CN114202091A

CN114202091A - Indian ocean dipole index prediction method

Info

Publication number: CN114202091A
Application number: CN202110749720.8A
Authority: CN
Inventors: 冯源; 李晨; 孙天颖; 洪锋; 刘超; 田苗苗; 李首成
Original assignee: Ocean University of China
Current assignee: Ocean University of China
Priority date: 2021-07-02
Filing date: 2021-07-02
Publication date: 2022-03-18

Abstract

The invention relates to the technical field of marine environment parameter calculation, in particular to an Indian ocean dipole index prediction method. The invention provides a prediction method of Indian ocean dipole index, which comprises the following steps: acquiring data characteristics at least including wind field data around a designated sea area and sea-atmosphere information to form a training data set; inputting the data set into a preset ConvLSTM model for training to obtain a prediction model capable of predicting the sea surface temperature of the designated sea area; utilizing the predictive models for the eastern region (90 ° E-110 ° E, 10 ° S-0 °) and the western region (50 ° E-70 ° E, 10 ° S-10 ° N) of the Indian ocean to obtain a sea surface temperature for the eastern and western regions for a particular period of time in the future; and (3) differentiating the average sea surface temperature of the west region from the average sea surface temperature of the east region to obtain the Indian dipole index.

Description

Indian ocean dipole index prediction method

Technical Field

The invention relates to the technical field of marine environment parameter calculation, in particular to an Indian ocean dipole index prediction method.

Background

Dipole phenomenon (IOD) is a large scale physical marine phenomenon occurring in the indian ocean basin and has an important role in predicting tropical pacific erno-south billows. Whether the IOD occurs or not is predicted, and the method has important significance for researching climate change and other major marine phenomena.

IOD is one of the important systems affecting asian climate anomalies. There was a scientist who developed the concept of IOD in 1999 and defined the average sea surface temperature in the western india (50 ° E-70 ° E, 10 ° S-10 ° N) area minus the average sea surface temperature in the eastern (90 ° E-110 ° E, 10 ° S-0 °) area as the IOD index. IODs have a remarkable feature of season phase locking, usually beginning to develop in summer, reaching a peak in autumn, and rapidly decaying in winter, and have attracted much attention from domestic and foreign scientists since 1999.

The IOD, although only manifested as an abnormality in the ocean temperature in india, has a greater effect on the interaction between the ocean and the atmosphere, resulting in climatic abnormalities in the regions around the indian ocean, in the middle of south america, in the south of africa, in the southeast of australia, in northeast asia, etc.

Disclosure of Invention

In view of the shortcomings of the prior art, it is an object of the present invention to provide a indian ocean dipole index prediction method.

The Indian dipole index prediction method provided by the invention comprises the following steps:

obtaining ocean-atmosphere information at least comprising data characteristics of a wind field around a designated sea area to form a training data set;

inputting the data set into a preset network model for training to obtain a prediction model capable of predicting the sea surface temperature of the designated sea area;

predicting the designated sea area by using the prediction model to obtain the sea surface temperature of the designated sea area in a specific period of time in the future;

dividing the designated sea area into an east area and a west area according to a preset dividing mode;

and subtracting the average sea surface temperature of the west region from the average sea surface temperature of the east region to obtain the dipole index of the designated sea area.

Further, the data characteristics of the wind field around the designated sea area comprise:

wind speeds of the first six months distributed at a plurality of points around the designated sea area

Further, the data of each month in the training data set contains 96 features;

the 96 features include, in the respective month:

atmospheric temperature, potential height, vertical speed, water vapor, east-west wind speed, south-north wind speed at 1000hPa, 850hPa, 500hPa and 300hPa respectively;

sea surface height of the central point of the sea area, sea surface temperature of the central point and sea surface temperature of 15 points around the central point;

temperature, east-west direction sea current, south-north direction sea current and salinity when sea area central points are at sea depths of 5m, 15m, 25m, 35m, 45m, 55m, 65m, 75m, 85m and 95m respectively;

wind speed of the first six months distributed at 15 points around the sea area

Further, the network model is a ConvLSTM model.

Further, 5 months of marine-atmospheric information was considered as a time step.

Further, the time of the marine atmosphere information is exaggerated for at least 10 years.

Further, the convolution mode in the ConvLSTM model is set as one-dimensional convolution.

Further, the convolutional layer in the ConvLSTM model is two layers.

Further, the fitting round in training the ConvLSTM model was set to 100.

Compared with the prior art, the method for adding the wind field information of 6 months on the basis of the original data is provided, and a new idea is provided for the development of marine physics. It is very difficult to select factors having an influence on physical phenomena from the huge marine data by means of only manpower. The embodiment of the application applies deep learning to the research of physical oceanography and provides help for multidisciplinary fusion. After the wind field is added, the correlation coefficient of the predicted year can be improved.

Drawings

Other features, objects and advantages of the invention will become more apparent upon reading of the detailed description of non-limiting embodiments with reference to the following drawings:

FIG. 1 is a graphical representation of RMSE prediction data provided herein;

FIG. 2 is a schematic illustration of ACC prediction data provided herein;

FIG. 3 is a line graph of predictive data provided herein;

4a, 4b, 4c, 4d are schematic diagrams of the resulting dipole index provided herein;

fig. 5a, 5b, 5c, 5d are schematic diagrams of the resulting dipole index provided herein;

fig. 6a, 6b, 6c, 6d, 6e, 6f and 6g are schematic diagrams of real sea temperature data in months 6-12 of 2015 provided by the present application, respectively;

fig. 7a, 7b, 7c, 7d, 7e, 7f, and 7g are schematic diagrams of predicted sea temperature data in months 6-12 of 2015 provided by the present application, respectively.

Detailed Description

The present invention will be described in detail with reference to specific examples. The following examples will assist those skilled in the art in further understanding the invention, but are not intended to limit the invention in any way. It should be noted that it would be obvious to those skilled in the art that various changes and modifications can be made without departing from the spirit of the invention. All falling within the scope of the present invention.

In the early 21 s, scientists also studied the remote influence of IOD on east asian climate, which mainly transmitted the abnormal signal of IOD to east asian area in wave train form, and resulted in abnormal temperature in south china and japan. In addition, scientists compare the difference of rainfall in south China under the positive phase and the negative phase of the IOD and find out the development stage of the positive phase of the IOD, the reverse cyclone type circulation of the Bengal gulf conveys water vapor to the south China through the south China sea, and the abnormal increase of the rainfall in summer is caused. The students also mentioned that the IOD has an influence on the precipitation of the Yangtze river basin in China through remote correlation. Some researchers also studied the influence of IOD and ENSO on rainfall in southern China. Therefore, the generation of the IOD can be accurately predicted, and help can be provided for flood prevention work before the coming season.

At present, many scholars at home and abroad research the IOD prediction continuously. The American national environment prediction coupling system (NCEP CFS) comprises a global ocean model, an atmospheric layer 3D circulation model and a land model, and the coupling system has more advantages in prediction 2-3 months ahead of prediction than in prediction 5-6 months ahead of prediction. Scientists also developed 24 collective forecasting tests in 1981-2017 by using the seasonal-seasonal climate forecasting system FGOALS-f2 of the institute of atmospheric physics of Chinese academy of sciences. According to the prediction results, the predicted fitness in advance of one month is 0.82, which decreases with the time of prediction, and the predicted fitness in advance of 5 months is 0.56. Throughout the research process of IOD prediction, from a simple coupling model, to a more complex coupling model, to an initialization strategy of full-field assimilation and abnormal-field assimilation, an optimal interpolation technology (EnOI) and an analysis increment updating technology (IAU) are introduced, and finally by using a seasonal modulation and ENSO forced random dynamics (SDM) model, the prediction duration and accuracy of the two models cannot meet the requirements of experimental research. These methods rely on manual screening and integration to extract complex data relationships from large volumes of data. With the increasing bulkiness of ocean data, new methods are needed to digest the data, and the current deep learning methods in the field of artificial intelligence provide responses to the data.

On the basis of physical oceanographic knowledge, a ConvLSTM (Convolutional LSTM Network) deep neural Network capable of mining space-time information is used for learning ocean information including sea surface temperature, underwater temperature, water flow velocity, salinity and sea surface height and atmospheric information including temperature, humidity, wind field and other information. And the model is intervened by taking the wind field signal as one of the factors which can cause the IOD, and the wind field signal is added to each time step for learning before data is input into the model, so that the model can consider the influence of the wind field on the IOD event. By means of the learning capacity of ConvLSTM on the space-time problem and the intervention of physical oceanographic knowledge, the IOD index fitting degree predicted by the embodiment of the application is high, and the change trend of the real IOD index can be fitted.

The embodiment of the application adopts a re-analysis data set with the spatial resolution of 1 degree multiplied by 1 degree provided by the American meteorological environment forecasting center (NCEP) to carry out quality control and assimilation processing on observation data of various sources (ground, ships, radio sounding, anemometry balloons, airplanes, satellites and the like). The data set has the characteristics of multiple contained elements, wide range and long extension time. The reanalyzed data set provided by NCEP is monthly data.

The present examples are directed to the interior of the indian ocean, which is the third major ocean of the world (30 ° E-135 ° E, 30 ° N-66.5 ° S). The embodiment of the application selects a sea area with the range of 40 DEG E-110 DEG E, -25 DEG S-25 DEG N from a reanalysis data set provided by NECP, and constructs the data set in a sliding window mode by using data from 1980 to 2018. The training set of the 2015 predictive model contained 1980-2014 data, with a sliding window sliding backwards one year at a time, for a total of 25 × 253363325 data, also referred to as the number of samples m. The test set contained 2533 pieces of data, 2005-2015 year data. The embodiment of the application predicts the selected sea area, selects the ocean-atmosphere data of 1-5 months in the continuous decade and the eleventh year, and predicts the sea surface temperature of 6-12 months in the eleventh year. The average sea surface temperature of the west region (50 ° E-70 ° E, 10 ° S-10 ° N) was subtracted from the average sea surface temperature of the east region (90 ° E-110 ° E, 10 ° S-0 °) per month to obtain the IOD index, and the change in the IOD index was observed by a line graph.

The marine-atmospheric data per month in the present example consisted of 81 features. The possible factors influencing the temperature of the data point sea surface are divided into three parts, namely atmosphere, sea surface and undersea parameters. The embodiment of the application selects the temperature, the potential height, the vertical speed, the water vapor, the east-west wind speed, the south-north wind speed, the data on different heights (1000, 850, 500 and 300hPa) and 24 atmospheric factors in total. The sea surface parameters include 17 Sea Surface Height (SSH) at the center point, Sea Surface Temperature (SST) at the center point and sea surface temperature at 15 points around the center point. The sub-sea parameters comprise temperature, east-west, north-south and south-north sea currents and salinity (SSS) of different sea depths (5,15,25,35,45, 55,65,75,85,95m) of the centre point, for a total of 40.

The method and the device introduce the wind field data into the characteristics, and explore the influence of the wind field data on the IOD index. A dataset with different characteristics was constructed by adding the wind speed of the first 6 months at 15 points around each month of the marine-atmospheric data (each month of data consists of 96 characteristics).

The data set adopted by the embodiment of the application is as follows:

TABLE 1

TABLE 2

TABLE 3

TABLE 4

The ConvLSTM method is adopted in the embodiment of the application, the spatial correlation is not considered in the traditional FC-lstm, only the temporal correlation is considered, the CNN is just opposite, the most important operation of the CNN is convolution, the spatial feature can be well extracted, and the time series information cannot be grasped. The antagonistic neural network (GAN) is trained in an antagonistic manner, is a generative model, and is not mature enough for the time series prediction problem. And ConvLSTM can better mine spatial correlation on the basis of the problem of time series. In the process of predicting the sea surface temperature, in order to better extract the sea-atmosphere features, the ConvLSTM method is adopted to convolve the sea-atmosphere data at each time step to extract the features.

i_t＝σ(W_xi*X_t+W_i*H_t-1+W_Ci℃_t-1+b_i) (1)

f_t＝σ(W_xf*X_t+W_f*H_t-1+W_Cf℃_t-1+b_f) (2)

C_t＝f_t℃_t-1+I_T°tan(W_xC*X_t+W_C*H_t-1+b_C) (3)

0_t＝σ(W_xo*X_t+W_o*H_t-1+W_CO℃_t+b_O) (4)

H_t＝O_t°tanh(C_t) (5)

The ConvLSTM structure is shown in the above formula, where i, f, C, O, H represent the input gate, the forgetting gate, the cell state, the output gate, and the output at a certain time, respectively. In contrast to LSTM, where X, c, h, I, o, f all become three-dimensional tensors, representing convolution operations,

which represents the multiplication of corresponding elements of the matrix, also known as Hadamard products. ConvLSTM uses three gate control structures of Lstm, can forget unimportant information in history and keep important information, has better performance on the problem of long-time sequence, and can effectively avoid the problems of RNN gradient disappearance and gradient explosion. The ConvLSTM method can convolve complex spatial features and provide strong memory for long-term sequences,the embodiment of the present application thus adopts this method.

The embodiment of the application aims to predict the sea surface temperature of 7-12 months per year, the time sequence is long, and a multi-step prediction model needs to be provided. When the model is designed, one month is not selected as a time step, and when the time step number is set to 125, gradient explosion occurs due to an excessive winding machine process. In order to better extract space-time information and avoid gradient problems, five months of ocean-atmosphere information is regarded as a time step, the number of the time steps is reduced, and data in the time step is increased. Because iod occurs over a longer period of time, the embodiment uses historical ten-year information to predict the training data for this embodiment to include a complete iod event period. In setting the ConvLSTM hyperparameters, it was found that ConvLSTM is more advantageous to process such long time sequences using one-dimensional convolution. The embodiment of the application sets the convolution mode in ConvLSTM to be one-dimensional convolution by changing the dimension of the sample. In the experimental process, it is found that due to the complex structure of ConvLSTM, the iteration times of the training set are set too large, so that the model parameter matrix is completely suitable for the data of the training set, and the test set cannot be fitted. This may cause an overfitting phenomenon. In consideration of the actual situation in the process of adjusting the parameters, the embodiment of the present application sets the epoch (round) to 100, so that the fitting degree of the model to the training set achieves a just good effect. The number of model layers is directly related to the prediction effect, and when the CNN model is used for predicting the sea surface temperature, three-five convolutional layers are generally adopted, but the model of the embodiment of the application only adopts two ConvLSTM layers for predicting the sea surface temperature. This is facilitated by ConvLSTM, which is a combination of LSTM and CNN neural networks, which has a strong mining capability for spatio-temporal data. In the experiment, adding more model layers leads to the occurrence of the overfitting phenomenon, and the ConvLSMT model with two ConvLSTM layers and one Dense layer has the best effect.

The model of the embodiment of the application uses adam as an optimizer, uses MSE as a loss function, and the size of the convolution kernel is (1, 3). Using two layers of ConvLSTM2D, the neuron number is set to 42, the output is returned at each time step as input for the next time step, and the convolution fill is set to same fill. The output of the ConvLSTM2d is directly mapped to the sea surface temperature of 7 months to be predicted in the embodiment through a full connection layer, and multi-step output is realized.

According to the embodiment of the application, ocean-atmosphere data of the past 125 months are used as a training set, and the sea surface temperature of 7 months in the future, namely 6-12 months, is predicted. And using 20% of the training set as a verification set, and adjusting model parameters to finally obtain the optimal model. Regarding the prediction of the IOD index, in order to represent the fitting of the real IOD and the predicted IOD index, the present embodiment uses the correlation to represent the prediction performance.

3.1 RMSE, ACC comparison

For the sea surface temperature prediction, the present embodiment uses Root Mean Square Error (RMSE) and Accuracy (ACC) as evaluation indexes. The RMSE is used to measure the deviation between the observed and true values, and can reflect the actual situation of the predicted value error, and the formula is as follows:

where m represents the number of sample points. X_pred，iRepresenting the predicted value, X, of the ith sample point_real，iThen represents the ith sample

The predicted value of this point. ACC, RMSE under windy and no wind conditions are shown in table 4 below, and in fig. 1, 2 and 3:

the IOD index is predicted based on sea surface temperature calculation, the accuracy of final iododidex prediction is determined to a certain extent through sea surface temperature prediction, and from Rmse, it is obvious that the errors of prediction in 2015 and 2016 of IOD are larger than those of prediction in 2018 of 2017 of normal years, and the prediction time is longer, and the Rmse is larger, so that the prediction is in line with expectation. The RMSE of this example for long-term prediction of sea surface temperature is between 0.4 ℃ and 0.6 ℃ overall, except that the individual months RMSE exceed 0.6 ℃ (12 months in 15 years, 12 months in 18 years, and 9 months in 2016), from RMSE it can be seen that the predicted data after wind farm addition performed better in IOD years. However, in normal years, the rmse of the wind field is not much different from that of the wind field.

From the aspect of ACC, the overall accuracy after the wind field is added is more than 98%, and the effect is better in 2015 and 2016 in two IOD years. Although the prediction accuracy of the wind farm is reduced in 2016 and 2017 in normal years, the difference is not large, and the wind farm is more stable as a whole. On the whole, the model for predicting the IOD index provided by the embodiment of the application has higher accuracy, and the higher accuracy is achieved before the wind field data is not added.

3.2 true IOD index and predicted IOD index

In this example, the sea surface temperatures were predicted by the trained ConvLSTM method using 2015, 2016, 2017, and 2018 as test sets, and the IOD index was calculated by subtracting the average sea surface temperature of the east (90 ° E-110 ° E, 10 ° S-0 °) from the average sea surface temperature of the (50 ° E-70 ° E, 10 ° S-10 ° N) region. Two sets of models are made in the embodiment, one set of models is characterized by 81-dimensional marine atmospheric data, and the other set of models is provided with a wind field 6 months ago on the basis of 81-dimensional characteristics at each time step. The IODindex obtained by the two models is shown in fig. 4a, 4b, 4c and 4 d; and fig. 5a, 5b, 5c, 5 d.

As can be seen from the figure, the IOD index is characterized by seasonal changes, rising in summer, peaking in fall and decaying rapidly in winter, which are especially evident in 2015 ultra-strong IOD. The solid red line in the figure is the curve of the change in the real IOD index in 2015, and it can be seen that the IOD index reaches a peak value close to 1.0 in 2015 10 months. The IOD phenomenon is not only manifested as a significant increase in the IOD index in autumn (referred to as a positive IOD phenomenon), but often a decrease in the IOD index to a valley occurs in the following year, which is referred to as a negative IOD phenomenon. The phenomenon of super-strong negative IOD appears in 2016, the IOD index of month 6 is in a decreasing trend, and is reduced to the lowest in month 7, which is close to-2.5. The IOD phenomenon tends to occur successively as a positive IOD and a negative IOD phenomenon. It can be seen from the figure that the IOD index sequence obtained from the predicted sea surface temperature was substantially characterized by the seasonal variation of IOD in 2015-2016. In addition, in a line graph added with wind field data, a blue line is closer to a red line, and the fitting degree of a true value is higher. The model error added into the wind field data is smaller, so that the calculated IOD index is closer to the true value, and the positive influence of the wind field data on the IOD index is proved. In general, the degree of fitting of the IOD index sequence calculated by the model predictive value of the embodiment of the application to the real IOD change curve is high, and the IOD index sequence has obvious peak values and valleys, so that whether the IOD occurs or not can be judged. Comparing the results of the two experiments in 2017, it can be seen that the errors of the IOD index in 6 and 12 months in 2017 are eliminated after the wind field data is added. By using a model of an original data set, the IOD index has a sudden lifting condition in 11 months, and after wind field data is added, the condition that the error of a certain month is abnormal is solved. This indicates that the model stability is higher after the wind field data is added. It can be seen from the comparison in 2018 that the IOD index is closer to the real IOD index in more months after the wind field is added. The model of this application embodiment has verified the priori knowledge of marine physics through the comparison to two sets of models, and many scholars think that the IOD phenomenon is influenced to the local wind field information 6 to 8 months ago or earlier. As can be seen from the figure, the IOD index calculated by the model after the wind field data is added is closer to the true value, and the important relation between the wind field data and the IOD phenomenon is verified.

The correlation coefficient may reflect a statistical indicator of the closeness of the correlation between the variables. The closer the value of the correlation coefficient is to 1, the higher the degree of correlation of the two sets of variables is proved. This example calculated the correlation coefficient of the predicted iod index at 7 months with the true iod index. The higher the correlation coefficient, the closer the predicted IOD index curve is to the trend of change in the true value. The correlation coefficients are shown in table 5 below:

as can be seen from the IOD index graph, after the wind field is added to the original data, the prediction precision of the IOD index in 2015-2017 is improved, the prediction strength of the super-strong positive IOD event in 2015 is improved, and the correlation is improved from 91.51% to 92.24%. For 2016 of super-strong negative IOD events, the prediction is also closer, and the correlation degree is improved from 69.50% to 72.24%. The correlation coefficients in 2015, 2016 and 2017 are obviously improved. The average relevance from 2015 to 2018 increased from 81.48% to 82.84%. However, it can be seen that the prediction strength of this embodiment is not high enough, and in the event of super-strong positive and negative IOD in 2015 and 2016, the prediction strength is not high enough. In the predictions of normal years (2017, 2018), the degree of fit of the IOD index is better. 2015-2017 year correlation coefficient is combined with an ACC line graph, 2017 year is the year with the lowest correlation coefficient, and the model is found to have the highest precision of more than 0.990 in 6 months in 17 years, but the precision fluctuation is large. The year 2018 is the year with the highest correlation coefficient, and it can be seen that the model ACC in the year 2018 is more stable than other years in the ACC line graph, and fluctuates around 0.985. This indicates that the stability of the model and the correlation coefficient are closely related.

Such as fig. 6a, 6b, 6c, 6d, 6e, 6f and 6g, and fig. 7a, 7b, 7c, 7d, 7e, 7f and 7g, respectively, show diagrams of real temperature and forecast data for months 6-12 in 2015. It can be seen that the indian ocean surface temperature exhibits east-high-west-low in 6 months, and the surface temperature is higher in the west sea area than in the east sea area in 10 months as a whole. The east and west sea areas affected by the IOD phenomenon show a trend of seasonal variation, and accord with the rule that the IOD index is reduced to the valley bottom in summer and reaches the peak value in autumn. Viewed in its entirety, the predicted thermodynamic diagram substantially simulates the distribution trend of the real sea temperature of the indian ocean. In a real thermodynamic diagram of 6-10 months, it can be seen that the temperatures of the arabian sea, the menglan bay and the sea area near the equator (7 ° N to-7 ° N and 50 ° E to 100 ° E) are higher, the temperatures of other sea areas are lower, and the corresponding predicted sea temperature diagram also has the characteristics, which shows that the model of the embodiment of the application has higher accuracy and smaller error. The predicted thermal distribution also conforms to the law of the IOD phenomenon, and the temperature difference between the west and east sea areas changes obviously. In the prediction graph, the east sea area temperature is slightly higher than the west in month 6, and as the month gradually enters the winter, the indian ocean shows a trend of overall cooling, and in month 10, the temperature value of the west sea area is increased overall, while the overall sea temperature of the east sea area is decreased.

The embodiment of the application verifies that the prior knowledge of marine physics, Iodindex is influenced by a wind field signal, and the average correlation degree from 2015 to 2018 is improved from 81.48% to 82.84%. The embodiment of the application verifies the existing marine knowledge through the combination of marine physical knowledge and deep learning. The model used in the embodiment of the application has small error, and in the prediction of the sea surface temperature, the average RMSE is 0.5493 and 0.5599 in 2015-2018, and after 0.462 and 0.4836 are added into the wind field, the average RMSE is 0.5124,0.5532,0.5121 and 0.4902, and the RMSE is obviously reduced, which indicates that the local wind field also has important influence on the sea surface temperature.

The IOD is an important factor influencing the change of global climate season to the year as a large-scale marine physical phenomenon, is influenced by various factors such as wind field, atmosphere, undersea heat transfer, ocean current and the like, and the experimental reliability predicted only by the surface temperature of the sea is poor. The IODindex is predicted in the medium-long term through multi-source multi-modal data of ocean-atmosphere, and correlation coefficients of 91.51%, 69.50%, 69.17% and 95.73% in 15-18 years are obtained. Especially accurate for 15 years of very strong IOD events and 18 years of forecasting. In the past, huge and complicated data processing needs to consume a large amount of labor and computing resources, and the ConvLSTM is used for processing multi-dimensional characteristic and long-term ocean information, so that a simple physical coupling system is accelerated and accurate. The prediction fitting degree of the atmospheric physics research institute FGOALS-f2 season internal-season climate prediction system of Chinese academy of sciences to 5 months in advance is 0.56, and the precision is reduced along with the advance of the prediction period. The model of the embodiment of the application predicts in advance for 7 months, and the difference between the IOD index and the true value is small overall.

The method for adding the wind field information of 6 months on the basis of the original data is provided, and a new idea is provided for the development of marine physics. It is very difficult to select factors having an influence on physical phenomena from the huge marine data by means of only manpower. The embodiment of the application applies deep learning to the research of physical oceanography and provides help for multidisciplinary fusion. After the wind field is added, the correlation coefficient is improved in 15-17 years, and the correlation coefficient is reduced in 18 years, which is related to the contingency of experimental results.

The embodiment of the application uses ConvLSTM neural network structure, and predicts iod index 6-12 months by ten years and eleven years of ocean atmosphere data of 125 months 1-5, and predicts 7 months ahead. And the degree of fitting between the predicted iod index curve and the real curve is represented by using the correlation coefficient, and the test result shows that the correlation coefficient added to the wind field characteristic reaches 82.84%.

The foregoing description of specific embodiments of the present invention has been presented. It is to be understood that the present invention is not limited to the specific embodiments described above, and that various changes and modifications may be made by those skilled in the art within the scope of the appended claims without departing from the spirit of the invention.

Claims

1. An Indian ocean dipole index prediction method is characterized by comprising the following steps:

acquiring data characteristics at least including wind field data around a designated sea area and sea-atmosphere information to form a training data set;

2. The indian ocean dipole index prediction method of claim 1 wherein the specified sea area ambient wind field data characteristics comprise:

wind speeds of the first six months distributed at a plurality of points around the designated sea area.

3. The indian ocean dipole index prediction method of claim 1 wherein each month of data in the training data set contains 96 features;

the 96 features include, in the respective month:

wind speeds of the first six months distributed at 15 points around the sea area.

4. The indian ocean dipole index prediction method of claim 1 wherein the network model is a ConvLSTM model.

5. The indian ocean dipole index prediction method of claim 4 wherein 5 months of ocean-atmosphere information is considered as a time step.

6. The indian ocean dipole index prediction method of claim 4 wherein the time of the ocean atmosphere information is exaggerated for at least 10 years.

7. The indian ocean dipole index prediction method of claim 4 wherein the convolution in the ConvLSTM model is set to one-dimensional convolution.

8. The indian ocean dipole index prediction method of claim 4 wherein the convolution layer in the ConvLSTM model is two layers.

9. The indian ocean dipole index prediction method of claim 4 wherein the fitting round in training the ConvLSTM model is set to 100.