Knowledge and data coupled global sea surface temperature prediction method and system
By constructing an ocean knowledge graph and combining it with a large language model, the problem of lack of oceanographic knowledge in deep learning methods is solved, achieving efficient and accurate sea surface temperature prediction and improving the physical interpretability and computational efficiency of the model.
Patent Information
- Application Number
- CN202511732589.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-24
- Publication Date
- 2026-02-10
AI Technical Summary
Existing deep learning-based sea surface temperature prediction models neglect structured knowledge in the field of oceanography, resulting in a lack of physical consistency and regional interpretability, as well as high computational costs and sensitivity to initial conditions.
A marine knowledge graph is constructed, and vectorized modeling is performed using the knowledge graph embedding algorithm TransE. The knowledge and data are fused together with the Large Language Model (LLM). Feature encoding and prediction optimization are performed using a multilayer perceptron and cross-attention mechanism to generate future sea surface temperature sequences.
It improves the accuracy, efficiency, and interpretability of sea surface temperature prediction, achieves a deep integration of physical mechanisms and data-driven approaches, and enhances the physical interpretability and computational efficiency of the model.
Smart Images
Figure CN121503618A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of artificial intelligence and ocean science, and particularly relates to a global sea surface temperature prediction method and system coupling knowledge and data. BACKGROUND
[0002] Sea surface temperature (SST) is an important indicator affecting the global climate system. Although the traditional SST prediction based on numerical physical models has certain physical interpretability, it has large computational overhead and is sensitive to initial conditions. In recent years, deep learning methods have been widely used in SST prediction tasks, but they generally ignore the large amount of structured knowledge accumulated in the field of oceanography (such as ocean currents, monsoons, climate zones, etc.), resulting in a lack of physical consistency and regional interpretability of the model. With the development of large language models (LLM), the model has the ability to understand and reason complex knowledge. However, there are modality differences and granularity mismatches between LLM and numerical ocean data. Therefore, how to effectively integrate ocean knowledge and numerical observation data has become a key challenge to improve the accuracy and interpretability of SST prediction. SUMMARY
[0003] The present application is carried out to solve the above problems, and aims to provide a global sea surface temperature prediction method and system coupling knowledge and data.
[0004] The application provides a knowledge and data coupled global sea surface temperature prediction method, which has the characteristics of comprising the following steps: S1: ocean knowledge graph construction step, constructing an ocean knowledge graph OKG, defining the semantic relationship type of entities in the ocean knowledge graph, determining the geographical boundary of the entity, mapping the sea surface temperature grid unit to the entity of the ocean knowledge graph, forming the spatial-to-semantic association relationship, and the entity includes: region, current, climate zone, monsoon system and special sea area; S2: knowledge graph embedding representation learning step, using the knowledge graph embedding algorithm TransE to vectorize the entity and the associated relationship, obtaining the structure embedding, performing k-hop neighborhood search on each region, extracting its local semantic description and obtaining semantic embedding through a large language model, and fusing the structure embedding and the semantic embedding through an adapter network to obtain a knowledge enhanced vector; S3: time series feature coding step, obtaining global historical sea surface temperature data, performing reversible instance normalization on the time series of each region, dividing the time series into fixed-length time series patches, and mapping each time series patch to a low-dimensional time series embedding space through a multilayer perceptron MLP to obtain a region-level time series feature matrix; S4: knowledge and data alignment step, based on the time series feature matrix, constructing a query vector for each region, and performing cross-attention calculation with the knowledge enhanced vector to form a context enhanced region representation, inputting the region representation into a frozen large language model for high-dimensional representation learning to obtain a high-dimensional representation; S5: prediction optimization step, inputting the high-dimensional representation into a Transformer decoder to learn the spatial and temporal dependence relationship through a multi-head attention mechanism, generating a predicted future τ step sea surface temperature sequence, and optimizing the generated result with a loss function to obtain a prediction result; and S6: output step: outputting the prediction result.
[0005] In the knowledge and data coupled global sea surface temperature prediction method provided by the application, the number of entities in S1 can be not less than 1,700 region nodes, 20 current nodes, 6 climate zone nodes, 5 monsoon system nodes and 80 special sea area nodes.
[0006] In the knowledge and data coupled global sea surface temperature prediction method provided by the application, the semantic relationship type in S1 can include located in, part of, influenced by and adjacent to.
[0007] In the knowledge and data coupled global sea surface temperature prediction method provided by the application, the method for obtaining the knowledge enhanced vector in S2 can be specifically:
[0008] TransE, a translation-based knowledge graph embedding algorithm, is used to represent entities and associated relationships in low-dimensional vectors, and a triple is defined as , , , , ,
[0009] , , ,
[0010] , , , , , , ,
[0011] , , , ,
[0012] , , ,
[0013] , ,
[0014] , , , , , ,
[0015] , , ,
[0016] , , ,
[0017] , , ,
[0018] In the global sea surface temperature prediction method provided by the knowledge and data coupling of the application, the forming method of the context-enhanced region representation in S4 can be specifically as follows:
[0019] Based on the time series feature matrix, for each region node , a query vector is constructed
[0020]
[0021] wherein, represents the query vector of the i-th region node, which is obtained by splicing the region static feature and the time series feature of the region and performing LayerNorm,
[0022] The cross-attention mechanism is used to fuse the knowledge-enhanced vector and the time series feature matrix, and cross-attention is calculated as Key / Value, and the formula is as follows:
[0023]
[0024] wherein, the cross-attention weight is used to depict the correlation between the query vector and the i-th entity representation in the marine knowledge graph, and the index score is normalized by softmax, wherein the denominator is the weighted sum of all entity vectors ,
[0025] The fusion output is as follows:
[0026]
[0027] wherein, the original time series feature and the knowledge-enhanced representation are spliced to obtain the fusion representation of the region , i.e. the context-enhanced region representation of the region.
[0028] The application provides a global sea surface temperature prediction system coupled with knowledge and data, which has the characteristics of comprising: an ocean knowledge graph construction module, which constructs an ocean knowledge graph OKG, defines semantic relationship types of entities in the ocean knowledge graph, determines geographical boundaries of the entities, maps sea surface temperature grid cells to entities of the ocean knowledge graph, forms a spatial-to-semantic association relationship, and the entities comprise: a region Region, a current Current, a climate zone Climate Zone, a monsoon system Monsoon and a special sea area Special Area; a knowledge graph embedding representation learning module, which uses a knowledge graph embedding algorithm TransE to vectorize modeling of the entities and the association relationship, obtains structural embedding, performs k-hop neighborhood retrieval on each region, extracts its local semantic description and obtains semantic embedding through a large language model, fuses the structural embedding and the semantic embedding through an adapter network to obtain a knowledge-enhanced vector; a time series feature encoding module, which obtains global historical sea surface temperature data, performs reversible instance normalization on the time series of each region, divides the time series into fixed-length time series patches, and maps each time series patch to a low-dimensional time series embedding space through a multilayer perceptron MLP to obtain a region-level time series feature matrix; a knowledge and data alignment module, which constructs a query vector for each region based on the time series feature matrix, and performs cross-attention calculation with the knowledge-enhanced vector to form a context-enhanced regional representation, inputs the regional representation into a frozen large language model for high-dimensional representation learning to obtain a high-dimensional representation; a prediction optimization module, which inputs the high-dimensional representation into a Transformer decoder to learn spatial and temporal dependency relationships through a multi-head attention mechanism, generates a predicted future τ-step sea surface temperature sequence, and optimizes the generated result with a loss function to obtain a prediction result; and an output module: outputting the prediction result.
[0029] Effects of the application
[0030] According to the knowledge and data coupled global sea surface temperature prediction method and system, the following beneficial effects are achieved:
[0031] The application combines structured knowledge, numerical data and large language model capabilities to form a new global sea surface temperature prediction method, which has made significant progress in precision, efficiency, interpretability and universality, and provides a more reliable and efficient intelligent solution for global sea surface temperature prediction, specifically:
[0032] The present application constructs a global ocean knowledge graph covering the global ocean area, and improves the universality of the present application. Taking the ocean area as a core node, the information of ocean current, climate zone, monsoon system and special ocean phenomenon is integrated to form a multi-level graph, and the SST grid unit is mapped to the knowledge graph entity to form the association relationship from space to semantics. The deep fusion of physical mechanism and data driving is realized, the structured ocean knowledge is provided for the model to provide domain prior knowledge, the prediction result conforms to the ocean science law, and the physical interpretability of the model is improved.
[0033] The present application adopts the TransE model to learn and obtain structure embedding, combines LLM to extract semantic embedding, and then obtains knowledge enhanced vector through an adapter network. Multi-granularity representation of multi-source knowledge is realized, and the problem of insufficient information expression of single embedding method is overcome.
[0034] The present application performs reversible instance normalization RIN on the time series of each region, reduces seasonal bias, and improves the ability of the model to capture long-term trends.
[0035] The present application adopts regional-level retrieval and cross-attention mechanism to realize dynamic fine fusion of knowledge embedding and observation data, adopts a frozen large language model to perform high-dimensional representation learning on a global input matrix, maintains general language understanding ability and improves the generalizability of prediction. In this stage, the large language model remains frozen, only the downstream adapter structure is trained, thereby reducing the calculation cost, enhancing the stability of the model, and improving the calculation efficiency. BRIEF DESCRIPTION OF DRAWINGS
[0036] Figure 1 is a flowchart of the knowledge and data coupled global sea surface temperature prediction method in the embodiment of the present application.
[0037] Figure 2 is a flowchart of the knowledge and data coupled global sea surface temperature prediction method in the embodiment of the present application.
[0038] Figure 3 is an overview diagram of the ocean knowledge graph in the embodiment of the present application.
[0039] Figure 4 is a schematic diagram of k-hop neighborhood retrieval in the embodiment of the present application. DETAILED DESCRIPTION
[0040] In order to make the technical means, creative features, purposes and effects realized by the present application easy to understand, the following embodiments combine the drawings to specifically describe the knowledge and data coupled global sea surface temperature prediction method and system of the present application.
[0041] Figure 1is a flowchart of a knowledge and data coupled global sea surface temperature prediction method in embodiments of the present invention. Figure 2 is a flowchart of a knowledge and data coupled global sea surface temperature prediction method in embodiments of the present invention.
[0042] As shown in Figures 1-2 , the knowledge and data coupled global sea surface temperature prediction method of the present embodiment comprises the following steps:
[0043] Initialization: Freeze large language model parameters , initialize the adapter and decoder parameters .
[0044] Step S1 is a marine knowledge graph construction step.
[0045] Figure 3 is an overview diagram of a marine knowledge graph in embodiments of the present invention.
[0046] As shown in Figure 3 , a marine knowledge graph OKG covering the global marine area is constructed, and the semantic relationship types of the entities in the marine knowledge graph are defined. The geographical boundaries of the entities are determined, and the sea surface temperature grid cells are mapped to the entities of the marine knowledge graph. The spatial-to-semantic association relationship is formed.
[0047] The entities mainly include: regions Region, ocean currents Current, climate zones Climate Zone, monsoon systems Monsoon and special sea areas Special Area. The number of entities is not less than 1,700 region nodes, more than 20 ocean current nodes, 6 climate zone nodes, 5 monsoon system nodes and more than 80 special sea area nodes.
[0048] The semantic relationship types include: located in, part of, influenced by and adjacent to.
[0049] Figure 4 is a schematic diagram of k-hop neighborhood retrieval in embodiments of the present invention.
[0050] As shown in Figure 4 , step S2 is a knowledge graph embedding representation learning step. The knowledge graph embedding algorithm TransE is used to vectorize the modeling of entities and associated relationships, and the structured embedding representation is learned by minimizing the ranking loss function based on positive and negative samples to obtain the structure embedding. On this basis, k-hop neighborhood retrieval is performed on each region, the local semantic description is extracted, and the semantic embedding is obtained through the large language model. The structure embedding and the semantic embedding are fused through a lightweight adapter network to obtain a knowledge-enhanced vector, specifically:
[0051] TransE, a translation-based knowledge graph embedding algorithm, is used to represent entities and associated relationships in low-dimensional vectors, and the triplets are defined as , , , , where ,
[0052] , , ,
[0053] , where is the set of positive samples, is the negative sample, is the interval hyperparameter, and the entity structure embedding matrix and the relationship embedding matrix are obtained.
[0054] Perform k-hop neighborhood search on each region to obtain related entities and relationships, and linearize the search results into text , and obtain semantic embedding through the word vector layer of the large language model LLM.
[0055] Fuse the entity structure embedding matrix and semantic embedding through the adapter fusion layer Adapter Layer to obtain the knowledge enhanced vector .
[0056] Step S3 is the time series feature encoding step. Global historical sea surface temperature data is obtained, and reversible instance normalization is performed on the time series of each region. The time series is divided into fixed-length time series patches, and each time series patch is mapped to a low-dimensional time series embedding space through a multi-layer perceptron MLP to obtain a regional-level time series feature matrix, which is:
[0057] Global sea surface temperature observation data is obtained , where is the number of regional nodes, is the time length. The data sources include NOAA OISST and ERA5 reanalysis data.
[0058] Reversible instance normalization RIN is performed on the time series of each region to reduce seasonal bias. The time series is divided into equal-length time series patches, each patch length is , which is defined as:
[0059]
[0060] Through linear transformation and MLP mapping to a low-dimensional time series embedding space, the time feature matrix is obtained.
[0061] After splicing all regional nodes, a global time series feature matrix is formed , that is, a time series feature matrix at the regional level.
[0062] Step S4 is a knowledge and data alignment step. Based on the time series feature matrix, a query vector is constructed for each region, and cross-attention calculation is performed with the knowledge enhanced vector to form a context enhanced regional representation. The regional representation is input into the frozen large language model for high-dimensional representation learning to obtain a high-dimensional representation, specifically:
[0063] The regional level retrieval and cross-attention mechanism are used to realize dynamic and fine fusion of knowledge embedding and observation data.
[0064] Based on the time series feature matrix, a query vector is constructed for each regional node , which is represented as:
[0065]
[0066] wherein, represents the query vector of the i-th regional node, which is obtained by splicing the regional static feature and the time series feature of the region and performing LayerNorm.
[0067] The cross-attention mechanism is used to realize fine fusion of the knowledge enhanced vector and the time series feature matrix, taking as Key / Value, calculating cross-attention, and the formula is:
[0068]
[0069] wherein, the cross-attention weight is used to depict the relevance between the query vector and the i-th entity representation in the marine knowledge graph, and is obtained by softmax normalization of the exponential score of , wherein the denominator is the weighted sum of all entity vectors .
[0070] The fusion output is:
[0071]
[0072] wherein, the original time series feature is spliced with the knowledge enhanced representation to obtain the fusion representation of the region , i.e., the region representation enhanced by the region context, as the final input feature of the region before entering the frozen large language model, is input into the frozen large language model LLM encoder.
[0073] The region representation is subjected to high-dimensional representation learning by using a frozen large language model (such as GPT, Llama, OceanGPT, or DeepSeek) to maintain the general language understanding ability and improve the generalizability of the prediction, and a high-dimensional representation is obtained (forward propagation only).
[0074] The LLM parameters in this stage remain frozen, and only the downstream adapter structure is trained, thereby improving efficiency and stability.
[0075] Step S5 is a prediction optimization step. The high-dimensional representation is input into the Transformer decoder to learn the spatial and temporal dependency relationship through the multi-head attention mechanism, generate a predicted future τ-step sea surface temperature sequence, and optimize the generated result with a loss function to obtain a prediction result, specifically:
[0076] A trainable Transformer decoder is used to capture the spatial dependency relationship through the multi-head attention mechanism to generate a predicted future τ-step sea surface temperature sequence, which is represented as:
[0077]
[0078] wherein the parameter represents all learnable weights in the decoder (including multi-head attention layers, feedforward networks, and projection layers), the input high-dimensional representation captures the spatial dependency relationship between regions through the multi-head attention mechanism, thereby generating a predicted future step sea surface temperature sequence, and a prediction loss function is defined as:
[0079]
[0080] wherein is the mean absolute error, is the true step sea surface temperature, is the corresponding model prediction value, is the prediction length. The loss is used to measure the deviation between the prediction and the true sequence, and drives the update of the decoder parameters .
[0081] The learnable parameters are updated using the learning rate , the model parameters are trained by the Adam optimizer, and the prediction result is obtained by iteration to convergence.
[0082] Step S6 is an output step. The prediction result is output .
[0083] The embodiment also provides a global sea surface temperature prediction system based on knowledge and data coupling, comprising:
[0084] A marine knowledge graph construction module is configured to implement step S1, i.e., constructing a marine knowledge graph OKG and defining semantic relationship types of entities in the marine knowledge graph. Geographical boundaries of the entities are determined, and sea surface temperature grid cells are mapped to the entities of the marine knowledge graph to form a spatial-to-semantic association. The entities include a region, a current, a climate zone, a monsoon system, and a special sea area.
[0085] A knowledge graph embedding representation learning module is configured to implement step S2, i.e., using a knowledge graph embedding algorithm TransE to vectorize modeling of the entities and the association, to obtain structural embedding. k-hop neighborhood retrieval is performed on each region, a local semantic description thereof is extracted, and a semantic embedding is obtained through a large language model. The structural embedding and the semantic embedding are fused through an adapter network to obtain a knowledge-enhanced vector.
[0086] A time series feature encoding module is configured to implement step S3, i.e., obtaining global historical sea surface temperature data, and performing reversible instance normalization on a time series of each region. The time series is divided into fixed-length time series patches, and each time series patch is mapped to a low-dimensional time series embedding space through a multilayer perceptron MLP to obtain a region-level time series feature matrix.
[0087] A knowledge and data alignment module is configured to implement step S4, i.e., based on the time series feature matrix, a query vector is constructed for each region, and cross-attention calculation is performed with the knowledge-enhanced vector to form a context-enhanced region representation. The region representation is input into a frozen large language model for high-dimensional representation learning to obtain a high-dimensional representation.
[0088] A prediction optimization module is configured to implement step S5, i.e., the high-dimensional representation is input into a Transformer decoder, spatial and temporal dependency relationships are learned through a multi-head attention mechanism, a predicted future τ-step sea surface temperature sequence is generated, and a loss function is used to optimize the generated result to obtain a prediction result.
[0089] An output module is configured to implement step S6, i.e., outputting the prediction result.
[0090] Effects of the embodiment
[0091] According to the global sea surface temperature prediction method and system related to knowledge and data coupling provided by the present application, the following beneficial effects are achieved:
[0092] The present application combines structured knowledge, numerical data and large language model capabilities to form a new global sea surface temperature prediction method, which has made significant progress in precision, efficiency, interpretability and universality, providing a more reliable and efficient intelligent solution for global sea surface temperature prediction.
[0093] The present application constructs a global ocean knowledge graph covering the global ocean area, improving the universality of the present application. Taking the ocean area as the core node, the information of ocean currents, climate zones, monsoon systems and special ocean phenomena is integrated to form a multi-level graph, and the SST grid unit is mapped to the knowledge graph entity to form a spatial to semantic association. The deep integration of physical mechanism and data-driven is realized, and the domain prior knowledge is provided to the model through structured ocean knowledge, so that the prediction result conforms to the ocean science law, and the physical interpretability of the model is improved.
[0094] The present application uses the TransE model to learn and obtain structure embedding, and combines LLM to extract semantic embedding, and then obtains knowledge enhanced vector through adapter network fusion. Multi-granularity representation of multi-source knowledge is realized, and the problem of insufficient information expression of single embedding method is overcome.
[0095] The present application performs reversible instance normalization RIN on the time series of each region, reduces the seasonal bias, and improves the ability of the model to capture long-term trends.
[0096] The present application uses regional-level retrieval and cross-attention mechanism to realize dynamic fine fusion of knowledge embedding and observation data, uses a frozen large language model to learn high-dimensional representation of the global input matrix, maintains general language understanding ability and improves the generalization of the prediction. In this stage, the large language model remains frozen, only the downstream adapter structure is trained, thereby reducing the calculation cost and enhancing the stability of the model, and improving the calculation efficiency.
[0097] Those skilled in the art should understand that the present application is not limited by the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application, and various changes and improvements can be made without departing from the spirit and scope of the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.
Claims
1. A knowledge- and data-coupled method for predicting global sea surface temperature, characterized in that, Includes the following steps: S1: Steps for constructing a marine knowledge graph: Construct a marine knowledge graph OKG, define the semantic relationship types of entities in the marine knowledge graph, determine the geographical boundaries of the entities, map sea surface temperature grid cells to the entities in the marine knowledge graph, forming a spatial-to-semantic association. The entities include: Region, Current, Climate Zone, Monsoon system, and Special Area. S2: Knowledge graph embedding representation learning steps, using the knowledge graph embedding algorithm TransE, to vectorize the entity and the relationship to obtain structural embedding, perform k-hop neighborhood retrieval for each region, extract its local semantic description and obtain semantic embedding through a large language model, and fuse the structural embedding and the semantic embedding through an adapter network to obtain a knowledge enhancement vector; S3: Temporal feature encoding step, acquire global historical sea surface temperature data, perform reversible instance normalization on the time series of each region, divide the time series into fixed-length temporal patches, and map each temporal patch to a low-dimensional temporal embedding space through a multilayer perceptron (MLP) to obtain a regional-level time series feature matrix. S4: Knowledge and data alignment step: Based on the time series feature matrix, construct a query vector for each region and perform cross-attention calculation with the knowledge enhancement vector to form a context-enhanced region representation. Input the region representation into the frozen large language model for high-dimensional representation learning to obtain a high-dimensional representation. S5: Prediction optimization step, input the high-dimensional representation into the Transformer decoder, learn the spatial and temporal dependencies through a multi-head attention mechanism, generate the predicted future τ-step sea surface temperature sequence, and optimize the generated result with a loss function to obtain the prediction result; S6: Output Step: Output the prediction result.
2. The global sea surface temperature prediction method based on knowledge and data coupling according to claim 1, characterized in that: in, In S1, the number of entities is: no less than 1,700 regional nodes, 20 ocean current nodes, 6 climate zone nodes, 5 monsoon system nodes, and 80 special sea area nodes.
3. The global sea surface temperature prediction method based on knowledge and data coupling as described in claim 1, Its features are: In S1, the semantic relationship types include: located in, part of, influenced by, and adjacent to.
4. The global sea surface temperature prediction method based on knowledge and data coupling according to claim 1, characterized in that: in, In S2, the method for obtaining the knowledge enhancement vector is as follows: The TransE algorithm, a translational knowledge graph embedding algorithm, is used to represent entities and relationships in a low-dimensional vectorized manner, defining triples. ,in , , Representing the head entity, association relationship, and tail entity respectively, the embedding learning objective is: ， in For the positive sample set, For negative samples, Using the interval hyperparameter, the entity structure embedding matrix is obtained. and relation embedding matrix , Perform k-hop neighborhood search on each region to obtain relevant entities and relationships, and then linearize the search results into text. Semantic embeddings are obtained through the word vector layer of the Large Language Model (LLM). , The entity structure embedding matrix and semantic embedding are fused through an adapter layer to obtain the knowledge enhancement vector. .
5. The global sea surface temperature prediction method based on knowledge and data coupling according to claim 1, characterized in that: in, In S3, the method for obtaining the regional-level time series feature matrix is as follows: Acquire global sea surface temperature observation data ,in This represents the number of nodes in the region. For each region's time series, reversible instance normalized RIN is performed to reduce seasonal bias. The time series is then divided into equal-length time series patches, each with a length of [length missing]. Defined as: ， By mapping the linear transformation to the MLP and then to a low-dimensional temporal embedding space, the temporal feature matrix is obtained. , A global time series feature matrix is formed by concatenating all regional nodes. That is, the regional-level time series feature matrix.
6. The global sea surface temperature prediction method based on knowledge and data coupling according to claim 5, characterized in that: in, In step S4, the method for forming the context-enhanced region representation is as follows: Based on the time series feature matrix, for each of the regional nodes Construct query vector , represented as: ， in, Indicates the first The query vector for each region node is derived from the region's static features. With respect to the time series characteristics of this region After splicing and performing LayerNorm, the result is obtained. A cross-attention mechanism is used to fuse knowledge-enhanced vectors with time-series feature matrices, in order to For Key / Value pairs, calculate cross-attention using the following formula: ， Among them, cross-attention weights Used to characterize query vectors In the ocean knowledge graph, the first Each entity represents The correlation between them, through the The exponential score is obtained by softmax normalization, where the denominator is the sum of all entity vectors. Weighted summation, The fused output is: ， Among them, the original time series features With knowledge enhancement representation The splicing results in a merged representation of the region. This is the context-enhanced representation of the region.
7. A global sea surface temperature prediction system that couples knowledge and data, characterized in that, include: The ocean knowledge graph construction module constructs an ocean knowledge graph OKG, defines the semantic relationship types of entities in the ocean knowledge graph, determines the geographical boundaries of the entities, and maps sea surface temperature grid cells to the entities in the ocean knowledge graph to form spatial-semantic associations. The entities include: Region, Current, Climate Zone, Monsoon system, and Special Area. The knowledge graph embedding representation learning module uses the knowledge graph embedding algorithm TransE to vectorize the entities and their relationships to obtain structural embeddings. It performs k-hop neighborhood retrieval on each region to extract its local semantic description and obtains semantic embeddings through a large language model. The structural embeddings and semantic embeddings are then fused through an adapter network to obtain knowledge-enhanced vectors. The time series feature encoding module acquires global historical sea surface temperature data, performs reversible instance normalization on the time series of each region, divides the time series into fixed-length time series patches, and maps each time series patch to a low-dimensional time series embedding space through a multilayer perceptron (MLP) to obtain a region-level time series feature matrix. The knowledge and data alignment module constructs a query vector for each region based on the time series feature matrix, and performs cross-attention calculation with the knowledge enhancement vector to form a context-enhanced region representation. The region representation is then input into a frozen large language model for high-dimensional representation learning to obtain a high-dimensional representation. The prediction optimization module inputs the high-dimensional representation into the Transformer decoder, learns the spatial and temporal dependencies through a multi-head attention mechanism, generates the predicted future τ-step sea surface temperature sequence, and optimizes the generated result with a loss function to obtain the prediction result. Output module: Outputs the prediction results.