El Nino / La Nina long-time prediction method based on machine learning

Abstract

The invention discloses an El Nino / La Nina long-time prediction method based on machine learning, and the method comprises the steps: constructing a long-time-sequence air-sea parameter data set, carrying out the calculation after removing abnormal values and carrying out the unified spatial-temporal scale preprocessing, obtaining an air-sea parameter distance flat field, constructing a prediction factor data set after the normalization and standardization processing, and carrying out the prediction of the air-sea parameter distance flat field. Generating a prediction factor library of tropical-subtropical latitude crossing and pacific-Atlantic ocean basin crossing; combining a root-mean-square error loss function to train the long short-term memory neural network, and optimizing parameter setting to obtain an optimal El Nino / La Nina prediction model; the model is used for predicting the El Nino / La Nina phenomenon, and the prediction index and the ocean Nino index of the climate prediction center are compared and evaluated. According to the method, the optimal ocean-atmosphere and tropical-subtropical prediction factor combination is constructed through the system, the contribution of the ocean-air interaction process in the El Nino / La Nina prediction model is enhanced, and the time and accuracy of the El Nino / La Nina prediction are optimized and improved.

Description

Technical Field

[0001] The present invention relates to reanalysis, multi-source heterogeneous data processing of ocean and atmospheric parameters of satellite remote sensing, El Nino / La Nina extreme weather prediction technology, and machine learning prediction fields, and in particular to a long-term prediction method for El Nino / La Nina based on machine learning. Background Art

[0002] At present, the occurrence and development of El Nino / La Nina events affect the global atmospheric temperature and extreme precipitation changes, which can cause global climate anomalies, resulting in severe droughts and floods, high temperature or low temperature ice and snow disasters in many regions, and causing huge losses to industrial and agricultural production and people's lives and property in many countries. For example, the occurrence and evolution of the El Nino / La Nina phenomenon has a close impact on the rainfall in North my country, and has important guiding significance for the prevention and control of drought and flood disasters in North China during the flood season. Global warming superimposed on the impact of El Nino / La Nina has led to more frequent, wider-ranging and stronger-intensity extreme weather and climate events.

[0003] Traditional statistical prediction (such as linear regression) has a good prediction effect on the linear evolution process of El Nino / La Nina phenomenon, but the complex ocean-atmosphere coupling model has a large amount of calculation and involves complex physical processes. It is difficult to well describe the nonlinear process in the interaction of multiple parameters between the ocean and the atmosphere, especially the prediction performance of 18 months and above needs to be further improved. How to achieve long-term and accurate prediction of the occurrence time, development process, and change characteristics of El Nino / La Nina is of great significance to the prevention and preparation of disaster prevention and mitigation during the flood season in my country. Summary of the invention

[0004] To this end, the present invention provides a long-term prediction method for El Nino / La Nina based on machine learning. The method improves the time and accuracy of long-term prediction of El Nino / La Nina by optimizing the combination of ocean and atmospheric parameters as prediction factors, and setting prediction factors for the combination of tropical and subtropical, Pacific and Atlantic, and ocean and atmospheric physical quantities.

[0005] In order to achieve the above object, the present invention provides the following technical solution: a long-term prediction method of El Nino / La Nina based on machine learning, comprising:

[0006] In the ocean and atmospheric parameter data preprocessing stage, a prediction factor normalization and standardization data set is constructed, and data preprocessing is performed on the historical ocean and atmospheric parameters to construct an ocean and atmospheric parameter data set. The historical ocean parameters include sea surface temperature and sea surface height, and the historical atmospheric parameters include meridional wind field, latitudinal wind field and outward longwave radiation physical quantities. The data preprocessing includes outlier removal and unified spatiotemporal scales to obtain a long-time series ocean and atmospheric parameter data set after data preprocessing; in the process of unified spatiotemporal scales of the data preprocessing, the ocean and atmospheric parameter data in the ocean and atmospheric parameter data set are uniformly processed into a spatial resolution of 1 degree × 1 degree and a monthly time resolution; the ocean and atmospheric parameter prediction factors in the ocean and atmospheric parameter data set after data preprocessing are subjected to anomaly calculation and normalization processing, and each feature of each data is zero-centered using a scaler, and its numerical range is normalized to within the range of [-1,1] to accelerate the convergence efficiency, and obtain 8 normalized ocean and atmospheric parameter data sets; at the same time, since the various historical ocean parameters and atmospheric parameters involved in the calculation have different dimensions, the ocean and atmospheric parameters in the ocean and atmospheric parameter data set after data preprocessing are subjected to data standardization, and then the ocean and atmospheric parameter data after data standardization are analyzed.

[0007] In the forecast factor combination setting stage, a forecast factor data set of ocean and atmospheric parameters is constructed, and key areas of ocean and atmospheric parameters that are sensitive to the occurrence and development of El Nino / La Nina are selected to construct tropical and subtropical El Nino / La Nina prediction factor combinations as one of the data input fields of the subsequent long short-term memory network El Nino / La Nina model; an El Nino / La Nina index forecast factor combination data set is constructed, and the tropical forecast factors in the forecast factor data set include 4 and 3 The sea surface temperature anomaly in the key area (160°E-90°W, 5°S-5°N), the sea surface height anomaly (160°E-90°W, 5°S-5°N), the meridional wind field anomaly (160°E-90°W, 5°S-5°N), the zonal wind field anomaly (160°E-90°W, 5°S-5°N), the outward longwave radiation prediction factor is the outward longwave radiation anomaly (160°E-160°W, 5°S-5°N), the subtropical prediction factors include the northwest Pacific sea surface temperature anomaly (90°E-150°W, 10°S-30°N), the South Pacific sea surface temperature anomaly (90°E-150°W, 10°S-30°N), and the tropical and North Atlantic sea surface temperature anomalies (EQ-15°N, 90°W-20°E). The forecast factor combination includes the sea surface temperature anomaly in the key sea areas of the tropical Pacific, the sea surface height anomaly, the zonal wind field anomaly, the meridional wind field anomaly, the tropical Pacific and northwest Pacific sea surface temperature anomaly combined forecast factor, the tropical Pacific and North Atlantic sea surface temperature anomaly combined forecast factor, and the tropical Pacific and South Pacific sea surface temperature combined forecast factor. Designed according to the El Nino / La Nina identification index method provided by the Climate Prediction Center, the forecast factor with the same time range as the forecast factor is calculated. The three-month moving average sea surface temperature anomaly index in the 3.4 region (120°W-170°W, 5°S-5°N) is used as the label data for diagnosing the occurrence of El Nino / La Nina events. The ocean and atmospheric parameter prediction factors and the El Nino / La Nina identification index are divided into training data sets and test data sets in proportion; the length of the training data set accounts for 85% of the total data, and the length of the test data set accounts for 15% of the total data. The label data of the ocean and atmospheric parameter prediction factors and the El Nino / La Nina index correspond to each other in the training data and the test data.

[0008] Construct a long-term prediction model of El Nino / La Nina with a multi-layer long short-term memory neural network, initialize the long short-term memory neural network prediction model and start training; divide the ocean atmospheric parameter prediction factor data set into a training data set and a test data set of the machine learning prediction model in proportion, and divide the label data of the El Nino / La Nina identification index into a training data set and a test data set in proportion, corresponding to the training data set and the test data set of the prediction factor; train the El Nino index prediction model through the root mean square error loss function; extract features of ocean and atmospheric physical quantities at multiple spatiotemporal scales, design a combination of ocean and atmospheric sensitive area prediction factors, and use the ocean and atmospheric prediction factors, tropical-subtropical prediction factors, 3.4 The three-month moving average labels of the sea surface temperature anomalies in the key areas are input into the trained El Niño / La Niña index prediction model respectively. The start and end time, change process, and intensity evolution are predicted through the El Niño / La Niña index prediction model. The physical mechanism constraints of the interaction between the ocean and the atmosphere are used to strengthen the consistency of the characteristics of different ocean and atmospheric physical quantities in the prediction of the El Niño / La Niña index, and the characteristics of the ocean-atmosphere interaction are used to predict the El Niño / La Niña index.

[0009] During the LSTM network training phase, the input layer of the El Nino / La Nina index prediction model consists of two parts: one is the ocean and atmospheric parameter prediction factors, the input nodes of the hidden layer are 361, the hidden nodes are 24, and the other is 3.4 El Nino / La Nina discrimination index of three-month moving average in key areas, with 18 output nodes. The ocean and atmospheric prediction factors are passed through a fully connected layer with activation function tanh, followed by a Dropout layer with a Dropout layer of 0.5 to prevent overfitting. The model loss function uses root mean square error for gradient descent iterative training. The input gate, forget gate, and output gate functions need to adjust the weight coefficient, which is 20 by default. The weight adjustment ratio for each iteration of network training is 0.01; the error threshold is 0.000001. Plot the error loss function result graph to verify that the prediction result gradually converges and stabilizes to the minimum value, which is the optimal model result.

[0010] The memory unit of the LSTM El Nino / La Nina model consists of three gates to collect control information, namely input gate i <t>< / t> 、Forget gate <t>< / t> , output gate o <t>< / t> The gated recurrent unit is mainly used to protect the unit from being affected by the information of other units. σ is the internal activation function, tanh is used; h is the target function, w is the weight matrix, and x is the input function. ⊙ represents the dot product symbol.

[0011] Forget gate: f <t>< / t> =σ(x1*W xf x <t>< / t> +x2*W hf x <t-1>< / t-1> +b f )

[0012] Input gate: i <t>< / t> =σ(x3*W xi x <t>< / t> +x4*W hi x <t-1>< / t-1> +b i )

[0013] Output gate: o <t>< / t> =σ(x5*W xox <t>< / t> +x6*W ho h <t-1>< / t-1> +b o )

[0014] Neuron: c <t>< / t> =f <t>< / t> ⊙c <t-1>< / t-1> +i <t>< / t> ⊙tanh(W xc x <t>< / t> +W hc h <t-1>< / t-1> +b c )

[0015] Objective function: h <t>< / t> =o <t>< / t> tanh⊙(c <t>< / t> )

[0016] In the formula, σ is the sigmoid function, the range is [0, 1], and it is the activation function of the control gate; tanh is the tanh function, the activation function when calculating the state, the range is [-1, 1], and it is used as the activation function of the standard neural network to avoid numerical problems during forward propagation and facilitate the rapid convergence of the model. <t>< / t> is the input data. <t-1>< / t-1> is the historical data calculated by the previous neuron. <t>< / t> The data calculated for this neuron. <t-1>< / t-1> is the memory data of the previous neuron. <t>< / t> The memory data calculated by this neuron. W xf , W hf is the weight matrix of the forget gate, b f is the bias term of the forget gate. xi , W hi is the weight matrix of the input gate, b i is the bias term of the input gate. W xo , W ho is the weight matrix of the output gate, b o is the bias term of the output gate. x1~x6 are adaptive coefficients, which are adjusted according to the amount of data and parameter configuration of the training field.

[0017] During the LSTM network testing phase, the test prediction factor data and label data are input into the El Nino / La Nina LSTM neural network prediction model for testing, and the optimal El Nino / La Nina prediction model and prediction results after the test are obtained; the optimal El Nino / La Nina prediction model is used in combination with the real-time monitored ocean parameters and atmospheric parameters to predict the El Nino / La Nina phenomenon in the designated area. The obtained prediction results are evaluated for accuracy with the Ocean Nino Index, and finally the El Nino / La Nina index prediction model for the next 18 months is obtained. The prediction index for the next 18 months obtained by the El Nino / La Nina prediction model will be used for diagnosis and accuracy evaluation with the Ocean Nino Index. The Ocean Nino Index is based on the ERSSTV5 product of the Climate Prediction Center of the National Oceanic and Atmospheric Administration of the United States, and is calculated by 3.4 El Nino / La Nina identification index obtained by the 3-month moving average of the sea surface temperature anomaly in the region. When the index is greater than 0.5℃ and lasts for more than 5 months, it is determined to be an El Nino event; on the contrary, when the index is less than -0.5℃ and lasts for more than 5 months, it is determined to be a La Nina event. This prediction model can predict the occurrence of El Nino / La Nina events for more than one year, and diagnose and analyze the possible regional or global climate impacts. The output prediction results are compared and evaluated with the Ocean Niño Index, and the correlation coefficient (CC), root mean square error (RMSE), bias (Bias), standard deviation (s), absolute error (MAE), etc. are calculated. When measuring the accuracy of the El Nino / La Nina prediction model, the lower the root mean square error and absolute error values, the higher the accuracy of the prediction model for the El Nino / La Nina identification index; the correlation coefficient result range is between 0 and 1, and the closer it is to 1, the higher the accuracy of the El Nino / La Nina identification index predicted by the prediction model.

[0018] The expression formula of the correlation coefficient is:

[0019]

[0020] The expression formula of root mean square error is:

[0021]

[0022] The deviation is expressed as:

[0023]

[0024] The standard deviation is expressed as:

[0025]

[0026] The absolute error expression formula is:

[0027]

[0028] Where N represents the total number of ocean and atmospheric parameter data grid points involved in the calculation, X r represents the El Nino / La Nina forecast index for the next 18 months obtained through the long short-term memory neural network forecasting model, X o The El Nino / La Nina identification index calculated by the method designed by the Climate Prediction Center, μ is the average.

[0029] The present invention has the following advantages: forming an ocean and atmospheric parameter data set by acquiring ocean parameter and atmospheric parameter data such as historical satellite remote sensing and reanalysis, wherein the historical ocean parameters include sea surface temperature and sea surface height, and the historical atmospheric parameters include meridional wind field, latitudinal wind field and outward long-wave radiation; performing data preprocessing on the historical ocean and atmospheric parameters, wherein the data preprocessing includes outlier removal and unification of spatiotemporal scales, and obtaining an ocean and atmospheric parameter anomaly data set after data preprocessing; performing normalization and standardization on the ocean and atmospheric parameter anomaly data after data preprocessing, and obtaining an ocean and atmospheric parameter data set after normalization and standardization; and performing data preprocessing on the ocean and atmospheric parameter anomaly data after data preprocessing. The method comprises the following steps: constructing prediction factors based on a dataset of ocean and atmosphere parameters, wherein the prediction factors include tropical prediction factors and subtropical prediction factors; dividing the dataset of ocean and atmosphere parameters after the prediction factors are constructed into a training dataset and a test dataset in proportion; training and adjusting the parameters of a long short-term memory neural network through the training dataset to obtain a primary El Nino / La Nina prediction model; testing the primary El Nino / La Nina prediction model after the training and adjustment through the test dataset to obtain an optimal El Nino / La Nina prediction model after the test; and predicting the El Nino / La Nina phenomenon in a designated area by using the optimal El Nino / La Nina prediction model in combination with the ocean parameters and atmospheric parameters monitored in real time. The present invention can obtain the El Nino / La Nina prediction results for the next 18 months, and compare and evaluate with the ocean El Nino / La Nina index, wherein the combination of the tropical Pacific Ocean sea surface temperature and the northwestern Pacific Ocean sea surface temperature anomaly prediction factor settings has better prediction skills in the next 18 months, and the correlation coefficient with the ocean El Nino / La Nina index reaches the highest and passes the significance test, and can monitor the nonlinear interaction characteristics of the ocean and atmosphere during the occurrence and development of El Nino / La Nina, thereby improving the accuracy of long-term prediction analysis of the El Nino / La Nina phenomenon. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for the implementation methods or the prior art descriptions. The structures, proportions, sizes, etc. of the drawings drawn in this specification are only used to match the contents disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the limiting conditions under which the present invention can be implemented, so they have no technical substantive significance. Any structural modification, change in proportional relationship or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical content disclosed by the present invention.

[0031] Figure 1 Flowchart of the long-term prediction method of El Niño / La Niña based on machine learning;

[0032] Figure 2 Network structure diagram constructed for the feature matrix of the El Nino / La Nina long-term prediction method based on machine learning;

[0033] Figure 3 A schematic diagram of the technical implementation roadmap for the long-term El Niño / La Niña prediction method based on machine learning;

[0034] Figure 4 Schematic diagram of the network structure used in the long short-term memory network method;

[0035] Figure 5 A diagram of the network model used for the long short-term memory network method;

[0036] Figure 6 This is a diagram of the prediction model structure used in the long-term El Niño / La Niña prediction method based on machine learning. DETAILED DESCRIPTION

[0037] The following is a description of the implementation of the present invention by specific embodiments. People familiar with the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] Example 1

[0039] See also Figure 1 , Figure 2 and Figure 3 Embodiment 1 of the present invention provides a method for long-term prediction of El Nino / La Nina based on machine learning, comprising the following steps:

[0040] 1) constructing an ocean prediction factor data set, which includes tropical and subtropical prediction factors, namely, tropical key sea area sea surface temperature anomaly, sea surface height anomaly, northwest Pacific sea surface temperature anomaly, North Atlantic sea surface temperature anomaly, South Pacific sea surface temperature anomaly, and reprocessing the above fields to generate an ocean prediction factor database;

[0041] 2) constructing an atmospheric prediction factor data set, which includes tropical and subtropical prediction factors, namely, tropical key sea area meridional wind field anomaly, zonal wind field anomaly, and outward longwave radiation anomaly, and reprocessing the above fields to generate an atmospheric prediction factor database;

[0042] 3) Construct an El Nino / La Nina index prediction model and train the El Nino / La Nina index prediction model using the root mean square error loss function;

[0043] 4) The ocean and atmosphere forecast factors, tropical-subtropical forecast factors, 3.4 The index labels calculated from the three-month moving average of the sea surface temperature anomalies in the key areas are input into the trained El Niño / La Niña index prediction model, and the start and end time, change process, and intensity evolution of the El Niño / La Niña event are predicted through the index prediction model;

[0044] 5) The forecast index for the next 18 months obtained through the El Nino / La Nina index forecast model will be used for diagnosis and accuracy assessment with the Ocean Nino index;

[0045] In this embodiment, historical ocean parameter and atmospheric parameter data related to the occurrence and development of El Nino / La Nina are collected, outliers are eliminated from the historical ocean and atmospheric parameter data, and the spatial and temporal scales are unified to a spatial resolution of 1 degree × 1 degree and a monthly temporal resolution; the anomaly fields of ocean and atmospheric parameters are calculated and normalized and standardized;

[0046] Among them, data standardization is to index the statistical data before data analysis. Data standardization processing includes two aspects: data homogeneity processing and dimensionless processing. Data homogeneity processing mainly solves the problem of data of different properties. Directly summing up indicators of different properties cannot correctly reflect the comprehensive results of different forces. It is necessary to first consider changing the properties of the inverse indicator data so that the forces of all indicators on the evaluation scheme are homogeneous, and then summing them up to get the correct result. Data dimensionless processing mainly solves the comparability of data. Data standardization can adopt related technologies such as "minimum-maximum standardization", "Z-score standardization", "maximum value standardization" and "sum standardization". After data standardization, the original data are converted into dimensionless indicator evaluation values, that is, the values ​​of each indicator are at the same quantitative level, and comprehensive evaluation and analysis can be carried out.

[0047] At the same time, since the various historical ocean parameters and historical atmospheric parameters involved in the calculation have different dimensions, the ocean and atmospheric parameters in the ocean and atmospheric parameter data set after data preprocessing are standardized, including extracting key areas of historical ocean parameters and atmospheric parameters and constructing a prediction factor data set; calculating 3.4 The sea surface temperature in key areas is averaged over three months and the El Niño / La Niña diagnostic index data are obtained as the label data set for the long and short-term memory El Niño / La Niña model.

[0048] Assistance Figure 4 , in the structure diagram of the long short-term memory model, is the multiplication sign, For the plus sign. Among them, in the long short-term memory machine learning model:

[0049] Forget gate: f <t>< / t> =σ(x1*W xf x <t>< / t> +x2*W hf x <t-1>< / t-1> +b f )

[0050] Input gate: i <t>< / t> =σ(x3*W xi x <t>< / t> +x4*W hi x <t-1>< / t-1> +b i )

[0051] Output gate: o <t>< / t> =σ(x5*W xo x <t>< / t> +x6*W ho h <t-1>< / t-1> +b o )

[0052] Neuron: c <t>< / t> =f <t>< / t> ⊙c <t-1>< / t-1> +i <t>< / t> ⊙tanh(W xc x <t>< / t> +W hc h <t-1>< / t-1> +b c )

[0053] Objective function: h <t>< / t> =o <t>< / t> tanh⊙(c <t>< / t> )

[0054] Where σ is the sigmoid function, ranging from [0, 1], and tanh is the tanh function, which is used as the activation function of the standard neural network. <t>< / t> is the input data. <t-1>< / t-1> is the historical data calculated by the previous neuron. <t>< / t> The data calculated for this neuron. <t-1>< / t-1> is the memory data of the previous neuron. <t>< / t> The memory data calculated by this neuron. W xf , W hf is the weight matrix of the forget gate, b f is the bias term of the forget gate. xi , W hi is the weight matrix of the input gate, b i is the bias term of the input gate. W xo , W ho is the weight matrix of the output gate, b o is the bias term of the output gate. x1~x6 are adaptive coefficients, which are adjusted according to the amount of data and parameter configuration of the training field.

[0055] Specifically, the input gate can calculate a value between 0 and 1 based on the current input and the output at the previous moment. When this value is close to 1, it means that the long short-term memory machine learning model should remember the current input information, so that important data can be marked to ensure that it is not easily forgotten in subsequent calculations.

[0056] Specifically, the forget gate calculates a value between 0 and 1 based on the current input and the output of the previous moment. If this value is close to 0, it means that the LSTM machine learning model will forget the previous memory. In this way, the LSTM machine learning model can clear out no longer important information in a timely manner, allowing the network to adapt to changes in different time periods.

[0057] Specifically, the output gate calculates a value between 0 and 1 based on the current input and the output at the previous moment. When this value is close to 1, it means that the long short-term memory machine learning model should output the current memory, so that information can be selectively passed on as needed.

[0058] In this embodiment, the ocean and atmospheric parameter data sets after the prediction factors are constructed are divided into training data sets and test data sets in proportion according to preset rules; wherein the training data set of ocean and atmospheric parameters including satellite remote sensing data covers the period from 1982 to 2013, and the test data set covers the period from 2014 to 2021. In addition, experiments are conducted on CMIP5 climate model data, with the training data set covering the period from 1951 to 2013 and the test data set covering the period from 2014 to 2021.

[0059] The training data set is input into the long short-term memory model for repeated iterations until the loss function is minimized and the correlation coefficient with the target prediction value, the ocean El Nino / La Nina prediction index, is the highest, thereby obtaining the optimal El Nino / La Nina prediction model.

[0060] In this embodiment, the obtained El Nino / La Nina prediction index is evaluated:

[0061] The performance of the diagnostic index was evaluated using the correlation coefficient, root mean square error, bias, and standard deviation:

[0062] The expression formula of the correlation coefficient is:

[0063]

[0064] The expression formula of root mean square error is:

[0065]

[0066] The deviation is expressed as:

[0067]

[0068] The standard deviation is expressed as:

[0069]

[0070] The absolute error expression formula is:

[0071]

[0072] Where N represents the total number of ocean and atmospheric parameter data grid points involved in the calculation, X r represents the El Nino / La Nina forecast index for the next 18 months obtained through the long short-term memory neural network forecasting model, X o The El Nino / La Nina identification index calculated by the method designed by the Climate Prediction Center, μ is the average.

[0073] See also Figure 5 and Figure 6 In this embodiment, the specific structure of the long short-term memory network model used is as follows:

[0074] The input layer consists of two parts, one is the key factors and combinations of ocean and atmosphere in key sea areas, and the other is the label vector of the ocean El Niño / La Niña index; the key factors of ocean and atmosphere in key sea areas are activated through the tanh function, and the fully connected layer is followed by a Dropout layer with a probability of 0.5.

[0075] Including sea surface temperature forecast factors for key sea areas, covering 4 and 3 The sea surface temperature anomaly in the key area (160°E-90°W, 5°S-5°N), the sea surface height prediction factor is the sea surface height anomaly (160°E-90°W, 5°S-5°N), the zonal wind prediction factor is the zonal wind anomaly (160°E-90°W, 5°S-5°N), the meridional wind prediction factor is the meridional wind anomaly (160°E-90°W, 5°S-5°N), and the outward longwave radiation prediction factor is the outward longwave radiation anomaly (1 The forecast factor for the North Atlantic is the tropical Pacific SST and the North Atlantic SST anomaly (EQ-15°N, 90°W-20°E), the forecast factor for the North Pacific is the tropical Pacific SST and the North Pacific SST anomaly (90°E-150°W, 10°S-30°N), and the forecast factor for the South Pacific is the tropical Pacific and South Pacific SST anomalies (150°E-75°W, 20°S-65°S).

[0076] Among them, the training data set is input into the long short-term memory network machine learning model for training and parameter debugging to obtain the optimal long short-term memory machine learning model after training; in order to deal with the overfitting problem, the dropout rate is 0.1, 0.2, 0.3, 0.4, and 0.5 respectively, the number of iterations is 100, 120, 150, 180, and 200 respectively, and the cell number is set to 24, 36, and 48. When the number of iterations is 100, the dropout rate is 0.5, and the cell number is 24, the El Niño / La Niña prediction model results are optimal.

[0077] In summary, the embodiment of the present invention obtains historical ocean and atmospheric parameter data, wherein the historical ocean parameters include sea surface temperature and sea surface height, and the historical atmospheric parameters include meridional wind field, latitudinal wind field and outward long wave radiation; performs data preprocessing on the ocean and atmospheric parameters, wherein the data preprocessing includes outlier removal and unification of spatiotemporal scales, and obtains anomaly calculation of the ocean and atmospheric parameter data set after data preprocessing; performs normalization and standardization on the ocean and atmospheric parameter anomaly data to obtain the normalized and standardized ocean and atmospheric parameter data set; performs prediction factor construction on the ocean and atmospheric parameter data set The prediction factors include tropical prediction factors and subtropical prediction factors; the ocean and atmospheric parameter data sets after the prediction factors are constructed are divided into training data sets and test data sets in proportion; the long short-term memory neural network is trained and adjusted by the training data set to obtain a primary El Nino / La Nina prediction model, and the prediction model is trained by the root mean square error loss function to obtain the El Nino / La Nina long short-term memory neural network prediction model after parameter adjustment optimization; the occurrence and development process of El Nino / La Nina in the designated area is predicted by combining the optimal El Nino / La Nina prediction model with the ocean parameters and atmospheric parameters monitored in real time. The present invention can obtain the El Nino / La Nina prediction results for the next 18 months, and compare and evaluate with the ocean Nino index, wherein the combination of the tropical Pacific sea surface temperature and the northwest Pacific sea surface temperature has better prediction skills in the next 18 months, and the correlation coefficient with the ocean El Nino / La Nina index is the highest. Through the correlation significance test, the nonlinear interaction characteristics of the ocean and the atmosphere during the occurrence and development of El Nino / La Nina can be monitored, and the accuracy of the long-term prediction analysis of the El Nino / La Nina phenomenon can be improved.

[0078] It should be noted that the method of the embodiment of the present disclosure can be performed by a single device, such as a computer or a server. The method of the present embodiment can also be applied in a distributed scenario and completed by multiple devices cooperating with each other. In the case of such a distributed scenario, one of the multiple devices can only perform one or more steps in the method of the embodiment of the present disclosure, and the multiple devices will interact with each other to complete the described method.

[0079] It should be noted that the above describes some embodiments of the present disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the above embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A long-term prediction method for El Nino / La Nina based on machine learning, characterized in that: include: 1) Construct a long-term prediction model of El Nino / La Nina with a multi-layer long short-term memory neural network, initialize the long short-term memory neural network prediction model, and train the model through the root mean square error loss function; extract the characteristics of ocean and atmospheric physical quantities at multiple spatiotemporal scales, design a combination of prediction factors for ocean and atmospheric sensitive areas, and use the physical mechanism constraints of the interaction between the ocean and the atmosphere to strengthen the consistency of the characteristics of different ocean and atmospheric physical quantities in predicting the El Nino / La Nina index, and use the characteristics of ocean-atmosphere interaction to predict the El Nino / La Nina index; 2) In the stage of ocean and atmospheric parameter data preprocessing, an ocean and atmospheric parameter data set is constructed. The historical ocean parameters include sea surface temperature and sea surface height, and the historical atmospheric parameters include meridional wind field, zonal wind field and outward long-wave radiation physical quantity. The data preprocessing includes outlier removal and unified spatiotemporal scale processing to obtain a long-time series ocean and atmospheric parameter data set after data preprocessing; the ocean and atmospheric parameter prediction factors in the ocean and atmospheric parameter data set after data preprocessing are subjected to anomaly calculation and normalization processing, and the features of each data are zero-centered using a scaler, and their numerical ranges are normalized to the range of [-1,1] to accelerate the convergence efficiency, and obtain the normalized ocean and atmospheric parameter prediction factor data set; on this basis, standardization processing is performed; 3) In the forecast factor combination setting stage, a forecast factor dataset of ocean and atmospheric parameters is constructed, and key areas of ocean and atmospheric parameters that are sensitive to El Niño / La Niña changes are extracted to construct El Niño / La Niña prediction factors; a forecast factor combination dataset of El Niño / La Niña index is constructed, in which tropical forecast factors include sea surface temperature anomaly, sea surface height anomaly, meridional wind field anomaly, zonal wind field anomaly and outward longwave radiation anomaly field, and subtropical forecast factors include northwest Pacific sea surface temperature anomaly, South Pacific sea surface temperature anomaly, and subtropical forecast factors. The forecast factor combination includes tropical Pacific sea surface temperature anomaly, sea surface height anomaly, outward longwave radiation anomaly, zonal wind field anomaly, meridional wind field anomaly, tropical Pacific and northwest Pacific sea surface temperature anomaly combined forecast factor, tropical Pacific and North Atlantic sea surface temperature anomaly combined forecast factor, tropical Pacific and South Pacific sea surface temperature anomaly combined forecast factor, designed according to the El Niño / La Niña identification index method provided by the Climate Prediction Center, and calculated the same time range as the forecast factor. 3.4 The three-month moving average sea surface temperature anomaly index in the region is used as a marker data for diagnosing the occurrence of El Niño / La Niña events; 4) In the long short-term memory network training phase, the ocean atmospheric parameter prediction factor data set is divided into a training data set and a test data set for the machine learning prediction model in proportion, and the label data of the El Nino / La Nina discrimination index is divided into a training data set and a test data set in proportion, corresponding to the training data set and the test data set of the prediction factor; The input layer of the El Nino / La Nina index prediction model consists of two parts: one is the ocean and atmospheric parameter prediction factors, the input nodes of the hidden layer are 361, and the hidden nodes are 24; the other is 3.4 El Nino / La Nina identification index of three-month moving average in key areas, the output node is 18, the ocean and atmosphere prediction factors are passed through a fully connected layer with activation function tanh, and the fully connected layer is followed by a Dropout layer with a Dropout layer of 0.5 to prevent overfitting. The model loss function uses the root mean square error for gradient descent iterative training. The input gate, forget gate, and output gate functions need to adjust the weight coefficient, which is 20 by default. The weight adjustment ratio for each iteration of network training is 0.01; the error threshold is 0.000001, and the error loss function result graph is drawn to verify that the prediction result gradually converges and stabilizes to the minimum value, which is the optimal model result; 5) In the LSTM network testing phase, the test data set is input into the El Nino / La Nina LSTM neural network prediction model for testing, and the optimal El Nino / La Nina prediction model and prediction results after the test are obtained; the optimal El Nino / La Nina prediction model is used in combination with the real-time monitored ocean parameters and atmospheric parameters to predict the El Nino / La Nina phenomenon in the designated area, and the optimized LSTM neural network model is used to predict the test data. The obtained prediction results are evaluated for accuracy with the ocean Niño index, and the correlation coefficient, root mean square error, bias, standard deviation, and absolute error are calculated to obtain the El Nino / La Nina index prediction model for the next 18 months.

2. The method for long-term prediction of El Nino / La Nina based on machine learning according to claim 1, characterized in that: Initialize the long short-term memory neural network prediction model and start training, and train the model through the root mean square error loss function; input the preprocessed historical ocean and atmospheric forecast factor data set and the label data of the El Nino / La Nina diagnostic index into the long short-term memory neural network El Nino / La Nina prediction model respectively; generalize the long short-term memory neural network prediction model, arrange the input and output of the data separately, transform the input and pass it into the prediction function for step-by-step prediction, inversely scale and inversely differentiate the obtained prediction value, and restore it to the original numerical range, traverse the data of the entire data set, and save the calculated prediction index data; predict the exponential change characteristics of the time, amplitude intensity, and development process of the El Nino / La Nina event through the long short-term memory neural network El Nino / La Nina prediction model.

3. The method for long-term prediction of El Nino / La Nina based on machine learning according to claim 2, characterized in that: In the unified spatiotemporal scale process of the ocean and atmospheric parameter data preprocessing, outliers are eliminated and the spatiotemporal scale is unified, the ocean atmospheric parameter data in the ocean and atmospheric parameter data set are uniformly processed into a spatial resolution of 1 degree × 1 degree and a monthly temporal resolution, and anomaly calculations and normalization operations are performed to obtain 8 ocean and atmospheric parameter data sets after normalization. At the same time, since the various historical ocean parameters and atmospheric parameters involved in the calculation have different dimensions, the ocean and atmospheric parameters in the ocean and atmospheric parameter data set after data preprocessing are standardized, and then the ocean atmospheric parameter data after data standardization are analyzed.

4. The method for long-term prediction of El Nino / La Nina based on machine learning according to claim 3, characterized in that: The forecast factor combination model design stage is characterized by analyzing the normalized and standardized ocean and atmospheric parameter data, including: extracting key areas sensitive to El Nino / La Nina changes from historical ocean and atmospheric parameters, where the definition and range of the sea surface temperature index of the tropical forecast factor is 4 and 3. Sea surface temperature anomalies in key areas. The sea surface height prediction factor is the sea surface height anomalies in the range of 160°E to 90°W and 5°N and S. The zonal wind prediction factor is the zonal wind anomalies in the range of 160°E to 90°W and 5°N and S. The meridional wind prediction factor is the meridional wind anomalies in the range of 160°E to 90°W and 5°N and S. The outward longwave radiation prediction factor is the outward longwave radiation anomalies in the range of 160°E to 160°W and 5°N and S. The North Atlantic prediction factor is the tropical Pacific SST anomalies and the North Atlantic SST anomalies. The temperature anomaly covers the equator to 15°N, 90°W to 20°E. The prediction factors for the Northwest Pacific are the tropical Pacific SST and the Northwest Pacific SST anomaly. The Northwest Pacific SST anomaly covers 90°E to 150°W and 10°S to 30°N. The prediction factors for the South Pacific are the tropical Pacific SST and the South Pacific SST anomaly. The South Pacific SST anomaly covers 150°E to 75°W and 20°S to 65°S. The El Nino / La Nina index of the interaction between the ocean and the atmosphere is statistically calculated to obtain the ocean and atmosphere prediction factor data set. 3.4 After calculating the anomaly and sliding three-month average of the sea surface temperature in the key area, the El Niño / La Niña diagnostic identification index data is obtained, which is used as the label data set for the long short-term memory neural network El Niño / La Niña prediction model.

5. The method for long-term prediction of El Nino / La Nina based on machine learning according to claim 5, characterized in that: The LSTM testing phase is characterized in that the ocean and atmospheric parameter data are divided into two parts, a training data set and a test data set, in proportion; the length of the training data set accounts for 85% of the total data, and the length of the test data set accounts for 15% of the total data; the label data of the ocean and atmospheric parameter prediction factors and the El Nino / La Nina index correspond to each other in the training data and the test data; the test data is input into the trained LSTM El Nino / La Nina index prediction model, and the prediction results are correlated with the ocean Nino index. The root mean square error, bias, standard deviation, and absolute error accuracy assessment are used to obtain the El Nino / La Nina index for the next 18 months. The index reflects the start and end time, change process, and intensity evolution of the El Nino / La Nina event. When the index is greater than 0.5℃ and lasts for more than 5 months, it is determined to be an El Nino event; on the contrary, when the index is less than -0.5℃ and lasts for more than 5 months, it is determined to be a La Nina event. The prediction model can predict the occurrence of El Nino / La Nina events for more than one year, and thereby diagnose and analyze the possible regional or global climate impacts. The expression formula of the correlation coefficient is: The expression formula of root mean square error is: The deviation is expressed as: The standard deviation is expressed as: The absolute error expression formula is: Where N represents the total number of ocean and atmospheric parameter data grid points involved in the calculation, X r represents the El Nino / La Nina forecast index for the next 18 months obtained through the long short-term memory neural network forecasting model, X o The El Nino / La Nina identification index calculated by the method designed by the Climate Prediction Center, μ is the average.

Images