Early warning method and device based on multi-ocean physical field fusion representation based on graph convolution

Abstract

The present invention discloses an early warning method and device based on graph convolution for fusion representation of multiple ocean physical fields. The method comprises: using a three-layer graph convolutional neural network to perform feature updates on multiple physical fields, mining the mutual relationships between different physical fields, so as to update the overall feature representation with different comprehensive characteristics of physical fields; using a double-layer temporal attention operation, inputting the overall feature representation with different comprehensive characteristics of physical fields and the position coding information corresponding to the overall feature representation into the double-layer temporal attention, and obtaining a coupled representation of the data features of multiple physical fields in a time period; using the coupled representation of the data features of multiple physical fields in a time period as a global feature, and mining the temporal information to predict the Nino3.4 index, and fusing the Nino3.4 index with the input typhoon observation data to predict the typhoon intensity. The device comprises: a processor and a memory. The present invention solves the problem in the prior art that the prediction of the typhoon intensity in the sea area near the communication base station in the coastal area is not accurate enough.

Description

Technical Field

[0001] The present invention relates to the field of early warning of the marine El Nino phenomenon, and in particular to an early warning method and device based on graph convolution-based fusion representation of multiple marine physical fields. Background Art

[0002] The intensity of typhoon generation is related to the time-frequency characteristics of typhoons in the past and present, and the ENSO (El Nino) phenomenon is also one of the influencing conditions for the change of typhoon generation intensity. With the continuous change of the earth's surface climate, the interaction of atmospheric and sea level oscillations in the equatorial Pacific region, referred to as the El Nino-Southern Oscillation (ENSO), is the strongest and most significant interannual climate signal on the earth, and is also an influencing factor for extreme weather phenomena in multiple regions in short-term forecasts. Therefore, the study of accurate prediction of ENSO is the key to improving the level of climate prediction and disaster prevention and mitigation in East Asia and the world. The data required to predict ENSO mainly include the change data of physical fields such as sea surface temperature anomaly, heat content anomaly, zonal wind anomaly, and meridional wind anomaly. These data contain a lot of hidden information, so the mining of correlations within time periods of multiple ocean physical fields has also attracted the attention of many researchers.

[0003] Traditional ENSO [1] The prediction methods rely on numerical climate models and mainly use some learning methods to solve non-convex optimization problems to minimize the error between the predicted value and the actual value in the historical data. With the rapid development of deep learning and the widespread application of neural networks, more and more deep learning models are applied to various machine learning tasks and have achieved good results. Therefore, this has led to a wave of using deep learning technology to make ENSO predictions, resulting in more skilled ENSO predictions, such as ConvLSTM [2] ,CNN [3] However, they only consider the time series information and ignore the complementary information between multiple physical fields in the time segment, while graph convolutional neural network [4] Being able to better couple information from multiple physical fields will help improve the prediction accuracy of Nino3.4.

[0004] Most existing methods only explore the ocean ENSO phenomenon from the time series information of a small amount of physical fields, and the prediction accuracy is low. If the low-accuracy Nino3.4 information is coupled with the frequency information of typhoon formation to predict the typhoon intensity, it is bound to fail to achieve a good prediction effect. In addition, the prediction error of the existing statistical model method for typhoon intensity may increase with the prediction time. Therefore, there is an urgent need for a method that can accurately predict the typhoon intensity in coastal areas and then determine the risk level of communication base stations to achieve early warning. Summary of the invention

[0005] The present invention provides an early warning method and device based on multi-ocean physical field fusion representation of graph convolution. The present invention improves the prediction accuracy of typhoon intensity, solves the problem of inaccurate prediction of typhoon intensity in the sea area near communication base stations in coastal areas in the prior art, and reduces natural disaster losses. See the following description for details:

[0006] In a first aspect, an early warning method based on multi-ocean physical field fusion representation of graph convolution, the method comprising:

[0007] A three-layer graph convolutional neural network is used to update features on multiple physical fields, mining the relationships between different physical fields to update the overall feature representation with comprehensive characteristics of different physical fields;

[0008] By using the double-layer temporal attention operation, the overall feature representation with different comprehensive characteristics of physical fields and the position encoding information corresponding to the overall feature representation are input into the double-layer temporal attention to obtain the coupled representation of multi-physical field data features in the time period;

[0009] The coupled representation of multi-physical field data features in different time periods is used as global features, and the time series information is mined to predict the Nino3.4 index. The Nino3.4 index is fused with the input typhoon observation data to predict the typhoon intensity.

[0010] The method of using a three-layer graph convolutional neural network to perform feature updates on multiple physical fields and to explore the relationships between different physical fields so as to update the overall feature representation with comprehensive characteristics of different physical fields is as follows:

[0011] The first-layer graph convolutional neural network is used to update the features on the spatial sub-region structure of a single ocean physical field, and to mine the mutual correlation between the spatial sub-regions of a single ocean physical field;

[0012] For multiple ocean physical fields, the second and third layers of graph convolutional neural networks are used to couple and associate the shallow and deep information on multiple physical fields to learn the overall feature representation with comprehensive characteristics of different physical fields.

[0013] Furthermore, the first-layer graph convolutional neural network is used to perform feature update on the spatial sub-region structure of a single ocean physical field, and to mine the mutual correlation between the spatial sub-regions of a single ocean physical field, specifically:

[0014] For the information of a single physical field, a convolutional neural network is used to extract sub-region features, which are recorded as And the sub-region features are used as graph structure nodes Edge information E between sub-regions, single physical field graph structure G 1 The adjacency matrix of is:

[0015]

[0016] in, n is the number of input sub-region features, represents the Euclidean distance between the feature vectors of two sub-regions in the i-th frame of the physical field x, Normalization(·) represents the normalization function, x_a represents the a-th sub-region of the physical field x, and x_b represents the b-th sub-region of the physical field x.

[0017] Among them, for multiple ocean physical fields, the second and third layers of the graph convolutional neural network are used to couple and associate the shallow and deep information on multiple physical fields to learn and obtain the overall feature representation with comprehensive characteristics of different physical fields. Specifically,

[0018] The second-layer graph structure is a shallow multi-physical field relationship fusion layer. The nodes of this layer of graph structure are consistent with the updated nodes of the first-layer graph structure. The difference between the second-layer graph structure and the first-layer is that the edge information corresponding to each physical place is exchanged;

[0019] The third-layer graph structure is a deep multi-physical field relationship fusion layer. The nodes of this layer of graph structure are obtained by splicing the nodes obtained by updating the x physical fields in the second layer, and the side information of the third-layer graph structure is also obtained by summing and averaging the side information of the second layer. Through the deep physical information coupling of the third-layer graph structure, the correlation between the physical fields is aggregated into an overall feature representation, which serves as the time series input of all physical field couplings on the i-th frame.

[0020] Furthermore, the dual-layer temporal attention is specifically:

[0021] The first layer of the two-layer temporal attention is a time period information fusion self-attention layer, and the second layer predicts by extracting the time period coupled physical field information sequence; after the multi-physical field information is aggregated and updated by the graph convolutional neural network, the final coupling correlation features of different physical fields at the same longitude and latitude on the same frame are obtained; the coupling correlation features containing temporal information are divided into a group of M features as the input of the time period fusion information layer;

[0022] The second layer of temporal self-attention makes predictions by extracting the time period coupled physical field information sequence, and its input is the time period coupled correlation features output by the first layer.

[0023] Wherein, the spatial sub-region of the single ocean physical field is:

[0024] The size of each frame of time series data is H×W. Each frame of data is divided into equal parts according to length and width, and the spatial sub-regions are selected with equal sizes based on the geographic center. For the x-th physical field, the i-th frame is divided into n spatial sub-regions, which are recorded as It is the data of the nth spatial sub-region in the i-th frame of the x-th physical field.

[0025] In the second aspect, an early warning device based on multi-ocean physical field fusion representation of graph convolution, the early warning device comprising:

[0026] The overall feature update module is used to update features on multiple physical fields using a three-layer graph convolutional neural network, mining the relationships between different physical fields to update the overall feature representation with comprehensive characteristics of different physical fields;

[0027] A feature coupling representation module is used to use a double-layer temporal attention operation to input the overall feature representation with different physical field comprehensive characteristics and the position encoding information corresponding to the overall feature representation into the double-layer temporal attention to obtain a feature coupling representation of multi-physical field data in a time period;

[0028] The typhoon intensity prediction module is used to represent the coupled characteristics of multi-physical field data in a time period as a global feature, and to mine time series information to predict the Nino3.4 index, and to fuse the Nino3.4 index with the input typhoon observation data to predict the typhoon intensity.

[0029] In the third aspect, an early warning device based on graph convolution and fusion representation of multiple ocean physical fields, the device comprises: a processor and a memory, the memory stores program instructions, and the processor calls the program instructions stored in the memory to enable the device to execute any one of the method steps described in the first aspect.

[0030] A fourth aspect, a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, the computer program includes program instructions, and when the program instructions are executed by a processor, the processor executes any one of the method steps described in the first aspect.

[0031] The beneficial effects of the technical solution provided by the present invention are:

[0032] 1. The present invention jointly explores the temporal dynamic change characteristics between the spatiotemporally coupled multi-ocean physical fields from the two aspects of ocean multi-physical field coupling and spatiotemporal sequence. In terms of ocean multi-physical field coupling, the graph convolutional neural network is used to extract the correlation comprehensive information between different ocean physical fields. After the ocean multi-physical field coupling, the temporal attention is used to extract the feature representation in the dynamic change of the time series. The comprehensive ocean multi-physical field information can learn to obtain a more comprehensive global feature representation, which is conducive to better ocean ENSO prediction, thereby increasing the accuracy of typhoon intensity prediction;

[0033] 2. The present invention uses graph convolutional neural networks to fully explore the mutual connections between different physical fields under multiple physical fields, and transmits information based on the similarities of different physical fields; the present invention uses temporal attention to fully explore the temporal information of the comprehensive characteristics of different physical fields after coupling multiple physical fields.

[0034] Therefore, the present invention can fully learn and mine the coupled time-series dynamic change characteristic information of multi-physical field data, integrate the predicted Nino3.4 index with the time-frequency characteristics of typhoon formation, improve the prediction accuracy of typhoon intensity, and then realize the risk level classification and early warning of communication base stations in coastal areas, thereby reducing natural disaster losses. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flowchart of the early warning method based on graph convolution for fusion representation of multiple ocean physical fields;

[0036] Figure 2 Schematic diagram of the early warning method based on graph convolution for fusion representation of multiple ocean physical fields;

[0037] Figure 3 It is a schematic diagram of the structure of the early warning device based on the fusion representation of multiple ocean physical fields based on graph convolution;

[0038] Figure 4 Another structural schematic diagram of the early warning device based on graph convolution for fusion representation of multiple ocean physical fields. DETAILED DESCRIPTION

[0039] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention are described in further detail below.

[0040] Example 1

[0041] A warning method based on multi-ocean physical field fusion representation based on graph convolution, see Figure 1 , the method comprises the following steps:

[0042] Step 101: using data information obtained from the ocean multi-physical field as data input, selecting a central sub-region, and constructing spatial sub-regions of various ocean physical fields based on the central sub-region;

[0043] For each physical field of the ocean, at the same longitude and latitude, each frame of physical data is equally divided according to the size of the fixed time series data and the geographic center to select the central sub-region, and construct the spatial sub-regions of each ocean physical field based on the central sub-region.

[0044] Step 102: using the first layer of graph convolutional neural network to perform feature update on the spatial sub-region structure of a single ocean physical field, and mining the mutual correlation between the spatial sub-regions of a single ocean physical field;

[0045] Step 103: for multiple ocean physical fields, use the second and third layers of the graph convolutional neural network to couple and correlate the shallow and deep information on the multiple physical fields, so as to learn and obtain an overall feature representation with comprehensive characteristics of different physical fields;

[0046] Step 104: Repeat steps 102-103, use a three-layer graph convolutional neural network to perform feature updates on multiple physical fields, and mine the relationships between different physical fields to update the overall feature representation with comprehensive characteristics of different physical fields;

[0047] Step 105: using a double-layer temporal attention operation, inputting each overall feature representation with different comprehensive characteristics of physical fields extracted from the three-layer graph convolutional neural network and the position encoding information corresponding to the overall feature representation into the double-layer temporal attention to obtain a coupled representation of multi-physical field data features in a time period;

[0048] Step 106: Utilize multiple linear layer operations to take the coupled representation of the multi-physical field data features of the time period as the global feature, and mine the time series information to predict the Nino3.4 index. Finally, fuse the Nino3.4 index with the input typhoon observation data to predict the typhoon intensity and classify the risk level of the communication base station, so as to issue an early warning.

[0049] Among them, the typhoon observation data is well-known data, which is well known to those skilled in the art and will not be elaborated in detail in the embodiments of the present invention.

[0050] In summary, the embodiment of the present invention improves the accuracy of event detection through the above steps 101 to 106 and meets various needs in practical applications.

[0051] Example 2

[0052] The scheme in Example 1 is further introduced below in combination with specific calculation formulas and examples, as described below for details:

[0053] 201: Get multiple sets of data from multi-physics fields as input, including: sea surface temperature anomaly, heat content anomaly, etc.;

[0054] Assuming that the size of each frame of time series data is H×W, divide each frame of data into equal parts (e.g., two equal parts) according to length and width, and select the central sub-region with equal size based on the geographic center. For the xth physical field, the i-th frame is divided into n sub-regions, denoted as It is the data of the nth sub-area in the i-th frame of the x-th physical field.

[0055] 202: For the information of a single physical field, a convolutional neural network is used to extract sub-region features, denoted as And the sub-region features are used as graph structure nodes The edge information E between sub-regions, the first-level graph structure of a single physical field is:

[0056]

[0057] Where i represents the i-th frame of the time series, represents the sub-region characteristics of a single physical field, Represents the side information matrix between each sub-region of a single physical field; the side information E of the first layer of graph structure is represented by the adjacency matrix.

[0058] In order to further explore the relationship between the sub-areas of a single physical field, a single physical field graph structure G 1 The adjacency matrix of is defined as:

[0059]

[0060] in, n is the number of input sub-region features, represents the Euclidean distance between the feature vectors of two sub-regions in the i-th frame of the physical field x, Normalization(·) represents the normalization function, x_a represents the a-th sub-region of the physical field x, and x_b represents the b-th sub-region of the physical field x. The nodes and edges obtained after the first-layer graph structure update are recorded as and

[0061] 203: For the information of multiple physical fields, based on the single physical field node and edge information updated in the first layer of graph structure, two layers of graph convolutional neural network are added to mine the correlation characteristics between physical fields;

[0062] Among them, the second-layer graph structure is a shallow multi-physical field relationship fusion layer. The nodes of this layer of graph structure are consistent with the updated nodes of the first-layer graph structure, that is, the sub-region characteristics of the same physical field. The difference between the second-layer graph structure and the first layer is that the edge information corresponding to each physical place is exchanged.

[0063] Then the second graph structure is defined as:

[0064]

[0065] Among them, x represents the physical field x, and y represents the physical field y.

[0066] From the above expression, it can be seen that the second-layer graph structure only exchanges the association information between different physical fields at the same longitude and latitude. For example, the adjacency matrix obtained by calculating the physical field sea surface temperature anomaly in the first layer is used as the side information of the physical field heat content anomaly in the second layer. The updated nodes and side information of the second-layer graph structure are recorded as and

[0067] The third-layer graph structure for information coupling of multiple physical fields is a deep multi-physical field relationship fusion layer, in which the nodes of this layer of graph structure are spliced ​​by the nodes obtained by updating the x physical fields in the second layer, which can be expressed as:

[0068]

[0069] in, It is the input feature of the third-layer graph structure, which is composed of the nodes obtained by updating x physical fields.

[0070] The side information of the third-layer graph structure is also obtained by summing and averaging the side information of the second layer, which is recorded as:

[0071]

[0072] in, The edge information input for the third-layer graph structure, Updated edge information for the second-level graph structure.

[0073] Then the third-layer graph structure can be expressed as:

[0074]

[0075] Through the deep physical information coupling of the third-layer graph structure, the correlation between the physical fields is aggregated into an overall feature representation Z i , which indicates that Z i Serves as the timing input for all physics couplings at the i-th frame.

[0076] 204: Using a three-layer graph convolutional neural network to update features in multiple physical fields and explore the relationships between different physical fields to learn the overall feature coupling representation containing rich physical field information;

[0077] According to the constructed multi-physics field coupling graph structure, this method uses the graph convolutional neural network algorithm [3] To update the node features in the graph structure, so that information can be transferred between nodes and similar nodes, and the correlation information between multiple physical fields can be better mined.

[0078] For the first layer of graph structure Design the following convolutional neural network GCN to update node information:

[0079]

[0080]

[0081] in, represents the initial data characteristics of the xth physical field, The updated feature of the sub-region node representing the xth physical field, E x Representing graph structure The adjacency matrix, I represents the identity matrix, D represents (I+E x ), σ represents the nonlinear activation function, θ x are learnable parameters.

[0082] For the second layer graph structure Design the following convolutional neural network GCN to update node information:

[0083]

[0084]

[0085] in, represents the characteristics of the xth physical field after the first layer of graph convolution, Represents the physical field sub-region characteristics of the second layer, E y Representing graph structure The adjacency matrix, I represents the identity matrix, D represents (I+E y ), ρ represents the nonlinear activation function, are learnable parameters.

[0086] For the third-level graph structure Design the following convolutional neural network GCN to update node information:

[0087]

[0088]

[0089] Among them, Z x Represents the feature concatenation representation of x physical fields after the second layer of graph convolution, Represents the comprehensive overall characteristics of all physical fields in the third layer, Represents the graph structure G 3 The adjacency matrix of The degree matrix of , δ represents the nonlinear activation function, ∈ x are learnable parameters.

[0090] This method uses a three-layer graph convolutional neural network to extract the coupling and correlation features of multi-physical field information.

[0091] 205: Using two-layer temporal self-attention [5] Operation, aggregate the information of multiple physical fields into overall correlation features in the graph convolutional neural network, and then use the overall correlation features as global features for ocean ENSO prediction.

[0092] The self-attention consists of two layers. The first layer is the time period information fusion self-attention layer, and the second layer predicts by extracting the time period coupled physical field information sequence. After the graph convolutional neural network aggregates and updates the multi-physical field information, the final coupling correlation features of different physical fields at the same longitude and latitude on the same frame are obtained.

[0093] In order to extract the time period fusion information, the coupled correlation feature sequence containing the time series information is divided into a group of M features as the input of the time period fusion information layer, and the position embedding information of each coupled correlation information is added. The position embedding information is:

[0094] PE (pos,2i) =sin(pos / 10000 2i / dmodel ) (13)

[0095] PE (pos,2i+1) =cos(pos / 10000 2i / dmodel ) (14)

[0096] Among them, d model It represents the feature dimension of coupled association information, pos represents the word embedding representation of coupled association information, and i represents the order of input features.

[0097] The M coupled correlation features and their corresponding position embedding information are input into the time period information fusion self-attention mechanism of the first layer, namely:

[0098]

[0099]

[0100] in, represents the projection matrix, g represents the result of time-period fusion of M coupled correlation features, and d model Represents the feature dimension of coupling correlation information, Z m is the comprehensive overall feature representation of all physical fields output by the graph convolutional neural network, q m is the query obtained by linear operation, T is the transpose operation of the matrix, k m is the key obtained by linear operation, vm The value obtained by the linear operation.

[0101] The second layer of temporal self-attention predicts by extracting the time period coupled physical field information sequence, and its input is the time period coupled correlation feature output by the first layer, namely:

[0102] q=W q g,k=W k g,v=W v g (17)

[0103]

[0104] Among them, W q , W k , W v represents the projection matrix, F represents the feature result of predicting the time period coupled physical field information sequence, q is the query obtained by the linear operation, k is the key obtained by the linear operation, and v is the value obtained by the linear operation. The two-layer temporal self-attention mines the multi-physical field coupling information in the time series and time period.

[0105] 206: Use multiple linear layer operations to process the multi-physics field overall feature F containing time period information and generate the Nino3.4 index. The Nino3.4 index exists in the form of a vector. Therefore, the similarity metric loss function is used to constrain the model training:

[0106]

[0107] Among them, Y i is the accurate Nino3.4 index corresponding to the coupled physical field information in the ith period, F i It is the predicted value of the Nino3.4 index output by multiple linear layer operations, and k represents the number of times a batch is input.

[0108] The output Nino3.4 index is fused with the typhoon observation data, denoted as S (i) =(Nino3.4 (i) ,Q (i) ), and then output the typhoon forecast intensity through the time recurrent neural network. The typhoon intensity obtained by the output is divided into risk levels and displayed in different colors, such as 6 to 8 corresponding to blue, 8 to 10 corresponding to yellow, 10 to 12 corresponding to orange, and greater than 12 corresponding to red. According to the different risk level colors, professionals are notified and warned in advance, and emergency responses are made to ensure the stability and safety of communication base stations.

[0109] Example 3

[0110] The feasibility of Examples 1 and 2 is verified by combining specific experiments, as described below:

[0111] For example, the present invention selects July 2016 as the target month, and m is set to 5. Then, the SST and HC graphs of November, December 2015 and January 2016 are required as parallel inputs and the inputs are divided into spatial sub-regions, and then the convolutional neural network is used to obtain the features of each spatial sub-region. Then, the features of the spatial sub-regions are input into a three-layer graph convolutional neural network, aggregated into an overall feature representation, and the overall feature representation is input into a double-layer temporal self-attention layer to predict the Nino3.4 index 5 months later. The output Nino3.4 index is fused with typhoon observation data to predict the value corresponding to the typhoon intensity, so that the risk level warning of the communication base station can be achieved 5 months in advance, reducing the occurrence of natural disasters on the communication base station, or the occurrence of casualties caused by the collapse of the base station.

[0112] Example 4

[0113] An early warning device based on multi-ocean physical field fusion representation based on graph convolution, see Figure 3 , the early warning device comprises:

[0114] The overall feature update module is used to update features on multiple physical fields using a three-layer graph convolutional neural network, mining the relationships between different physical fields to update the overall feature representation with comprehensive characteristics of different physical fields;

[0115] A feature coupling representation module is used to use a double-layer temporal attention operation to input the overall feature representation with different physical field comprehensive characteristics and the position encoding information corresponding to the overall feature representation into the double-layer temporal attention to obtain a feature coupling representation of multi-physical field data in a time period;

[0116] The typhoon intensity prediction module is used to represent the coupled characteristics of multi-physical field data in a time period as a global feature, and to mine time series information to predict the Nino3.4 index, and to fuse the Nino3.4 index with the input typhoon observation data to predict the typhoon intensity.

[0117] It should be pointed out here that the device description in the above embodiment corresponds to the method description in the embodiment, and the embodiment of the present invention will not be described in detail here.

[0118] In summary, the embodiment of the present invention improves the prediction accuracy of typhoon intensity through the above-mentioned module, solves the problem of inaccurate prediction of typhoon intensity in the sea area near communication base stations in coastal areas in the prior art, and reduces the losses caused by natural disasters.

[0119] Example 5

[0120] An early warning device based on multi-ocean physical field fusion representation based on graph convolution, see Figure 4 The device includes: a processor and a memory, wherein program instructions are stored in the memory, and the processor calls the program instructions stored in the memory to enable the device to execute the following method steps in Example 1:

[0121] A three-layer graph convolutional neural network is used to update features on multiple physical fields, mining the relationships between different physical fields to update the overall feature representation with comprehensive characteristics of different physical fields;

[0122] By using the double-layer temporal attention operation, the overall feature representation with different comprehensive characteristics of physical fields and the position encoding information corresponding to the overall feature representation are input into the double-layer temporal attention to obtain the coupled representation of multi-physical field data features in the time period;

[0123] The coupled representation of multi-physical field data features in different time periods is used as global features, and the time series information is mined to predict the Nino3.4 index. The Nino3.4 index is fused with the input typhoon observation data to predict the typhoon intensity.

[0124] Among them, a three-layer graph convolutional neural network is used to update features on multiple physical fields, and the relationship between different physical fields is mined to update the overall feature representation with comprehensive characteristics of different physical fields. Specifically:

[0125] The first-layer graph convolutional neural network is used to update the features on the spatial sub-region structure of a single ocean physical field, and to mine the mutual correlation between the spatial sub-regions of a single ocean physical field;

[0126] For multiple ocean physical fields, the second and third layers of graph convolutional neural networks are used to couple and associate the shallow and deep information on multiple physical fields to learn the overall feature representation with comprehensive characteristics of different physical fields.

[0127] Furthermore, the first-layer graph convolutional neural network is used to perform feature update on the spatial sub-region structure of a single ocean physical field, and to mine the mutual correlation between the spatial sub-regions of a single ocean physical field, specifically:

[0128] For the information of a single physical field, a convolutional neural network is used to extract sub-region features, which are recorded as And the sub-region features are used as graph structure nodes Edge information E between sub-regions, single physical field graph structure G 1 The adjacency matrix of is:

[0129]

[0130] in, n is the number of input sub-region features, represents the Euclidean distance between the feature vectors of two sub-regions in the i-th frame of the physical field x, Normalization(·) represents the normalization function, x_a represents the a-th sub-region of the physical field x, and x_b represents the b-th sub-region of the physical field x.

[0131] Among them, for multiple ocean physical fields, the second and third layers of the graph convolutional neural network are used to couple and associate the shallow and deep information on multiple physical fields to learn and obtain the overall feature representation with comprehensive characteristics of different physical fields. Specifically:

[0132] The second-layer graph structure is a shallow multi-physical field relationship fusion layer. The nodes of this layer of graph structure are consistent with the updated nodes of the first-layer graph structure. The difference between the second-layer graph structure and the first-layer is that the edge information corresponding to each physical place is exchanged;

[0133] The third-layer graph structure is a deep multi-physical field relationship fusion layer. The nodes of this layer of graph structure are obtained by splicing the nodes obtained by updating the x physical fields in the second layer, and the side information of the third-layer graph structure is also obtained by summing and averaging the side information of the second layer. Through the deep physical information coupling of the third-layer graph structure, the correlation between the physical fields is aggregated into an overall feature representation, which serves as the time series input of all physical field couplings on the i-th frame.

[0134] Furthermore, the two-layer temporal attention is specifically:

[0135] The first layer of the two-layer temporal attention is the time period information fusion self-attention layer, and the second layer predicts by extracting the time period coupled physical field information sequence; after the multi-physical field information is aggregated and updated by the graph convolutional neural network, the final coupling correlation features of different physical fields at the same longitude and latitude on the same frame are obtained; the coupling correlation features containing temporal information are divided into a group of M features as the input of the time period fusion information layer;

[0136] The second layer of temporal self-attention makes predictions by extracting the time period coupled physical field information sequence, and its input is the time period coupled correlation features output by the first layer.

[0137] Among them, the spatial sub-region of a single ocean physical field is:

[0138] The size of each frame of time series data is H×W. Each frame of data is divided into equal parts according to length and width, and the spatial sub-regions are selected with equal sizes based on the geographic center. For the x-th physical field, the i-th frame is divided into n spatial sub-regions, which are recorded as It is the data of the nth spatial sub-region in the i-th frame of the x-th physical field.

[0139] It should be pointed out here that the device description in the above embodiment corresponds to the method description in the embodiment, and the embodiment of the present invention will not be described in detail here.

[0140] The execution subjects of the above-mentioned processor 1 and memory 2 can be devices with computing functions such as computers, single-chip microcomputers, and microcontrollers. In specific implementation, the embodiments of the present invention do not limit the execution subjects and are selected according to the needs of actual applications.

[0141] The data signal is transmitted between the memory 2 and the processor 1 via the bus 3, which will not be described in detail in the embodiment of the present invention.

[0142] Based on the same inventive concept, an embodiment of the present invention further provides a computer-readable storage medium, the storage medium includes a stored program, and when the program is running, the device where the storage medium is located is controlled to execute the method steps in the above embodiment.

[0143] The computer-readable storage medium includes but is not limited to a flash memory, a hard disk, a solid-state drive, and the like.

[0144] It should be pointed out here that the description of the readable storage medium in the above embodiment corresponds to the description of the method in the embodiment, and the embodiment of the present invention will not be described in detail here.

[0145] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of the present invention are generated.

[0146] The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions may be stored in a computer-readable storage medium or transmitted via a computer-readable storage medium. The computer-readable storage medium may be any available medium that can be accessed by the computer or a data storage device such as a server or a data center that includes one or more available media. The available medium may be a magnetic medium or a semiconductor medium, etc.

[0147] References

[0148] [1] Yu Qianqian, Xie Dongmei, Chen Yongping, Zhu Ye. A preliminary study on the relationship between tropical cyclones affecting China's coastal areas and ENSO changes from 1979 to 2019[J]. Marine Bulletin, 2022, 41(01): 29-38.

[0149] [2]Mu B, Ma S, Yuan S, et al.Applying convolutional LSTM network toppredict El events:Transfer learning from the data of dynamical model and observation[C] / / 2020IEEE 10th International Conference on ElectronicsInformation and Emergency Communication(ICEIEC).IEEE,2020:215-219.

[0150] [3]Ham YG,Kim JH,Luo J J.Deep learning for multi-year ENSOforecasts[J].Nature,2019,573(7775):568-572.

[0151] [4]Defferrard M, Bresson

[0152] [5]Vaswani A,Shazeer N,Parmar N,et al.Attention is all you need[J].Advances in neural information processing systems,2017,30.

[0153] Those skilled in the art can understand that the accompanying drawing is only a schematic diagram of a preferred embodiment, and the serial numbers of the embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments.

[0154] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An early warning method based on multi-ocean physical field fusion representation based on graph convolution, It is characterized in that The method comprises: A three-layer graph convolutional neural network is used to update features on multiple physical fields, mining the relationships between different physical fields to update the overall feature representation with comprehensive characteristics of different physical fields; By using the double-layer temporal attention operation, the overall feature representation with different comprehensive characteristics of physical fields and the position encoding information corresponding to the overall feature representation are input into the double-layer temporal attention to obtain the coupled representation of multi-physical field data features in the time period; The coupled representation of multi-physics field data features in a time period is used as a global feature, and the time series information is mined to predict the Nino3.4 index. The Nino3.4 index is fused with the input typhoon observation data to predict the typhoon intensity. The method of using a three-layer graph convolutional neural network to perform feature updates on multiple physical fields and to explore the relationships between different physical fields so as to update the overall feature representation with comprehensive characteristics of different physical fields is as follows: The first-layer graph convolutional neural network is used to update the features on the spatial sub-region structure of a single ocean physical field, and to mine the mutual correlation between the spatial sub-regions of a single ocean physical field; For multiple ocean physical fields, the second and third layers of graph convolutional neural networks are used to couple and correlate the shallow and deep information on multiple physical fields to learn the overall feature representation with comprehensive characteristics of different physical fields.

2. According to claim 1, a warning method based on multi-ocean physical field fusion representation based on graph convolution, It is characterized in that The first-layer graph convolutional neural network is used to perform feature update on the spatial sub-region structure of a single ocean physical field, and to mine the mutual correlation between the spatial sub-regions of a single ocean physical field, specifically: For the information of a single physical field, a convolutional neural network is used to extract sub-region features, which are recorded as And the sub-region features are used as graph structure nodes Edge information E between sub-regions, single physical field graph structure G 1 The adjacency matrix of is: in, n is the number of input sub-region features, represents the Euclidean distance between the feature vectors of two sub-regions in the i-th frame of the physical field x, Normalization(·) represents the normalization function, x_a represents the a-th sub-region of the physical field x, and x_b represents the b-th sub-region of the physical field x.

3. According to claim 1, the early warning method based on multi-ocean physical field fusion representation of graph convolution, It is characterized in that For multiple ocean physical fields, the second and third layers of the graph convolutional neural network are used to couple and associate the shallow and deep information on multiple physical fields to learn and obtain the overall feature representation with different comprehensive characteristics of physical fields. Specifically, The second-layer graph structure is a shallow multi-physical field relationship fusion layer. The nodes of this layer of graph structure are consistent with the updated nodes of the first-layer graph structure. The difference between the second-layer graph structure and the first-layer is that the edge information corresponding to each physical place is exchanged; The third-layer graph structure is a deep multi-physical field relationship fusion layer. The nodes of this layer of graph structure are obtained by splicing the nodes obtained by updating the x physical fields in the second layer, and the side information of the third-layer graph structure is also obtained by summing and averaging the side information of the second layer. Through the deep physical information coupling of the third-layer graph structure, the correlation between the physical fields is aggregated into an overall feature representation, which serves as the time series input of all physical field couplings on the i-th frame.

4. According to claim 1, the early warning method based on multi-ocean physical field fusion representation of graph convolution, It is characterized in that The dual-layer temporal attention is specifically: The first layer of the dual-layer temporal attention is a time period information fusion self-attention layer, and the second layer predicts by extracting the time period coupled physical field information sequence; after the multi-physical field information is aggregated and updated by the graph convolutional neural network, the final coupling correlation features of different physical fields at the same longitude and latitude on the same frame are obtained; The coupled correlation features containing time series information are divided into a group of M features as the input of the time period fusion information layer; The second layer of temporal self-attention makes predictions by extracting the time period coupled physical field information sequence, and its input is the time period coupled correlation features output by the first layer.

5. According to claim 1, the early warning method based on multi-ocean physical field fusion representation of graph convolution, It is characterized in that The spatial sub-region of the single ocean physical field is: The size of each frame of time series data is H×W. Each frame of data is divided into equal parts according to length and width, and the spatial sub-regions are selected with equal sizes based on the geographic center. For the x-th physical field, the i-th frame is divided into n spatial sub-regions, which are recorded as It is the data of the nth spatial sub-region in the i-th frame of the x-th physical field.

6. An early warning device based on multi-ocean physical field fusion representation based on graph convolution, It is characterized in that The early warning device comprises: The overall feature update module is used to update features on multiple physical fields using a three-layer graph convolutional neural network, mining the relationships between different physical fields to update the overall feature representation with comprehensive characteristics of different physical fields; A feature coupling representation module is used to use a double-layer temporal attention operation to input the overall feature representation with different physical field comprehensive characteristics and the position encoding information corresponding to the overall feature representation into the double-layer temporal attention to obtain a feature coupling representation of multi-physical field data in a time period; The typhoon intensity prediction module is used to express the coupled characteristics of multi-physical field data in a time period as a global feature, and to mine time series information to predict the Nino3.4 index, and to fuse the Nino3.4 index with the input typhoon observation data to predict the typhoon intensity; The three-layer graph convolutional neural network is used to perform feature updates on multiple physical fields, and the relationships between different physical fields are mined to update the overall feature representation with comprehensive characteristics of different physical fields. Specifically, The first-layer graph convolutional neural network is used to update the features on the spatial sub-region structure of a single ocean physical field, and to mine the mutual correlation between the spatial sub-regions of a single ocean physical field; For multiple ocean physical fields, the second and third layers of graph convolutional neural networks are used to couple and correlate the shallow and deep information on multiple physical fields to learn the overall feature representation with comprehensive characteristics of different physical fields.

7. An early warning device based on multi-ocean physical field fusion representation based on graph convolution, It is characterized in that The device comprises: a processor and a memory, wherein program instructions are stored in the memory, and the processor calls the program instructions stored in the memory to enable the device to execute the method steps described in any one of claims 1-5.

8. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor is caused to perform the method steps described in any one of claims 1 to 5.

Images