El nino prediction method and system based on graph neural network

Abstract

The application discloses an El Nino prediction method and system based on a graph neural network, and the method comprises the following steps: obtaining sea surface temperature and heat content data of the previous m months and storing the data on a Grid graph; constructing a Mesh graph and a Grid2Mesh graph by using a graph construction module in the model; embedding node features of the Grid graph, edge features of the Grid2Mesh graph and node and edge features of the Mesh graph by using an embedding layer in the model, so as to obtain corresponding embedding representations; updating node embedding of the Mesh graph by using an encoder in the model according to node embedding of the Grid graph and the Mesh graph and edge embedding of the Grid2Mesh graph, and then updating the node embedding of the Mesh graph by using a processor in the model; and inputting seasonal encoding and node embedding of the Mesh graph into a predictor in the model to predict a marine Nino index. The model is constructed, and the accuracy of El Nino phenomenon prediction is improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of meteorological prediction and artificial intelligence, and particularly relates to an El Nino prediction method and system based on a graph neural network, a computer device, and a computer readable storage medium. BACKGROUND

[0002] In the field of climate science, the prediction of El Nino phenomenon has always been a hot research topic because it has a profound impact on global climate patterns and extreme weather events. Traditionally, El Nino prediction mainly relies on statistical models and atmosphere-ocean coupled models. Although these methods have made some progress, they still have problems such as low accuracy and limited prediction time window. Especially when dealing with the complex interaction between the ocean and the atmosphere, traditional models often fail to fully capture the nonlinear characteristics of these dynamic systems.

[0003] In addition, with the development of remote sensing technology and ocean monitoring technology, vast amounts of ocean and atmospheric data are collected and stored. Traditional prediction models often face the challenges of huge demand for computing resources and slow processing speed when dealing with these big data. These limitations not only affect the accuracy of the prediction, but also limit the practicality and timeliness of the prediction results.

[0004] Therefore, there is an urgent need to develop new prediction methods and technologies to improve the accuracy and efficiency of El Nino phenomenon prediction. Artificial intelligence technology, especially graph neural networks, has shown its powerful data processing and pattern recognition capabilities in multiple fields. Graph neural networks, by effectively representing and processing unstructured data, have shown the potential to handle complex climate data, providing new possibilities for improving the accuracy and efficiency of El Nino prediction. SUMMARY

[0005] Therefore, the present application provides an El Nino prediction method and system based on a graph neural network, a computer device, and a computer readable storage medium. By constructing a multi-scale Mesh graph, the spatial relationship between ocean data can be more accurately analyzed, and the prediction accuracy of El Nino phenomenon can be improved. By splicing the seasonal information and the learned graph representation into a new embedding input into the predictor, the accuracy of El Nino prediction is improved.

[0006] The first object of the present application is to provide an El Nino prediction method based on a graph neural network.

[0007] The second object of the present application is to provide an El Nino prediction system based on a graph neural network.

[0008] The third object of the present application is to provide a computer device.

[0009] The fourth object of the present application is to provide a computer readable storage medium.

[0010] The first object of the present application can be achieved by adopting the following technical solutions:

[0011] An El Nino prediction method based on a graph neural network, the method comprising:

[0012] According to the time of predicting the marine Nino index, the sea surface temperature and heat content data of the previous m months are obtained and stored in the form of latitude and longitude grid data on the Grid graph; m is a positive integer greater than or equal to 2;

[0013] The Mesh graph and the Grid2Mesh graph are constructed by using the graph construction module in the El Nino prediction model, comprising: regarding the earth as a regular unit sphere, a multi-scale Mesh graph is constructed on the surface of the sphere; if the distance between the nodes of the Grid graph and the nodes of the Mesh graph is a set condition, the Grid2Mesh graph is constructed by using the directed edges from the nodes of the Grid graph to the nodes of the Mesh graph; wherein the node features of the Grid graph include the sea surface temperature and heat content data, and the cosine value of the latitude and the sine and cosine values of the longitude; the node features of the Mesh graph include the cosine value of the corresponding latitude and the sine and cosine values of the corresponding longitude on the earth, and the edge features include the length of the edge and the displacement vector from the sending node to the receiving node, and the edge features of the Grid2Mesh graph include the length of the edge and the displacement vector from the sending node to the receiving node;

[0014] The node features of the Grid graph, the edge features of the Grid2Mesh graph, and the node features and edge features of the Mesh graph are embedded by using the embedding layer in the El Nino prediction model, respectively, to obtain the corresponding fixed-dimensional embedding representations;

[0015] According to the node embedding of the Grid graph, the edge embedding of the Grid2Mesh graph, and the node embedding of the Mesh graph, the node embedding of the Mesh graph is updated by using the encoder in the El Nino prediction model;

[0016] According to the edge embedding and the updated node embedding of the Mesh graph, the node embedding of the Mesh graph is updated by using the processor in the El Nino prediction model;

[0017] The season encoding and the node embedding of the Mesh graph are input into the predictor in the El Nino prediction model to predict the marine Nino index.

[0018] Further, the process of constructing a multi-scale Mesh graph is as follows:

[0019] An inscribed regular polyhedron is initialized on the sphere, and each face is a triangular Mesh graph; each face of the polyhedron is then refined and iteratively divided, and after multiple iterations, polyhedron Meshs of different coarse and fine scales are obtained;

[0020] The nodes of the coarse scale Mesh graph are a subset of the nodes of the fine scale Mesh graph, because the fine scale Mesh graph is divided on the basis of the coarse scale Mesh graph;

[0021] The edges of the Mesh graphs of different scales are superimposed on the Mesh graph of the finest scale to form a multi-scale Mesh graph.

[0022] Further, the nodes of the Mesh graph are uniformly distributed in the R-order refined Mesh grid M i , where R is a positive integer greater than or equal to 3;

[0023] The setting condition is less than or equal to 0.6 times the length of the edge in the Mesh grid M R .

[0024] Further, the node embedding of the Grid graph, the edge embedding of the Grid2Mesh graph and the node embedding of the Mesh graph are used to update the node embedding of the Mesh graph by using the encoder in the El Nino prediction model, which comprises:

[0025] The message passing is performed on the Grid2Mesh graph by using the graph neural network to update each edge of the Grid2Mesh graph:

[0026]

[0027] In the formula, represents the φ function in the graph network; represents the s node embedding of the Grid graph, represents the r node embedding of the Mesh graph; represents the node edge embedding of the Grid2Mesh graph; represents the updated edge embedding; For each node of the Mesh graph, the edge information of the Grid2Mesh graph with the node as the receiving node is aggregated to update the node:

[0028]

[0029] In the formula,

[0030] represents the i node embedding of the Mesh graph, represents the updated node embedding; represents the φ function in the graph network, and the summation operation represents the ρ function in the graph network.

[0031] ​Further, the node embedding of the Mesh graph is updated by using the processor in the El Nino prediction model according to the edge embedding and the updated node embedding of the Mesh graph, including:

[0032] The multi-step message passing is performed on the Mesh graph by stacking the multi-layer graph network model, so that the nodes of the Mesh graph are aggregated to the neighbor node information of a higher order;

[0033] For one-step message passing, the update process of each edge and each node in the Mesh graph is as follows:

[0034] The update of each edge is as follows:

[0035]

[0036] In the formula, respectively represent the s node embedding and the r node embedding in the Mesh graph, represent the node to the node ; represent the φ function in the graph network;

[0037] The update of each node is as follows:

[0038]

[0039] In the formula, represent the updated i node embedding in the Mesh graph, represent the updated node embedding; represent the φ function in the graph network, and the summation operation represents the ρ function in the graph network.

[0040] Further, the predictor includes a pooling layer and a multi-layer perception;

[0041] The predictor inputs the season encoding and the node embedding of the Mesh graph into the El Nino prediction model, and predicts the marine Nino index, including:

[0042] The node embedding of the Mesh graph is characterized by pooling into a graph embedding representation by using the pooling layer;

[0043] The graph embedding representation and the season feature vector are spliced and input into the multi-layer perception to obtain the predicted marine Nino index.

[0044] Further, the season feature vector is a one-hot encoding with a length of 12, and the season position corresponding to the initial Grid graph is taken as 1 and the remaining positions are taken as 0.

[0045] Further, the multi-layer perception is used to output a dimension of n, that is, to predict the ocean Nino index of n months; wherein n is a positive integer greater than or equal to 1.

[0046] The second object of the application can be achieved by adopting the following technical solutions:

[0047] The system comprises:

[0048] The acquisition module is configured to acquire the sea surface temperature and heat content data of the previous m months according to the time of predicting the ocean Nino index and store the data on the Grid graph in the form of latitude and longitude grid data; m is a positive integer greater than or equal to 2.

[0049] The construction module is configured to construct the Mesh graph and the Grid2Mesh graph by using the graph construction module in the El Nino prediction model, including: regarding the earth as a regular unit sphere, constructing a multi-scale Mesh graph on the surface of the sphere; if the distance between the nodes of the Grid graph and the nodes of the Mesh graph meets a set condition, constructing the Grid2Mesh graph by using the directed edges from the nodes of the Grid graph to the nodes of the Mesh graph; wherein the node features of the Grid graph include the sea surface temperature and heat content data, and the cosine value of the latitude and the sine and cosine values of the longitude of the node; the node features of the Mesh graph include the cosine value of the corresponding latitude and the sine and cosine values of the corresponding longitude of the node on the earth, and the edge features include the length of the edge and the displacement vector from the sending node to the receiving node; and the edge features of the Grid2Mesh graph include the length of the edge and the displacement vector from the sending node to the receiving node.

[0050] The embedding module is configured to embed the node features of the Grid graph, the edge features of the Grid2Mesh graph, and the node features and edge features of the Mesh graph by using the embedding layer in the El Nino prediction model, respectively, to obtain the corresponding fixed-dimension embedding representations.

[0051] The first update module is configured to update the node embedding of the Mesh graph by using the encoder in the El Nino prediction model according to the node embedding of the Grid graph, the edge embedding of the Grid2Mesh graph, and the node embedding of the Mesh graph.

[0052] The second update module is configured to update the node embedding of the Mesh graph by using the processor in the El Nino prediction model according to the edge embedding and the updated node embedding of the Mesh graph.

[0053] The prediction module is configured to input the seasonal encoding and the node embedding of the Mesh graph into the predictor in the El Nino prediction model to predict the ocean Nino index.

[0054] The third object of the application can be achieved by adopting the following technical solutions:

[0055] A computer device comprises a processor and a memory for storing a program executable by the processor, and the processor implements the above-mentioned El Nino prediction method based on a graph neural network when executing the program stored in the memory.

[0056] The fourth object of the present application can be achieved by adopting the following technical solution:

[0057] A computer readable storage medium stores a program, and the program is executed by a processor to implement the above-mentioned El Nino prediction method based on a graph neural network.

[0058] The present application has the following beneficial effects relative to the prior art:

[0059] The present application considers information of different spatial scales, constructs a multi-scale Mesh graph, solves the problem that the actual spatial size is different in the same latitude and longitude range in the traditional latitude and longitude grid, can more accurately process and analyze the spatial relationship between marine data, and improves the prediction accuracy of the El Nino phenomenon; by splicing the seasonal information as a feature vector and the learned graph representation into a new embedded input into the predictor, the accuracy of the El Nino prediction is improved. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from the structures shown in these drawings without creative labor.

[0061] Figure 1 A simple flowchart of the El Nino prediction method based on a graph neural network of the present application embodiment 1;

[0062] Figure 2 A detailed flowchart of the El Nino prediction method based on a graph neural network of the present application embodiment 1;

[0063] Figure 3 A flowchart of constructing a Mesh graph of the present application embodiment 1;

[0064] Figure 4 A structure diagram of a graph network of the present application embodiment 1;

[0065] Figure 5 A flowchart of updating node information and edge information in a graph by an encoder and a processor of the present application embodiment 1;

[0066] Figure 6A structural schematic diagram of a predictor of Embodiment 1 of the present application;

[0067] Figure 7 A structural block diagram of an El Nino prediction system based on a graph neural network of Embodiment 2 of the present application;

[0068] Figure 8 A structural block diagram of a computer device of Embodiment 3 of the present application. DETAILED DESCRIPTION

[0069] In order to make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in conjunction with the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application. It should be understood that the described specific embodiments are only used to explain the present application, and are not used to limit the present application.

[0070] Embodiment 1:

[0071] As shown in Figure 1 , 2 , the El Nino prediction method based on a graph neural network provided by the present embodiment comprises the following steps:

[0072] S101, acquiring data.

[0073] From the ocean data of the existing data set, 3-month sea surface temperature (SST) and ocean heat content (HC) data are acquired, and the SST and HC data are stored in the form of grid data. For example, a geographical area of 55°S-60°N, 0°-360°E is divided into 5°×5° latitude and longitude grids (Grid), and there are 24×72 Grids. For a certain month, each grid corresponds to the average SST and HC anomaly in the latitude and longitude area of the month.

[0074] S102, inputting the data stored in the Grid graph into an El Nino prediction model to predict the El Nino phenomenon.

[0075] The El Nino prediction model comprises a graph construction module, an embedding layer, an encoder, a processor and a predictor.

[0076] Further, step S102 specifically comprises the following steps:

[0077] (1) constructing a Mesh graph and a Grid2Mesh graph by using the graph construction module.

[0078] (1-1) constructing a Mesh graph.

[0079] For Grid data form, each Grid is regarded as a node of the graph, and the edges are connected between adjacent Grid nodes, thus a Grid graph is constructed. However, the Grid graph only determines whether there is an edge connection between nodes according to the geographical location of the neighborhood, that is, only local information in the ocean is considered and global ocean-atmosphere interaction is ignored. On the other hand, due to the definition of latitude and longitude, the actual coverage range of the Grid node at high latitude is smaller than that of the Grid node at low latitude, resulting in uneven distribution of nodes in geographical location.

[0080] In order to better utilize the modeling capability of GNN for random sparse interaction, a multi-scale mesh graph is constructed on the earth surface, as shown in Figure 3 .

[0081] The specific mapping process is as follows: the earth is regarded as a regular unit sphere, an inscribed regular icosahedron is initialized on the surface of the sphere, each face is a triangular mesh, and then the icosahedron is iteratively refined, as shown in Figure 1 . After 3 iterations, a multi-polyhedron mesh of different coarse and fine scales can be obtained. Since the fine-scale mesh is divided on the basis of the coarse-scale mesh, the coarse mesh nodes are a subset of the fine mesh nodes, so the edges of the mesh graphs of different scales can be superimposed on the mesh graph of the finest scale to form a multi-scale mesh graph, in which the edges of the coarse scale capture long-distance interaction information and the edges of the fine scale capture local interaction information. This graph structure connects positions of different spatial distances on the earth, unlike the Grid graph in which the grid is sparse at low latitudes and dense at high latitudes, the Mesh graph has uniform resolution in the global range, and the nodes constructed thereby have similar spatial coverage areas. Therefore, message passing on the Mesh graph can to some extent solve the defects of the Grid graph.

[0082] V M represents a set of nodes on the multi-scale mesh graph, where the i-th node is represented by . The Mesh nodes are uniformly distributed on the R-order refined Mesh grid M R inscribed in the unit sphere, M 0 represents the initial icosahedron with 12 vertices and 20 edges, of which two vertices are located at the two polar points of the earth. Each refinement process M r → M r+1 is to divide one triangular face of the polyhedron into four small triangular faces of equal size, and three new nodes are formed at the midpoints of the original three edges, and then the new nodes are mapped to the unit sphere. After 3 iterations of refinement, the number of nodes, edges and faces of polyhedrons of different scales are shown in Table 1. The initial features of the Mesh nodes ​ It includes three features: the cosine value of the node at its corresponding latitude on Earth and the sine and cosine values ​​of its corresponding longitude.

[0083] Table 1. Mesh graph structure information at different scales

[0084]

[0085] E M This represents the set of bidirectional edges connecting nodes in a multi-scale mesh graph. The edge lengths differ for different mesh graph scales. Because the thinning process preserves low-order mesh nodes, some nodes act as "hubs," connecting both coarse-scale and fine-scale edges, thus forming a multi-scale mesh graph. From nodes... To the node edge initial features It includes four features: the length of the edge and the displacement vector (three-dimensional displacement vectors x, y, z) from the sending node to the receiving node of the directed edge.

[0086] (1-2) Construct the Grid2Mesh graph.

[0087] To map the original Grid data onto a Mesh graph for message passing, a Grid2Mesh graph was constructed.

[0088] The specific construction rule is: if the distance between a node in the Grid graph and a node in the Mesh graph is less than or equal to the distance between the nodes in the Mesh graph... 3 If the length of the middle edge is 0.6 times, then there is a directed edge connecting the two nodes from the Grid to the Mesh, which also ensures that each Grid node is connected to at least one Mesh node.

[0089] The final Grid2Mesh graph is a bipartite graph with only directed edges from Grid nodes to Mesh nodes.

[0090] V G This represents the set of nodes on the Grid graph, where the i-th node is represented by... This indicates its initial characteristics. It includes SST and HC anomalies (a 3-month time window is used in this embodiment), as well as the cosine value of the latitude and the sine and cosine values ​​of the longitude of the node, for a total of 9 features.

[0091] E G2M This represents the set of one-way directed edges on the Grid2Mesh graph. initial features The calculation method is consistent with the edge features in the Mesh graph, and the position information of the Grid node is obtained by coordinate conversion of the corresponding latitude and longitude in the spherical coordinate system.

[0092] (2) The node features of the Grid graph, the edge features of the Grid2Mesh graph, and the node features and edge features of the Mesh graph are respectively embedded by using an embedding layer.

[0093] The original input data is mapped to the representation of the latent space for subsequent message passing on the graph by the encoder module and the processor module.

[0094] First, the node features of the Grid graph and the edge features of the Grid2Mesh graph, and the node and edge features of the Mesh graph in the original input are embedded, i.e., input into the MLP (Multi-Layer Perceptron) to map them into embedded representations with fixed feature dimension:

[0095]

[0096] (3) According to the node embedding of the Grid graph, the edge embedding of the Grid2Mesh graph, and the node embedding of the Mesh graph, the node embedding of the Mesh graph is updated by using the encoder.

[0097] After obtaining the embedded representation, the representation of each Grid node needs to be mapped from the Grid graph to the Mesh graph, which can be completed by applying a graph neural network on the Grid2Mesh graph for one-step message passing.

[0098] The basic model of the graph neural network applied in this embodiment is modified according to the graph network (GN) model, and the related operations representing the global information of the whole graph are removed in its complete architecture. The GN architecture is shown in Figure 4 φ e and φ v in the figure both represent multi-layer perceptrons.

[0099] For one-step message passing process, in the Grid2Mesh graph, the update process of each edge is as follows:

[0100]

[0101] wherein, φ

[0102] Each node is updated by aggregating the edge information of the node as the receiving node:

[0103]

[0104] wherein, The summation operation corresponds to the rho function in GN. Since the Grid2Mesh receiving nodes are only Mesh nodes, and the subsequent calculations of the model do not involve Grid nodes, no additional operations are required for Grid nodes.

[0105] (4) Update the node embedding of the Mesh graph using the processor according to the edge embedding and the updated node embedding of the Mesh graph.

[0106] The embedding representation mapped to the Mesh graph can be stacked with multiple GN blocks for multi-step message passing.

[0107] For a one-step message passing process, the update process of each edge in the Mesh graph is as follows:

[0108]

[0109] The update process of each node is as follows:

[0110]

[0111] By stacking multiple GN blocks that do not share information, the node information in the graph can be aggregated to higher-order neighbor nodes, capturing long-range dependencies.

[0112] The flow of the encoder and the processor updating the node information and edge information in the graph can be referred to in Figure 5 .

[0113] (5) Input the seasonal encoding and the node embedding of the Mesh graph into the predictor in the El Nino prediction model to predict the Oceanic Niño Index.

[0114] The classic architecture of the traditional GN block is the "encoder-processor-decoder" structure, and both the input and the output contain graph structure information, i.e., the information of nodes and edges. However, in the prediction task of the present embodiment, the learned graph embedding representation is used to predict the ONI, so the original decoder structure is not suitable for graph regression tasks. It is changed to a predictor, and its architecture can be referred to in Figure 6 .

[0115] The predictor specifically consists of a pooling layer and an MLP. The pooling layer takes the Mesh graph node embedding representation V M pooling the graph embedding representation G M , and inputs it into the MLP after being concatenated with the seasonal feature vector S to obtain the predicted ONI, i.e.:

[0116] ONI = MLP predictor([pool(V M ||S])

[0117] wherein the seasonal feature vector S is a one-hot encoding of length 12, the Grid map of the original input takes 1 at the seasonal position and 0 at the rest positions; || represents the vector splicing operation; pool represents the pooling operation, which can be maximum pooling, average pooling, etc.

[0118] The traditional ONI prediction is single output, i.e., predicting the ONI of a future month. The embodiment improves it to multi-output prediction, i.e., predicting the ONI of months 1-23 in the future, as shown in the following formula:

[0119]

[0120] wherein t w represents the length of the time window used to construct the training sample.

[0121] The embodiment changes the output dimension of the MLP in the predictor from 1 to 23, so as to simultaneously output the ONI values of 23 months.

[0122] The embodiment introduces the predicted seasonal information, which is input into the MLP after splicing the seasonal encoding and the graph embedding representation, so as to guarantee the accuracy of multi-step prediction. In the training process of the model, the mean square error between the predicted value and the true value of multiple months is used as the loss function to optimize the model.

[0123] The embodiment adopts the multi-step prediction strategy in the predictor, which not only reduces the time and resource consumption of the training of the El Nino prediction model, but also enables the prediction result to be generated faster, thereby improving the prediction efficiency.

[0124] Further, before inputting the data stored in the Grid map into the El Nino prediction model, the El Nino prediction model is trained, specifically including:

[0125] The training set in the embodiment is CMIP5 and SODA, and the test set is GODAS;

[0126] First, the SST and HC latitude and longitude grid data in the training set are extracted by the sea-land mask to obtain the data of the ocean part, and the data is divided into time windows with a length of 3 in chronological order, and the features of the Grid nodes, Mesh nodes, Grid2Mesh edges and Mesh edges are initialized according to the process of the graph construction module.

[0127] The El Nino prediction model uses the mean square error between the predicted value and the true value of the ONI of the future 23 months as the loss function:

[0128]

[0129] wherein, and Y i (t) respectively represent the ONI predicted value and the true value of the i-th sample at the pre-injection time t, and n represents the number of samples.

[0130] The training objective of the model is to minimize the value of the loss function.

[0131] In this embodiment, the Adam optimizer is used for optimization.

[0132] Those skilled in the art can understand that all or part of the steps in the method of the above-mentioned embodiments can be instructed by a program to relevant hardware to complete, and the corresponding program can be stored in a computer readable storage medium.

[0133] It should be noted that although the method operations of the above-mentioned embodiments are described in a specific order in the accompanying drawings, this does not require or imply that the operations must be performed in this specific order, or that all of the shown operations must be performed to achieve the desired result. On the contrary, the depicted steps can change the order of execution. Additionally or alternatively, some steps can be omitted, a plurality of steps can be combined into one step, and / or one step can be divided into a plurality of steps.

[0134] Embodiment 2:

[0135] As Figure 7 shown, the embodiment provides an El Nino prediction system based on a graph neural network, which comprises an acquisition module 701, a construction module 702, an embedding module 703, a first update module 704, a second update module 705 and a prediction module 706, wherein:

[0136] The acquisition module 701 is configured to acquire the sea surface temperature and heat content data of the previous m months according to the time of predicting the ocean Nino index and store the data on the Grid graph in the form of latitude and longitude grid data; m is a positive integer greater than or equal to 2.

[0137] The constructing module 702 is configured to construct a Mesh graph and a Grid2Mesh graph by using a graph constructing module in the El Nino prediction model, including: regarding the earth as a regular unit sphere, constructing a multi-scale Mesh graph on the surface of the sphere; if the distance between a node of a Grid graph and a node of the Mesh graph meets a set condition, constructing a Grid2Mesh graph by using a directed edge from the node of the Grid graph to the node of the Mesh graph; wherein the node features of the Grid graph include sea surface temperature and heat content data, and a cosine value of the latitude and sine and cosine values of the longitude of the node; the node features of the Mesh graph include a cosine value of the latitude and sine and cosine values of the longitude corresponding to the node on the earth, and the edge features include the length of the edge and a displacement vector from the sending node to the receiving node, and the edge features of the Grid2Mesh graph include the length of the edge and the displacement vector from the sending node to the receiving node;

[0138] The embedding module 703 is configured to respectively embed the node features of the Grid graph, the edge features of the Grid2Mesh graph, and the node features and edge features of the Mesh graph by using an embedding layer in the El Nino prediction model, to obtain corresponding fixed-dimension embedding representations.

[0139] The first updating module 704 is configured to update the node embedding of the Mesh graph by using an encoder in the El Nino prediction model according to the node embedding of the Grid graph, the edge embedding of the Grid2Mesh graph, and the node embedding of the Mesh graph.

[0140] The second updating module 705 is configured to update the node embedding of the Mesh graph by using a processor in the El Nino prediction model according to the edge embedding and the updated node embedding of the Mesh graph.

[0141] The prediction module 706 is configured to input the seasonal encoding and the node embedding of the Mesh graph into a predictor in the El Nino prediction model, to predict the marine El Nino index.

[0142] The specific implementation of each module in this embodiment can be referred to the above-described embodiment 1, which will not be repeated here; it should be noted that the system provided in this embodiment is only exemplified by the division of the above-described functional modules, and in actual application, the above-described functions can be completed by different functional modules according to needs, that is, the internal structure is divided into different functional modules to complete all or part of the functions described above.

[0143] Embodiment 3

[0144] The embodiment provides a computer device, which can be a computer, such as Figure 8As shown, the processor 802, the memory, the input device 803, the display 804 and the network interface 805 connected by the system bus 801, the processor is used to provide computing and control ability, the memory includes non-volatile storage medium 806 and internal memory 807, the non-volatile storage medium 806 stores operating system, computer program and database, the internal memory 807 provides environment for the running of operating system and computer program in non-volatile storage medium, when the processor 802 executes the computer program stored in the memory, the above-mentioned embodiment 1 based on the graph neural network of the El Nino prediction method is realized, as follows:

[0145] According to the time of predicting the ocean Nino index, the sea surface temperature and heat content data of the previous m months are obtained and stored in the form of latitude and longitude grid data on the Grid graph; m is a positive integer greater than or equal to 2;

[0146] The graph construction module in the El Nino prediction model is used to construct the Mesh graph and the Grid2Mesh graph, including: regarding the earth as a regular unit sphere, a multi-scale Mesh graph is constructed on the surface of the sphere; if the distance between the nodes of the Grid graph and the nodes of the Mesh graph is a set condition, then the Grid2Mesh graph is constructed by using the directed edge from the node of the Grid graph to the node of the Mesh graph; wherein the node features of the Grid graph include the sea surface temperature and heat content data, and the cosine value of the latitude and the sine and cosine values of the longitude of the node; the node features of the Mesh graph include the cosine value of the corresponding latitude and the sine and cosine values of the corresponding longitude of the node on the earth, the edge features include the length of the edge and the displacement vector from the sending node to the receiving node, and the edge features of the Grid2Mesh graph include the length of the edge and the displacement vector from the sending node to the receiving node;

[0147] The embedding layer in the El Nino prediction model is used to embed the node features of the Grid graph, the edge features of the Grid2Mesh graph, and the node features and edge features of the Mesh graph, respectively, to obtain the corresponding fixed-dimensional embedding representation;

[0148] According to the node embedding of the Grid graph, the edge embedding of the Grid2Mesh graph and the node embedding of the Mesh graph, the encoder in the El Nino prediction model is used to update the node embedding of the Mesh graph;

[0149] According to the edge embedding and the updated node embedding of the Mesh graph, the processor in the El Nino prediction model is used to update the node embedding of the Mesh graph;

[0150] The season encoding and the node embedding of the Mesh graph are input into the predictor in the El Nino prediction model to predict the ocean Nino index.

[0151] Embodiment 4:

[0152] The embodiment provides a computer readable storage medium storing a computer program, the computer program being executed by a processor to implement the El Nino prediction method based on a graph neural network in the above embodiment 1, as follows:

[0153] According to the time of predicting the oceanic Nino index, the sea surface temperature and heat content data of the previous m months are obtained and stored in the form of grid data on a Grid graph; m is a positive integer greater than or equal to 2.

[0154] The Mesh graph and the Grid2Mesh graph are constructed by using the graph construction module in the El Nino prediction model, including: regarding the earth as a regular unit sphere, a multi-scale Mesh graph is constructed on the surface of the sphere; if the distance between the nodes of the Grid graph and the nodes of the Mesh graph is a set condition, the Grid2Mesh graph is constructed by using the directed edges from the nodes of the Grid graph to the nodes of the Mesh graph; wherein the node features of the Grid graph include the sea surface temperature and heat content data, and the cosine value of the latitude and the cosine and sine values of the longitude of the node; the node features of the Mesh graph include the cosine value of the corresponding latitude and the cosine and sine values of the corresponding longitude of the node on the earth, the edge features include the length of the edge and the displacement vector from the sending node to the receiving node, and the edge features of the Grid2Mesh graph include the length of the edge and the displacement vector from the sending node to the receiving node.

[0155] The node features of the Grid graph, the edge features of the Grid2Mesh graph, and the node features and edge features of the Mesh graph are embedded by using the embedding layer in the El Nino prediction model, respectively, to obtain the corresponding fixed-dimensional embedding representations.

[0156] According to the node embedding of the Grid graph, the edge embedding of the Grid2Mesh graph and the node embedding of the Mesh graph, the node embedding of the Mesh graph is updated by using the encoder in the El Nino prediction model.

[0157] According to the edge embedding and the updated node embedding of the Mesh graph, the node embedding of the Mesh graph is updated by using the processor in the El Nino prediction model.

[0158] The season encoding and the node embedding of the Mesh graph are input into the predictor in the El Nino prediction model to predict the oceanic Nino index.

[0159] It should be noted that the computer readable storage medium of the embodiment can be a computer readable signal medium or a computer readable storage medium or any combination of the two. The computer readable storage medium may, for example, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or apparatus, or any combination of the above. More specific examples of computer readable storage medium can include, but are not limited to, an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0160] The above is only the preferred embodiment of the present application, but the protection scope of the present application is not limited to this. Any skilled person in the art can make equivalent replacement or change according to the technical scheme and the inventive concept of the present application within the scope disclosed by the present application, and it still belongs to the protection scope of the present application.

Claims

1. An El Niño prediction method based on graph neural networks, characterized in that, The method includes: Based on the predicted time of the Oceanic Niño index, sea surface temperature and heat content data for the m months prior to the prediction are obtained and stored on a grid map in the form of latitude and longitude grid data; m is a positive integer greater than or equal to 2; The El Niño prediction model utilizes a graph construction module to build Mesh and Grid2Mesh graphs, including: treating the Earth as a regular unit sphere and constructing a multi-scale Mesh graph on the sphere's surface; if the distance between nodes in the Grid graph and the Mesh graph is a set condition, then the directed edges from nodes in the Grid graph to nodes in the Mesh graph are used to construct the Grid2Mesh graph; wherein, the node features of the Grid graph include sea surface temperature and heat content data, as well as the cosine value of the latitude and the sine and cosine values ​​of the longitude of the node; the node features of the Mesh graph include the cosine value of the corresponding latitude and the sine and cosine values ​​of the corresponding longitude of the node on Earth, and the edge features include the edge length and the displacement vector from the sending node to the receiving node; the edge features of the Grid2Mesh graph include the edge length and the displacement vector from the sending node to the receiving node. The node features of the Grid graph, the edge features of the Grid2Mesh graph, and the node and edge features of the Mesh graph are embedded using the embedding layer in the El Niño prediction model to obtain the corresponding fixed-dimensional embedding representations. Based on the node embeddings of the Grid graph, the edge embeddings of the Grid2Mesh graph, and the node embeddings of the Mesh graph, the encoder in the El Niño prediction model is used to update the node embeddings of the Mesh graph. Based on the edge embeddings and updated node embeddings of the Mesh graph, the node embeddings of the Mesh graph are updated using the processor in the El Niño prediction model. Seasonal coding and Mesh graph nodes are embedded into the El Niño prediction model to predict the oceanic Niño index.

2. The El Niño prediction method according to claim 1, characterized in that, The process of constructing a multi-scale mesh graph is as follows: Initialize an inscribed regular polyhedron on the sphere, with each face being a triangle mesh; then refine and iterate each face of the polyhedron, obtaining polyhedron meshes of different coarseness scales after multiple iterations. Since the fine-scale mesh graph is partitioned based on the coarse-scale mesh graph, the nodes of the coarse-scale mesh graph are a subset of the nodes of the fine-scale mesh graph. By superimposing the edges of mesh graphs of different thicknesses onto the finest-scale mesh graph, a multi-scale mesh graph is formed.

3. The El Niño prediction method according to claim 2, characterized in that, The nodes of the mesh graph are uniformly distributed within an R-order refined mesh M inscribed in a unit sphere. R In the above, R is a positive integer greater than or equal to 3; The setting condition is less than or equal to the Mesh mesh M. R 0.6 times the length of the middle side.

4. The El Niño prediction method according to claim 1, characterized in that, The step of updating the node embeddings of the mesh graph using the encoder in the El Niño prediction model based on the node embeddings of the Grid graph, the edge embeddings of the Grid2Mesh graph, and the node embeddings of the mesh graph includes: Graph neural networks are used for message passing on the Grid2Mesh graph to update each edge of the graph: In the formula, Represents the φ function in a graph network; This indicates the embedding of node s in the Grid graph. This represents the embedding of r nodes in a Mesh graph; Nodes representing the Grid2Mesh graph To the node Edge embedding; This represents the updated edge embedding; For each node in the Mesh graph, update it by aggregating the edge information in the Grid2Mesh graph that has that node as the receiving node: In the formula, This represents the embedding of the i-node in the Mesh graph. express Updated node embedding; The φ function represents the graph network, and the summation operation represents the ρ function.

5. The El Niño prediction method according to claim 1, characterized in that, The step of updating the node embeddings of the mesh graph using the processor in the El Niño prediction model based on the edge embeddings and the updated node embeddings of the mesh graph includes: Multi-step message passing is performed on the Mesh graph by stacking multi-layer graph network models, so that the nodes of the Mesh graph can be aggregated to higher-order neighbor node information. For one-step message passing, the update process for each edge and each node in the Mesh graph is as follows: Update for each edge: In the formula, These represent the embedding of node s and node r in the Mesh graph, respectively. Represents a node To the node The edge; Represents the φ function in a graph network; Updates for each node: In the formula, This represents the updated embedding of the i-node in the Mesh graph. express Updated node embedding; The φ function represents the graph network, and the summation operation represents the ρ function.

6. The El Niño prediction method according to any one of claims 1 to 5, characterized in that, The predictor includes a pooling layer and a multilayer perceptron; The process of embedding seasonal codes and Mesh graph nodes into the El Niño prediction model to predict the oceanic Niño index includes: The node embedding representation of the Mesh graph is pooled into a graph embedding representation using a pooling layer; The graph embedding representation and seasonal feature vector are concatenated and then input into a multilayer perceptron to obtain the predicted ocean Niño index.

7. The El Niño prediction method according to claim 6, characterized in that, The seasonal feature vector is a one-hot encoded vector of length 12, with the seasonal position corresponding to the initial Grid graph set to 1 and the other positions set to 0.

8. The El Niño prediction method according to claim 6, characterized in that, Using a multilayer perceptron, the oceanic Niño index for n months is predicted by pre-outputting a dimension of n; where n is a positive integer greater than or equal to 1.

9. An El Niño prediction system based on graph neural networks, characterized in that, The system includes: The acquisition module is used to acquire sea surface temperature and heat content data for the m months prior to the prediction of the Oceanic Niño index and store them on a Grid map in the form of latitude and longitude grid data; m is a positive integer greater than or equal to 2. The construction module is used to construct Mesh and Grid2Mesh graphs using the graph construction module in the El Niño prediction model. This includes: treating the Earth as a regular unit sphere and constructing a multi-scale Mesh graph on the sphere's surface; if the distance between nodes in the Grid graph and the Mesh graph is a set condition, then constructing a Grid2Mesh graph using the directed edges from nodes in the Grid graph to nodes in the Mesh graph; wherein, the node features of the Grid graph include sea surface temperature and heat content data, as well as the cosine value of the node's latitude and the sine and cosine values ​​of its longitude; the node features of the Mesh graph include the cosine value of the node's corresponding latitude and the sine and cosine values ​​of its corresponding longitude on Earth, and the edge features include the edge length and the displacement vector from the sending node to the receiving node; the edge features of the Grid2Mesh graph also include the edge length and the displacement vector from the sending node to the receiving node. The embedding module is used to embed the node features of the Grid graph, the edge features of the Grid2Mesh graph, and the node and edge features of the Mesh graph using the embedding layer in the El Niño prediction model, respectively, to obtain the corresponding fixed-dimensional embedding representations. The first update module is used to update the node embedding of the Mesh graph using the encoder in the El Niño prediction model based on the node embedding of the Grid graph, the edge embedding of the Grid2Mesh graph, and the node embedding of the Mesh graph. The second update module is used to update the node embeddings of the Mesh graph using the processor in the El Niño prediction model based on the edge embeddings and the updated node embeddings of the Mesh graph. The prediction module is used to embed seasonal codes and nodes of the Mesh graph into the predictor of the El Niño prediction model to predict the oceanic Niño index.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the El Niño prediction method according to any one of claims 1-8.

Images