Ocean heat wave prediction method based on Transform network and graph neural network

Abstract

The invention relates to the technical field of ocean heat wave prediction, and discloses an ocean heat wave prediction method based on a Transform network and a graph neural network, and the method comprises the steps: obtaining the ocean meteorological data of a to-be-measured ocean region; extracting long-period features and short-period features of the marine meteorological data by using a Transform network and a time convolution network respectively; performing fusion processing on the long-period features and the short-period features to obtain fusion features; and sequentially inputting the fusion features into a graph neural network and a time convolutional network for prediction processing to obtain an ocean heat wave prediction result. According to the method, the space-time correlation and dependency of the marine meteorological data under different time granularities can be better captured, the dynamic change of the time sequence is captured, the spatial modeling capability of the graph neural network is fully utilized to predict the marine heat waves, the prediction accuracy is greatly improved, and the calculation cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of ocean heat wave prediction, and more specifically, to an ocean heat wave prediction method based on a Transformer network and a graph neural network. Background Art

[0002] Marine heatwaves are discrete abnormally warm water events, described by characteristics such as duration, intensity, evolution speed, and spatial range. In recent years, global warming has caused marine heatwave events to become more frequent and extreme, which has had a serious impact on marine ecosystems and the services they provide. Effective prediction of marine heatwaves is of great significance to reducing damage to marine ecosystems.

[0003] At present, marine heat wave prediction methods are mainly divided into two categories: those based on numerical models and those based on data-driven technologies. Numerical models rely on initial conditions, boundary conditions, and kinetic and thermodynamic equations, but the generation and change mechanisms of seawater parameters are difficult to describe clearly, which makes model construction complicated. Data-driven methods use the correlation of spatiotemporal data for prediction. However, existing technologies still have the defects of low accuracy and high computational cost when processing high-dimensional, spatiotemporally correlated and limited-range data. Summary of the invention

[0004] In order to overcome the defects of low accuracy and high computational cost in the existing marine heat wave prediction technology, the present invention proposes the following technical solutions:

[0005] In a first aspect, the present invention proposes a method for predicting ocean heat waves based on a Transformer network and a graph neural network, comprising:

[0006] Obtain the marine meteorological data of the ocean area to be measured.

[0007] The Transformer network and the temporal convolutional network are used to extract the long-term characteristics and short-term characteristics of the marine meteorological data respectively.

[0008] The long-period features and the short-period features are fused to obtain fused features.

[0009] The fusion features are sequentially input into the graph neural network and the temporal convolutional network for prediction processing to obtain the ocean heat wave prediction results.

[0010] As a preferred technical solution, the location information is embedded into the marine meteorological data using a location embedding method, and its expression is as follows:

[0011]

[0012] in, represents the marine meteorological data with location information, N represents the batch size, which means the number of samples processed simultaneously in each training iteration, and T input and D input Respectively represent the length and dimension of the input marine meteorological data X, D PE It represents the dimension of the location information PE, and concat(X,PE) represents the concatenation operation of the input marine meteorological data X and the location information PE. K represents the time step of an element in the sequence, and period represents the pre-set period.

[0013] The Transformer network is used to extract long-term features of marine meteorological data with location information.

[0014] As a preferred technical solution, the long-period feature and the short-period feature are fused to obtain a fused feature, the expression of which is as follows:

[0015]

[0016] Among them, F i represents the fusion feature of the i-th one, represents the i-th long-term feature extracted by the encoder of the Transformer network, T i represents the i-th short-term feature extracted by the temporal convolutional network, Fusion(·) represents the fusion process, Attention(·) is the attention function, and Y i Reason With T i The feature representation obtained by the concatenation operation concat(·).

[0017] As a preferred technical solution, the fusion features are sequentially input into the graph neural network and the temporal convolutional network for prediction processing to obtain the marine heat wave prediction results, including:

[0018] Extract fusion feature F through graph neural network i The spatial characteristics S in i , whose expression is as follows:

[0019] S i =GNN(F i ,A η )

[0020] Among them, GNN(·) represents the execution of graph neural network operations, A η is the adjacency matrix of the graph neural network.

[0021] According to the spatial features, the temporal convolutional network operation is performed to obtain the ocean heat wave prediction result, which is expressed as follows:

[0022] G i =TCN(S i ).

[0023] As a preferred technical solution, the steps of constructing the adjacency matrix of the graph neural network include:

[0024] Each latitude and longitude point of the ocean area to be measured is a spatial node, and the ocean area to be measured is divided into several sub-areas.

[0025] The adjacency matrix of adjacent spatial nodes and the adjacency matrix of adjacent sub-regions are constructed respectively, and the adjacency matrix of adjacent spatial nodes and the adjacency matrix of adjacent sub-regions are used to form the adjacency matrix of the graph neural network.

[0026] As a preferred technical solution, constructing an adjacency matrix of adjacent spatial nodes includes:

[0027] Compare whether the ocean meteorological data change trends of adjacent spatial nodes within a specific time step meet the preset conditions. If so, it is determined that the adjacent spatial nodes are associated.

[0028] Using the associated spatial nodes, an adjacency matrix of adjacent spatial nodes is constructed.

[0029] As a preferred technical solution, an adjacency matrix of adjacent sub-regions is constructed.

[0030] Compare whether the changing trends of the ocean meteorological data of adjacent sub-areas within a specific time step meet the preset conditions. If so, it is determined that the adjacent sub-areas are related.

[0031] Using the associated sub-regions, an adjacency matrix of adjacent sub-regions is constructed.

[0032] As a preferred technical solution, after obtaining the adjacency matrix of the graph neural network, the method further includes optimizing the adjacency matrix of the graph neural network, including:

[0033] Perform linear operations on the initial adjacency matrix A of the graph neural network to obtain the adjacency matrix A μ , whose expression is as follows:

[0034] A μ =ω μ ·A+b μ

[0035] Among them, ω μ is the preset weight matrix, b μ is the preset bias.

[0036] For the adjacency matrix A μAfter cosine transform, linear operation is performed to obtain the final adjacency matrix A η , whose expression is as follows:

[0037] A η =ω η ·cos(A μ )+b η

[0038] Where cos(·) represents cosine transform, ω η is the preset weight matrix, b η is the preset bias.

[0039] As a preferred technical solution, after constructing the adjacency matrix of the graph neural network, the method further includes:

[0040] The marine meteorological data carried by each sub-region is obtained as the marine meteorological data input to the graph neural network. The expression is as follows:

[0041]

[0042] Among them, X B represents the marine meteorological data of the sub-area associated with sub-area A, x i and d i Respectively represent the i-th spatial node data and the total number of spatial nodes in the associated sub-area, N A represents the set of sub-regions associated with sub-region A, is the cosine similarity corresponding to each associated sub-region, Represents the marine meteorological data that is finally input to the graph neural network.

[0043] As a preferred technical solution, the fusion features are sequentially input into the graph neural network and the time convolution network for prediction processing, and the expression of the objective function L of the ocean heat wave prediction result is obtained as follows:

[0044] L=λ1L GNN +λ2L Trans

[0045] L GNN =argmin||S i -S target ||

[0046]

[0047] Among them, λ1 and λ2 are used to balance L GNN With L Trans The hyperparameter, L GNN is the loss function of the graph neural network, L Transis the loss function of the decoder of the Transformer network, argmin represents the minimization operation, ||S i -S target || indicates that the output result S of the graph neural network is measured using the L1 norm i and target result S target The error between Indicates the use of L1 norm to measure the output of the i-th decoder and the target result T target The error between .

[0048] The beneficial effects of the present invention include at least:

[0049] The present invention provides a method for predicting ocean heat waves based on a Transformer network and a graph neural network. The long-period features and short-period features of ocean meteorological data are learned respectively by the Transformer network and the time convolutional network, and the two features are fused, so as to better capture the spatial-temporal correlation and dependency at different time granularities, thereby realizing medium- and long-term ocean heat wave predictions. The fused features are first input into the graph neural network, and the graph neural network is used to effectively model complex spatial correlation relationships. Then, the features processed by the graph neural network are input into the time convolutional network for prediction processing to obtain the ocean heat wave prediction results, which can not only capture the dynamic changes of the time series, but also make full use of the spatial modeling ability of the graph neural network, thereby greatly improving the accuracy of the prediction. Compared with using only a single deep learning model, this fusion method also greatly reduces the computational cost and improves the prediction efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a schematic flow chart of the ocean heat wave prediction method based on the Transformer network and graph neural network provided in Example 1.

[0051] Figure 2 This is a schematic diagram of the principle architecture of the ocean heat wave prediction method based on the Transformer network and graph neural network provided in Example 2.

[0052] Figure 3 This is a schematic diagram of the principle of arranging time series in different periods according to Example 2.

[0053] Figure 4 This is a schematic diagram of the structure of the Transformer network provided in Example 2.

[0054] Figure 5 (a) is a schematic diagram of dividing the ocean area to be measured into sub-areas provided in Example 2, Figure 5(b) is a schematic diagram showing the relationship between the spatial nodes of each sub-region provided in Example 2.

[0055] Figure 6 The distribution diagram of the duration of ocean heat waves at different time steps provided in Example 3.

[0056] Figure 7 This is a distribution diagram of the average intensity of ocean heat waves at different time steps provided in Example 3.

[0057] Figure 8 This is a comparison chart of different models provided in Example 3 at different prediction steps and the original data. DETAILED DESCRIPTION

[0058] The following will describe the embodiments of the present invention with reference to the accompanying drawings and preferred technical solutions. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred technical solutions are only for illustrating the present invention, not for limiting the scope of protection of the present invention.

[0059] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0060] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it is obvious to those skilled in the art that the embodiments of the present invention can be implemented without these specific details. In other embodiments, well-known structures and devices are shown in the form of block diagrams rather than in detail to avoid making the embodiments of the present invention difficult to understand.

[0061] Example 1

[0062] This embodiment proposes a method for predicting ocean heat waves based on a Transformer network and a graph neural network. Figure 1 As shown, Figure 1 A schematic diagram of a process flow of a method for predicting ocean heat waves based on a Transformer network and a graph neural network provided in this embodiment, the method comprising the following steps:

[0063] S1: Obtain the marine meteorological data of the ocean area to be measured.

[0064] S2: The long-period features and short-period features of the marine meteorological data are extracted using a Transformer network and a temporal convolutional network respectively.

[0065] S3: Fusing the long-period features and the short-period features to obtain fused features.

[0066] S4: Input the fusion features into the graph neural network and the temporal convolutional network in sequence for prediction processing to obtain the ocean heat wave prediction results.

[0067] As an example, the marine heat wave forecast results include indicators such as the occurrence time, duration and average intensity of marine heat waves in a certain period of time in the future. These results can intuitively reflect the temporal and spatial distribution of marine heat waves.

[0068] It is understandable that by learning the long-term and short-term features of marine meteorological data through the Transformer network and the time convolution network respectively, and fusing these two features, the space-time correlation and dependence at different time granularities can be better captured, thereby realizing medium- and long-term marine heat wave predictions. The fused features are first input into the graph neural network, and the graph neural network is used to effectively model complex spatial correlations. Then, the features processed by the graph neural network are input into the time convolution network for prediction processing to obtain the marine heat wave prediction results, which can not only capture the dynamic changes of the time series, but also make full use of the spatial modeling capabilities of the graph neural network, thereby greatly improving the accuracy of the prediction. Compared with using only a single deep learning model, this fusion method also greatly reduces the computational cost and improves the prediction efficiency. This embodiment verifies the effectiveness of the method of this embodiment by conducting simulation experiments using real data from the South China Sea.

[0069] Example 2

[0070] like Figure 2 As shown, this embodiment makes improvements on the ocean heat wave prediction method based on Transformer network and graph neural network proposed in the first embodiment, as shown in FIG. Figure 2 As shown, Figure 2 This is a principle framework diagram of the ocean heat wave prediction method based on the Transformer network and graph neural network provided in this embodiment.

[0071] It should be noted that long-term marine heat wave (MHW) prediction is a fundamental and thorny problem in current meteorological research. To solve this problem, this embodiment proposes a Transformer network that combines multiple time periods to extract long-term temporal relationships from different time periods. Based on the usual period, this embodiment extracts the corresponding "weekly" period, "monthly" period and "yearly" period. In order to capture the long-term and short-term periodic semantic information, the input sequence is rearranged according to their natural period, such as Figure 3 As shown, Figure 3 The left side shows the sorting status of a time series by day, and the right side shows the sorting status of the data at the same time point in different periods, including a weekly period, a monthly period, and a annual period.

[0072] In this embodiment, the position information is embedded into the marine meteorological data using a position embedding method, and its expression is as follows:

[0073]

[0074] in, represents the marine meteorological data with location information, N represents the batch size, which means the number of samples processed simultaneously in each training iteration, and T input and D input Respectively represent the length and dimension of the input marine meteorological data X, D PE It represents the dimension of the location information PE, and concat(X,PE) represents the concatenation operation of the input marine meteorological data X and the location information PE. K represents the time step of an element in the sequence, and period represents the pre-set period. It is connected with the sequence data through position embedding so that the model can obtain the position information of the data.

[0075] The Transformer network is used to extract long-term features of marine meteorological data with location information.

[0076] As an exemplary illustration, Figure 4 As shown, Figure 4 This is a schematic diagram of the structure of the Transformer network of this embodiment. This embodiment uses a Transformer network to extract long-term temporal relationships from data. The Transformer network consists of 3 encoders and 3 corresponding decoders. Each encoder consists of four main components: a multi-head attention mechanism, a fully connected feedforward network, layer normalization, and a residual connection. The expression of the i-th Transformer encoder is as follows:

[0077]

[0078] in, represents the output of the i-th Transformer encoder, NL(·) represents the normalization operation, Dropout(·) represents the regularization operation, FF(·) and Res represent the feedforward network and residual connection, respectively.

[0079] In this embodiment, a short-term feature is extracted through a temporal convolutional network, and its expression is as follows:

[0080] T i =TCN(T i-1 )

[0081] Among them, T i-1 represents the feature input of the i-th layer, T i Indicates the input result.

[0082] It should be noted that this embodiment uses a long-short cycle information fusion module Fusion(·) to achieve better information expression, and this module achieves fusion through the attention mechanism Attention(·).

[0083] In this embodiment, the long-period feature and the short-period feature are fused to obtain a fused feature, the expression of which is as follows:

[0084]

[0085] Among them, F i represents the fusion feature of the i-th one, represents the i-th long-term feature extracted by the encoder of the Transformer network, T i represents the i-th short-term feature extracted by the temporal convolutional network, Fusion(·) represents the fusion process, Attention(·) is the attention function, and Y i Reason With T i The feature representation obtained by the concatenation operation concat(·).

[0086] In this embodiment, the fusion features are sequentially input into the graph neural network and the temporal convolutional network for prediction processing to obtain the ocean heat wave prediction results, including:

[0087] Extract fusion feature F through graph neural network i The spatial characteristics S in i , whose expression is as follows:

[0088] S i =GNN(F i ,A η )

[0089] Among them, GNN(·) represents the execution of graph neural network operations, A η is the adjacency matrix of the graph neural network.

[0090] According to the spatial features, the temporal convolutional network operation is performed to obtain the ocean heat wave prediction result, which is expressed as follows:

[0091] G i =TCN(S i ).

[0092] In this embodiment, the steps of constructing the adjacency matrix of the graph neural network include:

[0093] Each latitude and longitude point of the ocean area to be measured is a spatial node, and the ocean area to be measured is divided into several sub-areas.

[0094] The adjacency matrix of adjacent spatial nodes and the adjacency matrix of adjacent sub-regions are constructed respectively, and the adjacency matrix of adjacent spatial nodes and the adjacency matrix of adjacent sub-regions are used to form the adjacency matrix of the graph neural network.

[0095] In this embodiment, constructing an adjacency matrix of adjacent spatial nodes includes:

[0096] Compare whether the ocean meteorological data change trends of adjacent spatial nodes within a specific time step meet the preset conditions. If so, it is determined that the adjacent spatial nodes are associated.

[0097] Using the associated spatial nodes, an adjacency matrix of adjacent spatial nodes is constructed.

[0098] In this embodiment, in order to reduce the computational cost, for each spatial node a, the nearest k spatial nodes P are selected. a As its neighbors, as shown below:

[0099] P a =argtop-k(d(a,P))

[0100] Where P represents the set of other spatial nodes connected to spatial node a, d(a,P) represents the distance between spatial node a and other spatial nodes, and argtop-k(·) represents the selection of the top k spatial nodes with the smallest distance. While retaining the weights of connected spatial nodes, the non-connected spatial nodes -P a The weight A[a,-P a ] is set to zero, as shown in the following formula:

[0101] A[a,-P a ]=0

[0102] In this embodiment, an adjacency matrix of adjacent sub-regions is constructed.

[0103] Compare whether the changing trends of the ocean meteorological data of adjacent sub-areas within a specific time step meet the preset conditions. If so, it is determined that the adjacent sub-areas are related.

[0104] Using the associated sub-regions, an adjacency matrix of adjacent sub-regions is constructed.

[0105] It is understandable that, in terms of spatial information, for the sea temperature monitoring data of the South China Sea, each longitude and latitude is regarded as a spatial node, so that information can be transmitted through the relationship between nodes. When constructing the adjacency matrix required for the graph neural network, this embodiment introduces co-occurrence changes as a basis for construction. By tracking the changes of two spatial nodes (except land nodes because they have no data records) within the same time step, it is determined whether there is a correlation between them. Specifically, set within a quarter (3 months), if the change trend of spatial node a is similar to that of node b, it is determined that there is a correlation. It is worth noting that the change trend here can be two similar positive growths or negative growths, or a positive growth and a negative growth. If a pair of data has never co-occurred in training, no associated edges will be constructed for them. In this way, an initial adjacency matrix is ​​obtained, which is then converted into the final required adjacency matrix through linear calculation, cosine similarity and linear scaling.

[0106] In this embodiment, after obtaining the adjacency matrix of the graph neural network, the method further includes optimizing the adjacency matrix of the graph neural network, including:

[0107] Perform linear operations on the initial adjacency matrix A of the graph neural network to obtain the adjacency matrix A μ , whose expression is as follows:

[0108] A μ =ω μ ·A+b μ

[0109] Among them, ω μ is the preset weight matrix, b μ is the preset bias.

[0110] For the adjacency matrix A μ After cosine transform, linear operation is performed to obtain the final adjacency matrix A η , whose expression is as follows:

[0111] A η =ω η ·cos(A μ )+b η

[0112] Where cos(·) represents cosine transform, ω η is the preset weight matrix, b η is the preset bias.

[0113] In this embodiment, in order to better utilize spatial information, the ocean area to be measured is divided into multiple sub-areas at equal intervals, such as Figure 5 As shown, Figure 5 The schematic diagram of dividing the ocean area to be measured provided in this embodiment is as follows: Figure 5 (a) is a schematic diagram of dividing the ocean area to be measured into sub-areas in this embodiment, Figure 5 (b) is a schematic diagram showing the relationship between the spatial nodes of each sub-region. The ocean area to be tested is divided into 6*6 sub-regions. There is a certain correlation between these sub-regions, and there is a certain correlation between the nodes in each sub-region. Figure 5 (b) In each sub-region, the mean temperature of all points in the sub-region is first calculated. If all points in a sub-region are land, the sub-region does not participate in the subsequent calculation. Next, it is determined whether there is co-occurrence between the mean temperature changes of different sub-regions, and based on this, an adjacency matrix between sub-regions with weight coefficients is constructed, as shown in Figure 5 (a). This adjacency matrix describes the spatial correlation between different sub-regions. When performing graph neural network calculations, each sub-region carries not only its own feature information, but also the information of other sub-regions related to it. This ensures that the spatial correlation of the entire region can be fully considered during modeling, which helps to improve the accuracy of the prediction. This sub-region-based approach can not only make better use of spatial information, but also greatly reduce the computational complexity. Compared with directly performing graph neural network calculations across the entire region, this partitioning method can significantly improve computational efficiency and provide feasibility for practical applications. The marine meteorological data carried by each sub-region is used as the marine meteorological data input to the graph neural network, and its expression is as follows:

[0114]

[0115] Among them, X B represents the marine meteorological data of the sub-area associated with sub-area A, x i and d i Respectively represent the i-th spatial node data and the total number of spatial nodes in the associated sub-area, N A represents the set of sub-regions associated with sub-region A, is the cosine similarity corresponding to each associated sub-region, Represents the marine meteorological data that is finally input to the graph neural network.

[0116] It should be noted that in order to make the prediction result G i Can get closer to the target data G target , the mean absolute error minimization is applied in the loss function of the graph neural network. Similarly, the output of the last Transformer decoder is With T target It is also necessary to keep the gap as small as possible to retain more temporal information.

[0117] In this embodiment, the fusion features are sequentially input into the graph neural network and the temporal convolutional network for prediction processing, and the expression of the objective function L of the ocean heat wave prediction result is obtained as follows:

[0118] L=λ1L GNN +λ2L Trans

[0119] L GNN =argmin||G i -G target ||

[0120]

[0121] Among them, λ1 and λ2 are used to balance L GNN With L Trans The hyperparameter, L GNN is the loss function of the graph neural network, L Trans is the loss function of the decoder of the Transformer network, argmin represents the minimization operation, ||G i -G target || indicates that the output result G of the graph neural network is measured using the L1 norm i and target result G target The error between Indicates the use of L1 norm to measure the output of the i-th decoder and the target result T target The error between .

[0122] Example 3

[0123] Based on Example 2, this example verifies the effectiveness of the ocean heat wave prediction method based on Transformer network and graph neural network of the present application by studying the South China Sea region.

[0124] In this embodiment, the selected ocean area to be measured is the South China Sea, and the selected data set is the OISST-V2 data set provided by the National Oceanic and Atmospheric Administration of the United States. The OISST-V2 data set is based on the sea surface temperature (SST) remotely sensed by the high-resolution radiometer (AVHRR), and the global high-resolution SST estimated by the optimal interpolation (OI) method. The resolution of the OISST data set is 0.25°×0.25°. ERA5 is the fifth generation of atmospheric reanalysis conducted by the European Center for Medium-Range Weather Forecasts. ERA5 provides estimates of atmospheric, land and ocean variables from 1979 to the present. The variables used in this embodiment include the temperature at 2 meters above the sea surface (T2M), the sea surface pressure (P), the 10-meter zonal wind (U10) and the 10-meter meridional wind (V10), where U and V describe the movement of the wind. The resolution of the ERA5 data set is 0.25°×0.25°. Table 1 summarizes the observation and reanalysis data used in this embodiment.

[0125] Table 1 Description of variables in observational and reanalysis data

[0126]

[0127]

[0128] In this embodiment, the mean variance normalization method is used to preprocess the training set of marine meteorological data, and its expression is as follows:

[0129]

[0130] Among them, X max is the maximum value of the sample, X min is the minimum value of the sample, X scale is the normalized value.

[0131] In this embodiment, the mean absolute error (MAE) and the root mean square error (RMSE) are used, and their expressions are as follows:

[0132]

[0133] Among them, y i is the true value, Represents the predicted value.

[0134] SST is the basis for determining MHW. First, the gap between the predicted SST values ​​and the true values ​​at different time steps is compared. At the same time, on this basis, a single variable mode experiment is added for comparison. Specifically, the single variable prediction mode refers to using only one meteorological variable data as input, keeping other operations unchanged, and performing multi-step predictions. This paper selects the basic variable SST of MHW for single variable prediction, and the experimental results are shown in Tables 2 and 3.

[0135] Table 2 MAE and RMSE (single variable) at different time steps of the present invention

[0136]

[0137] Table 3 MAE and RMSE (multivariate) at different time steps of the present invention

[0138]

[0139] In this embodiment, in order to more intuitively evaluate the prediction effect of the model at different time steps, the prediction results are converted into a distribution diagram of the duration and average intensity of MHW (marine heat wave) for display. Figure 6 The distribution of MHW duration at different time steps is shown. The vertical direction is the distribution effect of the number of days of each method at different time steps, and the horizontal direction is the comparison of different methods at the same time step. Figure 7 The distribution of the average intensity of MHW at different time steps is shown in the figure. The vertical direction is the average intensity distribution predicted by the same method at different time steps. The horizontal comparison is the average intensity distribution predicted by different methods at the same time step. Specifically, the predicted data replaces the original data, and then the Hobday method is used to determine the duration of the marine heat wave. Taking the 10-day time step as an example, the prediction results include 10 days of data from June 30, 2022 to July 9, 2022. On the basis of retaining the original data before June 29, 2022, the subsequent original data are replaced with the predicted 10-day data. Then, the occurrence and end of MHW are determined by calculating the threshold and climate state from 1982 to July 9, 2022. Finally, only the distribution map corresponding to the predicted days is shown, which can more clearly reflect the prediction effect without being obscured by the overall data. Through this visualization method, the performance of different prediction models at different time scales can be more intuitively compared. Figure 8The comparison between different models and original data at different prediction steps is shown. The location is 13.375°N, 100.125°E, starting from June 30, 2022. Four different methods are selected for prediction performance comparison, namely, the long short-term memory network (LSTM) based on recurrent neural network (RNN), the temporal convolutional network (TCN) based on convolutional neural network (CNN), the informer based on transformer, and the graph convolutional network (GCN) based on graph neural network. The black dotted line is the original data, the dark blue line is the prediction model constructed on the basis of the method proposed in this embodiment, the green line is the LSTM model, the red line is the TCN model, the light blue line is the informer model, and the purple line is the GCN model.

[0140] It can be understood that the present invention provides a method for predicting ocean heat waves based on Transformer networks and graph neural networks, which can effectively extract the long- and short-cycle characteristics of marine meteorological data and use these characteristics to achieve accurate prediction of ocean heat waves.

[0141] In terms of time information, the present invention sorts time series into different periods such as "week", "month", and "year", providing a basis for capturing long-term dependencies at different time granularities. The Transformer network can effectively extract deep correlations in the time dimension.

[0142] In terms of spatial information, the present invention divides the South China Sea region into multiple sub-regions at equal intervals, analyzes the multivariate data in these sub-regions, learns the changing trends and similarities between adjacent sub-regions, and thus establishes accurate spatial interdependencies. By introducing two different adjacency matrices, it is possible to ensure that while calculating local information, key information of the overall region can be focused on and integrated.

[0143] Finally, the graph neural network combined with the temporal convolutional network is used to analyze and predict the collected comprehensive information. The effectiveness of the invention is verified by simulation experiments using real data from the South China Sea, providing a new and valuable solution for predictive analysis in the fields of ocean and climate change.

[0144] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms are not necessarily directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0145] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "N" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0146] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present application belong.

[0147] It should be understood that the various parts of the present application can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiment, the N steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array, a field programmable gate array, etc.

[0148] A person skilled in the art may understand that all or part of the steps in the above-mentioned embodiment method may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0149] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A method for predicting marine heat waves based on Transformer network and graph neural network, characterized in that: include: Obtaining marine meteorological data of the ocean area to be measured; The long-period characteristics and short-period characteristics of the marine meteorological data are extracted using a Transformer network and a temporal convolutional network respectively; The long-period features and the short-period features are fused to obtain fused features; The fusion features are sequentially input into the graph neural network and the temporal convolutional network for prediction processing to obtain the ocean heat wave prediction results.

2. The method for predicting ocean heat waves based on Transformer network and graph neural network according to claim 1, characterized in that: The Transformer network is used to extract the long-term characteristics of the marine meteorological data, including: The location information is embedded into the marine meteorological data using the location embedding method, and its expression is as follows: in, represents the marine meteorological data with location information, N represents the batch size, which means the number of samples processed simultaneously in each training iteration, and T input and D input Respectively represent the length and dimension of the input marine meteorological data X, D PE represents the dimension of the location information PE, and concat(X,PE) represents the concatenation operation of the input marine meteorological data X and the location information PE; K represents the time step of an element in the sequence, and period represents the pre-set period; The Transformer network is used to extract long-term features of marine meteorological data with location information.

3. The method for predicting ocean heat waves based on Transformer network and graph neural network according to claim 1, characterized in that: The long-period features and the short-period features are fused to obtain fused features. Its expression is as follows: Among them, F i represents the fusion feature of the i-th feature, represents the i-th long-term feature extracted by the encoder of the Transformer network, T i represents the i-th short-term feature extracted by the temporal convolutional network, Fusion(·) represents the fusion process, Attention(·) is the attention function, and Y i Reason With T i The feature representation obtained by the concatenation operation concat(·).

4. The method for predicting ocean heat waves based on Transformer network and graph neural network according to claim 3, characterized in that: The fusion features are sequentially input into the graph neural network and the temporal convolutional network for prediction processing to obtain the ocean heat wave prediction results, including: Extract fusion feature F through graph neural network i The spatial characteristics S in i , whose expression is as follows: S i =GNN(F i ,A η ) Among them, GNN(·) represents the execution of graph neural network operations, A η is the adjacency matrix of the graph neural network; According to the spatial features, the temporal convolutional network operation is performed to obtain the ocean heat wave prediction result, which is expressed as follows: G i =TCN(S i )。 5. The method for predicting ocean heat waves based on Transformer network and graph neural network according to claim 4, characterized in that: The steps to construct the adjacency matrix of a graph neural network include: Each latitude and longitude point of the ocean area to be measured is a spatial node, and the ocean area to be measured is divided into several sub-areas; The adjacency matrix of adjacent spatial nodes and the adjacency matrix of adjacent sub-regions are constructed respectively, and the adjacency matrix of adjacent spatial nodes and the adjacency matrix of adjacent sub-regions are used to form the adjacency matrix of the graph neural network.

6. The method for predicting ocean heat waves based on Transformer network and graph neural network according to claim 5, characterized in that: Constructing the adjacency matrix of adjacent spatial nodes includes: Compare whether the change trend of the ocean meteorological data of adjacent spatial nodes within a specific time step meets the preset conditions. If so, it is determined that the adjacent spatial nodes are associated; Using the associated spatial nodes, an adjacency matrix of adjacent spatial nodes is constructed.

7. The method for predicting ocean heat waves based on Transformer network and graph neural network according to claim 5, characterized in that: Construct an adjacency matrix of adjacent sub-regions; Compare whether the change trend of the ocean meteorological data of adjacent sub-areas within a specific time step meets the preset conditions. If so, it is determined that the adjacent sub-areas are related; Using the associated sub-regions, an adjacency matrix of adjacent sub-regions is constructed.

8. The method for predicting ocean heat waves based on Transformer network and graph neural network according to claim 5, characterized in that: After obtaining the adjacency matrix of the graph neural network, the method further includes optimizing the adjacency matrix of the graph neural network, including: Perform linear operations on the initial adjacency matrix A of the graph neural network to obtain the adjacency matrix A μ , whose expression is as follows: A μ =ω μ ·A+b μ Among them, ω μ is the preset weight matrix, b μ is the preset bias; For the adjacency matrix A μ After cosine transform, linear operation is performed to obtain the final adjacency matrix A η , whose expression is as follows: A η =ω η ·cos(A μ )+b η Where cos(·) represents cosine transform, ω η is the preset weight matrix, b η is the preset bias.

9. The method for predicting ocean heat waves based on Transformer network and graph neural network according to claim 5, characterized in that: After constructing the adjacency matrix of the graph neural network, the method further includes: The marine meteorological data carried by each sub-region is obtained as the marine meteorological data input to the graph neural network. The expression is as follows: Among them, X B represents the marine meteorological data of the sub-area associated with sub-area A, x i and d i Respectively represent the i-th spatial node data and the total number of spatial nodes in the associated sub-area, N A represents the set of sub-regions associated with sub-region A, is the cosine similarity corresponding to each associated sub-region, Represents the marine meteorological data that is finally input to the graph neural network.

10. The method for predicting ocean heat waves based on Transformer network and graph neural network according to claim 5, characterized in that: The fusion features are sequentially input into the graph neural network and the temporal convolutional network for prediction processing, and the expression of the objective function L of the ocean heat wave prediction result is obtained as follows: L=λ1L GNN +λ2L Trans L GNN =argmin||S i -S target || Among them, λ1 and λ2 are used to balance L GNN With L Trans The hyperparameter, L GNN is the loss function of the graph neural network, L Trans is the loss function of the decoder of the Transformer network, argmin represents the minimization operation, ||S i -S target || indicates that the output result S of the graph neural network is measured using the L1 norm i and target result S target The error between Indicates the use of L1 norm to measure the output of the i-th decoder and the target result T target The error between .

Images