CN114330132B

CN114330132B - ENSO diversity prediction method based on artificial intelligence

Info

Publication number: CN114330132B
Application number: CN202111659938.0A
Authority: CN
Inventors: 黄平; 王听雨
Original assignee: Institute of Atmospheric Physics of CAS
Current assignee: Institute of Atmospheric Physics of CAS
Priority date: 2021-12-30
Filing date: 2021-12-30
Publication date: 2022-07-01
Anticipated expiration: 2041-12-30
Also published as: CN114330132A

Abstract

The invention discloses an ENSO diversity forecasting method based on artificial intelligence, which extracts the first three main modes from equatorial Pacific SSTA observation data by using an EOF decomposition method, and projects CMIP6 historical simulation data on the three main modes to obtain three groups of PC values; three groups of PC values are used as forecast values, SSTA of an initial month and Tendency items of sea temperature data are used as input values of training, and a CMIP6 mode is used for training VGG-11; and inputting observation data into the trained model to obtain PC values at three future moments, and combining the PC values with 3 EOF main modes to reconstruct the SSTA space form of the equatorial Pacific region at the future moments. The method improves the forecasting skill of the middle Pacific type Erleno and breaks through the bottleneck of forecasting in the middle Pacific area in the previous dynamic mode. The method improves the forecasting skill of the ENSO, is beneficial to forecasting and early warning of climate disasters, and is beneficial to reducing personnel and property loss.

Description

ENSO diversity prediction method based on artificial intelligence

Technical Field

The invention relates to the technical field of climate prediction, in particular to an ENSO diversity forecasting method based on artificial intelligence.

Background

Early nino-southern billow (ENSO) has a significant impact on global climate and can cause serious flooding disasters. Therefore, the improvement of the ENSO forecasting skill is beneficial to disaster prevention and reduction in various countries. The existing artificial intelligence-based forecasting model refers to the el nino phenomenon on the entire equatorial pacific by forecasting the nino3.4 index, but such forecasting techniques do not adequately solve the el nino forecasting problem. Since el nino appears as an anomaly in Sea Surface Temperature (SST) on the equatorial pacific, there is spatial diversity in temperature anomalies, but the nino3.4 index cannot exhibit spatial diversity. Spatial diversity of erlinuo in addition to the common eastern pacific type erlinuo, there is a medium pacific type erlinuo, both types having distinct effects on global climate; even in the existing dynamic numerical model capable of predicting the spatial type, the prediction capability of the middle pacific type erlinuo is not high, so that the problem of predicting the erlinuo spatial type is not solved well.

Disclosure of Invention

The invention aims to provide an ENSO diversity forecasting method based on artificial intelligence, so as to solve the problems in the prior art.

In order to achieve the purpose, the technical scheme adopted by the invention is as follows:

an ENSO diversity forecasting method based on artificial intelligence comprises the following steps:

s1, extracting three EOF main modes from the surface temperature anomaly observation data of the equatorial pacific ocean by using an EOF decomposition method: a weft-wise consistent type, a weft-wise inconsistent type and a middle warming type;

s2, projecting the historical simulation data of the CMIP6 mode on three EOF main modes to respectively obtain three groups of historical simulation data PC values;

s3, training an improved deep learning model VGG-11 by using the three groups of historical simulation data PC values obtained in the step S2 as forecast values and the sea surface temperature abnormal value of the initial month and the Tendency term as input values;

s4, inputting the observed new data as forecast input values including sea surface temperature abnormal values and Tendency terms into the trained improved deep learning model VGG-11 to obtain three forecast PC values, and combining the three forecast PC values with the three EOF main modes obtained in the step S1 to obtain the predicted SSTA space form of the equatorial pacific region.

Preferably, the EOF decomposition method in step S1 is a principal component analysis method for decomposing spatiotemporal data into a time dimension PC value and a space dimension EOF mode.

Preferably, the projection process in step S2 is to obtain a PC value corresponding to the EOF mode at a certain time by using a matrix multiplication between the spatial data and the EOF mode at the certain time.

Preferably, the matrix multiplication uses a longitude and latitude grid point matrix of the EOF mode and a longitude and latitude grid point with abnormal sea surface temperature at a certain moment, a new matrix is obtained according to a rule of corresponding multiplication of grid points with the same longitude and latitude, and all values in the obtained new matrix are added to obtain a corresponding PC value at the moment.

Preferably, the training parameters of the improved deep learning model VGG-11 in step S3 are: small batch number: 128 using a rom-up training strategy, learning rate set to 0.0001, random gradient descent optimizer, batch norm method for normalization at each layer, and root mean square error of model truth and predicted values as loss function.

Preferably, the merging process of the three forecasted PC values in step S4 and the 3 EOF main modes obtained in step S1 specifically includes: and multiplying each PC value by the corresponding EOF main mode, and adding the obtained products to reconstruct the space form of the SSTA.

The invention has the beneficial effects that:

the invention discloses an ENSO diversity prediction method based on artificial intelligence, which improves the prediction skill of the middle Pacific type Erleno and breaks through the prediction bottleneck of the previous dynamic mode in the middle Pacific area. Compared with a deep learning model for forecasting the Nino3.4 index in the past, the technology can forecast more spatial form details of the SST, and is beneficial to forecasting and analyzing the specific change of future Ernino; the method is beneficial to disaster prevention and reduction, ENSO is an important factor influencing the annual change of global climate, the method improves the forecasting skill of ENSO, is beneficial to forecasting and early warning of climate disasters, and is beneficial to reducing the loss of personnel and property.

Drawings

FIG. 1 is a flowchart of the artificial intelligence based ENSO diversity prediction method provided in example 1;

FIG. 2 is a schematic diagram of three principal modes obtained in example 1 by decomposing the observed equatorial Pacific SSTA data using EOF;

FIG. 3 is a schematic structural diagram of an improved deep learning model VGG-11 provided in embodiment 1;

fig. 4 is a result of calculating a warm pool index WPI in example 2 to compare the prediction result of the present invention with the power pattern prediction that is internationally common.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail below with reference to the accompanying drawings. It should be understood that the detailed description and specific examples, while indicating the invention, are intended for purposes of illustration only and are not intended to limit the scope of the invention.

Example 1

The embodiment provides an artificial intelligence based ENSO diversity forecasting method, as shown in FIG. 1, including the following steps:

s1, extracting the first three main modes from the observation data of the surface temperature anomaly of the equatorial pacific ocean by using an EOF decomposition method: a weft-wise uniform type, a weft-wise nonuniform type and a middle warming type;

s2, projecting the CMIP6 historical simulation data on the three main mode EOF main modes to respectively obtain three groups of historical simulation data PC values;

s3, using three sets of historical simulation data PC values as forecast values, using SSTA of the initial month and two sea temperature data of Tendency items as training input values, and training an improved deep learning model (VGG-11) by using a CMIP6 mode.

S4, inputting the new observation data as forecast input values into a trained model, wherein the new observation data comprises sea surface temperature abnormal values and a Tendency item, inputting a trained improved deep learning model VGG-11 to obtain three forecast PC values, and combining the three forecast PC values with the three EOF main modes obtained in the step S1 to obtain the predicted SSTA space form of the equatorial pacific region.

It should be noted that the initial month in step S3 means that when a future time is to be forecasted, an initial condition needs to be input into the model to drive the model to calculate future data.

Any month in the historical data can be used as the input data of the initial month to obtain 'future' data, and the 'future' data is compared with the true value, so that the pattern prediction skill can be evaluated.

Since the latest observed month can be regarded as the initial month, with the serial number 0, denoted SSTA (0), the term Tendency is defined as the historical value of the initial month minus 2 months ahead of the initial month, and represents the trend of the change in sea temperature in these two months, namely Tendency — SSTA (0) -SSTA (-2).

In this embodiment, the EOF decomposition method in step S1 is a principal component analysis method, and is used to decompose the spatiotemporal data into a time dimension PC value and a space dimension EOF mode.

In this embodiment, the projection process in step S2 is to use matrix multiplication to multiply the spatial data at a certain time with the EOF mode, so as to obtain the PC value corresponding to the EOF mode at the certain time.

The matrix multiplication content is as follows: a longitude and latitude lattice point matrix of an EOF mode is used, an E (M x N) matrix is assumed, M represents the number of latitude lattice points, N represents the number of longitude lattice points, and the longitude and latitude matrix with abnormal sea surface temperature at a certain moment is set as an S (M x N) matrix, the lattice points with the same longitude and latitude are correspondingly multiplied to obtain a new matrix P (M x N), and all values in the obtained new matrix are added to obtain a corresponding PC value at the moment.

The improved deep learning model VGG-11 adopted in step S3 in this embodiment is shown in fig. 3, and the parameters of deep learning are:

small batch number: 128 using a rom-up training strategy, learning rate set to 0.0001, random gradient descent optimizer, batch norm method for normalization at each layer, and root mean square error of model truth and predicted values as loss function.

The merging process of the three forecasted PC values in step S4 and the 3 EOF main modes obtained in step S1 specifically includes: and multiplying the PC value by the corresponding EOF mode, and adding to reconstruct the spatial form of the SSTA.

Example 2

This example provides a prediction of early-onset events in the mid-pacific early-onset region using data from the equatorial pacific surface of 1984-2017.

The Warm Pool Index (WPI) was used in this example to evaluate the prediction of the pacific-type erlinuo in comparison.

1. Decomposing the observed equatorial Pacific SSTA data by using EOF to obtain the first three main modes (EOF 1-3, as shown in figure 2)

2. The PC values of each month in 1948-2014 can be obtained by projecting SSTA of 39 CMIP6 historical modes (time: 1948-2014) into three EOF main modes; while the SSTA and trending entries per month are available from the 39 history patterns.

3. Inputting a deep learning model by using SSTA and Tendency of an initial month, and outputting 3 PC values predicted in advance by N months (N is 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11); because of a large amount of data of 39 modes, after deep learning training, the model can learn ocean memory signals existing in the initial month (namely the input SSTA, Tendency of the initial month), and the signals can output 3 PC values which are advanced by N months after passing through the deep learning model;

4. by using the model, a set of results of a reward test (hindcast) can be obtained by combining the 3 predicted PC values with the corresponding EOF modes in the observation data (time: 1984-2017) (see step S4, FIG. 2), namely, the surface temperature anomaly of the equatorial pacific sea in 1984-2017 is predicted. We can compare the prediction results of the present invention with those of a plurality of power patterns commonly used internationally by calculating the Warm Pool Index (WPI), and compare their prediction skills in the middle pacific type of erlinuo, see fig. 4. The marks on the right indicate all lines in turn, the first line being the forecasting technique of the present invention, the other lines being the forecasting techniques of multiple power modes.

By adopting the technical scheme disclosed by the invention, the following beneficial effects are obtained:

the invention discloses an ENSO forecasting method based on artificial intelligence, which improves the forecasting skill of the middle Pacific type Erleno and breaks through the forecasting bottleneck of the previous dynamic mode in the middle Pacific area. 2. Compared with a deep learning model for forecasting the Nino3.4 index in the past, the technology can forecast more spatial form details of the SST, and is beneficial to forecasting and analyzing the specific change of future Ernino; the method is beneficial to disaster prevention and reduction, ENSO is an important factor influencing the annual change of global climate, the method improves the forecasting skill of ENSO, is beneficial to forecasting and early warning of climate disasters, and is beneficial to reducing the loss of personnel and property.

The foregoing is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, various modifications and improvements can be made without departing from the principle of the present invention, and such modifications and improvements should also be considered within the scope of the present invention.

Claims

1. An ENSO diversity forecasting method based on artificial intelligence is characterized by comprising the following steps:

s1, extracting three EOF main modes from the surface temperature anomaly observation data of the equatorial pacific ocean by using an EOF decomposition method: a weft-wise uniform type, a weft-wise nonuniform type and a middle warming type;

s4, inputting the new observed data as forecast input values including sea surface temperature abnormal values and Tendency terms into a trained improved deep learning model VGG-11 to obtain three forecast PC values, and combining the three forecast PC values with the three EOF main modes obtained in the step S1 to obtain a forecasted SSTA spatial form of the equatorial pacific region;

the Tendency term is defined as the initial month minus the historical value of the initial month 2 months ahead, representing the trend of the change in sea temperature over these two months;

the projection process in step S2 is to obtain the historical simulation data PC value corresponding to the EOF mode at a certain time by using matrix multiplication between the spatial data and the EOF mode at the certain time;

the matrix multiplication uses a longitude and latitude grid point matrix of an EOF mode and longitude and latitude grid points with abnormal sea surface temperature at a certain moment, a new matrix is obtained according to the rule of corresponding multiplication of the grid points with the same longitude and latitude, and all values in the obtained new matrix are added to obtain the corresponding historical simulation data PC value at the moment;

the three forecasted PC values in step S4 and the 3 EOF main modes obtained in step S1 are specifically: and multiplying each forecasted PC value by the corresponding EOF main mode, and adding the obtained products to reconstruct the space form of the SSTA.

2. The artificial intelligence based ENSO diversity prediction method of claim 1, wherein the EOF decomposition method in step S1 is a principal component analysis method for decomposing spatiotemporal data into a time dimension PC value and a space dimension EOF modality.

3. The artificial intelligence based ENSO diversity prediction method according to claim 1, wherein the training parameters of the improved deep learning model VGG-11 in step S3 are as follows: small batch number: 128 using a rom-up training strategy, learning rate set to 0.0001, random gradient descent optimizer, batch norm method for normalization at each layer, and root mean square error of model truth and predicted values as loss function.