A Method for Correcting 3D Sea Surface Temperature Numerical Prediction Based on Physically Constrained Spatiotemporal Graph Networks

Abstract

This invention relates to the field of ocean temperature forecasting, specifically a three-dimensional sea surface temperature (SST) numerical forecast correction method based on a physically constrained spatiotemporal graph network. It aims to address the problems of shallow algorithm models, limited spatiotemporal dimensions, and poor adaptability in existing SST numerical forecast correction methods. The method first utilizes Kriging interpolation to spatially reconstruct and spatiotemporally align multi-source SST data. Second, it constructs a fully coupled spatiotemporal-depth model, combining graph convolutional networks, spatiotemporal attention mechanisms, and LSTM to extract multidimensional features and temporal patterns, and uses a deep embedding module to characterize seawater structure. During model training, the three-dimensional heat diffusion equation and vertical temperature gradient smoothing constraints are explicitly embedded into the loss function, constructing an optimization objective driven by both data and physics. Finally, the optimal model is used to predict residuals and generate the final correction result. This method can correct systematic biases in numerical forecast products and improve the accuracy of SST forecasts in complex sea areas.

Description

Technical Field

[0001] This invention relates to the field of ocean temperature forecasting, specifically to a method for correcting three-dimensional ocean temperature numerical forecasts based on physically constrained spatiotemporal graph networks. Background Technology

[0002] The current hydrological environment is complex, with frequent ocean dynamic processes such as hydrothermal processes, oceanic fronts, mesoscale eddies, and internal waves. These dynamic processes interact and influence each other, resulting in complex and drastically changing spatiotemporal distribution characteristics of the three-dimensional sea surface temperature (SST) field in this sea area, posing a significant challenge to the accuracy of three-dimensional SST numerical forecasts. Existing ocean numerical forecasts are limited by imperfect approximate calculation assumptions and parameterization schemes, making it difficult to accurately characterize the evolution trends of these complex processes. This leads to significant systematic biases in the forecast results, necessitating the introduction of efficient correction techniques to eliminate model errors.

[0003] However, most existing sea surface temperature correction techniques rely on traditional empirical statistics, simple formulas, or basic machine learning algorithms, and their application scope is limited to the correction of surface sea surface temperature in small areas of the sea. They lack effective correction methods for deep sea water and large-scale three-dimensional sea surface temperature fields under complex dynamic environments. For example, patent CN114611608A discloses a method for correcting the numerical prediction bias of sea surface height based on a deep learning model, which mainly uses principal component analysis and LSTM module to extract temporal features for fitting. Patent CN116028475A proposes to use bilinear interpolation to unify the grid points and then use a convolutional recurrent neural network (ConvLSTM) to obtain the spatiotemporal features of marine elements for error correction. Such deep learning methods still have significant limitations: (1) Most existing networks are limited to processing two-dimensional sea surface data and lack the ability to encode and structure the physical state of deep sea water in a high dimension, making it difficult to perform global collaborative optimization of three-dimensional sea surface temperature fields. (2) The aforementioned deep learning models are essentially data-driven, and during training, they only aim to reduce statistical errors without incorporating prior knowledge of ocean dynamics into the models. This leads to the fact that in extreme sea conditions or data-sparse areas, the model's predictions are prone to violating the basic laws of thermodynamics, resulting in poor physical consistency and reliability.

[0004] Therefore, existing sea surface temperature (SST) correction methods mainly face three major technical bottlenecks: first, the model structure is shallow and mostly limited to surface data; second, they fail to effectively integrate temporal evolution and three-dimensional spatial features; and third, their prediction accuracy is poor when facing complex marine dynamic environments. To address these issues, this invention proposes an intelligent correction method for three-dimensional SST numerical forecasts. This method can correct the systematic bias of numerical forecast products and significantly improve the accuracy of SST forecasts in complex sea areas. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a three-dimensional sea surface temperature (SST) numerical forecast correction method based on a physically constrained spatiotemporal graph network. This method solves the problems of shallow algorithm models, limited spatiotemporal dimensions, and poor adaptability to complex environments found in existing SST numerical forecast correction methods.

[0006] To achieve the above objectives, the present invention provides a three-dimensional sea surface temperature numerical forecast correction method based on a physically constrained spatiotemporal graph network, which specifically includes the following steps: (1) Acquire numerical forecast data, reanalysis data, and measured data; construct a three-dimensional spatial grid covering the target sea area, and interpolate and register the numerical forecast data and the reanalysis data to the grid points of the three-dimensional spatial grid based on the terrain determination results to complete the spatiotemporal alignment of the above two types of data; The measured data is combined with the reanalysis data, and interpolation is performed using the Kriging interpolation method to obtain the processed reanalysis data. (2) The numerical forecast data is used as input features, and the processed reanalysis data is used as truth labels; the input features and truth labels are mapped to the three-dimensional spatial grid points respectively to construct a spatiotemporal graph structure and generate graph structure data; (3) Construct a physical constraint spatiotemporal graph network model and design a loss function that embeds physical information; (4) Use the graph structure data to train the physical constraint spatiotemporal graph network model until iterative convergence, and save the optimal model parameters; (5) Input the data to be corrected into the optimal model to generate the final sea surface temperature numerical correction result.

[0007] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0008] This invention proposes a physically constrained spatiotemporal graph network model, constructing an end-to-end intelligent architecture with full-dimensional coupling of spatiotemporal and depth dimensions. In the temporal dimension, by integrating LSTM and spatiotemporal attention mechanisms, it achieves accurate capture of long-term evolution patterns. In the spatial dimension, by constructing an explicit ocean grid graph structure and flexibly defining the topological relationships between nodes using an adjacency matrix, it effectively improves the spatial adaptability and feature extraction capabilities for complex sea areas. In the depth dimension, an innovative deep embedding module is introduced, integrating seawater depth as an explicit feature vector into the network, successfully achieving vertical and global collaborative optimization of the three-dimensional sea surface temperature field. Based on this full-dimensional coupling mechanism, this model can accurately capture the three-dimensional features of complex sea areas, significantly improving the accuracy and continuity of three-dimensional sea surface temperature prediction.

[0009] This invention directly embeds the three-dimensional thermal diffusion equation and the vertical temperature gradient smoothing constraint into the model's loss function, forcing the model output to follow the fundamental laws of thermodynamics. This mechanism enhances the physical consistency and reliability of the prediction results and ensures the rationality of the sea surface temperature field in thermodynamic evolution. Attached Figure Description

[0010] Figure 1 This is a flowchart of the present invention; Figure 2 This is a schematic diagram of the physical constraint spatiotemporal graph network model of the present invention. Detailed Implementation

[0011] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0012] like Figure 1 As shown, the method for correcting three-dimensional sea surface temperature numerical forecasts based on physically constrained spatiotemporal graph networks includes the following steps: S1: Import Data Import numerical weather prediction data, reanalysis data, and measured data, and extract their latitude, longitude, and sea surface temperature information. The three types of data must include at least five categories of information: time, depth, longitude, latitude, and sea surface temperature. The reanalysis data is daily average grid data, and the numerical weather prediction data is grid data with a forecast lead time of N days.

[0013] S2: Terrain Determination and Data Interpolation Since numerical forecast data, reanalysis data, and measured data typically have different original resolutions, and the three-dimensional sea surface temperature numerical forecast correction model constructed in this invention requires all input data and output results to maintain strict consistency on the spatial grid, this step first constructs a unified target three-dimensional spatial grid and registers all data onto this grid. The specific processing flow includes two stages: terrain assessment and data interpolation. (1) Terrain determination: The purpose of terrain assessment is to determine the land and sea attributes of each grid point in the target 3D spatial grid, in order to eliminate invalid land points and determine the effective computational domain. Given that the intelligent correction range (total interpolation range A) is slightly larger than the coverage area of ​​the numerical prediction data, this invention adopts a regional assessment strategy: The total interpolation range A is divided into two sub-regions, A1 and A2. The spatial range of sub-region A1 is completely consistent with the numerical weather prediction data. The topographic attribute data of the target grid points in this region are directly extracted as the basis for interpolation. Sub-region A2 is the remaining region after deducting A1 from the total interpolation range A. Based on the reanalysis data, the target grid points in A2 are determined to be either land or sea: when the resolution of the reanalysis data is lower than the resolution of the target interpolation grid, the reanalysis pixels within the coverage area of ​​the target grid are counted. If the proportion of land pixels exceeds 50%, the target point is determined to be land; otherwise, it is determined to be a point to be interpolated.

[0014] (2) Data interpolation: After determining the effective ocean grid points, spatial statistical methods were used to fuse and interpolate the measured data and reanalysis data to generate a high-precision true field.

[0015] In this process, measured data, serving as high-precision real observation constraints, participates in spatial reconstruction along with reanalysis data, aiming to eliminate potential local biases in the reanalysis data. This invention employs the Kriging interpolation method to spatially reconstruct sea surface temperature (SST) after incorporating measured data. This method utilizes the spatial autocorrelation characteristics of the SST field, estimating SST at unknown locations through a weighted linear combination. Let the combined grid and observation point set be... The corresponding sea surface temperature is then analyzed. Then the target location The estimated sea surface temperature can be expressed as: , in The Kriging weights are calculated using a semi-variogram model and satisfy the constraint that the sum of the weights is 1. This method can yield a reanalysis truth field with good spatial continuity.

[0016] S3: Building the dataset First, the processed reanalysis data and numerical weather prediction data are preprocessed. Specifically, this includes standardizing both types of data to eliminate the influence of dimensions, masking land areas and anomalous missing data regions, and constructing a sliding window along the time axis to transform continuous spatiotemporal data into tensor sequences suitable for network input. Next, a dataset is constructed: numerical weather prediction data is used as input, and reanalysis data is used as label data. Finally, the dataset is divided into training, validation, and test sets in an 8:1:1 ratio.

[0017] S4: Construction of Spatiotemporal Graph Networks under Physical Constraints like Figure 2As shown, a spatiotemporal graph network model under physical constraints is constructed at the output end. Its hierarchical structure includes: a graph convolutional network, a spatiotemporal attention module, a deep embedding module, an LSTM network, and a residual correction output end. Based on this, a physical information loss function is designed, embedding the residuals of the ocean diffusion equation and the vertical temperature gradient conservation constraint as regularization terms into the total loss function to impose physical constraints.

[0018] In 3D sea surface temperature (SST) forecasting, traditional convolutional neural networks (CNNs) can typically only handle regular grid data. However, in the complex marine environment, to more flexibly characterize the spatial neighborhood dependencies of the SST field, this algorithm introduces graph convolutional networks to model spatial features. The detailed design and computation process is as follows: First, to transform the pre-processed data into a computationally computable graph structure, the algorithm uses the numerical forecast data as input features and the processed reanalysis ground truth data as ground truth labels. Then, the input features and ground truth labels are mapped to pre-constructed 3D spatial grid points covering the target sea area, completing the binding of features to physical space. Based on this mapping, the algorithm abstracts the ocean grid of the target sea area into a graph structure G=(V,E) to generate graph structure data.

[0019] Node set V: Each node in the diagram corresponds to a specific grid point in the ocean space, and the node has already carried the numerical forecast features and reanalysis ground truth labels mapped to it.

[0020] Edge set E: The edges in the graph represent the spatial topological relationships between adjacent grid points. In this way, the algorithm constructs an adjacency matrix describing node connections. This matrix clarifies how the various ocean grid points in space are interconnected and influence each other, laying the foundation for subsequent feature aggregation.

[0021] When defining the spatial graph structure, this invention uses a Gaussian kernel function based on geographic distance to construct a static adjacency matrix. Set grid nodes. With nodes The spatial distance between them is Elements in the adjacency matrix Defined as: ; in, The standard deviation parameter is used to control the spatial distribution pattern; This is the spatial connectivity threshold; when the distance between two nodes is greater than... When it is assumed that there is no direct spatial interaction, the edge weights are reset to 0 to ensure the sparsity of the graph, thereby reducing computational complexity and avoiding the introduction of long-distance noise.

[0022] After defining the spatial graph structure, the model aggregates information from surrounding neighboring nodes by performing graph convolution operations, thereby extracting the spatial structural features of the sea surface temperature field.

[0023] The basic forward propagation update formula for each layer of a graph convolutional network is as follows: , in: Indicates the first The node feature matrix of the layer, that is, the spatial features of sea surface temperature input to the current network layer. Representing the The trainable weight parameter matrix of the layer is used to perform linear transformations on the features. This represents the adjacency matrix after adding self-connections. This is the original adjacency matrix. It is an identity matrix. Self-connection is added to preserve the node's own characteristic information from the previous layer while aggregating information from surrounding neighboring nodes. The corresponding degree matrix is ​​used to normalize the adjacency matrix (i.e., ...). This prevents nodes with higher degrees from dominating the entire feature representation during feature aggregation. It is a non-linear activation function used to introduce non-linear expressive power into the model.

[0024] In terms of specific network architecture configuration, the graph convolutional network in this algorithm is set as a 2-layer hidden layer structure. To ensure the depth and breadth of feature extraction, each hidden layer contains 128 neurons.

[0025] Through this stacking operation of multi-layer graph convolution, the model can gradually aggregate information from a broader neighborhood, thereby accurately and efficiently extracting high-order spatial topological features of the sea surface temperature field in complex sea areas.

[0026] After obtaining static spatial topological features through graph convolutional networks, considering that the evolution of the ocean dynamic environment (such as ocean currents and mesoscale eddies) is a process of deep coupling between spatial interaction and temporal memory, the model introduces a spatiotemporal attention mechanism to further explore the dynamic evolution patterns in sea surface temperature changes. Traditional graph networks or time series models usually treat time and space in isolation, while the spatiotemporal attention module of this invention, by jointly modeling the "time series dimension" and the "spatial topological dimension," can quantify the cross-influence of the state of a certain sea area at a historical moment on the target sea area at the current moment, thereby achieving true spatiotemporal collaborative optimization.

[0027] Specifically, the computational process and physical significance of the spatiotemporal attention mechanism are as follows: First, the multi-timestep feature sequences output by the graph convolutional network are spatiotemporally flattened and concatenated. Let the time step size of the input sequence be... The number of spatial grid nodes is The feature dimension of a single node is Then the spatiotemporal joint characteristic matrix can be expressed as Each row in this matrix represents the physical state of a specific spatial grid point at a specific historical moment, achieving a preliminary fusion of time coordinates and spatial coordinates in the data structure.

[0028] Based on the self-attention framework, the model integrates spatiotemporal joint features through three independent learnable linear transformation matrices. Mapped to query matrices respectively (Query), Key Matrix (Key) and value matrix (Value). At this time , , Each contains global spatiotemporal coordinate information. The formula for calculating the spatiotemporal attention weights is as follows: , The core step in the formula lies in calculating the query matrix. AND key matrix transpose The inner product of matrices. A dimension of The global spatiotemporal correlation matrix. Any element in this matrix has a clear oceanographic physical meaning: it quantitatively characterizes the... Spatiotemporal nodes (i.e., historical moments) grid points The sea surface temperature fluctuations of the first Spatiotemporal nodes (i.e., the current moment) grid points The hysteresis and spatial radiation effects of ( ). For example, if the upstream sea area is Abnormal temperature rises occur frequently, and are caused by heat conduction or ocean currents. The matrix will assign a high similarity score at the corresponding spatiotemporal index, as it constantly impacts the downstream target sea area. (Scaling factor) Nonlinear activation functions used to prevent gradient divergence. The physical influence of each spatiotemporal node is then probabilistically normalized, and this weight is ultimately used to assess the physical characteristics. Perform weighted aggregation.

[0029] Because ocean systems exhibit both high-frequency local abrupt changes (such as dramatic diurnal sea surface responses) and low-frequency global long-range dependencies (such as large-scale seasonal circulation), this invention employs a multi-head spatio-temporal attention mechanism to avoid the limitations of capturing information from a single perspective. Specifically, the model is configured with four attention heads, each with a feature dimension of 32.

[0030] Each attention head performs the aforementioned spatiotemporal correlation measurement in parallel within an independent subspace, and the outputs of multiple heads are ultimately fused. Through this multi-head collaborative mechanism, the model can adaptively decouple and capture the complex dynamic features of different dimensions in sea surface temperature evolution.

[0031] The feature vectors reconstructed by this module not only retain the original ocean grid topology but also explicitly integrate dynamic features across time and space, thereby effectively improving the accuracy and generalization ability of the forecast model under complex sea conditions.

[0032] In real ocean dynamics, ocean temperature does not change linearly and uniformly in the vertical (depth) direction, but rather exhibits a significant hierarchical structure and highly nonlinear characteristics. Traditional deep learning methods are often limited to processing two-dimensional surface data. Even when three-dimensional data is introduced, seawater depth is often simply input into the network as a regular scalar value. This shallow approach cannot fully represent the physical state of deep seawater. To capture the physical continuity in the depth dimension, this model introduces a deep embedding module.

[0033] This module abandons the traditional approach of treating depth as a single numerical value, and instead encodes and structures depth information in a high dimension. The specific forward computation process is as follows: First, using the deep coding function Seawater depth in scalar form Mapped to high-dimensional feature vectors Its mathematical expression is: , In the specific network configuration, the model explicitly maps seawater depth into a 64-dimensional embedding vector using this function. This high-dimensional mapping gives the physical quantity of "depth" a more expressive mathematical space.

[0034] The depth coding function A linear position coding system based on multi-frequency sine and cosine functions is employed. For a given seawater depth... Its corresponding 3D deep embedding vector The and The formula for calculating bit components is as follows: , , in, This serves as the feature dimension index. Through this periodic function mapping, the one-dimensional depth coordinates are extended to a high-dimensional manifold space, enabling the model to automatically learn the nonlinear thermodynamic relationships between different water layers.

[0035] Subsequently, to enable clear "water layer perception" capabilities for spatial and temporal features, the algorithm embeds the generated depth embedding vectors. Features of the original input in network flow To splice and merge: , Through the aforementioned deep feature embedding and concatenation mechanism, seawater depth is no longer an isolated coordinate label, but rather an explicit feature vector deeply integrated into the network's internal structure. This enables the model to model data from different depth layers within a unified framework. The algorithm adaptively extracts and links the thermodynamic correlations between different water layers, achieving vertical and global collaborative optimization of the three-dimensional sea surface temperature field, thereby constructing a spatiotemporal and depth-coupling architecture.

[0036] The evolution of ocean temperature is not instantaneous and isolated, but is profoundly influenced by previous hydrological conditions (such as the continuous transport of ocean currents and the slow movement of mesoscale eddies), exhibiting high time dependence and nonlinear dynamic characteristics. Traditional recurrent neural networks (RNNs) are prone to gradient vanishing or exploding problems when processing long-sequence data, making it difficult to maintain the memory of physical laws spanning long periods. To address this, this invention introduces a Long Short-Term Memory (LSTM) network to construct a time series model after fusing spatiotemporal and depth features, in order to deeply mine and characterize complex evolutionary patterns over long time scales.

[0037] The computation process of an LSTM unit includes a forget gate, an input gate, and an output gate, and its calculation formula is as follows: , , , , , , in, The input features at the current time, This indicates that the hidden state from the previous moment will be restored. With current input To splice; , and These represent the output vectors of the forget gate, input gate, and output gate, respectively, which are used to control the retention and updating of information. and Representing the cell states at the previous and current moments, respectively, they characterize the long-term temporal memory of sea surface temperature. The formula is based on... Calculated candidate cell states ; , , , These represent the trainable weight matrices in each gating and state update step, respectively, and are concatenated with the vector. Connecting them indicates that matrix multiplication is being performed. , , , This is the corresponding trainable bias vector; This represents the Sigmoid activation function, which maps values ​​to the (0,1) interval to act as a gating switch for information flow. This represents the hyperbolic tangent activation function, which maps numerical values ​​to the (-1, 1) interval to nonlinearize the features and prevent numerical divergence in long-term computations. Ultimately, this represents the current hidden state that will be output to the next time step or the next layer of the network.

[0038] To ensure the model possesses sufficient temporal nonlinear fitting capability, this algorithm employs a deep network architecture in the time series modeling part, specifically a four-layer LSTM network, with each layer containing 256 hidden units. The LSTM network captures the evolution of sea surface temperature over time, providing crucial temporal features for subsequent corrections.

[0039] S5: Training Process and Iterative Convergence Mechanism of Spatiotemporal Graph Networks under Physical Constraints The spatiotemporal graph network model under physical constraints employs a random initialization method for parameter initialization. In terms of training configuration, the model uses the Adam optimizer for parameter updates, with an initial learning rate of 0.001, a batch size of 64, and a maximum training epoch of 300. To improve training stability, a learning decay strategy is introduced: when the validation set error does not significantly decrease over 10 consecutive training epochs, the learning rate is decayed by a factor of 0.5. Simultaneously, to prevent overfitting, Dropout regularization with a ratio of 0.2 is applied during training, along with a weight decay coefficient of 1×10⁻⁶. -4 The L2 regularization term. The above hyperparameter settings are only a preferred embodiment of the present invention in a specific sea area. In practical applications, the above parameters can be adaptively adjusted within a reasonable range according to the data volume and complexity of the target sea area.

[0040] To ensure that the model prediction results meet both data fitting accuracy and marine physical laws, this invention introduces a Physics-Informed Loss function during the model training phase. This loss function embeds the ocean thermodynamic conservation equations and vertical gradient constraints into the traditional data error term, constructing a data-physical dual-driven optimization objective, thereby significantly improving the physical consistency and reliability of the prediction field.

[0041] First, we define the prediction error loss between the model's predicted values ​​and the actual observations. This loss is constructed using the mean squared error and is used to measure the goodness of fit between the model's predicted values ​​and the actual observations. The calculation formula is as follows: , in This represents the sea surface temperature value predicted by the model. This represents the observed sea surface temperature value at the corresponding location. Indicates the number of samples.

[0042] Considering that the ocean temperature field typically exhibits a stable stratified structure in the vertical direction, meaning that sea temperature usually decreases or changes smoothly with increasing depth, this invention introduces a vertical temperature gradient constraint, the calculation formula of which is as follows: , in, Indicates depth Predicted sea surface temperature at that location Indicates the depth level.

[0043] Furthermore, based on the principles of ocean thermodynamics, the spatiotemporal evolution of the sea surface temperature field follows a thermal diffusion mechanism, and its governing equations can be approximately described as a three-dimensional thermal diffusion equation: , in is the thermal diffusivity.

[0044] Based on the above physical relationships, a diffusion constraint loss is constructed. This loss is based on the residual between the time-varying rate of change of the predicted temperature field and the spatial diffusion term, and its calculation expression is as follows: , Finally, the physical information loss function constructed in this invention is defined as: , in and These are weighting coefficients used to balance the influence between data fitting error and physical constraints.

[0045] By introducing the aforementioned physical information constraints during model training, it is possible to effectively ensure that the predicted sea surface temperature field conforms to the laws of ocean physics in terms of spatial structure and vertical distribution, thereby improving the stability and prediction accuracy of the three-dimensional intelligent sea surface temperature correction model.

[0046] After completing the feature extraction and model training described above, this invention uses residual correction to revise the initial prediction results. Specifically, the physically constrained spatiotemporal graph network model learns the error patterns between the numerically predicted sea surface temperature (SST) and the actual SST, and corrects the numerical prediction data accordingly.

[0047] The final correction result is expressed as follows: , in The initial sea surface temperature prediction field (i.e., the "numerical forecast data" imported in step S1 of the instruction manual and preprocessed in S2 and S3) These are the residual correction values ​​obtained from the learning of the physically constrained spatiotemporal graph network model. Residual correction can effectively reduce systematic errors in numerical weather prediction, thereby obtaining a more accurate three-dimensional sea surface temperature field.

[0048] During training, the model uses a random grid search method to optimize the network and update its parameters. Training terminates when the model meets the termination criteria, which include the following two: (1) If the optimal loss value on the validation set is not updated after 30 consecutive iterations during model training, the model is considered to have basically converged, and the training process is terminated in advance. (2) When the number of training iterations reaches the preset maximum number of iterations (300 times), stop the model training and output the final model parameters.

[0049] Subsequently, the models saved during training are evaluated on the validation set based on evaluation metrics, and the optimal correction network model is selected. This optimal model is then used for forward inference on the test set to predict the residual correction value between the numerically predicted sea surface temperature (SST) and the actual SST. This residual correction value is then added to the initial SST prediction field to output the final three-dimensional SST numerical correction result.

[0050] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for correcting three-dimensional sea surface temperature numerical forecasts based on physically constrained spatiotemporal graph networks, characterized in that, Includes the following steps: S1: Acquire numerical forecast data, reanalysis data, and measured data; construct a three-dimensional spatial grid covering the target sea area, and interpolate and register the numerical forecast data and the reanalysis data to the grid points of the three-dimensional spatial grid based on the terrain determination results, thereby completing the spatiotemporal alignment of the above two types of data; The measured data is combined with the reanalysis data, and interpolation is performed using the Kriging interpolation method to obtain the processed reanalysis data. S2: The numerical forecast data is used as input features, and the processed reanalysis data is used as ground truth labels; the input features and ground truth labels are mapped to the three-dimensional spatial grid points respectively to construct a spatiotemporal graph structure and generate graph structure data; S3: Construct a physical constraint spatiotemporal graph network model and design a loss function that embeds physical information; The specific implementation process of the physically constrained spatiotemporal graph network model includes: (1) Input the input features in the graph structure data into the graph convolutional network to extract spatial features, and strengthen key spatiotemporal information through a spatiotemporal attention mechanism; (2) The deep embedding module is used to model the vertical structure of the ocean, and the spatiotemporal dependence of sea temperature change is captured by the LSTM network; (3) The numerical forecast sea surface temperature is intelligently corrected by the residual correction module, and high-precision three-dimensional sea surface temperature correction results are output; The deep embedding module performs high-dimensional encoding and structured representation of deep information, specifically including the following steps: (1) Depth of seawater Mapped to high-dimensional feature vectors: ， in, For depth coding functions; (2) embedding the deep embedding vector Compared with the original input features To merge: ； S4: Use the graph structure data to train the physical constraint spatiotemporal graph network model until iterative convergence, and save the optimal model parameters; S5: Input the data to be corrected into the optimal model to generate the final sea surface temperature numerical correction result.

2. The method for correcting three-dimensional sea surface temperature numerical forecasts based on physically constrained spatiotemporal graph networks according to claim 1, characterized in that, The acquisition of the terrain determination result in S1 specifically includes: (1) Divide the total interpolation range A into two sub-regions A1 and A2; (2) The spatial range of sub-region A1 is completely consistent with the numerical forecast data. The terrain attribute data of the target grid points in this region corresponding to latitude and longitude are directly extracted as the basis for interpolation. (3) Sub-region A2 is the remaining area after deducting A1 from the total interpolation range A. Based on the reanalysis data, the target grid points in A2 are determined to be land or sea: when the resolution of the reanalysis data is lower than the resolution of the target interpolation grid, the reanalysis pixels within the coverage area of ​​the target grid are counted. If the number of land pixels accounts for more than 50%, the target point is determined to be land; otherwise, it is determined to be a point to be interpolated.

3. The method for correcting three-dimensional sea surface temperature numerical forecasts based on physically constrained spatiotemporal graph networks according to claim 1, characterized in that, The physical information loss function in S3 includes: prediction error loss, vertical smoothing constraint loss, and diffusion constraint loss.

4. The method for correcting three-dimensional sea surface temperature numerical forecasts based on physically constrained spatiotemporal graph networks according to claim 3, characterized in that, The prediction error loss is constructed using the mean squared error, and is used to measure the goodness of fit between the model's predicted values ​​and the actual observed values. Its calculation formula is as follows: ， in This represents the sea surface temperature value predicted by the model. This represents the observed sea surface temperature value at the corresponding location. Indicates the number of samples; The vertical smoothing constraint loss is constructed by penalizing temperature gradient changes between adjacent depth layers to prevent drastic, non-physical oscillations in the prediction results. Its calculation formula is as follows: ， in, Indicates depth Predicted sea surface temperature at that location This indicates the predicted value of the next layer. Indicates the depth level; The diffusion constraint loss is constructed based on the residual between the predicted time rate of change of the temperature field and the spatial diffusion term, and its calculation expression is as follows: ， in, denoted as the thermal diffusivity, T as the sea surface temperature, and x, y, z as the longitude, latitude, and depth coordinates, respectively.

Images