A data-driven method, apparatus, and medium for predicting ENSO multivariate three-dimensional fields

Abstract

The application discloses a data-driven ENSO multivariate three-dimensional field prediction method, equipment and medium, belongs to the field of artificial intelligence climate prediction technology, and the method comprises the steps of acquiring ENSO multivariate three-dimensional field data in a set time period, using an ENSO intelligent prediction model to rollingly predict control prediction three-dimensional field data in a future time period; acquiring a plurality of orthogonal disturbance basis modes of a strong El Nino trigger; judging whether the sea surface temperature anomaly field in the control prediction three-dimensional field data exceeds a set threshold value, if not, outputting the control prediction three-dimensional field data, if yes, constructing a disturbance sea surface temperature field, and using the same to predict three-dimensional field data in a future time period to obtain an ensemble prediction, and outputting the ensemble prediction and the control prediction three-dimensional field data. The application combines the CNOP ensemble prediction method with the ENSO intelligent prediction model, improves the prediction accuracy of extreme ENSO events, and overcomes the problems of non-development of random initial disturbance and insufficient dispersion of ensemble members in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence climate prediction technology, specifically relating to a data-driven ENSO multivariate three-dimensional field prediction method, device, and medium. Background Technology

[0002] El Niño-Southern Oscillation (ENSO) is an ocean-atmosphere coupling phenomenon occurring in the tropical Pacific Ocean and is the strongest interannual variability signal in the Earth system. El Niño and the Southern Oscillation describe the east-west out-of-phase oscillations of sea surface temperature (SST) and sea surface pressure in the tropical Pacific, respectively. They are different manifestations of the interannual variation of the tropical Pacific ocean-atmosphere system in the ocean and atmosphere, and are therefore collectively referred to as the El Niño-Southern Oscillation phenomenon. Although ENSO occurs in the tropical Pacific, it can have a significant impact on my country's weather and climate through atmospheric teleconnections and oceanic processes. Therefore, it is necessary to use climate models to conduct ENSO simulations and forecasts.

[0003] In recent years, data-driven artificial intelligence algorithms have provided a new paradigm for ENSO modeling and prediction due to their powerful nonlinear modeling capabilities. For example, traditional convolutional neural networks (CNNs) and recurrent neural networks (RNNs), with their inductive bias characteristics in spatial and sequential dimensions, have become benchmark models for spatiotemporal field prediction. However, due to the inherent limitations of traditional neural networks, research on multivariate high-dimensional spatiotemporal field prediction in the field of ENSO prediction remains limited.

[0004] Furthermore, since most ENSO intelligent prediction models are deterministic models trained under supervision, these models weaken the prediction accuracy for extreme values ​​in order to reduce the overall prediction loss. In the field of numerical models, there are already some methods that can be used to improve the prediction and simulation of extreme events. For example, ensemble forecasting methods are often used to reduce model forecast uncertainty and enhance the predictive ability for extreme events, but when transferred to deep learning models, they face problems such as the lack of development of ensemble perturbations and insufficient dispersion of ensemble members. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by providing a data-driven ENSO multivariate three-dimensional field prediction method, device, and medium. It combines ensemble forecasting methods with the ENSO intelligent prediction model to improve the accuracy of extreme predictions and overcome the problems of non-development of ensemble perturbations and insufficient dispersion of ensemble members in existing technologies.

[0006] This invention provides the following technical solution:

[0007] Firstly, a data-driven method for predicting ENSO multivariate three-dimensional fields is provided, including:

[0008] obtain ENSO multivariate three-dimensional field data of a starting time and a preceding period, the three-dimensional field data comprising: sea surface zonal wind stress anomaly field, meridional wind stress anomaly field, and sea temperature anomaly field within a set depth from the sea surface;

[0009] based on the preprocessed multivariate three-dimensional field data, use an ENSO intelligent prediction model to rollingly predict control prediction three-dimensional field data of a future period;

[0010] for a number of strong El Niño events observed in history, use a CNOP method to obtain a number of orthogonal disturbance basis modes triggering strong El Niño;

[0011] determine whether the sea temperature anomaly field in the control prediction three-dimensional field data exceeds a set threshold, if not, directly output the control prediction three-dimensional field data as a result, if yes, for each orthogonal disturbance basis mode, add it to the sea temperature anomaly field of the starting time to obtain a disturbed sea temperature field, and predict three-dimensional field data of the future period based on the disturbed sea temperature field to obtain an ensemble prediction, and output the ensemble prediction and the control prediction three-dimensional field data as a result together.

[0012] Optionally, the CNOP method is used to obtain a number of orthogonal disturbance basis modes triggering strong El Niño for a number of strong El Niño events observed in history, and specifically comprises:

[0013] Step a: for each strong El Niño event observed in history, generate a number of initial sea temperature disturbance fields satisfying energy constraints from climate model simulation results;

[0014] Step b: in the current disturbance space, use an ENSO intelligent prediction model to predict each initial sea temperature disturbance field, select the first M number of initial sea temperature disturbance fields with the maximum target function value, and put them into a disturbance candidate set;

[0015] Step c: constantly update the disturbance candidate set by multiple mean calculations and elimination of minimum values, and perform mean calculation on the last updated disturbance candidate set to obtain orthogonal disturbance basis modes of the current disturbance space and save them;

[0016] Step d: determine whether the number of disturbance samples in the saved orthogonal disturbance basis modes reaches a set number, if yes, stop and output all saved orthogonal disturbance basis modes, if not, use Schmidt projection method to subtract the projection of the initial sea temperature disturbance field on the orthogonal disturbance basis modes to update the disturbance space, and return to execute step b.

[0017] Optionally, in step a, the energy constraint condition of the initial sea temperature disturbance field is:

[0018] ; ;

[0019] wherein, is the initial sea surface temperature perturbation field, specifically the initial sea surface temperature anomaly field, is the initial sea surface temperature perturbation field, specifically the initial sea surface temperature anomaly field, is the initial sea surface temperature perturbation field, specifically the initial sea surface temperature anomaly field is obtained after energy constraint, is the initial sea surface temperature perturbation field, specifically the initial sea surface temperature anomaly field is the initial sea surface temperature perturbation field, specifically the initial sea surface temperature anomaly field is the initial sea surface temperature perturbation field, specifically the initial sea surface temperature anomaly field is the total energy of the initial sea surface temperature perturbation field, , and are the number of zonal, meridional grid points and the vertical layer number of the sea surface temperature anomaly field respectively, is the average sea surface temperature anomaly perturbation norm of all marine grid points.

[0020] Optionally, the objective function used for predicting each initial sea surface temperature perturbation field in step b is:

[0021] ;

[0022] wherein, is the initial sea surface temperature perturbation field, specifically the initial sea surface temperature anomaly field is the initial sea surface temperature perturbation field, specifically the initial sea surface temperature anomaly field is the initial sea surface temperature perturbation field, specifically the initial sea surface temperature anomaly field is the length of the predicted future time period, is the starting month of the prediction, and are the sea surface temperature anomaly values of the Niño3.4 region obtained by the perturbation experiment and the non-perturbation control experiment respectively at the target month of the prediction.

[0023] Optionally, step c specifically comprises the following steps:

[0024] Step c1: performing ensemble average and energy constraint on all perturbation fields in the perturbation candidate set to obtain an average perturbation field, and performing prediction by using the ENSO intelligent prediction model to obtain the objective function value of the average perturbation field;

[0025] Step c2: comparing the objective function values of the average perturbation field and all perturbation fields in the perturbation candidate set, discarding the perturbation field with the minimum objective function value, and updating the perturbation candidate set;

[0026] Step c3: judging whether the updating times of the perturbation candidate set reach a set threshold, if not, returning the updated perturbation candidate set to step c1 for processing, and if so, performing averaging on all perturbation fields in the last updated perturbation candidate set to obtain the orthogonal perturbation basis mode of the current perturbation space, and saving.

[0027] ​Optionally, in step d, the projection of the initial sea temperature disturbance field on the orthogonal disturbance basis mode is removed by using the Schmidt projection method, and specifically, the initial sea temperature disturbance field is updated by using the following formula:

[0028] ;

[0029] wherein, is the initial sea temperature disturbance field ; is the updated sea temperature disturbance field in the disturbance space is the orthogonal disturbance basis mode, ; represents the inner product of and ; represents the two-norm.

[0030] Optionally, the ENSO intelligent prediction model comprises an encoder preprocessing module, a decoder preprocessing module, an encoder, a decoder and an output layer.

[0031] The encoder preprocessing module embeds spatial position encoding and time position encoding after block processing and linear layer processing of model input data, to obtain embedded features; the model input data is ENSO multivariate three-dimensional field data of a starting time and a previous period.

[0032] The encoder is configured to analyze the embedded features based on a spatiotemporal attention algorithm, to capture the spatiotemporal relationship between each variable of the model input data, and to obtain multi-level features; the encoder comprises a plurality of encoding layers, each encoding layer comprises a multi-head spatiotemporal attention sublayer and a fully connected sublayer in sequence, and each sublayer is provided with a residual structure and layer normalization operation.

[0033] The decoder preprocessing module processes the model input data into a matrix composed of a vector sequence, and adds position encoding for processing by the decoder.

[0034] The decoder is configured to predict three-dimensional field data of a future period according to the multi-level features output by the encoder, and the decoder comprises a plurality of decoding layers, each decoding layer comprises a multi-head spatiotemporal attention sublayer, a temporal attention sublayer and a fully connected sublayer in sequence, and each sublayer is provided with a residual structure and layer normalization operation.

[0035] The output layer is connected to the fully connected sublayer of the last decoding layer of the decoder, and is configured to reorganize the three-dimensional field prediction data output by the decoder into the original spatial dimension after linear layer processing, to obtain the final predicted three-dimensional field data of the future period.

[0036] Optionally, the multi-head spatiotemporal attention sublayer specifically operates as follows:

[0037] The multi-head spatiotemporal attention sublayer input feature is split into several heads, the time attention weight of each head is obtained, and the split heads are spliced based on the time attention weight to obtain a first feature matrix; the first feature matrix is taken as input, and the several heads are split again, the spatial attention weight of each head is obtained, and the split heads are spliced based on the spatial attention weight to obtain a second feature matrix; the second feature matrix is connected in residual with the multi-head spatiotemporal attention sublayer input feature, and a layer normalization operation is performed to obtain an output feature matrix of the multi-head spatiotemporal attention sublayer;

[0038] The full connection sublayer of the encoding layer deeply integrates the output feature matrix of the multi-head spatiotemporal attention sublayer in the encoding layer, and outputs a feature matrix of the current encoding layer;

[0039] The time attention sublayer of the decoding layer, the query matrix comes from the output feature matrix of the multi-head spatiotemporal attention sublayer, and the key and value matrices come from the multi-level features output by the encoder; the time attention sublayer of the decoding layer splits the output feature matrix of the multi-head spatiotemporal attention sublayer into several heads and performs a time attention mechanism to obtain an encoder-decoder interaction attention matrix, performs multi-head splicing based on the encoder-decoder interaction attention matrix to obtain a third feature matrix, and obtains an output feature matrix of the current time attention sublayer after residual connection and layer normalization operation of the third feature matrix and the input feature of the time attention sublayer;

[0040] The full connection sublayer of the decoding layer deeply integrates the output feature matrix of the time attention sublayer in the decoding layer, and outputs a feature matrix of the current decoding layer.

[0041] In a second aspect, a computer device is provided, comprising a processor and a memory; wherein the processor implements the steps of the data-driven ENSO multivariate three-dimensional field prediction method according to any one of the first aspect when executing the computer program stored in the memory.

[0042] In a third aspect, a computer readable storage medium is provided for storing a computer program; the computer program is executed by a processor to implement the steps of the data-driven ENSO multivariate three-dimensional field prediction method according to any one of the first aspect.

[0043] Compared with the prior art, the beneficial effects of the present application are:

[0044] (1) The orthogonal disturbance basis mode triggering strong El Nino is obtained by using the CNOP method, the nonlinear development process of the extreme event can be directly considered for the specific extreme event, the captured disturbance is more in line with the energy development path in the real atmosphere / ocean, the probability distribution of the ensemble prediction is more accurate, the ensemble prediction result and the control prediction three-dimensional field data of the ENSO intelligent prediction model are output together, the problem that the traditional ENSO intelligent prediction model generally underestimates the extreme event can be greatly improved, and thus more reliable extreme event occurrence probability and risk warning can be provided.

[0045] (2) The ENSO intelligent prediction model of the application enhances the analysis and modeling capability of the time sequence dependence relationship and non-local spatial correlation through the multi-head spatiotemporal cross attention mechanism; and the coupling relationship between the ocean and the atmospheric variables is represented through the use of the rolling prediction strategy. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 is a flowchart of the data-driven ENSO multivariate three-dimensional field prediction method of the application;

[0047] Figure 2 is a multivariate anomaly correlation coefficient distribution diagram calculated according to the GODAS test set from 1980 to 2023 and the prediction field of the corresponding period;

[0048] Figure 3 is a result visualization diagram of the ENSO intelligent prediction model of the application;

[0049] Figure 4 is a model architecture diagram of the ENSO intelligent prediction model of the application;

[0050] Figure 5 The orthogonal disturbance basis mode acquisition flowchart of the application. DETAILED DESCRIPTION

[0051] The application will be further described below with reference to the accompanying drawings. The following examples are only used to more clearly illustrate the technical solutions of the application, and cannot be used to limit the protection scope of the application. It should be noted that the terms "comprise" and any variations thereof in the specification and claims of the application and the above-mentioned drawings are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device comprising a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0052] As shown in Figure 1 , a data-driven ENSO multivariate three-dimensional field prediction method is provided, comprising the following steps:

[0053] S1: obtaining ENSO multivariate three-dimensional field data of a previous period and a reporting time, the three-dimensional field data including: sea surface zonal wind stress anomaly field, meridional wind stress anomaly field, and sea surface temperature anomaly field within a set depth;

[0054] S2: based on the pre-processed multivariate three-dimensional field data, using an ENSO intelligent prediction model to rollingly predict control prediction three-dimensional field data of a future period;

[0055] S3: for a number of strong El Niño events observed in history, using the CNOP method to obtain a number of orthogonal disturbance basis modes triggering strong El Niño;

[0056] S4: judging whether the sea temperature anomaly field in the control prediction three-dimensional field data exceeds a set threshold, if not, directly outputting the control prediction three-dimensional field data as a result, if yes, adding each orthogonal disturbance basis mode to the sea temperature anomaly field of the reporting time to obtain a disturbed sea temperature field, and predicting three-dimensional field data of the future period based on the disturbed sea temperature field to obtain an ensemble prediction, and outputting the ensemble prediction and the control prediction three-dimensional field data as a result together.

[0057] Step S1 further includes pre-processing of the three-dimensional field data, for removing linear trend of each variable spatial grid point.

[0058] The data pre-processing method includes: interpolation and normalization, cutting out a region (0°-360°, 20°S-20°N, S and N respectively represent southern latitude and northern latitude) from the original range of the anomaly field of the variable; normalizing the multivariate anomaly field by dividing it by the standard deviation of its spatial average; and setting all land values and missing values to 0.

[0059] In the embodiment, the data interpolation method is: interpolating the multivariate anomaly field from the original spatial resolution to a rectangular grid with a zonal resolution of 2° and a meridional resolution of 0.5° (1°) within (outside) 5°N / S; in addition to horizontal interpolation, the three-dimensional sea temperature anomaly field also needs to be interpolated to 5, 20, 40, 60, 90, 120, and 150 meters in vertical depth.

[0060] In the embodiment, the normalization method is: first, calculating the standard deviation of the anomaly field of the variable at each ocean grid point to obtain a standard deviation spatial field within the range (0°-358°E, 20°S-20°N, E represents eastern longitude); then, spatially averaging the standard deviation field of each variable to obtain a field average standard deviation value; subsequently, dividing the anomaly field of each variable by its field average standard deviation to obtain the normalized multivariate field; finally, concatenating the multivariate field along the layer axis to 9 layers of data for subsequent input into the ENSO intelligent prediction model.

[0061] For example, Figure 4As shown in (a) in step S2, the ENSO intelligent prediction model includes: encoder preprocessing module, decoder preprocessing module, encoder, decoder and output layer.

[0062] I. Encoder Preprocessing Module.

[0063] This method is used to segment and linearly process the model input data, then embed spatial and temporal location codes to obtain embedded features. The model input data consists of ENSO multivariate 3D field data for the reporting start time and preceding time periods. It is worth noting that the reporting start time and preceding time periods in this application are typically measured in months.

[0064] like Figure 4 As shown in (b), the input data for the ENSO intelligent prediction model It is four-dimensional data, in which, The duration of the input data is taken as [time length] in this embodiment. Months The number of variables is 9 in this application. and These represent the number of spatial grid points in the meridional and latitudinal directions of the variable field, respectively. In this application... , .

[0065] Input variables The monthly 3D variable field will be divided into multiple non-overlapping segments of size [missing information] in the encoder preprocessing module. cube blocks, and These represent the initial number of spatial grid points in the longitudinal and latitudinal directions of the variable field, respectively. Therefore, the number of cubes that can be divided into the three-dimensional variable field each month is... The three-dimensional field of variables after partitioning is Next, After the matrix is ​​expanded in 3D, it is processed through a linear layer to a length of The vector.

[0066] To utilize the sequence's order information, spatial and temporal location encoding information needs to be added to the initial 3D field after linear layer processing to obtain embedded features. for:

[0067] ;

[0068] in, Indicates a linear layer. and Encoding matrix of spatial information and time information respectively. In the embodiment, the position encoding method can refer to the prior art, and alternatively, a learnable random encoding method is adopted, and the time encoding selects a trigonometric function encoding scheme: , , , , , is the vector length of the input variable.

[0069] II. Encoder

[0070] is used for analyzing the embedded features based on the spatio-temporal attention algorithm, capturing the spatio-temporal relationship between each variable of the model input data, and obtaining multi-level features.

[0071] As shown in (c) of FIG. 1, Figure 4 the encoder includes a plurality of encoding layers, each encoding layer includes a multi-head spatio-temporal attention sub-layer and a fully connected sub-layer, and each sub-layer is provided with a residual structure and a layer normalization operation. In the embodiment, the number of encoding layers is set to , each layer includes two sub-layers: the first sub-layer is a multi-head spatio-temporal attention sub-layer (the number of heads in the multi-head attention is set to ), and the second sub-layer is a fully connected (FNN) sub-layer.

[0072] In the embodiment, the first multi-head spatio-temporal attention sub-layer of the encoder outputs a feature matrix is obtained by performing time and space attention calculation on the query (Query), key (key), and value (value) vectors respectively. When performing time attention calculation, the calculation method of the query , key , and value is as follows:

[0073] ; ; ;

[0074] wherein, is the output value of the previous encoding layer in the encoder, when , , represents a vector splitting operation, and the purpose of the vector splitting operation is to divide the output vector of into equal parts in the encoding dimension, so as to facilitate subsequent multi-head attention calculation, and in this way, the dimension used for multi-head attention calculation will be changed to .

[0075] Once the query, key, and value are calculated, the dot product scaling attention algorithm can be used to perform time-consuming self-attention computation: .in, , The Softmax activation function is used. For the head Time attention weights , and Head respectively Query, key and value matrix, Subsequently, the vectors previously separated for multi-head attention computation are concatenated again along the encoding dimension, and after a linear transformation, the first feature matrix of the encoding layer is obtained: .

[0076] After performing the time attention calculation, This will be used in subsequent spatial attention calculations. The method for spatial attention calculation is similar to that for temporal attention calculation, except that the calculations are performed in the spatial dimension: , For the head Spatial attention weights, , , It is the query, key, and value matrix obtained by performing spatial attention computation. The second feature matrix of the encoding layer Specifically through Calculated This is a LayerNorm operation.

[0077] The matrix is ​​the feature matrix obtained after computation by the multi-head spatiotemporal attention sub-layer. It will then be fed into the second sub-layer in the coding layer, namely the fully connected layer, to obtain the first feature matrix. Feature matrix of the layer coding layer: .

[0078] The encoder repeats the above operation at each layer, after which... After the calculation of the coding layers, the final output multi-level features of the encoder are obtained. (To be continued) Abbreviated as ). The matrix contains continuous data after encoder processing. The matrix, containing multivariate field information from the past month, will be continuously fed into the decoder for subsequent predictions.

[0079] III. Decoder Preprocessing Module.

[0080] The model input data is processed into a matrix consisting of a sequence of vectors, and position encoding is added for processing by the decoder. The decoder preprocessing module can refer to the prior art.

[0081] IV. Decoder

[0082] For predicting three-dimensional field data of a future time period according to multi-level features output by the encoder.

[0083] As shown in (d) of FIG. 1, Figure 4 The decoder includes a plurality of decoding layers, and each decoding layer includes a multi-head space-time attention sub-layer, a time attention sub-layer, and a full connection sub-layer, and each sub-layer is provided with a residual structure and a layer normalization operation. The second sub-layer time attention sub-layer calculates the time attention coefficient of the output matrix of the encoding layer and the output matrix of the first decoder sub-layer, denoted as the encoder-decoder interaction attention matrix.

[0084] In the second sub-layer, the query matrix comes from the output of the previous decoder sub-layer (the output of the multi-head space-time attention sub-layer), and the key and value matrices are the output matrices of the encoder . The decoder can simultaneously focus on the correlation between the input variable and the predicted variable in each decoding layer. Therefore, three attention calculations are performed in each decoding layer to obtain a time attention matrix of a decoding layer, a space attention matrix of a decoding layer, and an encoder-decoder interaction attention matrix of a decoding layer. The calculation methods of these three attention matrices and the calculation methods of the output feature matrices after each attention matrix (denoted as the first feature matrix , the second feature matrix , and the third feature matrix of a decoding layer, respectively) are as follows:

[0085] ;

[0086] ;

[0087] wherein, , , , denotes the output feature matrix of the previous decoding layer, .

[0088] ;

[0089] ;

[0090] wherein, , , , , is the time length of the output data;

[0091] ;

[0092] ;

[0093] wherein, , , .

[0094] the output feature matrix of the decoding layer after the above two sub-layer processing after the fully connected sub-layer and through the residual structure and layer normalization, the output feature matrix of the first decoding layer is obtained .The output value of the first decoding layer will be continuously sent to the second decoding layer for the same processing, until the final output matrix of the decoder is obtained after all the decoding layers are processed . .

[0095] V. Output layer

[0096] connected with the fully connected sub-layer of the last decoding layer of the decoder, used for reorganizing the three-dimensional field prediction data output by the decoder into the original spatial dimension after linear layer processing, to obtain the final predicted three-dimensional field data of the future time period .

[0097] ;

[0098] wherein, represents the matrix dimensionality reduction and reorganization operation.

[0099] During the training of the ENSO intelligent prediction model, the loss function adopts the joint root mean square error (RMSE) of the multivariate field and the SST anomaly in the Niño3.4 region, and the prediction process adopts the rolling prediction scheme.

[0100] The training details of the ENSO intelligent prediction model are as follows: the batch size of the training data is set to 8, the Adam optimizer is selected to adjust the randomly initialized weights of the model, the learning rate of the model is selected using the warm-up adjustment strategy, and the maximum learning rate is 1.5×10 -4 ​​​After each training round, the model with the smallest loss function value on the validation set is calculated and saved. To avoid overfitting, training is stopped when the model's loss value on the validation set does not decrease for 5 consecutive rounds.

[0101] The rolling forecast process is as follows: First, the variable fields for the 12 months preceding the reporting month are used as the initial field of the model, and the multivariate field for the following month is predicted based on these initial conditions. Then, the predicted field for the next month is concatenated with the variable fields for the following 11 months from the original input field to obtain a new 12-month input field, and the forecast for the second month is then performed. This process is repeated until the predicted field for the next 20 months is obtained.

[0102] The loss function used in training the ENSO intelligent prediction model is a combination of the spatial average RMSE of the multivariate field and the Niño 3.4 exponential RMSE, representing the deviation between the predicted and target fields. ;

[0103] In the formula, For the predicted length of the future time period, To predict the target month, this application is for Months; and These are the model input field and target field, which contain the sea surface wind stress anomaly field and the upper sea surface temperature anomaly field, respectively. and These represent the total number of grid points in the longitude and latitudinal directions, respectively. For the total number of variables, For the first One latitudinal grid point, For the first A meridian grid point, For the first One variable, and These are calculated based on the SST anomalies in the input and target fields, respectively. index.

[0104] The data sources for training the ENSO intelligent prediction model are as follows: the training set includes historical simulation data from 23 CMIP6 climate models, covering the period from 1850 to 2014; the validation data consists of reanalysis data from SODA and ORAS5, covering the periods from 1871 to 1979 and 1958 to 1979, respectively; and the test set is GODAS reanalysis data from 1980 to 2023. Figure 2 As shown, the reanalysis data of the GODAS test set is presented. Figure 2(a), (b), (c) and (d) in FIG. 1 respectively represent the anomaly correlation coefficient (ACC) fields of the zonal wind stress, the meridional wind stress, the SST, and the sea temperature anomaly along the equatorial zonal profile 9 months in advance, and the observed fields. Figure 2 (e), (f), (g) and (h) in FIG. 2 respectively represent the anomaly correlation coefficient (ACC) fields of the zonal wind stress, the meridional wind stress, the SST, and the sea temperature anomaly along the equatorial zonal profile 9 months in advance, and the observed fields. Figure 2 (i), (j), (k) and (l) in FIG. 3 respectively represent the anomaly correlation coefficient (ACC) fields of the zonal wind stress, the meridional wind stress, the SST, and the sea temperature anomaly along the equatorial zonal profile 9 months in advance, and the observed fields.

[0105] The visualization diagram of the prediction result of the ENSO intelligent prediction model can be referred to FIG. 1, and specifically, Figure 3 Figure 3 (a) in FIG. 1 is the evolution of the SST anomaly in the Niño3.4 region predicted by the model at different prediction lead times, the blue curve on the right side is the observed value of the Niño3.4 index, and the orange line is the prediction value 12 months in advance. Figure 3 (b) and Figure 3 (c) in FIG. 1 are respectively the prediction correlation coefficient (red line) and the root mean square error (blue line) of the Niño3.4 index before and after 2000 varying with the prediction lead time.

[0106] As shown in FIG. 1, step S3 specifically comprises the following steps: Figure 5

[0107] Step a: For each historical observed strong El Niño event, a plurality of initial sea temperature disturbance fields satisfying the energy constraint are generated from the simulation results of the climate model.

[0108] In the present application, the extreme El Niño is selected as the four events of 1982 / 1983, 1996 / 1997, 2015 / 2016 and 2023 / 2024, wherein the three events of 1982 / 1983, 1996 / 1997 and 2023 / 2024 are used to calculate the CNOP initial disturbance base, and the 2015 / 2016 event is used to test the effect of the ensemble prediction.

[0109] The initial disturbance source is the upper 150-meter sea temperature anomaly field of the tropical Pacific region (120°E-80°W, 20°S-20°N, W represents west longitude); the initial sea temperature disturbance field is selected from 23 CMIP6 model simulation fields and SODA reanalysis data sets, and 100 monthly averaged sea temperature anomaly fields (a total of 2400) are randomly selected from each data set.

[0110] The energy constraint condition of the initial sea temperature disturbance field is:​​

[0111] ; ;

[0112] wherein, is the initial sea surface temperature perturbation field, specifically the initial sea surface temperature anomaly field , and is the initial sea surface temperature anomaly field after energy constraint, and is the initial sea surface temperature anomaly field , and is the original energy of the initial sea surface temperature anomaly field , and , and are the number of zonal, meridional grid points and the vertical layer number of the sea surface temperature anomaly field respectively, is the average sea surface temperature anomaly perturbation norm of all marine grid points.

[0113] Step b: in the current perturbation space, the ENSO intelligent prediction model is used to predict each initial sea surface temperature perturbation field, and the first M initial sea surface temperature perturbation fields with the maximum target function value are selected and put into the perturbation candidate set.

[0114] The CNOP defined in the present application is: the initial sea surface temperature perturbation that makes the Niño3.4 grow fastest in the control prediction from the ENSO intelligent prediction model after the March report within the 9-month prediction period (i.e. the target month is December of the current year); under the premise of energy constraint, an initial sea surface temperature perturbation field is found to make the target function maximum at the prediction target time.

[0115] The target function is:

[0116] ;

[0117] wherein, is the initial sea surface temperature perturbation field corresponding to the target function value, i.e. the target function value calculated according to the prediction result of the initial sea surface temperature perturbation field , is the length of the future time period of prediction, is the starting month of prediction, and are the sea surface temperature anomaly values of the Niño3.4 region obtained by the perturbation experiment and the non-perturbation control experiment respectively at the prediction target month .

[0118] ​For the three super El Niño events in 1982 / 1983, 1996 / 1997 and 2023 / 2024, firstly, the input field without added disturbance is used to predict the Niño3.4 region sea surface temperature anomaly in December from March, which is defined as the undisturbed control test. Then the 2400 energy-constrained sea temperature disturbance fields are added to the initial sea temperature anomaly field in the last month, and the Niño3.4 region sea surface temperature anomaly prediction value in December is obtained from March, which is defined as the disturbance test.

[0119] Step c: constantly update the disturbance candidate set by multiple means and eliminate the minimum value, and take the average of the last updated disturbance candidate set to obtain the orthogonal disturbance basis mode of the current disturbance space and save it.

[0120] The number M in step b can be 20, that is, step c is: the set average and energy constraint of the 20 disturbance fields obtain the 21st candidate disturbance. Then the target function value of adding the 21st disturbance is calculated and compared again, and the disturbance field with the minimum target function value is discarded; then the above "average-disturbance-comparison-elimination" step is repeated 30 times, and finally the set average of the 20 disturbance fields is taken as the orthogonal disturbance basis mode of the current disturbance space.

[0121] Step c specifically includes the following steps:

[0122] Step c1: take the set average and energy constraint of all disturbance fields in the disturbance candidate set to obtain the average disturbance field, and use the ENSO intelligent prediction model to obtain the target function value of the average disturbance field;

[0123] Step c2: compare the target function values of the average disturbance field and all disturbance fields in the disturbance candidate set, discard the disturbance field with the minimum target function value, and update the disturbance candidate set;

[0124] Step c3: determine whether the update number of the disturbance candidate set reaches the set threshold, if not, return the updated disturbance candidate set to step c1 for processing, if yes, take the average of all disturbance fields in the last updated disturbance candidate set to obtain the orthogonal disturbance basis mode of the current disturbance space, and save it.

[0125] Step d: determine whether the number of disturbance samples in the saved orthogonal disturbance basis mode reaches the set number, if yes, stop and output all saved orthogonal disturbance basis modes, if not, use the Schmidt projection method to subtract the projection of the initial sea temperature disturbance field on the orthogonal disturbance basis mode to update the disturbance space, and return to step b.

[0126] Step d is to calculate higher order CNOP, because it is necessary to subtract its projection on the orthogonal disturbance basis mode of the current disturbance space from 2400 initial disturbance fields by using the Schmidt projection method, that is, to deduct the cumulative component of all disturbance basis modes already existing in the initial sea temperature disturbance by using the Schmidt projection method ; wherein, is the initial sea temperature disturbance field in the disturbance space the sea temperature disturbance field updated for the time, is the orthogonal disturbance basis mode, denotes the inner product of and denotes the two-norm; then steps b-d are repeated multiple times to sequentially calculate CNOP disturbance fields of orders n = 2-5.

[0127] In step S4, the control prediction three-dimensional field data sea temperature anomaly field threshold calculation method can be determined based on expert experience. In the present application, the ENSO intelligent prediction model can be used to start prediction for 12 months from March of each year from 1980 to 2023, extract the spatially averaged sea temperature anomaly (T150) of the upper 150 meters in the equatorial eastern Pacific region (5°S-5°N, 140°W-80°W) from October to December each year, and calculate the standard deviation thereof as the threshold value; in real-time application, the new prediction value of each spring is compared with the threshold value, and if the predicted T150 anomaly in the equatorial eastern Pacific region from October to December exceeds the threshold value, it is determined that a strong El Niño event may occur in the current year, and the CNOP ensemble prediction scheme is started immediately.

[0128] For three extreme El Niño events, 5 CNOP orthogonal disturbance basis modes are obtained for each extreme event, and a total of 15 CNOP orthogonal disturbance basis modes are obtained. In subsequent ensemble prediction, the 15 CNOP-based sea temperature initial disturbances are sequentially added to the input sea temperature anomaly field of the last month (the starting time of the report), and 15 predictions are performed to obtain 15 prediction results forming an ensemble prediction, which is output together with the control prediction three-dimensional field data.

[0129] Embodiment 2

[0130] A specific example of applying the ENSO multivariate three-dimensional field prediction method of the present application for prediction is given, which specifically includes the following steps:

[0131] 1. Collect monthly average data of ocean and atmospheric variables.

[0132] The study region covers (92°E-330°E, 20°S-20°N), and the atmospheric variables include the sea surface zonal and meridional wind stress anomaly fields, and the ocean variables include the three-dimensional sea temperature anomaly field in the upper 150 meters.

[0133] 2. Process the original sea-air variable data.

[0134] The linear trend of each grid point is calculated and removed using the variable field data in the past 30 years. The climatological field of each variable is calculated, and the anomaly field is obtained by subtracting the climatological field of the corresponding month from the monthly mean data. The anomaly data in the selected latitude and longitude region is interpolated to a regular grid. The zonal resolution is 2°, and the meridional resolution is 1° (0.5°) outside (inside) 5°S / N. The sea temperature anomaly field is vertically interpolated to depths of 5, 20, 40, 60, 90, 120, and 150 meters. The standard deviation of each grid point is calculated, and the variable is normalized by dividing the anomaly field by the average standard deviation of the entire field. All land grids and missing values are assigned a value of zero. Finally, all normalized data is connected along the layer axis to construct a dataset containing 9 layers, which is provided as input to the model.

[0135] 3. Model construction.

[0136] An ENSO intelligent prediction model based on attention mechanism is established, and the basic architecture of the model includes the following: an encoder preprocessing module, a decoder preprocessing module, an encoder, a decoder, and an output layer.

[0137] 4. Model training.

[0138] The input field of the model contains 12 consecutive months of multivariate fields, and then outputs the multivariate anomaly field for the next 20 months in a rolling prediction manner. The training data is 23 CMIP6 model simulation data, and the validation data is reanalysis data before 1980. The early stopping mechanism is used, and when the loss value of the model on the validation set does not decrease for 5 consecutive rounds, the training is stopped.

[0139] 5. Model evaluation.

[0140] The reanalysis data from 1980 to 2023 is used to test the model performance, and the prediction performance of the multivariate field, Niño3.4 index, ENSO type, and development indicators is evaluated. The corresponding ensemble prediction base mode is calculated according to the CNOP calculation method. The extreme event warning threshold is calculated based on the return data.

[0141] 6. Real-time prediction

[0142] First, control forecast is carried out, if the control forecast shows that the eastern Pacific is neutral or La Nina state in the next year, single model prediction is adopted, and when the eastern Pacific T150 exceeds the warning threshold, i.e. an extreme El Nino may occur, the CNOP-based ensemble prediction is started.

[0143] Embodiment 3

[0144] The present application provides a computer device, comprising a processor and a memory; wherein the processor implements the steps of the above-mentioned data-driven ENSO multivariate three-dimensional field prediction method when executing the computer program stored in the memory.

[0145] The more specific process of the above-mentioned method can refer to the corresponding content disclosed in the foregoing embodiments, which will not be repeated here.

[0146] Embodiment 4

[0147] The present application provides a computer-readable storage medium for storing a computer program; the computer program is executed by a processor to implement the steps of the above-mentioned data-driven ENSO multivariate three-dimensional field prediction method.

[0148] The more specific process of the above-mentioned method can refer to the corresponding content disclosed in the foregoing embodiments, which will not be repeated here.

[0149] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the devices and storage media disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part.

[0150] Those skilled in the art can clearly understand that the technology in the embodiments of the present application can be realized by means of software and necessary general hardware platforms. Based on such understanding, the technical solutions in the embodiments of the present application can be embodied in the form of a software product, which can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the method described in the embodiments or some parts of the embodiments.

[0151] The above is only the preferred embodiment of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solution falling within the concept of the present application shall fall within the protection scope of the present application. It should be noted that, for ordinary skilled in the art, some improvements and refinements without departing from the principles of the present application shall be considered as the protection scope of the present application.

Claims

1. A data-driven ENSO multivariate three-dimensional field prediction method, characterized in that, The application relates to an ENSO intelligent prediction method. The method comprises the following steps: acquiring ENSO multivariate three-dimensional field data of a starting time and a preceding period, wherein the three-dimensional field data comprises sea surface zonal wind stress anomaly field, meridional wind stress anomaly field and sea surface temperature anomaly field within a set depth; based on the preprocessed multivariate three-dimensional field data, a control prediction three-dimensional field data of a future period is predicted by using an ENSO intelligent prediction model; for a plurality of strong El Nino events observed in history, a plurality of orthogonal disturbance base modes triggering the strong El Nino are acquired by using a CNOP method; it is judged whether the sea surface temperature anomaly field in the control prediction three-dimensional field data exceeds a set threshold value, if not, the control prediction three-dimensional field data is directly output as a result, if yes, for each orthogonal disturbance base mode, the orthogonal disturbance base mode is added to the sea surface temperature anomaly field of the starting time to obtain a disturbed sea surface temperature field, and the three-dimensional field data of the future period is predicted based on the disturbed sea surface temperature field to obtain an ensemble prediction, and the ensemble prediction and the control prediction three-dimensional field data are jointly output as the result; the method for acquiring the plurality of orthogonal disturbance base modes triggering the strong El Nino by using the CNOP method comprises the following steps: Step b: In the current disturbance space, the ENSO intelligent prediction model is used to predict each initial sea temperature disturbance field, and the first M initial sea temperature disturbance field with the maximum target function value is selected and put into the disturbance candidate set. step a: for each strong El Nino event observed in history, a plurality of initial sea surface temperature disturbance fields satisfying energy constraints are generated from climate model simulation results; step c: the disturbance candidate set is constantly updated by repeatedly calculating the mean value and discarding the minimum value, and the orthogonal disturbance base mode of the current disturbance space is obtained by calculating the mean value of the last updated disturbance candidate set and is saved; step d: it is judged whether the number of disturbance samples in the saved orthogonal disturbance base mode reaches a set number, if yes, the process is stopped, and all the saved orthogonal disturbance base modes are output, if not, the projection of the initial sea surface temperature disturbance field on the orthogonal disturbance base mode is deducted by using the Schmidt projection method to update the disturbance space, and the step b is returned to be executed; ； wherein, is the initial sea surface temperature perturbation field is the initial sea surface temperature perturbation field is the initial sea surface temperature perturbation field is the initial sea surface temperature perturbation field is the initial sea surface temperature perturbation field is the initial sea surface temperature perturbation field represents and represents the inner product represents the two-norm.

2. The data-driven ENSO multivariate three-dimensional field prediction method of claim 1, wherein, in the step d, the projection of the initial sea surface temperature disturbance field on the orthogonal disturbance base mode is deducted by using the Schmidt projection method, and the initial sea surface temperature disturbance field is updated by using the following formula: ； ； where, is the initial sea surface temperature anomaly field, which is specified by the th initial sea surface temperature perturbation field, is the th sea surface temperature anomaly field after energy constraint, is the initial sea surface temperature anomaly field, is the original energy of the th sea surface temperature anomaly field, is the total energy of the initial sea surface temperature perturbation field, and are the number of zonal, meridional grid points and the vertical layers of the sea surface temperature anomaly field, respectively, is the average sea surface temperature anomaly perturbation norm of all ocean grid points.

3. The data-driven ENSO multivariate three-dimensional field prediction method of claim 1, wherein, in the step a, the energy constraint condition of the initial sea surface temperature disturbance field is: ； wherein, is the initial sea surface temperature perturbation field for the corresponding objective function value, is the predicted future time period length, is the predicted starting month, and are the predicted target months respectively, the sea surface temperature anomaly values in the Niño 3.4 region obtained from the perturbation experiment and the unperturbed control experiment.

4. The data-driven ENSO multivariate three-dimensional field prediction method of claim 1, wherein, in the step b, the objective function used for predicting each initial sea surface temperature disturbance field is: the step c specifically comprises the following steps: step c1: all the disturbance fields in the disturbance candidate set are averaged and energy-constrained to obtain an average disturbance field, and the objective function value of the average disturbance field is obtained by using the ENSO intelligent prediction model for prediction; step c2: the objective function values of the average disturbance field and all the disturbance fields in the disturbance candidate set are compared, the disturbance field with the minimum objective function value is discarded, and the disturbance candidate set is updated; step c3: it is judged whether the updating times of the disturbance candidate set reach a set threshold value, if not, the updated disturbance candidate set is returned to the step c1 for processing, if yes, all the disturbance fields in the last updated disturbance candidate set are averaged to obtain the orthogonal disturbance base mode of the current disturbance space, and the orthogonal disturbance base mode is saved.

5. The data-driven ENSO multivariate three-dimensional field prediction method of claim 1, wherein, The ENSO intelligent prediction model comprises an encoder preprocessing module, a decoder preprocessing module, an encoder, a decoder and an output layer. The encoder preprocessing module embeds spatial position encoding and time position encoding after block processing and linear layer processing of model input data, to obtain embedded features; the model input data is ENSO multivariate three-dimensional field data of a previous period and a reporting time; The encoder is configured to analyze the embedded features based on a spatio-temporal attention algorithm, to capture the spatio-temporal relationship between each variable of the model input data, and to obtain multi-level features; the encoder comprises a plurality of encoding layers, each encoding layer comprises, in sequence, a multi-head spatio-temporal attention sublayer and a fully connected sublayer, and each sublayer is provided with a residual structure and layer normalization operation; The decoder preprocessing module processes the model input data into a matrix composed of a vector sequence, and adds position encoding, so as to be processed by the decoder; The decoder is configured to predict three-dimensional field data of a future period according to the multi-level features output by the encoder; the decoder comprises a plurality of decoding layers, each decoding layer comprises, in sequence, a multi-head spatio-temporal attention sublayer, a time attention sublayer and a fully connected sublayer, and each sublayer is provided with a residual structure and layer normalization operation; The output layer is connected to the fully connected sublayer of the last decoding layer of the decoder, and is configured to reorganize the three-dimensional field prediction data output by the decoder into the original spatial dimension after linear layer processing, to obtain the final predicted three-dimensional field data of the future period.

6. The data-driven ENSO multivariate three-dimensional field prediction method of claim 5, wherein, The specific operation of the multi-head spatio-temporal attention sublayer is as follows: The multi-head spatio-temporal attention sublayer input feature is divided into a plurality of heads, time attention weights of each head are obtained, and the divided heads are spliced based on the time attention weights to obtain a first feature matrix; the first feature matrix is taken as input, and is again divided into a plurality of heads, spatial attention weights of each head are obtained, and the divided heads are spliced based on the spatial attention weights to obtain a second feature matrix; the second feature matrix is residual connected with the multi-head spatio-temporal attention sublayer input feature, and layer normalization operation is performed, to obtain an output feature matrix of the multi-head spatio-temporal attention sublayer; The fully connected sublayer of the encoding layer deepens and integrates the output feature matrix of the multi-head spatio-temporal attention sublayer in the encoding layer, to output a feature matrix of the current encoding layer; The time attention sublayer of the decoding layer obtains a query matrix from the output feature matrix of the multi-head spatio-temporal attention sublayer, and obtains a key matrix and a value matrix from the multi-level features output by the encoder; The time attention sublayer of the decoding layer divides the output feature matrix of the multi-head spatio-temporal attention sublayer into a plurality of heads and performs time attention mechanism, to obtain an encoder-decoder interaction attention matrix; the encoder-decoder interaction attention matrix is spliced based on the multi-head splicing, to obtain a third feature matrix; after residual connection and layer normalization operation of the third feature matrix and the input feature of the time attention sublayer, an output feature matrix of the current time attention sublayer is obtained; The fully connected sublayer of the decoding layer deepens and integrates the output feature matrix of the time attention sublayer in the decoding layer, to output a feature matrix of the current decoding layer.

7. A computer device, characterized by A computer program product comprising a computer readable medium storing a computer program, wherein the computer program is executable by a processor to implement the steps of the data-driven ENSO multivariate three-dimensional field prediction method of any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, A computer program product comprising a computer readable medium storing a computer program, wherein the computer program is executable by a processor to implement the steps of the data-driven ENSO multivariate three-dimensional field prediction method of any one of claims 1-6.

Images