within 10 rounds); it decays to a lower rate later. To avoid gradient oscillations, a fixed high learning rate can lead to training failure due to gradient explosion.

[0060] hyperparameters Impact: When At that time, physical consistency (density deviation) and prediction accuracy improve simultaneously; if The model overfits the physical constraints, and the MSE increases by more than 30%.

[0061] This embodiment uses root mean square error (RMSE) and outlier correlation coefficient (ACC) as evaluation metrics for model performance, and visually compares the proposed model (TCN / PG-ConvLSTM in the figure) with three benchmark models (ConvLSTM, ConvLSTM-TCN, and PG-ConvLSTM): In this model, the rationality and temporal modeling capabilities are significantly enhanced through physical constraints and TCN. The three-layer ConvLSTM-TCN structure captures spatial correlations and multi-scale temporal dependencies through convolution and dilation operations, and exhibits more stable convergence (verification loss shows no oscillation). Combined with the optimized training strategy, the model's root mean square error (RMSE) is significantly reduced, and prediction accuracy is further improved. The dynamic learning rate strategy adapts to the entire training process, balancing convergence speed and parameter precision, thus avoiding problems such as gradient explosion, insufficient generalization ability, or sudden drops in learning rate that may occur with other models using a fixed learning rate, ensuring the continuity and stability of training. Simultaneously, the number of parameters in the model is reduced to one-third of the baseline model, significantly improving the training speed.

[0062] Experimental results show that the ConvLSTM model, which embeds a TCN and physical constraints, can effectively capture the spatiotemporal dynamics and multi-scale time dependence of sea surface temperature (SST). Specifically, the density equation constraint significantly improves the physical plausibility of the prediction results, avoiding unreasonable outputs that might arise from purely data-driven models. The model achieves a balance between the flexibility of data-driven methods and the interpretability of physical models, and its multi-scale time modeling capability is suitable for long-term SST prediction.

[0063] like Figure 7 One embodiment of this application provides a sea surface temperature prediction device based on physical constraints, the device comprising: The data acquisition module 1001 is used to acquire the historical temperature and salinity spatiotemporal sequence of the ocean area collected in historical time periods; the historical temperature and salinity spatiotemporal sequence consists of the historical sea surface temperature spatiotemporal sequence and the corresponding historical salinity spatiotemporal sequence; The model training module 1002 is used to build a sea surface temperature prediction model. The historical temperature and salinity spatiotemporal sequence is input into the sea surface temperature prediction model so that the corresponding spatiotemporal features are extracted from the historical temperature and salinity spatiotemporal sequence according to the sea surface temperature prediction model. Based on the spatiotemporal features, the predicted sea surface temperature and salinity values ​​at the first moment are predicted. The first moment is after the historical time period. Determine the constraint terms, and construct the loss function of the sea surface temperature prediction model based on the constraint terms, the predicted sea surface temperature and salinity values, and the actual sea surface temperature and salinity values ​​of the ocean area at the first moment; the constraint terms are the seawater density differences caused by the sea surface temperature and salinity predicted by the sea surface temperature prediction model. The sea surface temperature prediction model is optimized by backpropagation based on the loss function to obtain the trained sea surface temperature prediction model. The model prediction module 1003 is used to respond to the temperature prediction command of the target ocean area by inputting the spatiotemporal sequence of temperature and salinity collected in the target ocean area into the trained sea surface temperature prediction model to obtain the sea surface temperature of the target ocean area at future time output by the sea surface temperature prediction model.

[0064] It should be noted that the sea surface temperature prediction device based on physical constraints provided in this embodiment and the sea surface temperature prediction method based on physical constraints described above are based on the same inventive concept. Therefore, the content of the sea surface temperature prediction method based on physical constraints described above is also applicable to the content of the sea surface temperature prediction device based on physical constraints in this embodiment, and will not be repeated here.

[0065] like Figure 8 One embodiment of this application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned method for predicting sea surface temperature based on physical constraints. The electronic device includes: At least one battery; At least one memory; At least one processor; At least one program; The program is stored in memory, and the processor executes at least one program to implement the above-described physical constraint-based sea surface temperature prediction method of this disclosure.

[0066] This electronic device can be any smart terminal, including mobile phones, tablets, personal digital assistants (PDAs), and in-vehicle computers.

[0067] The electronic devices according to embodiments of this application will now be described in detail.

[0068] The processor 1600 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this disclosure. The memory 1700 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1700 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1700 and is called and executed by the processor 1600 to execute a sea surface temperature prediction method based on physical constraints according to an embodiment of this disclosure.

[0069] The input / output interface 1800 is used to implement information input and output. The communication interface 1900 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 2000 transmits information between various components of the device (e.g., processor 1600, memory 1700, input / output interface 1800, and communication interface 1900); The processor 1600, memory 1700, input / output interface 1800 and communication interface 1900 are connected to each other within the device via bus 2000.

[0070] This disclosure also provides a storage medium, which is a computer-readable storage medium storing computer-executable instructions for causing a computer to execute the aforementioned physical constraint-based sea surface temperature prediction method.

[0071] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0072] The embodiments described in this disclosure are for the purpose of more clearly illustrating the technical solutions of this disclosure and do not constitute a limitation on the technical solutions provided by this disclosure. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by this disclosure are also applicable to similar technical problems.

[0073] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this disclosure, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0074] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0075] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0076] The terms "first," "second," "third," "fourth," etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having" are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0077] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0078] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0079] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0080] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0081] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0082] The above is a detailed description of the preferred embodiments of this application. However, the embodiments of this application are not limited to the above-described implementation methods. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the embodiments of this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of the embodiments of this application.

Claims

1. A method for predicting sea surface temperature based on physical constraints, characterized in that, The method includes: Historical temperature and salinity spatiotemporal sequences of marine areas collected during historical periods are obtained; the historical temperature and salinity spatiotemporal sequences consist of historical sea surface temperature spatiotemporal sequences and corresponding historical salinity spatiotemporal sequences. A sea surface temperature prediction model is constructed, and the historical temperature and salinity spatiotemporal sequence is input into the sea surface temperature prediction model. Based on the sea surface temperature prediction model, corresponding spatiotemporal features are extracted from the historical temperature and salinity spatiotemporal sequence, and based on the spatiotemporal features, the predicted sea surface temperature and salinity values ​​at a first moment are predicted; the first moment is after the historical time period. Determine the constraint terms, and construct the loss function of the sea surface temperature prediction model based on the constraint terms, the predicted sea surface temperature and salinity values, and the actual sea surface temperature and salinity values ​​of the ocean area at the first moment; the constraint terms are the seawater density difference generated by the sea surface temperature and salinity predicted by the sea surface temperature prediction model. The sea surface temperature prediction model is optimized by backpropagation based on the loss function to obtain the trained sea surface temperature prediction model. In response to a temperature prediction command for a target ocean area, the spatiotemporal sequence of temperature and salinity collected in the target ocean area is input into the trained sea surface temperature prediction model to obtain the sea surface temperature of the target ocean area at a future time, as output by the sea surface temperature prediction model.

2. The sea surface temperature prediction method based on physical constraints according to claim 1, characterized in that, The sea surface temperature prediction model is constructed based on a convolutional long short-term memory network. The step of inputting the historical temperature-salinity spatiotemporal sequence into the sea surface temperature prediction model, extracting corresponding spatiotemporal features from the historical temperature-salinity spatiotemporal sequence according to the sea surface temperature prediction model, and predicting the sea surface temperature and salinity values ​​at the first moment based on the spatiotemporal features, includes: The historical temperature, salinity, and spatiotemporal sequence is input into the convolutional long short-term memory network to extract corresponding spatiotemporal features from the historical temperature, salinity, and spatiotemporal sequence based on the convolutional long short-term memory network, and to predict the sea surface temperature and salinity at the first moment based on the spatiotemporal features.

3. The sea surface temperature prediction method based on physical constraints according to claim 2, characterized in that, The sea surface temperature prediction model is constructed based on a convolutional long short-term memory network and a temporal convolutional network. The step of inputting the historical temperature-salinity spatiotemporal sequence into the sea surface temperature prediction model, extracting corresponding spatiotemporal features from the historical temperature-salinity spatiotemporal sequence according to the sea surface temperature prediction model, and predicting the sea surface temperature and salinity values ​​at the first moment based on the spatiotemporal features, includes: The historical temperature-salinity spatiotemporal sequence is input into the convolutional long short-term memory network to extract corresponding spatiotemporal features from the historical temperature-salinity spatiotemporal sequence based on the convolutional long short-term memory network. The spatiotemporal features are input into the temporal convolutional network to predict the sea surface temperature and salinity at the first moment based on the temporal convolutional network.

4. The sea surface temperature prediction method based on physical constraints according to claim 3, characterized in that, The loss function includes: ； ； in, For loss function, Let be the mean square error function. It is the first Predicted values ​​of the training samples. It corresponds to the actual value. The training samples consist of a set of historical sea surface temperature spatiotemporal data and corresponding historical salinity spatiotemporal data from the historical sea surface temperature spatiotemporal sequence and the corresponding historical salinity spatiotemporal sequence. For constraint terms The weight hyperparameters, The mapping function for averaging. This is the ReLU function.

5. The sea surface temperature prediction method based on physical constraints according to claim 4, characterized in that, Constraints include: ； in, To use the first set of training samples After inputting the sea surface temperature prediction model, the model outputs the predicted sea surface temperature and salinity values ​​for the first set of training samples. For the second set of training samples After inputting into the sea surface temperature prediction model, the sea surface temperature prediction and salinity prediction values ​​of the second set of training samples output by the sea surface temperature prediction model are: Based on To obtain the density of seawater, Based on The obtained seawater density.

6. The sea surface temperature prediction method based on physical constraints according to claim 5, characterized in that, The seawater density is calculated according to the seawater state equation of the TEOS-10 standard.

7. The sea surface temperature prediction method based on physical constraints according to claim 1, characterized in that, Prior to acquiring the historical temperature, salinity, and spatiotemporal sequences of the ocean area collected during historical periods, the method further includes: The historical sea surface temperature spatiotemporal sequence and the corresponding historical salinity spatiotemporal sequence are labeled, outlier and missing value processing is performed, normalization is performed, and data sequence reshaping is performed.

8. A sea surface temperature prediction device based on physical constraints, characterized in that, The device includes: The data acquisition module is used to acquire historical temperature and salinity spatiotemporal sequences of marine areas collected in historical time periods; the historical temperature and salinity spatiotemporal sequences consist of historical sea surface temperature spatiotemporal sequences and corresponding historical salinity spatiotemporal sequences; The model training module is used to build a sea surface temperature prediction model. The historical temperature and salinity spatiotemporal sequence is input into the sea surface temperature prediction model to extract corresponding spatiotemporal features from the historical temperature and salinity spatiotemporal sequence according to the sea surface temperature prediction model, and based on the spatiotemporal features, predict the sea surface temperature and salinity values ​​at a first moment; the first moment is after the historical time period. Determine the constraint terms, and construct the loss function of the sea surface temperature prediction model based on the constraint terms, the predicted sea surface temperature and salinity values, and the actual sea surface temperature and salinity values ​​of the ocean area at the first moment; the constraint terms are the seawater density difference generated by the sea surface temperature and salinity predicted by the sea surface temperature prediction model. The sea surface temperature prediction model is optimized by backpropagation based on the loss function to obtain the trained sea surface temperature prediction model. The model prediction module is used to respond to the temperature prediction command of the target ocean area by inputting the spatiotemporal sequence of temperature and salinity collected in the target ocean area into the trained sea surface temperature prediction model to obtain the sea surface temperature of the target ocean area at a future time output by the sea surface temperature prediction model.

Images