Sea surface temperature spatiotemporal prediction method and system based on static and dynamic graph learning network
By using a method based on static and dynamic graph learning networks, static and dynamic graphs between sea surface temperature (SST) observation stations are generated. Feature extraction is performed using graph convolutional networks, which solves the problem of low accuracy in existing SST prediction methods and achieves more efficient SST prediction.
Patent Information
- Application Number
- CN202211255298.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-10-13
- Publication Date
- 2025-12-30
- Estimated Expiration
- 2042-10-13
AI Technical Summary
Existing sea surface temperature prediction methods are limited by network structure, making it difficult to effectively capture the complex spatiotemporal correlations between sea surface temperature data, resulting in low prediction accuracy, especially when dealing with irregular spatial data and anomalies.
We employ a method based on static and dynamic graph learning networks. By generating static and dynamic graphs between sea surface temperature (SST) observation stations, we extract node features using graph convolutional networks and combine multilayer perceptrons and residual networks to predict SST, thus constructing an end-to-end prediction framework.
It effectively captures the spatiotemporal correlation of multivariate data, improves the accuracy of sea surface temperature prediction, and can uncover long-term and short-term patterns, providing more accurate sea surface temperature forecasts.
Smart Images
Figure CN115546609B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a spatiotemporal prediction method and system for sea surface temperature based on static and dynamic graph learning networks, belonging to the field of artificial intelligence technology. Background Technology
[0002] Sea surface temperature (SST), as a crucial physical property of the world's oceans, influences virtually every aspect of the Earth's atmosphere. Changes in SST profoundly impact local and global climate. For example, variations in SST, especially anomalous changes, inevitably cause dramatic fluctuations in atmospheric water vapor levels, resulting in various precipitation patterns and even extreme weather events such as torrential rains, severe droughts, and tropical cyclones. These extreme weather events often lead to severe socio-economic impacts, such as power outages, property damage, and loss of life. Furthermore, SST changes significantly affect biological activity, particularly the plants, animals, and microorganisms within marine ecosystems.
[0003] Accurate sea surface temperature (SST) forecasting can help governments and environmental agencies plan ahead, such as for precipitation monitoring, marine life conservation, tourism, and fisheries. However, predicting SST is no easy task. The global ocean system is a typical complex and open system, and many factors influence SST, such as sunlight absorption, human activities, and local geological structures. Furthermore, global warming has significantly increased the frequency of SST anomalies, posing new challenges to SST forecasting. The development of SST forecasting technology has gone through three stages. Initially, researchers used statistical time series models to accomplish this task. Subsequently, machine learning techniques were introduced into the field. Recently, deep learning techniques, due to their flexibility and good performance, have become the preferred choice for many existing studies, such as Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM), and Graph Neural Networks (GNNs). However, limited by their respective network structures, especially when dealing with irregular spatial data, RNNs and LSTMs struggle to capture and represent the complex spatiotemporal correlations between SST data. At the same time, GNN-based methods often rely on predefined graph matrices, which may not be flexible enough to handle anomalies in SST data. Therefore, the existing sea surface temperature prediction methods have low prediction accuracy. Summary of the Invention
[0004] This invention provides a spatiotemporal prediction method and system for sea surface temperature based on static and dynamic graph learning networks, which can solve the problem of low prediction accuracy of existing sea surface temperature prediction methods.
[0005] On one hand, this invention provides a spatiotemporal prediction method for sea surface temperature based on static and dynamic graph learning networks, the method comprising:
[0006] S1. Obtain sea surface temperature data of the target sea area observed by multiple sea surface temperature observation stations, and generate a static map between the multiple sea surface temperature observation stations based on the sea surface temperature data;
[0007] S2. Generate a dynamic diagram among the multiple sea surface temperature observation stations based on the sea surface temperature data and the static diagram;
[0008] S3. Use a graph convolutional network to obtain the node features of the static graph and the dynamic graph, and obtain the predicted sea surface temperature value of the target sea area based on the node features and the sea surface temperature data.
[0009] Optionally, S1 specifically includes:
[0010] S11. Obtain sea surface temperature data observed by multiple sea surface temperature observation stations, and generate an adaptive adjacency matrix based on the sea surface temperature data;
[0011] S12. Generate a static graph between the multiple sea surface temperature observation stations based on the adaptive adjacency matrix.
[0012] Optionally, S1 further includes:
[0013] S13. Optimize the static graph using graph regularization to obtain an optimized static graph.
[0014] Optionally, S11 specifically includes:
[0015] Sea surface temperature data from multiple sea surface temperature observation stations are acquired, and an adaptive adjacency matrix of the sea surface temperature data is calculated using a node embedding method.
[0016] Optionally, S2 specifically includes:
[0017] S21. Using a multilayer perceptron, the sea surface temperature data is mapped to a space with the same dimension as the adaptive adjacency matrix to obtain the mapping result;
[0018] S22. Using a gating mechanism, the mapping result and the adaptive adjacency matrix are fused to obtain a dynamic graph between the multiple sea surface temperature observation stations.
[0019] Optionally, S22 specifically includes:
[0020] The dynamic diagram between the multiple sea surface temperature observation stations is obtained based on the first set of formulas;
[0021] The first set of formulas is:
[0022]
[0023]
[0024]
[0025]
[0026] in, All are learnable parameters, M s For an adaptive adjacency matrix, X T For the mapping result, r T Z is used to reset the door to save past information. T The door is used to update newly added information. To hide feature information in a dynamic graph, h T This is the merged animated image.
[0027] Optionally, the graph convolutional network includes a first graph convolutional network and a second graph convolutional network;
[0028] S3 specifically includes:
[0029] S31. Input the static graph into the first graph convolutional network to obtain static graph node features, and input the dynamic graph into the second graph convolutional network to obtain dynamic graph node features;
[0030] S32. Based on the static graph node features, the dynamic graph node features, and the sea surface temperature data, obtain the predicted sea surface temperature value for the target sea area.
[0031] Optionally, S32 specifically includes:
[0032] S321. Obtain the sum of the static graph node features and the dynamic graph node features, and record it as the final feature;
[0033] S322. Input the sum of the final feature and the sea surface temperature data into the output module to obtain the predicted sea surface temperature value of the target sea area.
[0034] Optionally, the output module includes a residual network and a multilayer perceptron.
[0035] On the other hand, the present invention provides a sea surface temperature spatiotemporal prediction system based on static and dynamic graph learning networks, the system comprising:
[0036] The static map generation module is used to acquire sea surface temperature data of a target sea area observed by multiple sea surface temperature observation stations, and generate a static map between the multiple sea surface temperature observation stations based on the sea surface temperature data.
[0037] The dynamic graph generation module is used to generate a dynamic graph between the multiple sea surface temperature observation stations based on the sea surface temperature data and the static graph.
[0038] The sea surface temperature prediction module is used to obtain the node features of the static graph and the dynamic graph using a graph convolutional network, and to obtain the predicted sea surface temperature value of the target sea area based on the node features and the sea surface temperature data.
[0039] The beneficial effects that this invention can produce include:
[0040] The sea surface temperature (SST) spatiotemporal prediction method based on static and dynamic graph learning networks provided by this invention effectively captures the spatiotemporal correlations of multivariate data, enabling the extraction of long-term patterns and constantly changing short-term patterns, thus significantly improving prediction accuracy. Simultaneously, this invention provides an end-to-end SST prediction framework that integrates static and dynamic graph learning, temporal convolution, and graph convolution, offering promising prediction capabilities for multivariate time series data (such as SST). Attached Figure Description
[0041] Figure 1 The flowchart illustrates the spatiotemporal prediction method for sea surface temperature based on static and dynamic graph learning networks provided in this embodiment of the invention. Detailed Implementation
[0042] The present invention will be described in detail below with reference to the embodiments, but the present invention is not limited to these embodiments.
[0043] This invention provides a spatiotemporal prediction method for sea surface temperature based on static and dynamic graph learning networks, such as... Figure 1 As shown, the method includes:
[0044] S1. Obtain sea surface temperature data of the target sea area observed by multiple sea surface temperature observation stations, and generate a static map between multiple sea surface temperature observation stations based on the sea surface temperature data.
[0045] Specifically, it includes:
[0046] S11. Obtain sea surface temperature data from multiple sea surface temperature observation stations and generate an adaptive adjacency matrix based on the sea surface temperature data.
[0047] S12. Generate a static graph of multiple sea surface temperature observation stations based on the adaptive adjacency matrix. After obtaining the adaptive adjacency matrix M... s Then, based on the adaptive adjacency matrix M s The spatial dependence weights between ocean temperature observation stations are derived, and we call the spatial dependence weights a static graph.
[0048] In practical applications, after obtaining sea surface temperature data X from multiple sea surface temperature observation stations, the adaptive adjacency matrix M of the sea surface temperature data can be calculated using the node embedding method. sThen, using the adaptive adjacency matrix M s This is used to infer a static graph to obtain a fixed long-term pattern in the data. The node embedding method can employ existing network methods, such as feedforward neural networks; this embodiment of the invention does not limit this approach.
[0049] Furthermore, S1 also includes:
[0050] S13. Optimize the static graph using graph regularization to obtain the optimized static graph.
[0051] The quality of static graph generation can be improved by utilizing graph regularization methods.
[0052] S2. Based on sea surface temperature data and static maps, generate dynamic maps between multiple sea surface temperature observation stations.
[0053] Specifically, it includes:
[0054] S21. Using a multilayer perceptron, the sea surface temperature data is mapped to a space with the same dimension as the adaptive adjacency matrix to obtain the mapping result;
[0055] S22. By using a gating mechanism to fuse the mapping results and the adaptive adjacency matrix, a dynamic graph between multiple sea surface temperature observation stations is obtained.
[0056] Specifically: a dynamic graph between multiple sea surface temperature observation stations can be obtained based on the first set of formulas;
[0057] The first set of formulas is:
[0058]
[0059]
[0060]
[0061]
[0062] in, All are learnable parameters, M s For an adaptive adjacency matrix, X T As the mapping result, r T Z is used to reset the door to save past information. T The door is used to update newly added information. To hide feature information in a dynamic graph, h T This is the merged animated image.
[0063] The multilayer perceptron in S21 and the gating mechanism in S22 can be collectively referred to as the Dynamic Learning Layer. This invention models the dynamic dependencies of data in the short term by constructing a Dynamic Learning Layer, mining the spatiotemporal dependencies between sea surface temperature data from different observation stations within a time window, and ultimately generating a dynamic graph.
[0064] S3. Use graph convolutional networks to obtain node features of static and dynamic graphs, and obtain sea surface temperature prediction values for the target sea area based on node features and sea surface temperature data.
[0065] The graph convolutional network includes a first graph convolutional network and a second graph convolutional network.
[0066] S3 specifically includes:
[0067] S31. Input the static graph into the first graph convolutional network to obtain the static graph node features, and input the dynamic graph into the second graph convolutional network to obtain the dynamic graph node features.
[0068] The first graph convolutional network and the second graph convolutional network are the same graph convolutional network, but their parameters are different.
[0069] The static graph obtained from S1 and the dynamic graph obtained from S2 are captured by two graph convolutional networks to capture the node representations of the static and dynamic graphs respectively, namely the static graph node features and the dynamic graph node features.
[0070] S32. Based on the static graph node characteristics, dynamic graph node characteristics, and sea surface temperature data, obtain the predicted sea surface temperature value for the target sea area.
[0071] Specifically, it includes:
[0072] S321. Obtain the sum of the static graph node features and the dynamic graph node features, and denote it as the final feature.
[0073] The feature transformation result is obtained by adding the static graph node features and the dynamic graph node features, denoted as the final feature. The specific mathematical expression is shown in the second formula:
[0074] Z f =Z static +Z dynamic ;
[0075] Among them, Z f For the final feature, z static Z is a static graph node feature. dynamic Features of dynamic graph nodes.
[0076] S322. Input the sum of the final features and sea surface temperature data into the output module to obtain the predicted sea surface temperature value for the target sea area. The output module includes a residual network and a multilayer perceptron.
[0077] The final feature Z obtained from S321 f After being added to the sea surface temperature data, the data is input into the residual network and the multilayer perceptron, and the final output is obtained, which is the predicted sea surface temperature value of the target sea area in the future period, thereby capturing the time trend of the input sea surface temperature data.
[0078] This invention simultaneously uncovers stable long-term patterns and dynamic short-term patterns hidden in multivariate ocean temperature data, thereby significantly improving the spatiotemporal prediction performance of ocean temperature.
[0079] The sea surface temperature (SST) spatiotemporal prediction method based on static and dynamic graph learning networks provided by this invention effectively captures the spatiotemporal correlations of multivariate data, enabling the extraction of long-term patterns and constantly changing short-term patterns, thus significantly improving prediction accuracy. Simultaneously, this invention provides an end-to-end SST prediction framework that integrates static and dynamic graph learning, temporal convolution, and graph convolution, offering promising prediction capabilities for multivariate time series data (such as SST).
[0080] Another embodiment of the present invention provides a spatiotemporal prediction system for sea surface temperature based on static and dynamic graph learning networks, the system comprising:
[0081] The static map generation module is used to acquire sea surface temperature data of the target sea area observed by multiple sea surface temperature observation stations, and generate a static map between the multiple sea surface temperature observation stations based on the sea surface temperature data.
[0082] The dynamic graph generation module is used to generate dynamic graphs between multiple sea surface temperature observation stations based on sea surface temperature data and static graphs.
[0083] The sea surface temperature prediction module is used to obtain node features of static and dynamic graphs using graph convolutional networks, and to obtain the predicted sea surface temperature value of the target sea area based on the node features and sea surface temperature data.
[0084] For a detailed description of each module in the above prediction system, please refer to the description of each step in the prediction method. It will not be repeated here. The above prediction system can achieve the same function as the prediction method.
[0085] The above description is merely a few embodiments of this application and is not intended to limit this application in any way. Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any changes or modifications made by those skilled in the art without departing from the scope of the technical solution of this application using the disclosed technical content are equivalent to equivalent implementation cases and fall within the scope of the technical solution.
Claims
1. A method for spatiotemporal prediction of sea surface temperature based on static and dynamic graph learning networks, characterized in that, The method comprises: S1, obtaining sea temperature data of a target sea area observed by a plurality of sea temperature observation stations, and generating a static graph between the plurality of sea temperature observation stations according to the sea temperature data; S2, generating a dynamic graph between the plurality of sea temperature observation stations according to the sea temperature data and the static graph; S3, obtaining node features of the static graph and the dynamic graph by using a graph convolution network, and obtaining a sea temperature prediction value of the target sea area according to the node features and the sea temperature data; The S1 specifically comprises: S11, obtaining sea temperature data observed by a plurality of sea temperature observation stations, and generating an adaptive adjacency matrix according to the sea temperature data; S12, generating a static graph between the plurality of sea temperature observation stations according to the adaptive adjacency matrix; The S2 specifically comprises: S21, mapping the sea temperature data to a space with the same dimension as the adaptive adjacency matrix by using a multi-layer perceptron, to obtain a mapping result; S22, obtaining a dynamic graph between the plurality of sea temperature observation stations according to a first formula group; The first formula group is: ； ； ； ； wherein, , , , are all learnable parameters, is an adaptive adjacency matrix, is a mapping result, is a reset gate to save past information, is an update gate to update newly added information, is a dynamic graph hidden feature information, is a fused dynamic graph; The graph convolution network comprises a first graph convolution network and a second graph convolution network; and the S3 specifically comprises: S31, inputting the static graph into the first graph convolution network to obtain static graph node features, and inputting the dynamic graph into the second graph convolution network to obtain dynamic graph node features; S32, obtaining a sea temperature prediction value of the target sea area according to the static graph node features, the dynamic graph node features and the sea temperature data.
2. The method of claim 1, wherein, The S1 further comprises: S13, optimizing the static graph by using graph regularization to obtain an optimized static graph.
3. The method of claim 1, wherein, The S11 specifically comprises: Obtaining sea temperature data observed by a plurality of sea temperature observation stations, and calculating an adaptive adjacency matrix of the sea temperature data by using a node embedding method.
4. The method of claim 1, wherein, The S32 specifically comprises: S321, obtaining a sum value of the static graph node features and the dynamic graph node features, denoted as a final feature; S322, inputting a sum value of the final feature and the sea temperature data into an output module to obtain a sea temperature prediction value of the target sea area.
5. The method of claim 4, wherein, The output module comprises a residual network and a multi-layer perceptron.
6. A static and dynamic graph learning network based spatiotemporal sea surface temperature prediction system, characterized in that, The system comprises: A static graph generation module, configured to obtain sea temperature data of a target sea area observed by a plurality of sea temperature observation stations, and generate a static graph between the plurality of sea temperature observation stations according to the sea temperature data; A dynamic graph generation module, configured to generate a dynamic graph between the plurality of sea temperature observation stations according to the sea temperature data and the static graph; A sea temperature prediction module, configured to obtain node features of the static graph and the dynamic graph by using a graph convolution network, and obtain a sea temperature prediction value of the target sea area according to the node features and the sea temperature data; The static graph generation module is specifically configured to obtain sea temperature data observed by a plurality of sea temperature observation stations, generate an adaptive adjacency matrix according to the sea temperature data, and generate a static graph between the plurality of sea temperature observation stations according to the adaptive adjacency matrix; The dynamic graph generation module is specifically configured to map the sea temperature data into a space with the same dimension as the adaptive adjacency matrix by using a multi-layer perceptron to obtain a mapping result; and obtain the dynamic graph between the plurality of sea temperature observation stations according to a first formula group. The first formula group is: ； ； ； ； wherein, , , , are all learnable parameters, is an adaptive adjacency matrix, is a mapping result, is a reset gate to save past information, is an update gate to update newly added information, is dynamic graph hidden feature information, is a fused dynamic graph; The graph convolution network includes a first graph convolution network and a second graph convolution network; the sea temperature prediction module is specifically configured to input the static graph into the first graph convolution network to obtain static graph node features, and input the dynamic graph into the second graph convolution network to obtain dynamic graph node features; and obtain a sea temperature prediction value of a target sea area according to the static graph node features, the dynamic graph node features, and the sea temperature data.