A Sea Surface Temperature Prediction Method and System Based on Supervised Assisted Graph Learning and Conditional Diffusion

By constructing an adaptive fusion of geographic maps and implicit structure maps, and combining multi-scale conditional priors and Brownian bridge diffusion prediction, the high computational cost and structural instability of sea surface temperature prediction in existing technologies are solved, and high-precision and stable future sea surface temperature prediction is achieved.

CN123116135APending Publication Date: 2026-11-17OCEAN UNIV OF CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611275555.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing sea surface temperature prediction methods are computationally expensive and susceptible to model errors and parameter uncertainties in complex nearshore areas and high-resolution prediction scenarios. They are also difficult to effectively learn structural supervision signals for future evolution, resulting in global structural instability and insufficient regional evolution.

Method used

A method based on supervised auxiliary graph learning and conditional diffusion is adopted. By constructing a geographic map and an implicit structure graph, an adaptive fusion map is generated using a side-level adaptive fusion mechanism. Combined with multi-scale conditional priors and Brownian bridge diffusion prediction, the guidance intensity at each scale is dynamically adjusted to achieve the gradual generation of future sea surface temperature.

Benefits of technology

It improves the numerical accuracy, spatial structure consistency, and long-term stability of sea surface temperature prediction, reduces information interference between historically similar but future diverging regions, and alleviates the scale mixing problem caused by the one-time fusion of multi-scale information.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116135A_ABST
    Figure CN123116135A_ABST
Patent Text Reader

Abstract

This invention relates to the field of sea surface temperature (SST) prediction, and discloses a method and system for SST prediction based on supervised auxiliary graph learning and conditional diffusion. The method includes: acquiring historical SST sequences and valid ocean nodes; constructing and fusing a geographic map and an implicit structure map to obtain an adaptive fused map, which is then input into a graph encoder to obtain a historical spatiotemporal representation. During the training phase, backpropagation is performed through supervised auxiliary graph learning with future consistency oversight. The historical spatiotemporal representation is input into a time-series prediction network to generate multi-scale conditional priors; and the conditional priors at different scales are injected into a denoising network for multi-scale prior-guided Brownian bridge diffusion prediction to obtain future SST prediction results. This improves the numerical accuracy, spatial structure consistency, and long-term stability of SST prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of sea surface temperature prediction, and specifically relates to a sea surface temperature prediction method and system based on supervised graph learning and conditional diffusion. Background Technology

[0002] Sea surface temperature (SST) is a crucial physical variable characterizing the thermal state of the ocean, air-sea energy exchange, and changes in the marine environment. Accurate prediction of future SST fields provides vital data support for applications such as marine environmental monitoring, typhoon development forecasting, marine heat wave early warning, climate anomaly identification, fisheries resource management, and nearshore ecological environmental protection. Therefore, achieving high-precision, structurally consistent SST field prediction with good spatiotemporal continuity is a significant technical challenge in the fields of marine remote sensing monitoring and intelligent marine environmental forecasting.

[0003] Existing sea surface temperature (SST) prediction methods mainly include numerical modeling methods and data-driven methods. However, in complex nearshore areas and high-resolution prediction scenarios, numerical modeling methods typically rely on accurate initial fields, boundary conditions, external forcing, and parameterization schemes, resulting in high computational costs. Furthermore, they are susceptible to model errors and parameter uncertainties in locally complex marine environments.

[0004] With the continuous accumulation of satellite remote sensing observation data and reanalysis data, data-driven methods based on deep learning are gradually being applied to sea surface temperature (SST) prediction tasks. Methods such as convolutional neural networks, recurrent neural networks, ConvLSTM, and Transformer can automatically learn nonlinear spatiotemporal evolution patterns from historical SST sequences, achieving certain results in short-term prediction tasks.

[0005] To further characterize the spatial interactions between ocean regions, SST prediction methods based on graph neural networks have emerged in recent years. These methods typically represent ocean grids or observation stations as graph nodes, and construct spatial relationships between nodes based on geographical proximity, historical statistical correlation, or adaptive similarity learned from observation sequences. This approach breaks through the limitations of fixed convolutional neighborhoods, establishing connections between geographically close, historically correlated, or potentially similar ocean regions, thereby enhancing the model's ability to express complex ocean spatial dependencies to some extent.

[0006] However, the model lacks a structural supervision signal for future evolution during training, making it unable to explicitly guide the model in learning which regions will exhibit more consistent temperature changes in the future. This makes it difficult to further enhance existing graph structural relationships from "historical correlations" to "predictive structural priors for future evolution." Simultaneously, it is difficult to effectively align multi-scale conditional information with different stages of the diffusion generation process, easily leading to problems such as global structural instability, insufficient regional evolution, or inadequate recovery of local details.

[0007] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0008] To address or at least mitigate one or more of the above problems, a sea surface temperature prediction method and system based on supervised auxiliary graph learning and conditional diffusion is provided. This method can reduce information interference between historically similar but future divergent regions, alleviate the scale mixing problem caused by the one-time fusion of information at different scales, and improve the numerical accuracy, spatial structure consistency, and long-term stability of sea surface temperature prediction.

[0009] To achieve the above objectives, according to a first aspect of the present invention, a sea surface temperature prediction method based on supervised auxiliary graph learning and conditional diffusion is provided, comprising the following steps:

[0010] S1. Auxiliary graph learning based on future consistency supervision:

[0011] Historical sea surface temperature sequences and valid ocean nodes are obtained to construct a geographic map and an implicit structure map. An edge-level adaptive fusion mechanism is then used to fuse the geographic map and the implicit structure map to obtain an adaptive fused map. The adaptive fused map and the historical features of each valid ocean node are input into a graph encoder, and a historical spatiotemporal representation is obtained through graph message propagation.

[0012] During the training phase, real future sea surface temperature sequences are acquired, and future sea surface temperature change trajectories and future consistency positive sample sets for each effective ocean node are constructed. Based on the graph-encoded node representations of each effective ocean node obtained by the graph encoder, the relationship score between nodes is calculated, and the future consistency supervision loss is calculated using a multi-positive sample contrast learning objective function. The trainable parameters participating in auxiliary graph learning are updated using the future consistency supervision loss through backpropagation.

[0013] S2, Multi-scale conditional prior generation:

[0014] The historical spatiotemporal representation is input into the temporal prediction network to predict the future potential state of each effective ocean node at multiple future moments within the prediction period. The future potential feature map is obtained through mapping operations. The future potential feature map is input into convolutional projection branches of different scales, and conditional priors of different scales are generated through multi-scale feature transformation operations.

[0015] S3. Multi-scale prior asymptotically guided Brownian bridge diffusion prediction:

[0016] During the training phase, the denoising network is trained. During the inference phase, conditional priors of different scales are injected into the feature layers corresponding to the spatial resolution of the denoising network. The guidance strength of the conditional priors of different scales is dynamically determined according to the defined normalized inverse denoising progress variable. After conditional enhancement fusion based on the guidance strength, the inverse denoising update is repeatedly executed to obtain the future sea surface temperature prediction result.

[0017] To achieve the above objectives, according to a second aspect of the present invention, a sea surface temperature prediction system based on supervised graph learning and conditional diffusion is provided, comprising:

[0018] An auxiliary graph learning module based on future consistency supervision is used for auxiliary graph learning with future consistency supervision.

[0019] Historical sea surface temperature sequences and valid ocean nodes are obtained to construct a geographic map and an implicit structure map. An edge-level adaptive fusion mechanism is then used to fuse the geographic map and the implicit structure map to obtain an adaptive fused map. The adaptive fused map and the historical features of each valid ocean node are input into a graph encoder, and a historical spatiotemporal representation is obtained through graph message propagation.

[0020] During the training phase, real future sea surface temperature sequences are acquired, and future sea surface temperature change trajectories and future consistency positive sample sets for each effective ocean node are constructed. Based on the graph-encoded node representations of each effective ocean node obtained by the graph encoder, the relationship score between nodes is calculated, and the future consistency supervision loss is calculated using a multi-positive sample contrast learning objective function. The trainable parameters participating in auxiliary graph learning are updated using the future consistency supervision loss through backpropagation.

[0021] The multi-scale conditional prior generation module is used for generating multi-scale conditional priors.

[0022] The historical spatiotemporal representation is input into the temporal prediction network to predict the future potential state of each effective ocean node at multiple future moments within the prediction period. The future potential feature map is obtained through mapping operations. The future potential feature map is input into convolutional projection branches of different scales, and conditional priors of different scales are generated through multi-scale feature transformation operations.

[0023] A multi-scale prior-guided incremental Brownian bridge diffusion prediction module is used for multi-scale prior-guided incremental Brownian bridge diffusion prediction.

[0024] During the training phase, the denoising network is trained. During the inference phase, conditional priors of different scales are injected into the feature layers corresponding to the spatial resolution of the denoising network. The guidance strength of the conditional priors of different scales is dynamically determined according to the defined normalized inverse denoising progress variable. After conditional enhancement fusion based on the guidance strength, the inverse denoising update is repeatedly executed to obtain the future sea surface temperature prediction result.

[0025] Furthermore, the multi-scale prior asymptotically guided Brownian bridge diffusion prediction module includes:

[0026] The denoising network training unit is used to construct a Brownian bridge reference endpoint by repeating the sea surface temperature field of the last moment of the historical sea surface temperature sequence along the prediction time dimension during the training phase; constructing a Brownian bridge intermediate state and a corresponding Brownian bridge prediction target based on the real future sea surface temperature sequence, the Brownian bridge reference endpoint, the Brownian bridge interpolation coefficients corresponding to the current diffusion time step, and random noise; inputting the Brownian bridge intermediate state, the current diffusion time step, and the multi-scale conditional prior into the denoising network to obtain the Brownian bridge prediction result; calculating the diffusion training loss based on the error between the Brownian bridge prediction result and the Brownian bridge prediction target, and updating the network parameters of the denoising network through backpropagation.

[0027] The inverse denoising prediction unit, during the inference phase, uses the Brownian bridge reference endpoint constructed based on the historical sea surface temperature sequence to be predicted as the initial state of the inverse denoising process. Following a descending order of diffusion time steps, it progressively updates the current Brownian bridge state using the trained denoising network. At each inverse denoising time step, coarse-scale, meso-scale, and fine-scale conditional priors are injected into the feature layers corresponding to the spatial resolution of the denoising network. The guiding strength of each scale's conditional prior is dynamically determined based on the normalized inverse denoising progress variable, and conditional enhancement fusion is performed. This process is repeated until the future sea surface temperature field sequence for the prediction period is obtained.

[0028] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0029] In the training phase, this invention constructs future consistent positive samples using real future SST change trajectories. Contrastive learning is employed to make the graph encoding representation approximate future evolution trends, enhancing the ability of historical driving features to represent future states. Furthermore, the real future SST serves only as a supervisory signal during the training phase, preventing information leakage during inference. In the inference phase, future potential states are predicted based on historical graph encoding, generating coarse-scale, meso-scale, and fine-scale conditional priors. The guidance intensity at each scale is dynamically adjusted using a Brownian bridge reverse denoising process, achieving a progressive generation from global structure to local details. This method effectively reduces information interference between historically similar but future-differentiated regions, alleviates the scale mixing problem caused by the one-time fusion of multi-scale information, and significantly improves the numerical accuracy, spatial structure consistency, and long-term stability of sea surface temperature prediction.

[0030] Specifically, this invention constructs future consistent positive samples based on the real future SST change trajectory during the training phase, and improves the ability of the history-driven graph representation to represent the future SST evolution by comparing and learning the graph encoding representations of the corresponding nodes to be closer to the real future SST change trajectory. The real future SST is only used for training supervision, and the graph relationships and node representations in the inference phase are generated from the historical SST sequence, so there is no leakage of future information.

[0031] Specifically, this invention predicts future potential states based on graph-encoded historical spatiotemporal representations and generates coarse-scale, meso-scale, and fine-scale conditional priors. Subsequently, the guiding strength of each scale prior is dynamically adjusted according to the progress of Brownian bridge reverse denoising, so that the coarse-scale prior is used to establish the global structure, the meso-scale prior is used to refine the regional evolution pattern, and the fine-scale prior is used to recover local boundaries, fronts, and detailed information, thereby realizing the gradual generation of future SST fields from coarse to fine.

[0032] Specifically, the present invention can reduce information interference between historically similar but future diverging regions, alleviate the scale mixing problem caused by the one-time fusion of information at different scales, and improve the numerical accuracy, spatial structure consistency and long-term stability of sea surface temperature prediction.

[0033] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0034] The accompanying drawings, as part of this invention, are provided to further illustrate the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention, but do not constitute an undue limitation thereof. Clearly, the drawings described below are merely some embodiments; those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0035] In the attached diagram:

[0036] Figure 1 This is a schematic diagram of the auxiliary graph learning module with future consistency supervision in this specific embodiment;

[0037] Figure 2 This is a schematic diagram of the structure of the multi-scale conditional prior generation and stage-aligned Brownian bridge diffusion prediction module in this specific embodiment.

[0038] Figure 3 This is a system architecture diagram of the sea surface temperature prediction system based on supervised graph learning and conditional diffusion in this specific embodiment. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0040] In the following description, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0041] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0042] This invention provides a sea surface temperature prediction method based on supervised graph learning and conditional diffusion, comprising the following steps:

[0043] S1. Auxiliary graph learning based on future consistency supervision:

[0044] Historical sea surface temperature sequences and valid ocean nodes are obtained to construct a geographic map and an implicit structure map. An edge-level adaptive fusion mechanism is then used to fuse the geographic map and the implicit structure map to obtain an adaptive fused map. The adaptive fused map and the historical features of each valid ocean node are input into a graph encoder, and a historical spatiotemporal representation is obtained through graph message propagation.

[0045] During the training phase, real future sea surface temperature sequences are acquired, and future sea surface temperature change trajectories and future consistency positive sample sets for each effective ocean node are constructed. Based on the graph-encoded node representations of each effective ocean node obtained by the graph encoder, the relationship score between nodes is calculated, and the future consistency supervision loss is calculated using a multi-positive sample contrast learning objective function. The trainable parameters participating in auxiliary graph learning are updated using the future consistency supervision loss through backpropagation.

[0046] S2, Multi-scale conditional prior generation:

[0047] The historical spatiotemporal representation is input into the temporal prediction network to predict the future potential state of each effective ocean node at multiple future moments within the prediction period. The future potential feature map is obtained through mapping operation. The future potential feature map is input into the convolutional projection branch at different scales. Through multi-scale feature transformation operation, conditional priors at different scales are generated respectively.

[0048] S3. Multi-scale prior asymptotically guided Brownian bridge diffusion prediction:

[0049] During the training phase, the denoising network is trained. During the inference phase, conditional priors of different scales are injected into the feature layers corresponding to the spatial resolution of the denoising network. The guidance strength of the conditional priors of different scales is dynamically determined according to the defined normalized inverse denoising progress variable. After conditional enhancement fusion based on the guidance strength, the inverse denoising update is repeatedly executed to obtain the future sea surface temperature prediction result.

[0050] In some embodiments, constructing a geographic map and an implicit structure map includes:

[0051] An original two-dimensional spatial grid is formed based on the sea surface temperature field. Based on the original two-dimensional spatial grid, an effective ocean area mask and effective ocean nodes are defined, and historical sea surface temperature sequences are obtained.

[0052] The latitude and longitude coordinates of each valid ocean node are mapped to a three-dimensional unit sphere, and the spherical distance between any two valid ocean nodes is calculated. For each valid ocean node, a set of valid ocean nodes with the smallest spherical distances are selected from other valid ocean nodes (excluding itself) to form a set of geographic neighbors. Based on the spherical distances between the current valid ocean node and each valid ocean node in the set of geographic neighbors, and a preset geographic temperature parameter, the corresponding geographic relationship value is calculated. The geographic relationship value corresponding to each valid ocean node is subjected to row-by-row softmax normalization to obtain the geographic graph edge weights. For valid ocean nodes that do not belong to the set of geographic neighbors, the corresponding geographic graph edge weights are set to zero, thus forming a geographic graph.

[0053] Historical feature representations of each valid ocean node are extracted from historical sea surface temperature sequences using a historical spatiotemporal encoder. Based on these historical feature representations, the structural similarity between different valid ocean nodes is calculated using cosine similarity. For each valid ocean node, several valid ocean nodes with the highest structural similarity are selected from other valid ocean nodes (excluding itself) as structural neighbor nodes, forming a corresponding structural neighbor set. The structural relationship value is calculated based on the structural similarity between the current valid ocean node and its structural neighbor nodes. The structural relationship value corresponding to each valid ocean node is normalized using a mask row to obtain the edge weights of the structural graph. For other valid ocean nodes that do not belong to the structural neighbor set, the corresponding edge weights of the structural graph are set to zero, and an implicit structural graph is formed by all valid ocean nodes and their corresponding edge weights.

[0054] Specifically, the latitude range is selected as N to N, Longitude range is E to The sea surface temperature field of E is arranged according to the latitude and longitude sampling locations of the sea surface temperature data to form an original two-dimensional spatial network, and the spatial size of the original two-dimensional spatial network is denoted as . ,in and These represent the number of grids in the latitudinal and longitudinal directions of the original two-dimensional spatial network, respectively.

[0055] Based on the land-sea distribution in the original two-dimensional spatial grid and the data validity, an effective ocean region mask is defined, represented as:

[0056] ;

[0057] in, Indicates an effective marine area mask. Represents the position in the original two-dimensional spatial grid. The corresponding mask, Indicates position For effective ocean grid location, Indicates position This refers to a land location.

[0058] Effective ocean nodes are defined based on the effective ocean region mask and are represented as follows:

[0059] ;

[0060] ;

[0061] in, For the number of effective ocean nodes, Represents the set of valid ocean nodes;

[0062] The historical sea surface temperature sequence is obtained and represented as follows:

[0063] ;

[0064] in, Represents a historical sea surface temperature sequence. This indicates the time length of a historical sea surface temperature sequence.

[0065] Map the latitude and longitude of each valid ocean node to a three-dimensional unit sphere, and calculate the spherical distance between any two valid ocean nodes; for each valid ocean node, select the one with the smallest spherical distance from valid ocean node i, excluding the current valid ocean node i. A set of valid ocean nodes constitutes the geographical neighbor set of valid ocean node i. ;

[0066] Based on the spherical distances between the current valid ocean nodes and each valid ocean node in the set of geographic neighbors, and the preset geographic temperature parameter, the corresponding geographic relationship value is calculated. The geographic relationship value is negatively correlated with the spherical distance, as expressed below:

[0067] ;

[0068] in, This represents the unnormalized geographic relationship value between the valid ocean node i and its geographic neighbor node j; This represents the spherical distance between the valid ocean node i and its geographic neighbor node j; This represents the geographic distance attenuation parameter, used to adjust the degree of influence of spherical distance on geographic relationship values.

[0069] For each valid ocean node, perform row-wise softmax normalization on the geographic relation value to obtain the geographic graph edge weights:

[0070] ;

[0071] in, Represents the geographic graph edge weight from valid ocean node i to geographic neighbor node j; This represents an exponential function with the natural constant e as its base. Represents the set of geographical neighbors of the valid ocean node i. The geographic neighbor node index used at that time; This represents the unnormalized geographic relationship value between a valid ocean node i and its geographic neighbor node m; for other valid ocean nodes that do not belong to the set of geographic neighbors of valid ocean node i, the corresponding geographic graph edge weight between them and valid ocean node i is set to zero.

[0072] Specifically, geographic maps are mainly used to represent the local spatial continuity between adjacent ocean regions. However, sea surface temperature evolution is also affected by factors such as ocean current transport, air-sea interaction, and local dynamic processes. Non-local correlations may exist between different sea areas that extend beyond a fixed geographic neighborhood. Therefore, this invention further constructs an implicit structure map based on historical sea surface temperature sequences.

[0073] Specifically, the historical sea surface temperature sequence is encoded using a historical spatiotemporal encoder to obtain the historical feature matrix of all valid ocean nodes, represented as:

[0074] ;

[0075] in, Represents a historical spatiotemporal encoder; Represents a historical sea surface temperature sequence; This represents the historical feature matrix of all valid ocean nodes;

[0076] The historical feature representations of the current valid ocean node i and any other valid ocean node q besides valid ocean node i are extracted from the historical feature matrix and are respectively represented as:

[0077] ;

[0078] ;

[0079] in, Indicates a valid ocean node Historical characteristics indicate that Indicates a valid ocean node Historical characteristics indicate that Representation of historical feature matrix The OK, Representation of historical feature matrix The OK;

[0080] Based on the historical characteristics of each valid ocean node, the structural similarity between valid ocean node i and other valid ocean nodes q is calculated using cosine similarity, and expressed as:

[0081] ;

[0082] in, Indicates a valid ocean node With effective ocean nodes Structural similarity between them; The inner product between the historical feature representations of valid ocean node i and valid ocean node q is represented. The second norm of a vector;

[0083] For each valid ocean node, select the one with the highest structural similarity from the other valid ocean nodes q besides the current valid ocean node i. Each valid ocean node constitutes a structural neighbor set of valid ocean node i. ;

[0084] The unnormalized structural relationship value is calculated based on the structural similarity between the current effective ocean node i and its structural neighbor node k, and is expressed as follows:

[0085] ;

[0086] in, This represents the structural similarity between an effective ocean node i and its structural neighbor node k. Parameters used to control the sharpness of the distribution of structural relationships. This represents the structural relationship value between a valid ocean node i and its structural neighbor node k, where k represents the set of structural neighbors for traversing the valid ocean node i. The structural neighbor node index used at that time;

[0087] The structural relation values ​​corresponding to each valid ocean node are normalized using a masking process to obtain the edge weights of the structural graph, which are represented as follows:

[0088] ;

[0089] in, Let represent the edge weight of the structural graph pointing from a valid ocean node i to a structural neighbor node k; n represents the neighbor node index used when traversing the structural neighbor set. This represents the structural relationship value between a valid ocean node i and its neighbor node n; the edge weights of the structural graph corresponding to other valid ocean nodes that do not belong to the set of structural neighbors of valid ocean node i are set to zero.

[0090] In practical applications, a side-level adaptive fusion mechanism is used to fuse the geographic map and the implicit structure map to obtain an adaptive fused map, including:

[0091] For each valid ocean node, merge the geographic neighbor set and structural neighbor set of the current valid ocean node and remove duplicate nodes to obtain a candidate neighbor set. For each valid ocean node, determine the candidate neighbor set by the union of the geographic neighbor set and the implicit structural neighbor set of the valid ocean node, and determine the candidate edge by the connection between the valid ocean node and each node in the candidate neighbor set. Obtain the geographic graph edge weight and the structural graph edge weight of each candidate edge in the geographic graph and the implicit structural graph respectively. If the current candidate edge does not exist in the geographic graph, set the geographic graph edge weight corresponding to the current candidate edge to zero; if the current candidate edge does not exist in the implicit structural graph, set the structural graph edge weight corresponding to the current candidate edge to zero.

[0092] Based on the absolute differences between the geographic graph edge weights, structural graph edge weights, geographic graph edge weights, and structural graph edge weights corresponding to the candidate edges, and the sum of the geographic graph edge weights and structural graph edge weights, edge-level fusion features are constructed.

[0093] Edge-level fusion features are input into a multilayer perceptron, and edge-level adaptive fusion coefficients are generated using a sigmoid activation function. The edge-level adaptive fusion coefficients and their complements are then weighted and summed with respect to the geographic graph weights and structural graph weights corresponding to the candidate edges, respectively, to obtain the fusion candidate edge weights. Specifically, when a candidate edge does not exist in the geographic graph, the corresponding geographic graph weight is set to zero before being included in the weighted summation; when a candidate edge does not exist in the implicit structural graph, the corresponding structural graph weight is set to zero before being included in the weighted summation. The initial fusion graph is constructed from these fusion candidate edge weights.

[0094] The initial fused graph and its transpose are averaged to achieve symmetry, and self-connections are added to the symmetric fused graph. Based on the sum of all outgoing edge weights corresponding to each valid ocean node, the fused graph after adding self-connections is normalized row by row to obtain an adaptive fused graph. Graph message propagation is performed based on the adaptive fused graph to update the node representation of each valid ocean node layer by layer.

[0095] For example, in order to simultaneously utilize geographical continuity and the cross-regional structural associations implicit in historical sequences, the present invention employs a border-level adaptive fusion mechanism to fuse the geographical map and the implicit structural map into an adaptive fusion map.

[0096] For example, for each valid ocean node Effective marine nodes The geographic neighbor set and structural neighbor set are merged and duplicate nodes are removed to obtain the valid ocean nodes. The corresponding set of candidate neighbors is represented as:

[0097] ;

[0098] in, Indicates the currently valid ocean node The corresponding set of candidate neighbors, Indicates a valid ocean node The set of geographical neighbors; Indicates a valid ocean node The set of structural neighbors.

[0099] For example, for the candidate neighbor set Any candidate neighbor node r in the list will be an effective ocean node. The connection between a candidate neighbor node r and the candidate edge is determined as a candidate edge. ,r). Obtain candidate edges respectively ( The edge weights of the geographic map in the geographic map. and the edge weights of the structure graph in the implicit structure graph If candidate edge ( If r) does not exist in the geographic map, then the corresponding geographic map edge weights will be adjusted. Set to zero; if candidate edge ( If r) does not exist in the implicit structure graph, then the corresponding edge weights of the structure graph will be... Set to zero.

[0100] For any candidate edge ( Based on the geographic graph edge weights, structural graph edge weights, the absolute differences between geographic graph edge weights and structural graph edge weights, and the sum of geographic graph edge weights and structural graph edge weights corresponding to the candidate edges, edge-level fusion features are constructed, expressed as follows:

[0101] ;

[0102] in, The edge-level fusion feature corresponding to candidate edge (i, r) is represented; r represents the effective ocean node. The corresponding candidate neighbor node index; Indicate candidate edges ( The corresponding geographic graph edge weights for r); Indicate candidate edges ( The edge weights of the structure graph corresponding to r);

[0103] Candidate edges ( The edge-level fusion features corresponding to r) The input is a multilayer perceptron, and the corresponding edge-level adaptive fusion coefficients are generated through the Sigmoid activation function, as follows:

[0104] ;

[0105] in, MLP(.) represents the Sigmoid activation function, and MLP(.) represents a multilayer perceptron. Indicate candidate edges ( ,r) corresponds to the edge-level adaptive fusion coefficient;

[0106] For example, candidate edges ( The edge-level adaptive fusion coefficients and their complements are used to weight and sum the edge weights of the geographic graph and the structural graph, respectively, to obtain the fusion candidate edge weights, expressed as:

[0107] ;

[0108] in, Represent any candidate edge Candidate edge weights for fusion are obtained by weighted summation of the edge weights of the geographic graph and the edge weights of the structural graph using the edge-level adaptive fusion coefficients and their complements, respectively. This represents the complement of the edge-level adaptive fusion coefficients.

[0109] For example, the initial fused graph and its transpose are averaged to achieve symmetry, and self-joins are added to the symmetric fused graph. Based on the sum of all outgoing edge weights corresponding to each valid ocean node, the fused graph after adding self-joins is normalized row by row to obtain an adaptive fused graph, represented as follows:

[0110] ;

[0111] ;

[0112] in, This represents the initial adaptive fusion graph composed of the initial fusion edge weights of each candidate edge; This represents the transpose of the initial adaptive fusion graph; It is an identity matrix used to add self-connections to each valid ocean node; This represents the normalized adaptive fusion graph; Let represent the edge weight from valid ocean node i to valid ocean node j in the adaptive fusion graph; This represents the edge weight between valid ocean node i and valid ocean node j in the adaptive fusion graph; Represents the set of nodes consisting of all valid ocean nodes; Represents traversing a set of nodes The valid ocean node index used at that time; Represents the effective ocean node i and the effective ocean node The boundary weight between; This represents a default positive number used to avoid a denominator of 0.

[0113] The adaptive fusion graph is input into the graph encoder, which then performs graph message propagation among interconnected valid ocean nodes based on the adaptive fusion graph, updating the node representation of each valid ocean node layer by layer; the graph encoder from the lth layer to the th layer... The node representation update process of a layer is represented as follows:

[0114] ;

[0115] Where l represents the network layer index of the graph encoder, and the superscript... and These represent the first and second parts of the graph encoder. Layer and first layer; Let represent the node feature matrix of all valid nodes in the l-th layer, and let the i-th row of the node feature matrix represent the node feature matrix of valid ocean node i in the l-th layer. This represents the learnable feature transformation matrix of the l-th layer of the graph encoder. Representation layer normalization, This represents the SiLU activation function.

[0116] In some other embodiments, the training phase of step S1 includes:

[0117] The true future sea surface temperature sequence following the historical observation window is obtained from the sea surface temperature observation data. For each valid ocean node, the sea surface temperature value at the last moment of the historical observation window is used as a reference to calculate the change in the true future sea surface temperature sequence at each future moment relative to the reference sea surface temperature value, thus forming the future sea surface temperature change trajectory of the corresponding valid ocean node.

[0118] For each valid ocean node, Euclidean distance is used to calculate the difference in future trajectory between the current valid ocean node and other valid ocean nodes excluding itself. Several valid ocean nodes with the smallest difference in future trajectory are selected from the other valid ocean nodes to form a set of future consistent positive samples corresponding to the current valid ocean node. The adaptive fusion graph and the historical features of each valid ocean node are input into the graph encoder, and the graph encoded node representation corresponding to each valid ocean node is obtained through graph message propagation. The graph encoded node representation is normalized, and the relationship score between each valid ocean node is calculated based on the inner product between the normalized graph encoded node representations and the preset contrast learning temperature parameter.

[0119] Based on the set of future consistent positive samples and relationship scores corresponding to each valid ocean node, the future consistency supervision loss is calculated using a multi-positive sample contrastive learning objective function. Through backpropagation, the trainable parameters in the graph encoder, the historical spatiotemporal encoder for extracting historical feature representations, and the multilayer perceptron for generating edge-level adaptive fusion coefficients are updated using the future consistency supervision loss.

[0120] Specifically, the adaptive fusion graph can describe the local continuity and potential cross-regional associations within the historical observation window. However, relying solely on the final sea surface temperature prediction error for optimization cannot clearly define whether the constraint graph encoding representation preferentially associates ocean regions with similar future change processes. To address this, the present invention introduces future consistency supervision during the training phase. Instead of directly specifying or replacing the edge weights in the adaptive fusion graph, it scores the relationships between the constraint graph encoding node representations based on the actual future sea surface temperature change trajectory.

[0121] Specifically, obtain the true future sea surface temperature sequence located after historical observation windows from sea surface temperature observation data. For each valid ocean node i, using the sea surface temperature value of valid ocean node i at the last moment of the historical observation window as a reference, the change in the future sea surface temperature sequence relative to the reference sea surface temperature value at each future moment is calculated, forming the future sea surface temperature change trajectory of the corresponding valid ocean node i, as shown below:

[0122] ;

[0123] in, This represents the time length of a true future sea surface temperature sequence; This represents the true sea surface temperature value of the effective ocean node i at time t+1 in the future; express The actual sea surface temperature value of effective ocean node i at a future moment; This represents the sea surface temperature value of the last valid ocean node i in the historical sea surface temperature sequence; The future sea surface temperature change trajectory of effective ocean node i is represented. The smaller the difference in the future change trajectory, the more similar the sea surface temperature change trends of the two effective ocean nodes are during the prediction period.

[0124] For any valid ocean node b other than valid ocean node i, the difference in future trajectories between valid ocean node i and valid ocean node b is calculated using Euclidean distance, and is expressed as:

[0125] ;

[0126] in, This represents the difference in future trajectories between effective ocean node i and effective ocean node b. This represents the future sea surface temperature change trajectory of effective ocean node b; The L2 norm of a vector is used to represent the vector's L2 norm.

[0127] Specifically, for each valid ocean node i, from all valid ocean nodes other than i, the one with the smallest difference in future trajectory is selected. A set of valid ocean nodes constitutes a future consistent positive sample set, represented as:

[0128] ;

[0129] in, This represents the set of future consistent positive samples corresponding to valid ocean node i; This represents the set of nodes obtained after excluding valid ocean node i from the set of all valid ocean nodes; TopK(.) represents sorting the nodes in ascending order of differences in their future trajectories and returning the top K. Operations on the index of each node; This represents the number of future consistent positive sample nodes selected for each valid ocean node.

[0130] Specifically, the adaptive fusion graph and the historical features of each valid ocean node are input into the graph encoder for graph message propagation to obtain the graph-coded node representation of each valid ocean node; the graph-coded node representation of valid ocean node i is:

[0131] ;

[0132] in, This represents the graph-coded node representation corresponding to a valid ocean node i. The feature dimension represented by the graph encoding node.

[0133] The graph-coded node representations of each valid ocean node are normalized using the L2 norm to obtain the normalized graph-coded node representations:

[0134] ;

[0135] in, This represents the normalized graph-coded node representation of the valid ocean node i;

[0136] Based on the inner product between the normalized graph-coded node representations and the preset contrastive learning temperature parameter, the relationship score between effective ocean node i and other effective ocean nodes b is calculated and expressed as:

[0137] ;

[0138] in, and Let i and b represent the normalized graph-coded node representations of valid ocean node i and valid ocean node b, respectively. The relationship score between valid ocean node i and valid ocean node b is calculated based on graph-coded node representation. This indicates the temperature parameter for comparison and learning.

[0139] Specifically, based on the set of future consistent positive samples and the relationship score corresponding to each valid ocean node, a future consistency supervision loss is calculated using a multi-positive sample contrastive learning objective function, which increases the similarity between valid ocean nodes and their future consistent positive sample nodes in the representation space, as expressed as:

[0140] ;

[0141] in, Represents the future consistency supervision loss; p represents the index of the future consistent positive sample node used when traversing the set of future consistent positive samples; The score represents the relationship between valid ocean node i and future consistent positive sample node p; This represents the set of nodes obtained after excluding valid ocean node i from all valid ocean nodes; Represents the effective ocean node i and the effective ocean node The relationship between the two is scored; exp(⋅) represents the exponential function with the natural constant e as the base; log(⋅) represents the natural logarithm function.

[0142] Specifically, by minimizing the future consistency supervision loss, the relationship score between effective ocean nodes and their future consistent positive sample nodes is improved, so that effective ocean nodes with similar future sea surface temperature change trajectories have higher similarity in the graph encoding representation space. The future consistency supervision loss directly constrains the relationship score between graph encoding node representations, rather than directly specifying or replacing candidate edges or fusion edge weights in the adaptive fusion graph; through backpropagation, the future consistency supervision loss is used to update the trainable parameters in the graph encoder, the historical spatiotemporal encoder for constructing the implicit structure graph, and the multilayer perceptron for generating edge-level adaptive fusion coefficients.

[0143] Specifically, the real future sea surface temperature sequence, future sea surface temperature change trajectory, differences in future change trajectories, and the set of positive samples for future consistency are only used to construct future consistency supervision during the model training phase; they are unavailable and do not participate in computation during the inference phase. During the inference phase, the geographic map is constructed based on the spatial location information of each valid ocean node, the implicit structure graph is constructed based on historical feature representations extracted from historical sea surface temperature sequences, the adaptive fusion graph is obtained by edge-level adaptive fusion of the geographic map and the implicit structure graph, and the graph-encoded node representation is obtained by a graph encoder from the adaptive fusion graph and the historical features of each valid ocean node. Therefore, future consistency supervision does not cause future information leakage but rather improves the ability of the graph-encoded node representation to represent the evolution of future sea surface temperature.

[0144] In some embodiments, step S2 includes:

[0145] The historical spatiotemporal representation is input into the time series prediction network to predict the potential states of each effective ocean node at multiple future moments within the prediction period. Based on the positional correspondence between each effective ocean node and the original two-dimensional spatial grid, the future potential states of each effective ocean node are mapped to the corresponding grid positions in the original two-dimensional spatial grid to obtain the future potential feature map. The future potential feature map is then input into the fine-scale convolutional projection branch, the meso-scale convolutional projection branch, and the coarse-scale convolutional projection branch, which have mutually independent parameters, respectively. Through channel mapping and different degrees of spatial scale transformation, the coarse-scale conditional prior, the meso-scale conditional prior, and the fine-scale conditional prior used as input to the denoising network are obtained.

[0146] Specifically, the historical spatiotemporal representation output by the graph encoder in step S1 is denoted as:

[0147] ;

[0148] in, This represents the historical spatiotemporal representation obtained by constructing a geographic map and an implicit structure graph based on historical sea surface temperature sequences, followed by edge-level adaptive fusion and graph message propagation; B represents the batch size. This represents the number of historical moments included in the historical observation window, where N represents the total number of valid ocean nodes. The characteristic dimensions representing historical spatiotemporal representation.

[0149] The historical spatiotemporal representation is input into the temporal prediction network. A node-level gated recurrent unit encoder extracts the temporal evolution features of each valid ocean node within the historical observation period. These features are then input into the node-level gated recurrent unit decoder to generate the future potential states of each valid ocean node at multiple future moments within the prediction period, expressed as follows:

[0150] ;

[0151] in, and These represent the node-level gated loop unit encoder and the node-level gated loop unit decoder, respectively. This represents the potential future state of each effective ocean node at multiple future moments within the prediction period; Feature dimensions representing potential future states;

[0152] Based on the positional correspondence between each effective ocean node determined by the effective ocean region mask and the original two-dimensional spatial grid, the future potential state of each effective ocean node is mapped to the corresponding spatial grid position, and the invalid spatial positions are set to preset fill values ​​to obtain the future potential feature map:

[0153] ;

[0154] in, This indicates that node-to-mesh mapping operations are performed based on an effective ocean area mask; This represents the effective ocean region mask, used to record the correspondence between the effective ocean nodes and the original two-dimensional spatial grid positions; Represents a potential feature map for the future.

[0155] The future latent feature maps are input into three parameter-independent convolutional projection branches: fine-scale, medium-scale, and coarse-scale. Through channel mapping and spatial scale transformations of varying degrees, fine-scale, medium-scale, and coarse-scale conditional priors are generated, respectively, as follows:

[0156] ;

[0157] ;

[0158] ;

[0159] in, , as well as These represent the fine-scale convolution projection branch, the medium-scale convolution projection branch, and the coarse-scale convolution projection branch, respectively, whose parameters are independent of each other. , as well as These represent the conditional priors for fine-scale, meso-scale, and coarse-scale, respectively. This represents the channel dimension of the prior conditions at each scale.

[0160] Specifically, the time dimension and channel dimension in the conditional priors of different scales are arranged to obtain the conditional prior at scale s, which is represented as:

[0161] ;

[0162] in, and They are respectively equal to and , This represents the s-th scale conditional prior after the arrangement.

[0163] Specifically, the three scales of conditional priors provide complementary structural guidance for the generation of the future sea surface temperature field. Among them, the coarse-scale conditional priors mainly characterize the global warm and cold background structure and low-frequency temperature distribution; the meso-scale conditional priors mainly characterize the regional organization pattern and meso-scale variation structure; and the fine-scale conditional priors mainly characterize local boundaries, temperature fronts, and fine-grained spatial information.

[0164] In some embodiments, step S3 further includes training the denoising network during the training phase, including:

[0165] During the training phase, the real future sea surface temperature sequence corresponding to the training samples is obtained; the sea surface temperature field at the last moment in the historical sea surface temperature sequence is repeated multiple times along the future prediction time dimension to construct a Brownian bridge reference endpoint with the same dimension as the real future sea surface temperature sequence.

[0166] Random noise is generated from diffusion time step t; based on the Brownian bridge interpolation coefficients and noise scale corresponding to diffusion time step t, the real future sea surface temperature sequence and the Brownian bridge reference endpoint are weighted and combined, and random noise is superimposed to obtain the Brownian bridge intermediate state corresponding to diffusion time step t.

[0167] Based on the difference between the actual future sea surface temperature sequence and the Brown bridge reference endpoint, the Brown bridge interpolation coefficient, and random noise, construct the Brown bridge prediction target corresponding to the diffusion time step t;

[0168] The intermediate state of the Brown bridge, the diffusion time step t, and the prior conditions at different scales are input into the denoising network, which outputs the Brown bridge prediction result. The diffusion training loss is calculated based on the error between the Brown bridge prediction result and the Brown bridge prediction target, and the network parameters of the denoising network are updated through backpropagation.

[0169] Specifically, during the training phase, the real future sea surface temperature sequences corresponding to the training samples are obtained. C represents the number of channels for the sea surface temperature variable; the sea surface temperature field at the last moment in the historical sea surface temperature sequence is denoted as: ;

[0170] Repeat the sea surface temperature field of the last moment in the historical sea surface temperature sequence along the future prediction time dimension. Next, a Brownian bridge reference endpoint with the same dimensions as the actual future sea surface temperature sequence is constructed, denoted as:

[0171] ;

[0172] in, Indicates the reference endpoint of the Brown bridge; This represents the operation of repeatedly inputting the sea surface temperature field at the last moment of the historical sea surface temperature sequence along the future prediction time dimension; This represents the sea surface temperature field at the last moment in the historical sea surface temperature sequence.

[0173] Specifically, a diffusion time step t is selected from a preset diffusion time step range, and random noise is generated. Based on the Brownian bridge interpolation coefficients and noise scale corresponding to diffusion time step t, the real future sea surface temperature field sequence and the Brownian bridge reference endpoint are weighted and combined, and the random noise is superimposed to obtain the intermediate state of the Brownian bridge corresponding to diffusion time step t, which is expressed as:

[0174] ;

[0175] in, This represents the intermediate state of the Brownian bridge corresponding to diffusion time step t; This represents the Brownian bridge interpolation coefficients corresponding to the diffusion time step t; This represents the noise scale corresponding to the diffusion time step t; Indicates and Random Gaussian noise of the same dimension; This represents a standard Gaussian distribution with zero mean and an identity matrix. When When approaching 0, the intermediate state of the Brown Bridge closely approximates the true future sea surface temperature field sequence; when When the value approaches 1, the intermediate state of the Brown bridge is close to the reference endpoint of the Brown bridge.

[0176] Specifically, based on the difference between the actual future sea surface temperature sequence and the Brownian bridge reference endpoint, the Brownian bridge interpolation coefficients, and random noise, the Brownian bridge prediction target corresponding to the diffusion time step t is constructed, expressed as:

[0177] ;

[0178] in, Indicates the target predicted by Brown Bridge. The noise intensity coefficient at diffusion time step t is represented by a monotonically increasing function related to t, used to control random Gaussian noise. The injection noise is stronger when the injection intensity and diffusion time step t are larger.

[0179] Specifically, the intermediate state of the Brown bridge, the current diffusion time step t, and the conditional priors at different scales are jointly input into the denoising network to predict the Brown bridge prediction result, as shown below:

[0180] ;

[0181] in, This represents a 3D denoising network. This indicates the prediction result of the Brown Bridge.

[0182] In other embodiments, step S3, the inference phase, includes:

[0183] The reference endpoint of the Brown bridge, constructed based on the historical sea surface temperature sequence to be predicted, is used as the initial state for the reverse denoising process. Reverse denoising is performed in descending order of reverse denoising time steps.

[0184] Coarse-scale, meso-scale, and fine-scale conditional priors are injected into the feature layers corresponding to the spatial resolution in the denoising network. A normalized inverse denoising progress variable is determined based on the current inverse denoising time step and the total number of diffusion time steps. A third-order Bernstein basis function is constructed based on the normalized inverse denoising progress variable, and the guiding strength of the coarse-scale, meso-scale, and fine-scale conditional priors is determined based on the third-order Bernstein basis function, so that the guiding strength of each scale conditional prior is non-negative and the sum of the three is 1.

[0185] At the current reverse denoising time step, the original feature representation of each scale feature layer is combined with the conditional prior of the corresponding scale to perform conditional attention calculation. The conditional attention calculation result is multiplied by the guidance intensity of the corresponding scale and added to the residual of the original feature representation of the corresponding scale to obtain the conditional enhanced feature representation of the corresponding scale.

[0186] The conditionally enhanced feature representation continues to participate in the subsequent calculation of the denoising network to obtain the Brown bridge prediction result corresponding to the current reverse denoising time step; the reverse denoising update corresponding to the current reverse denoising time step is performed according to the Brown bridge prediction result to obtain the Brown bridge state of the next reverse denoising time step; the above process of determining the guiding intensity, injecting the conditional prior, and reversing the denoising update is repeated in descending order of the reverse denoising time steps until the future sea surface temperature prediction result is obtained.

[0187] Specifically, the reference endpoint of the Brown bridge constructed based on the historical sea surface temperature sequence to be predicted is used as the initial state of the reverse denoising process. The current Brown bridge state is updated step by step using the trained denoising network in descending order of diffusion time step.

[0188] Specifically, to match the conditional priors at different scales with the different stages of the reverse denoising process, coarse-scale, meso-scale, and fine-scale conditional priors are injected into the feature layers corresponding to the spatial resolutions of the denoising network, respectively. A normalized reverse denoising progress variable is determined based on the current reverse denoising time step t and the total number of diffusion time steps. This normalized reverse denoising progress variable is set to a value close to 0 at the beginning of reverse denoising and gradually increases to 1 as the reverse denoising process progresses, expressed as:

[0189] ;

[0190] in, This represents the normalized inverse denoising progress variable, where u=0 indicates the start of the inverse denoising stage and u=1 indicates the end of the inverse denoising stage. This represents the total diffusion time step in the Brownian bridge diffusion process.

[0191] A third-order Bernstein basis function is constructed based on the normalized inverse denoising progress variable, and the guiding strength of the coarse-scale conditional prior, meso-scale conditional prior, and fine-scale conditional prior are determined based on the third-order Bernstein basis function, as expressed as:

[0192] ;

[0193] ;

[0194] ;

[0195] ;

[0196] in, , , and These represent the progress of normalized inverse denoising, respectively. The corresponding four cubic Bernstein basis functions.

[0197] The guiding strength of each scale conditional prior is made non-negative and the sum of the three is 1, thereby maintaining the stability of the overall conditional guiding strength during the inverse denoising process and achieving a smooth transition between conditional priors of different scales, expressed as:

[0198] ;

[0199] ;

[0200] ;

[0201] ;

[0202] in, , as well as These represent the guiding strengths of coarse-scale conditional priors, meso-scale conditional priors, and fine-scale conditional priors, respectively.

[0203] Specifically, the coarse-scale conditional priors, meso-scale conditional priors, and fine-scale conditional priors after dimensionality rearrangement are injected into the feature layers corresponding to the spatial resolution in the denoising network. At the reverse denoising time step t corresponding to the current reverse denoising process, conditional attention is calculated between the feature representation at scale s in the denoising network and the conditional priors at the corresponding scales. The result of the conditional attention calculation is multiplied by the guidance strength at the corresponding scale and then added to the residual of the original feature representation at the corresponding scale to obtain the conditionally enhanced feature representation, expressed as:

[0204] ;

[0205] Where s represents the scale type of the conditional prior, and c, m and f represent coarse scale, meso scale and fine scale, respectively; This represents the original feature representation of the denoising network at diffusion time step t, in the s-th scale feature layer; This represents the conditional prior at the s-th scale; This indicates a conditional attention operation, used to conditionally enhance denoised features at the corresponding scale based on conditional priors at the s-th scale. This indicates the progress of the s-th scale conditional prior in the current inverse denoising. The guiding strength below; This represents the conditionally enhanced feature representation obtained after injecting the s-th scale conditional prior.

[0206] Specifically, in the early stage of reverse denoising, coarse-scale conditional priors have a large guiding strength, which is used to improve the guiding strength of coarse-scale conditional priors and establish the global warm and cold background structure of the future sea surface temperature field; in the middle stage of reverse denoising, meso-scale conditional priors play a major role, which is used to improve the guiding strength of meso-scale conditional priors and refine the regional organization pattern and meso-scale change structure; in the later stage of reverse denoising, fine-scale conditional priors gradually become dominant, which is used to improve the guiding strength of fine-scale conditional priors and restore local boundaries, temperature fronts and fine-grained spatial information.

[0207] Thus, the Brown Bridge reverse denoising process forms a staged generation method corresponding to the multi-scale spatial structure, enabling the future sea surface temperature sequence to complete the global structure establishment, regional pattern refinement, and local detail restoration in a coarse-to-fine order.

[0208] Specifically, in the actual prediction stage, the real future sea surface temperature sequence is not included in the calculation. The model constructs a Brownian bridge reference endpoint based on the historical sea surface temperature sequence and obtains the historical spatiotemporal representation based on the historical sea surface temperature sequence. The historical spatiotemporal representation is input into the time series prediction network to obtain the future potential state, and conditional priors of different scales are generated based on the future potential state. Subsequently, the future sea surface temperature prediction result is gradually obtained through the Brownian bridge reverse denoising process.

[0209] Example 2:

[0210] Please see Figures 1-3 Based on the same inventive concept as in Example 1, the present invention also provides a sea surface temperature prediction system based on supervised auxiliary graph learning and conditional diffusion, comprising:

[0211] The auxiliary graph learning module based on future consistency supervision is used for auxiliary graph learning under future consistency supervision: acquiring historical sea surface temperature sequences and valid ocean nodes, constructing a geographic graph and an implicit structure graph, and using an edge-level adaptive fusion mechanism to fuse the geographic graph and the implicit structure graph to obtain an adaptive fused graph; inputting the adaptive fused graph and the historical features of each valid ocean node into a graph encoder, and obtaining a historical spatiotemporal representation through graph message propagation;

[0212] During the training phase, real future sea surface temperature sequences are acquired, and future sea surface temperature change trajectories and future consistency positive sample sets for each effective ocean node are constructed. Based on the graph-encoded node representations of each effective ocean node obtained by the graph encoder, the relationship score between nodes is calculated, and the future consistency supervision loss is calculated using a multi-positive sample contrast learning objective function. The trainable parameters participating in auxiliary graph learning are updated using the future consistency supervision loss through backpropagation.

[0213] A multi-scale conditional prior generation module is used for multi-scale conditional prior generation: the historical spatiotemporal representation is input into the temporal prediction network to predict the future potential state of each effective ocean node at multiple future moments within the prediction period, and the future potential feature map is obtained through mapping operation. The future potential feature map is input into convolutional projection branches of different scales, and different scale conditional priors are generated through multi-scale feature transformation operation.

[0214] A multi-scale prior progressively guided Brownian bridge diffusion prediction module is used for multi-scale prior progressively guided Brownian bridge diffusion prediction: During the training phase, the denoising network is trained; during the inference phase, the guidance intensity of conditional priors at different scales is dynamically determined according to the defined normalized inverse denoising progress variable, and the conditional priors at different scales are injected into the feature layers corresponding to the spatial resolution in the denoising network. After conditional enhancement fusion based on the guidance intensity, the inverse denoising update is repeatedly executed to obtain the future sea surface temperature prediction result.

[0215] Specifically, the multi-scale prior-guided progressive Brownian bridge diffusion prediction module includes: a denoising network training unit, used during the training phase to construct a Brownian bridge reference endpoint by repeating the sea surface temperature field of the last moment of the historical sea surface temperature sequence along the prediction time dimension; constructing a Brownian bridge intermediate state and a corresponding Brownian bridge prediction target based on the real future sea surface temperature sequence, the Brownian bridge reference endpoint, the Brownian bridge interpolation coefficients corresponding to the current diffusion time step, and random noise; inputting the Brownian bridge intermediate state, the current diffusion time step, and the multi-scale conditional prior into the denoising network to obtain the Brownian bridge prediction result; calculating the diffusion training loss based on the error between the Brownian bridge prediction result and the Brownian bridge prediction target, and updating the network parameters of the denoising network through backpropagation;

[0216] The inverse denoising prediction unit is used in the inference phase to take the Brownian bridge reference endpoint constructed based on the historical sea surface temperature sequence to be predicted as the initial state of the inverse denoising process. The current Brownian bridge state is updated step by step using the trained denoising network in descending order of diffusion time steps. At each inverse denoising time step, coarse-scale conditional priors, meso-scale conditional priors, and fine-scale conditional priors are injected into the feature layers corresponding to the spatial resolution in the denoising network. The guiding strength of each scale conditional prior is dynamically determined according to the normalized inverse denoising progress variable. After conditional enhancement fusion, the inverse denoising update is repeated until the future sea surface temperature field sequence within the prediction period is obtained.

[0217] Example 3: Model Training

[0218] This invention obtains daily sea surface temperature data from NOAA OISST through publicly available data services, and extracts data from it for the period from 2013 to 2022, with a latitude range of [missing data]. N to N, Longitude range is E to E's sea area data, the spatial resolution of the data is... .

[0219] The acquired sea surface temperature data were sorted and cropped according to observation date. An effective ocean region mask was constructed based on land-sea distribution information. This mask determined the valid ocean nodes participating in graph construction and model calculation; land areas and invalid data areas were excluded from neighbor retrieval, graph message propagation, and loss calculation. The data was divided chronologically, with data from 2013 to 2021 used as training data and data from 2022 used as test data to avoid future observations within the test period from being used in model training. A sliding time window method was used to construct training samples, with each training sample including consecutive... Sea surface temperature field at a historical moment and the subsequent continuous This example uses a sequence of real sea surface temperature fields at future moments, where the historical sea surface temperature fields of 10 consecutive days are used as the historical input sequence. And based on the predicted task Set them to 1, 3, 5, 7, 10 or 14 respectively to form training samples with the corresponding prediction duration.

[0220] The model is implemented using Python and the PyTorch deep learning framework and trained on an NVIDIA Tesla V100 GPU. The Adam optimizer is used to update the trainable parameters in the model, with a batch size of 8 and an initial learning rate of [value missing]. The maximum number of training rounds is set to 200, and the number of geographical neighbors is [not specified]. and number of structural neighbors All parameters are set to 6, the contrastive learning temperature parameter is set to 0.2, the weighting coefficient for the auxiliary future latent state prediction loss is set to 1, and the weighting coefficient for the future consistency supervision loss is set to... .

[0221] During training, a batch of historical sea surface temperature sequences and temporally continuous real future sea surface temperature field sequences are read from the training sample set. A geographic map is constructed based on the spatial location information of the effective ocean nodes. Historical feature representations of each effective ocean node are extracted based on the historical sea surface temperature sequences, and an implicit structure graph is constructed. An edge-level adaptive fusion mechanism is used to fuse the geographic map and the implicit structure graph to obtain a normalized adaptive fusion graph. The normalized adaptive fusion graph and the historical features of each effective ocean node are input into the graph encoder, and the historical spatiotemporal representation is obtained through graph message propagation.

[0222] During the training phase, based on the real future sea surface temperature field sequence and with the sea surface temperature value at the last moment of the historical observation window as a reference, the future sea surface temperature change trajectory of each effective ocean node is constructed. The difference in future change trajectory between different nodes is calculated. For each anchor node, several nodes with the smallest difference in future change trajectory are selected as positive samples for future consistency. The future consistency supervision loss is calculated based on the relationship between the graph-encoded node representations. The historical spatiotemporal representation output by the graph encoder is input into the node-level gated recurrent unit encoder and decoder to predict the future potential state of each effective ocean node at multiple future moments within the prediction period. Based on the positional correspondence between the effective ocean nodes and the original two-dimensional spatial grid, the future potential state is mapped to the corresponding spatial grid position to obtain the future potential feature map. This map is then input into the fine-scale convolutional projection branch, the mesoscale convolutional projection branch, and the coarse-scale convolutional projection branch to generate the fine-scale conditional prior, the mesoscale conditional prior, and the coarse-scale conditional prior. At the same time, the auxiliary future sea surface temperature field sequence is obtained by auxiliary decoding based on the future potential state, and the auxiliary future potential state prediction loss is calculated.

[0223] Subsequently, the sea surface temperature field at the last moment of the historical observation window is repeated along the predicted time dimension to construct a Brownian bridge reference endpoint. The current diffusion time step is selected from a preset diffusion time step range, and random noise is generated. Based on the actual future sea surface temperature sequence, the Brownian bridge reference endpoint, the Brownian bridge interpolation coefficients, and the random Gaussian noise, intermediate states of the Brownian bridge and corresponding Brownian bridge prediction targets are constructed. The intermediate states of the Brownian bridge, the current diffusion time step, and conditional priors at different scales are input into a denoising network to obtain the Brownian bridge prediction results. The diffusion training loss is calculated based on the error between the Brownian bridge prediction results and the Brownian bridge prediction targets. The total training loss of the model is a weighted sum of the diffusion training loss, the auxiliary future potential state prediction loss, and the future consistency supervision loss, expressed as:

[0224] ;

[0225] in, Indicates the training loss due to diffusion. This indicates that the loss is predicted by assisting in the prediction of potential future states. Indicates future consistency monitoring losses, and These represent the corresponding loss weight coefficients.

[0226] Backpropagation is performed on the total training loss, and the trainable parameters in the implicit structure graph construction, edge-level adaptive fusion, graph encoder, node-level gated recurrent unit encoder and decoder, multi-scale convolutional projection branch, and denoising network are jointly updated through the Adam optimizer. One traversal of all training samples is counted as one training round, and the above training process is repeated until the preset maximum number of training rounds is reached or the preset convergence condition is met. The trained model parameters are then saved. During the inference phase, real future sea surface temperature sequences are not acquired or used, and no future change trajectories, future consistent positive sample sets, or training loss are constructed. Only the historical sea surface temperature sequence to be predicted is input into the trained model to generate a normalized adaptive fusion graph and multi-scale conditional priors. The Brownian bridge reference endpoint constructed from the sea surface temperature field at the last moment of the historical observation window is used as the initial state for backpropagation denoising. The future sea surface temperature field sequence within the prediction period is obtained through progressive backpropagation denoising.

[0227] Example 4: Comparative Experiment:

[0228] To verify the effectiveness of the sea surface temperature prediction method of the present invention, the method of the present invention is compared with nine representative spatiotemporal prediction methods. The comparative methods include: spatiotemporal prediction methods based on convolutional, recurrent networks, and Transformers, namely Convolutional Long Short-Term Memory Network (ConvLSTM), Shifted Window Transformer (Swin-Transformer), Spatio-Temporal U-Network (ST-UNet), and Dense Dilated Convolutional Long Short-Term Memory Network (D2CL); methods based on neural operators and spectral modeling, namely U-Net Enhanced Fourier Neural Operator (U-FNO), Fourier Forecasting Neural Network (FourCastNet), and Latent Spectral Model (LSM); and graph neural network-based methods, namely Adaptive Spatiotemporal Graph Recurrent Network. The methods described above cover different technical approaches, including recurrent neural networks (ASTGRN), three-dimensional dynamic spatial-temporal graph neural networks (3D-DSTGN), and three-dimensional dynamic spatial-temporal graph neural networks (3D-DSTGN). These methods can comprehensively verify the sea surface temperature prediction performance of the methods in this invention.

[0229] Among them, ConvLSTM introduces convolution operations into the input-to-state transition and state-to-state transition of the Long Short-Term Memory network, while extracting the spatial features and temporal dependencies of the sea surface temperature sequence; Swin-Transformer models the correlation between different spatial regions through a shift window self-attention mechanism; ST-UNet uses a U-shaped encoder-decoder structure to extract features at different temporal and spatial scales; D2CL expands the range of spatiotemporal feature perception through dense connections and dilated convolutions; U-FNO combines Fourier neural operators with U-shaped network branches to jointly extract global information in the frequency domain and local information in space; FourCastNet uses adaptive Fourier neural operators for high-resolution spatiotemporal field prediction; LSM uses spectral domain modeling in the latent space to solve high-dimensional spatiotemporal evolution problems; ASTGRN dynamically models the spatiotemporal dependencies between different ocean regions through an adaptive graph structure and recurrent networks; and 3D-DSTGN models the dynamic correlation of the ocean temperature field in the spatial, temporal, and vertical dimensions through a three-dimensional dynamic spatiotemporal graph structure. The names and main structures of ConvLSTM, Swin-Transformer, U-FNO, and FourCastNet are derived from their respective original methods.

[0230] To ensure fairness in the comparative experiments, all methods used the same dataset partitioning, data preprocessing, historical input length, prediction duration, and effective ocean area mask. All methods used the first 10 consecutive days of sea surface temperature (SST) observations as the historical input sequence and predicted the SST sequences for the subsequent 1, 3, 5, 7, 10, and 14 days, respectively. The 3-day, 5-day, 7-day, 10-day, and 14-day prediction tasks represent predictions of multiple consecutive future moments within the corresponding prediction period, rather than just predicting the 3rd, 5th, 7th, 10th, or 14th day. Each evaluation metric was the average of the evaluation results for all future moments within the corresponding prediction period. The comparative experimental results are shown in Tables 1 to 6. The root mean squared error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and structural similarity index (SSIM) were used to evaluate the sea surface temperature prediction performance of different methods. RMSE measures the average level of the squared error between the predicted and actual sea surface temperature values, and is more sensitive to larger prediction errors. MAE measures the average level of the absolute error between the predicted and actual values. MAPE measures the average percentage of the prediction error relative to the actual sea surface temperature value and is calculated only at valid ocean grid locations. SSIM measures the spatial structural similarity between the predicted and actual sea surface temperature fields. Smaller values ​​for RMSE, MAE, and MAPE indicate lower prediction errors, while a larger SSIM value indicates greater spatial structural similarity between the predicted and actual fields. Comparative experimental results are shown in Tables 1-6.

[0231] Table 1: Results of the 1-Day Prediction Comparison Experiment

[0232]

[0233] Table 2: Results of the 3-Day Prediction Comparison Experiment

[0234]

[0235] Table 3: Results of the 5-Day Prediction Comparison Experiment

[0236]

[0237] Table 4: Results of the 7-Day Prediction Comparison Experiment

[0238]

[0239] Table 5: Results of the 10-Day Prediction Comparison Experiment

[0240]

[0241] Table 6: Results of the 14-day prediction comparison experiment

[0242]

[0243] The method of this invention achieves the lowest RMSE, MAE, and MAPE at all prediction timeframes and the highest SSIM at most timeframes, indicating that the proposed future-aware spatial-structure prior and multi-scale progressive diffusion guidance can simultaneously improve the accuracy of point value prediction and the ability to preserve spatial structure.

[0244] Obviously, the accompanying drawings are merely some examples or embodiments of the present invention. Those skilled in the art can apply the present invention to other similar situations based on these drawings without any creative effort. Furthermore, it is understood that although the work done in this development process may be complex and lengthy, certain design, manufacturing, or production modifications made according to the technical content disclosed in the present invention are merely conventional technical means and should not be considered as insufficient disclosure of the present invention.

[0245] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-described technical content to create equivalent embodiments without departing from the scope of the present invention. The implementation schemes in the above embodiments can also be further combined or replaced. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A sea surface temperature prediction method based on supervised graph learning and conditional diffusion, characterized in that, Includes the following steps: S1. Auxiliary graph learning based on future consistency supervision: Historical sea surface temperature sequences and valid ocean nodes are obtained to construct a geographic map and an implicit structure map. An edge-level adaptive fusion mechanism is then used to fuse the geographic map and the implicit structure map to obtain an adaptive fused map. The adaptive fused map and the historical features of each valid ocean node are input into a graph encoder, and a historical spatiotemporal representation is obtained through graph message propagation. During the training phase, real future sea surface temperature sequences are acquired, and future sea surface temperature change trajectories and future consistency positive sample sets for each effective ocean node are constructed. Based on the graph-encoded node representations of each effective ocean node obtained by the graph encoder, the relationship score between nodes is calculated, and the future consistency supervision loss is calculated using a multi-positive sample contrast learning objective function. The trainable parameters participating in auxiliary graph learning are updated using the future consistency supervision loss through backpropagation. S2, Multi-scale conditional prior generation: The historical spatiotemporal representation is input into the temporal prediction network to predict the future potential state of each effective ocean node at multiple future moments within the prediction period. The future potential feature map is obtained through mapping operations. The future potential feature map is input into convolutional projection branches of different scales, and conditional priors of different scales are generated through multi-scale feature transformation operations. S3. Multi-scale prior asymptotically guided Brownian bridge diffusion prediction: During the training phase, the denoising network is trained. During the inference phase, conditional priors of different scales are injected into the feature layers corresponding to the spatial resolution of the denoising network. The guidance strength of the conditional priors of different scales is dynamically determined according to the defined normalized inverse denoising progress variable. After conditional enhancement fusion based on the guidance strength, the inverse denoising update is repeatedly executed to obtain the future sea surface temperature prediction result.

2. The method according to claim 1, characterized in that, Obtain historical sea surface temperature sequences and valid ocean nodes to construct geographic maps and implicit structure maps, including: An original two-dimensional spatial grid is formed based on the sea surface temperature field. Based on the original two-dimensional spatial grid, an effective ocean area mask and effective ocean nodes are defined, and historical sea surface temperature sequences are obtained. The latitude and longitude coordinates of each valid ocean node are mapped to a three-dimensional unit sphere, and the spherical distance between any two valid ocean nodes is calculated. For each valid ocean node, a set of valid ocean nodes with the smallest spherical distances are selected from other valid ocean nodes (excluding itself) to form a set of geographic neighbors. Based on the spherical distances between the current valid ocean node and each valid ocean node in the set of geographic neighbors, and a preset geographic temperature parameter, the corresponding geographic relationship value is calculated. The geographic relationship value corresponding to each valid ocean node is normalized row by row using softmax to obtain the geographic graph edge weights. For valid ocean nodes that do not belong to the set of geographic neighbors, the corresponding geographic graph edge weights are set to zero, thus forming a geographic graph. Historical feature representations of each valid ocean node are extracted from historical sea surface temperature sequences using a historical spatiotemporal encoder. Based on these historical feature representations, the structural similarity between different valid ocean nodes is calculated using cosine similarity. For each valid ocean node, several valid ocean nodes with the highest structural similarity are selected from other valid ocean nodes (excluding itself) as structural neighbor nodes, forming a corresponding structural neighbor set. The structural relationship value is calculated based on the structural similarity between the current valid ocean node and its structural neighbor nodes. The structural relationship value corresponding to each valid ocean node is normalized using a mask row to obtain the structural graph edge weights. For other valid ocean nodes not belonging to the structural neighbor set, the corresponding structural graph edge weights are set to zero, and an implicit structural graph is formed by all valid ocean nodes and their corresponding structural graph edge weights.

3. The method according to claim 2, characterized in that, A side-level adaptive fusion mechanism is used to fuse the geographic map and the implicit structure map to obtain an adaptive fused map, including: For each valid ocean node, merge the geographic neighbor set and structural neighbor set of the current valid ocean node and remove duplicate nodes to obtain a candidate neighbor set. For each valid ocean node, determine the candidate neighbor set by the union of the geographic neighbor set and the implicit structural neighbor set of the valid ocean node, and determine the candidate edge by the connection between the valid ocean node and each node in the candidate neighbor set. Obtain the geographic graph edge weight and the structural graph edge weight of each candidate edge in the geographic graph and the implicit structural graph respectively. If the current candidate edge does not exist in the geographic graph, set the geographic graph edge weight corresponding to the current candidate edge to zero; if the current candidate edge does not exist in the implicit structural graph, set the structural graph edge weight corresponding to the current candidate edge to zero. Based on the absolute differences between the geographical graph edge weights, structural graph edge weights, and geographical graph edge weights and structural graph edge weights corresponding to the candidate edges, and the sum of the geographical graph edge weights and structural graph edge weights, edge-level fusion features are constructed. Edge-level fusion features are input into a multilayer perceptron, and edge-level adaptive fusion coefficients are generated using a sigmoid activation function. The edge-level adaptive fusion coefficients and their complements are then weighted and summed with respect to the geographic graph weights and structural graph weights corresponding to the candidate edges, respectively, to obtain the fusion candidate edge weights. Specifically, when a candidate edge does not exist in the geographic graph, the corresponding geographic graph weight is set to zero before being included in the weighted summation; when a candidate edge does not exist in the implicit structural graph, the corresponding structural graph weight is set to zero before being included in the weighted summation. The initial fusion graph is constructed from these fusion candidate edge weights. The initial fused graph and its transpose are averaged to achieve symmetry, and self-connections are added to the symmetric fused graph. Based on the sum of all outgoing edge weights corresponding to each valid ocean node, the fused graph after adding self-connections is normalized row by row to obtain an adaptive fused graph. Graph message propagation is performed based on the adaptive fused graph to update the node representation of each valid ocean node layer by layer.

4. The method according to claim 1, characterized in that, The training phase of step S1 includes: The true future sea surface temperature sequence following the historical observation window is obtained from the sea surface temperature observation data. For each valid ocean node, the sea surface temperature value at the last moment in the historical observation window is used as a reference. The change in the true future sea surface temperature sequence at each future moment relative to the reference sea surface temperature value is calculated to form the future sea surface temperature change trajectory of the corresponding valid ocean node. For each valid ocean node, the difference in future trajectory between the current valid ocean node and other valid ocean nodes (excluding itself) is calculated using Euclidean distance. Several valid ocean nodes with the smallest future trajectory differences are selected from the other valid ocean nodes to form a set of future consistent positive samples corresponding to the current valid ocean node. The adaptive fusion graph and the historical features of each valid ocean node are input into the graph encoder, and the graph-coded node representations corresponding to each valid ocean node are obtained through graph message propagation. Normalization is performed on each graph-coded node representation, and the relationship score between each valid ocean node is calculated based on the inner product between the normalized graph-coded node representations and a preset contrastive learning temperature parameter. Based on the set of future consistent positive samples corresponding to each valid ocean node and the relationship score, the future consistency supervision loss is calculated using the multi-positive sample contrastive learning objective function; through backpropagation, the trainable parameters in the graph encoder, the historical spatiotemporal encoder for extracting historical feature representations, and the multilayer perceptron for generating edge-level adaptive fusion coefficients are updated using the future consistency supervision loss.

5. The method according to claim 1, characterized in that, Step S2 includes: The historical spatiotemporal representation is input into a time-series prediction network to predict the potential states of each effective ocean node at multiple future moments within the prediction period. Based on the positional correspondence between each effective ocean node and the original two-dimensional spatial grid, the future potential states of each effective ocean node are mapped to the corresponding grid positions in the original two-dimensional spatial grid to obtain a future potential feature map. The future potential feature map is then input into three parameters that are independent: a fine-scale convolutional projection branch, a medium-scale convolutional projection branch, and a coarse-scale convolutional projection branch. Coarse-scale conditional priors, medium-scale conditional priors, and fine-scale conditional priors are obtained through channel mapping and different degrees of spatial scale transformation.

6. The method according to claim 1, characterized in that, Step S3 trains the denoising network during the training phase, including: During the training phase, the real future sea surface temperature sequence corresponding to the training samples is obtained; the sea surface temperature field at the last moment in the historical sea surface temperature sequence is repeated multiple times along the future prediction time dimension to construct a Brownian bridge reference endpoint with the same dimension as the real future sea surface temperature sequence. Random noise is generated from diffusion time step t; based on the Brownian bridge interpolation coefficients and noise scale corresponding to diffusion time step t, the real future sea surface temperature sequence and the Brownian bridge reference endpoint are weighted and combined, and random noise is superimposed to obtain the Brownian bridge intermediate state corresponding to diffusion time step t. Based on the difference between the actual future sea surface temperature sequence and the Brown bridge reference endpoint, the Brown bridge interpolation coefficient, and random noise, construct the Brown bridge prediction target corresponding to the diffusion time step t; The intermediate state of the Brown bridge, the diffusion time step t, and the prior conditions at different scales are input into the denoising network, which outputs the Brown bridge prediction result. The diffusion training loss is calculated based on the error between the Brown bridge prediction result and the Brown bridge prediction target, and the network parameters of the denoising network are updated through backpropagation.

7. The method according to claim 6, characterized in that, Step S3, the reasoning phase, includes: The reference endpoint of the Brown bridge, constructed based on the historical sea surface temperature sequence to be predicted, is used as the initial state for the reverse denoising process. Reverse denoising is performed in descending order of reverse denoising time steps. Coarse-scale, meso-scale, and fine-scale conditional priors are injected into the feature layers corresponding to the spatial resolution in the denoising network. A normalized inverse denoising progress variable is determined based on the current inverse denoising time step and the total number of diffusion time steps. A third-order Bernstein basis function is constructed based on the normalized inverse denoising progress variable, and the guiding strength of the coarse-scale, meso-scale, and fine-scale conditional priors is determined based on the third-order Bernstein basis function, so that the guiding strength of each scale conditional prior is non-negative and the sum of the three is 1. At the current reverse denoising time step, the original feature representation of each scale feature layer is combined with the conditional prior of the corresponding scale to perform conditional attention calculation. The conditional attention calculation result is multiplied by the guidance intensity of the corresponding scale and added to the residual of the original feature representation of the corresponding scale to obtain the conditional enhanced feature representation of the corresponding scale. The conditionally enhanced feature representation continues to participate in the subsequent calculation of the denoising network to obtain the Brown bridge prediction result corresponding to the current reverse denoising time step; the reverse denoising update corresponding to the current reverse denoising time step is performed according to the Brown bridge prediction result to obtain the Brown bridge state of the next reverse denoising time step; the above process of determining the guiding intensity, injecting the conditional prior, and reversing the denoising update is repeated in descending order of the reverse denoising time steps until the future sea surface temperature prediction result is obtained.

8. A sea surface temperature prediction system based on supervised graph learning and conditional diffusion, characterized in that, include: The auxiliary graph learning module based on future consistency supervision is used for auxiliary graph learning under future consistency supervision: acquiring historical sea surface temperature sequences and valid ocean nodes, constructing a geographic graph and an implicit structure graph, and using an edge-level adaptive fusion mechanism to fuse the geographic graph and the implicit structure graph to obtain an adaptive fused graph; inputting the adaptive fused graph and the historical features of each valid ocean node into a graph encoder, and obtaining a historical spatiotemporal representation through graph message propagation; During the training phase, real future sea surface temperature sequences are acquired, and future sea surface temperature change trajectories and future consistency positive sample sets for each effective ocean node are constructed. Based on the graph-encoded node representations of each effective ocean node obtained by the graph encoder, the relationship score between nodes is calculated, and the future consistency supervision loss is calculated using a multi-positive sample contrast learning objective function. The trainable parameters participating in auxiliary graph learning are updated using the future consistency supervision loss through backpropagation. A multi-scale conditional prior generation module is used for multi-scale conditional prior generation: the historical spatiotemporal representation is input into the temporal prediction network to predict the future potential state of each effective ocean node at multiple future moments within the prediction period, and the future potential feature map is obtained through mapping operation. The future potential feature map is input into convolutional projection branches of different scales, and different scale conditional priors are generated through multi-scale feature transformation operation. A multi-scale prior-guided progressive Brownian bridge diffusion prediction module is used for multi-scale prior-guided progressive Brownian bridge diffusion prediction: the denoising network is trained during the training phase. During the inference phase, conditional priors of different scales are injected into the feature layers corresponding to the spatial resolution in the denoising network. The guidance strength of conditional priors of different scales is dynamically determined according to the defined normalized inverse denoising progress variable. After conditional enhancement fusion based on the guidance strength, the inverse denoising update is repeatedly executed to obtain the future sea surface temperature prediction result.

9. The system according to claim 8, characterized in that, The multi-scale prior asymptotically guided Brownian bridge diffusion prediction module includes: The denoising network training unit is used to construct a Brownian bridge reference endpoint by repeating the sea surface temperature field of the last moment of the historical sea surface temperature sequence along the prediction time dimension during the training phase; constructing a Brownian bridge intermediate state and a corresponding Brownian bridge prediction target based on the real future sea surface temperature sequence, the Brownian bridge reference endpoint, the Brownian bridge interpolation coefficients corresponding to the current diffusion time step, and random noise; inputting the Brownian bridge intermediate state, the current diffusion time step, and the multi-scale conditional prior into the denoising network to obtain the Brownian bridge prediction result; calculating the diffusion training loss based on the error between the Brownian bridge prediction result and the Brownian bridge prediction target, and updating the network parameters of the denoising network through backpropagation. The inverse denoising prediction unit is used in the inference phase to take the Brownian bridge reference endpoint constructed based on the historical sea surface temperature sequence to be predicted as the initial state of the inverse denoising process. The current Brownian bridge state is updated step by step using the trained denoising network in descending order of diffusion time steps. At each inverse denoising time step, coarse-scale conditional priors, meso-scale conditional priors, and fine-scale conditional priors are injected into the feature layers corresponding to the spatial resolution in the denoising network. The guiding strength of each scale conditional prior is dynamically determined according to the normalized inverse denoising progress variable. After conditional enhancement fusion, the inverse denoising update is repeated until the future sea surface temperature field sequence within the prediction period is obtained.