ENSO intelligent prediction method, device, equipment and storage medium

Abstract

The present invention relates to an intelligent ENSO prediction method, apparatus, device, and storage medium for long-term ENSO forecasting. The method includes the construction, training, and prediction processes of an intelligent ENSO prediction model based on a segmented deep neural network, as well as the description of an intelligent ENSO prediction device. By separately modeling the dominant characteristics of different ENSO evolutionary stages during the deep learning modeling process, the method achieves prediction accuracy superior to traditional numerical models and greater stability and interpretability than traditional deep learning models. Compared with existing technologies, the method offers advantages such as ease of construction, high accuracy, and resource conservation. It can be used for operational ENSO forecasting, effectively improving both the accuracy and duration of ENSO predictions.

Description

Technical Field

[0001] The present invention relates to the field of meteorological forecasting technology, and in particular to an ENSO intelligent forecasting method, device, equipment and storage medium. Background Art

[0002] El Niño-Southern Oscillation The ENSO (Environmental Southern Oscillation) is a coupled ocean-atmosphere phenomenon on an interannual timescale in the tropical Pacific Ocean. It exhibits a quasi-periodic oscillation lasting 2 to 7 years and is the strongest interannual variability signal in Earth's climate system. It has been shown to interact with other climate phenomena, such as the Persian Horizontal Oscillation (PDO), Indo-Pacific Horizontal Oscillation (IDO), and the Meteoric Horizontal Oscillation (MJO), significantly impacting weather and climate anomalies in my country. Accurate, timely, and effective prediction of the onset, development, and evolution of ENSO is of great scientific and practical significance.

[0003] Numerical models are currently the primary technology for ENSO prediction both domestically and internationally. Numerical models generally rely on a description of physical processes; as long as they can reasonably describe these processes, they can make reasonable ENSO forecasts. Therefore, ENSO prediction research based on numerical models often relies on improvements in meteorological research, such as accurate model initialization, parameterization of physical processes, and air-sea coupling models. However, due to the current lack of clarity among meteorologists regarding the physical mechanisms of ENSO-related atmospheric and oceanic processes and their interactions, as well as the fact that the accuracy of numerical model predictions is heavily dependent on initial environmental conditions and errors in model equations, errors in the simulation of relevant physical processes can occur in numerical models. Consequently, after a certain integration period, the prediction skill of ENSO degrades significantly, predictability is lost, and prediction accuracy is suboptimal. Summary of the Invention

[0004] The purpose of the present invention is to provide an ENSO intelligent prediction method, device, equipment and storage medium in order to overcome the defects of the above-mentioned prior art.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] The present invention relates to an ENSO intelligent prediction method based on a segmented deep neural network, which comprises the following steps:

[0007] S1: Based on the unbalanced characteristics of the ENSO evolution process, an ENSO intelligent prediction model based on a segmented deep neural network is constructed.

[0008] S2: Train the constructed ENSO intelligent prediction model based on segmented deep neural network.

[0009] S3: Determine the model input data and preprocess the input data, and use the preprocessed data to predict ENSO through the ENSO intelligent prediction model based on the segmented deep neural network.

[0010] Furthermore, the specific steps of building an ENSO intelligent prediction model based on a segmented deep neural network include:

[0011] S11: Select the required ENSO prediction length and divide the ENSO prediction into two sections: the front and the back according to the segmentation positioning model;

[0012] S12: Based on the different evolution processes and ocean-atmosphere characteristics of the first and second phases of ENSO, a dedicated encoder and decoder for the first and second phases are designed based on a deep learning model for spatiotemporal series prediction.

[0013] S13: Based on the different evolution processes and ocean-air characteristics of the first and second phases of ENSO, a weighted feature fusion model is constructed for the first and second phases respectively based on the deep learning model of spatiotemporal series prediction;

[0014] S14: Create an ENSO intelligent prediction model based on a segmented deep neural network based on the structure built in S12 and S13.

[0015] The expression of the segmentation and positioning model is:

[0016]

[0017] Where x t is the ENSO prediction dataset, t seg The dividing position of the front and back sections, is the floor rounding function, seg(·) is the model used to calculate the front and back segment segmentation positions, f seg is the slope of seg(·).

[0018] The expression of the ENSO intelligent prediction model based on the segmented deep neural network is:

[0019]

[0020]

[0021]

[0022] In the formula, concat(·) means concatenating the features of different segments together, encoder i (·) and decoder i (·) are respectively the encoder and decoder designed based on the spatiotemporal sequence prediction deep learning model in S12, i(·) Weighted feature fusion designed for S13, is the output of this segment decoder, is the weighted fusion feature obtained by the weighted feature fusion device, and They are the encoder outputs of this segment and another segment respectively.

[0023] Furthermore, the specific steps of S2 include:

[0024] S21: Based on the ocean-air characteristics of ENSO, select variables that play a key role in ENSO prediction and construct an ENSO prediction dataset;

[0025] S22: Perform quality control and data preprocessing on different physical variables in the ENSO prediction dataset;

[0026] S23: Divide the preprocessed dataset into a training set, a validation set, and a test set, and use the training set to train the constructed ENSO intelligent prediction model.

[0027] Furthermore, the specific steps of S3 include:

[0028] S31: Select the reporting start time based on actual business needs, obtain the initial time data, and perform quality control and data preprocessing. The processed results serve as the input of the ENSO intelligent prediction model based on the segmented deep neural network;

[0029] S32: Input the processed initial moment data into the front-end and back-end encoders of the segmented deep neural network ENSO intelligent prediction model for feature encoding to determine the segment features of ESNO;

[0030] S33: The features after the front and rear segment encoding are spliced ​​and input into their respective weighted feature fusion devices for feature fusion to obtain weighted fusion features;

[0031] S34: The weighted fusion features of the front and back segments are respectively input into the front and back segment decoders of the segmented deep neural network ENSO intelligent prediction model for decoding, and data post-processing is performed to obtain the ENSO grid prediction results. Long-term ENSO prediction is performed using a non-iterative prediction method.

[0032] S35: Calculate various ENSO-related indices and evaluation indicators based on the prediction results of the ENSO intelligent prediction model based on the segmented deep neural network.

[0033] The calculation formula of the weighted feature fusion is:

[0034]

[0035] Where, is the weighted fusion feature, and They are the encoder outputs of this segment and another segment respectively, and each segment constructs its own weighted feature fusion merge th (·).

[0036] Another aspect of the present invention provides an ENSO intelligent prediction device based on a segmented deep neural network, the device comprising:

[0037] Initial data acquisition module, which obtains the initial forecast time, initial data and forecast duration of ENSO;

[0038] The data preprocessing module converts the acquired initial moment data into the input format specified by the model, and the converted data is used as input data;

[0039] The prediction module transmits input data to the ENSO intelligent prediction model based on the segmented deep neural network, runs the model, and obtains the prediction results;

[0040] The result acquisition module processes the model output and converts the output data into meteorological grid data and various ENSO indices and quantitative evaluation indicators.

[0041] A third aspect of the present invention provides a computer device, comprising: a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein when the computer program is executed by the processor, the steps of the above-mentioned ENSO intelligent prediction method based on a segmented deep neural network are implemented.

[0042] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-mentioned segmented deep neural network ENSO intelligent prediction method.

[0043] The ENSO intelligent prediction method, apparatus, device, and storage medium provided by the present invention have at least the following advantages over the prior art:

[0044] 1) Convenient model building: Instead of relying on physical equations to build models, the model is built using structured deep learning modules, which simplifies the model building process and saves model building time.

[0045] 2) Save time and computational costs: This paper avoids using traditional numerical models. Instead, it uses models based on deep learning technology and graphics computing equipment to complete long-term ENSO forecasts, which have the characteristics of high computational efficiency and strong real-time performance.

[0046] 3) Highly accurate prediction results: The segmented deep neural network designed in this paper is tailored to the different main characteristics of different ENSO stages. It can use different network structures for feature extraction in different stages. In addition, the weighted feature fuser in the ENSO intelligent prediction model can also supplement the secondary features of each segment, which can efficiently and comprehensively extract the evolution characteristics of ENSO in different segments, providing stable and reliable prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Schematic diagram of the process of the ENSO intelligent prediction method based on the segmented deep neural network in the embodiment;

[0048] Figure 2 Schematic diagram of the architecture of the ENSO intelligent prediction model based on a segmented deep neural network in the embodiment;

[0049] Figure 3 Schematic diagram of the structure of an ENSO intelligent prediction device based on a segmented deep neural network in an embodiment;

[0050] Figure 4 Schematic diagram of “multi-input and multi-output” prediction using an ENSO intelligent prediction device based on a segmented deep neural network in an embodiment;

[0051] Figure 5 Schematic diagram of the structure of the computer device in the embodiment. DETAILED DESCRIPTION

[0052] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments described are only a portion of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

[0053] Example

[0054] ENSO is a significant signal of interannual and interdecadal climate change worldwide and has been shown to interact with other atmospheric oscillations, such as the Indian Ocean Dipole and the Tropical Intraseasonal Oscillation. Therefore, it is considered a key influencing factor in climate forecasting. Effectively predicting the occurrence and evolution of ENSO is of great scientific and practical significance. Currently, meteorological researchers generally use numerical models to predict ENSO. Numerical models calculate the values ​​of various variables using various known physical formulas and then integrate them over time to achieve predictions. However, this approach has gradually reached a bottleneck. Problems include poor predictive lead times and difficulty in predicting extreme ENSO events. The fundamental reason is that the ability of numerical models to predict ENSO is limited by theoretical research on ENSO's mechanisms. ENSO is a highly complex climate phenomenon, and much remains to be understood about its mechanisms in the meteorological field.

[0055] As black-box models, deep neural networks (DNNs) possess the ability to automatically extract complex relationships from data. Driven by sufficient data, they can identify structures and patterns in physical systems with complex interference. In recent years, with the advancement of data collection capabilities from satellites, radar, and other sources, as well as subsequent improvements in data analysis and processing techniques, DNNs have been widely used in ENSO forecasting. DNN models overcome some of the shortcomings of numerical model prediction methods, improving the skill and accuracy of ENSO forecasts and demonstrating broad application prospects.

[0056] Based on this, the present invention provides an ENSO intelligent prediction method based on a segmented deep neural network. This method constructs a data-driven ENSO intelligent prediction model based on a segmented deep neural network. The ENSO intelligent prediction model designed in this embodiment of the application is specifically designed to address the imbalances in the ENSO evolution process, thereby enhancing the deep learning model's ability to accurately mine the dominant characteristics of different ENSO phases, thereby improving the accuracy of operational forecasts.

[0057] The imbalance in ENSO's evolution refers to the constantly changing role and importance of various physical processes and characteristics in the onset and extinction of ENSO. A physical process or characteristic that dominates in one phase may have a minimal impact in another; even processes and characteristics that promote evolution in one phase may inhibit it in another. For example, one of the most important physical processes in ENSO's onset is the Bjerknes positive feedback. This occurs when westerly wind anomalies in the equatorial Pacific weaken the flow of warm water into the western Pacific. This weakens the upward compensating flow in the eastern Pacific, raising sea temperatures in the eastern Pacific, weakening the east-west sea temperature gradient, and further weakening the Walker circulation. This leads to the development of an EINo. During ENSO's extinction, delayed oscillations play a major role. When westward-propagating Rossby waves encounter the western Pacific boundary, they are reflected as eastward-propagating Kelvin waves. These waves propagate to the eastern Pacific, causing a phase shift in ENSO. Of course, many other physical processes play varying roles in the onset and extinction of ENSO. It can be seen that the evolution of ENSO is unbalanced and uneven, with different stages subject to different physical mechanisms. Therefore, the embodiments of this application utilize the above-mentioned unbalanced characteristics of the ENSO evolution process to construct a segmented deep neural network based on a spatiotemporal series prediction model, and then construct an ENSO intelligent prediction model and device, which can effectively improve the ENSO prediction skills and accuracy. The physical mechanisms and processes described here are some embodiments of this application, not all embodiments.

[0058] The following describes some embodiments of the present application in detail with reference to the accompanying drawings. In the absence of conflict, the following embodiments and features therein may be combined with each other.

[0059] Figure 1 A schematic diagram of a process flow of an ENSO intelligent prediction method based on a segmented deep neural network provided in an embodiment of the present application, which includes the following three contents (A, B and C):

[0060] A. The construction of the ENSO intelligent prediction model includes the following steps:

[0061] A1) Select the desired ENSO forecast length and divide the ENSO forecast into two segments: the front and the back, according to the following segmentation and positioning model;

[0062] like Figure 2 Module A.1 is a segmentation and positioning model module provided in the embodiment of the present application. The overall calculation formula of this module is:

[0063]

[0064] Among them, x t (t is the different step lengths of ENSO prediction, x is a data set composed of variables such as sea surface temperature, surface salinity, meridional wind, zonal wind, rainfall, and thermocline depth. The selected variables are part of the embodiments of this application, not all of them) is the ENSO prediction data set, and its output value is, t seg The dividing position of the front and back sections, is a floor function, and seg(·) is a model for calculating the position of the front and back segments, f seg is the slope of seg(·), Figure 2 The embodiment in provides a method for constructing a segmentation and positioning model seg(·), namely, using stacked ConvLSTM layers to build it.

[0065] The calculation formula of ConvLSTM is as follows:

[0066] f conv =f t (f c (input), f h (input))

[0067] Among them, f c Represents the convolution operation unit used to calculate memory information, f h Represents the convolution operation unit for feature extraction, f t The ConvLSTM model represents the information transfer and update process. It accumulates memory information through continuous iteration over time and continuously integrates this memory information into extracted features to obtain features at different times. This allows it to sensitively capture the shifts in dominant features during ENSO evolution, making it ideal for use as a segmentation and positioning model for ENSO forecasting.

[0068] A2) Based on the different evolution processes and ocean-atmosphere characteristics of the early and late stages of ENSO, dedicated encoders for the early and late stages are designed based on a deep learning model for spatiotemporal series prediction.

[0069] Preferably, the present invention uses multi-layer STLSTM to construct the encoder and decoder of the front end, and uses multi-layer CauslLSTM to construct the encoder and decoder of the back end.

[0070] STLSTM uses a unique serial structure to enhance the acquisition of spatiotemporal memory information. The calculation formula of this module is as follows:

[0071] f st =f h (f c (input), f m (input))

[0072] Among them, f c Represents the convolution operation unit used to calculate the time memory state, f m Represents the convolution operation unit used to calculate the spatial memory state, f h The STLSTM represents an arithmetic unit for integrating temporal and spatial memory. By capturing both temporal and spatial memory states, it can fully retain memory information and is more suitable for short-term prediction. Therefore, this work uses it as the front-end encoder and decoder.

[0073] CauslLSTM has made some improvements based on STLSTM. Its calculation formula is as follows:

[0074] f causal =f h (f s (f c (input), f m (input)))

[0075] Among them, f causal That is the output value of CauslLSTM, f c 、f m With f h The sparse_scaling operation (fs) represents the same operational unit as STLSTM, while the sparse_scaling operation (fs) is used to remove interference introduced during long-term memory transfer and retain valid memory information. This sparse_scaling operation significantly enhances long-term prediction accuracy, and is therefore chosen for the encoder and decoder in the latter stages.

[0076] A3) construct weighted feature fusion devices for the front and back segments respectively;

[0077] like Figure 2 Module A.3 is a weighted feature fusion module provided in the embodiment of the present application. The overall calculation formula of this module is:

[0078]

[0079] in, is the weighted fusion feature, and They are the encoder outputs of this segment and another segment respectively, and each segment constructs its own weighted feature fusion merge th (·), the embodiment of the present application is built using a deep learning module.

[0080] Here α is a weighted matrix generated based on the features themselves, which enables the feature fusion process to screen for effectiveness. The calculation formula of the weighted matrix is ​​as follows:

[0081]

[0082] Preferably, after processing according to step A2) and step A3), a model is constructed; that is, the overall calculation formula of the segmented deep neural network model composed of step A2) and step A3) is:

[0083]

[0084]

[0085]

[0086] Among them, concat(·) means to splice the features of different segments together, encoder i (·) and decoder i (·) are the encoder and decoder designed based on the spatiotemporal sequence prediction deep learning model in step A2), is the output of this segment decoder, i (·) is the weighted feature fusion device designed in step A3).

[0087] B. Training of the ENSO intelligent prediction model includes the following steps:

[0088] B1) Based on the ocean-air characteristics of ENSO, select variables that play a key role in ENSO prediction and construct an ENSO prediction dataset.

[0089] The data used can include multiple sources, including satellite remote sensing data, numerical model data, and reanalysis data. For example, these data cover key physical variables (initial meteorological data) such as sea surface temperature, sea surface salinity, meridional winds, zonal winds, rainfall, and thermocline depth. Data enhancement and quality control algorithms are designed to address the challenges of insufficient ENSO data and low-quality related data.

[0090] The initial meteorological data may be observed meteorological data or initial meteorological data obtained through a third-party service, for example, initial meteorological data predicted by the Global Forecasting System (GFS), such as OISST_v2 sea surface temperature data, TMI rainfall data, NCEP / NCAR wind field data, etc.

[0091] Preferably, the range of the physical element grid data in the embodiment of the present application is (90°N-90°S, 0°-180°) with a resolution of 1°.

[0092] B2) Perform quality control and data preprocessing on different physical variables in the ENSO prediction dataset;

[0093] Preferably, the embodiment of the present application uses the following formula to perform data preprocessing on different sea-air datasets:

[0094]

[0095] Among them, x min with x max are the minimum and maximum values ​​of the grid data of each sea and air element in the initial meteorological data, respectively. * This is the result after data preprocessing.

[0096] B3) Use the first 80% of the data in the dataset as the training set, the middle 10% of the data as the validation set, and the last 10% of the data as the test set.

[0097] Preferably, the embodiment of the present application uses the following formula as the cost function l for model training to guide the model to be fully trained:

[0098]

[0099] Among them, (i, j)∈Ω represents each grid point of the physical variable field, is the prediction result of the corresponding variable at the grid point (i, j), s i，j is the actual observation data at the grid point (i, j), N is the number of selected physical variables, MSE is the mean square error, and MAE is the mean absolute error.

[0100] C. The prediction of the ENSO intelligent prediction model includes the following steps:

[0101] C1) Select the reporting start time based on actual operational needs, obtain initial time data, and perform the same quality control and data preprocessing as in step B2). The processed results serve as input to the segmented deep neural network ENSO intelligent prediction model;

[0102] C2) Inputting the processed initial moment data into the front-end and back-end encoders of the segmented deep neural network ENSO intelligent prediction model for feature encoding to determine the segmental features of ESNO;

[0103] C3) The features after the front and rear segment coding are spliced ​​and input into their respective weighted feature fusion devices for feature fusion to obtain weighted fusion features;

[0104] C4) The weighted fusion features of the front and back segments are input into the decoders of the front and back segments of the segmented deep neural network ENSO intelligent prediction model for decoding, and data post-processing is performed to obtain the ENSO grid prediction results. Long-term ENSO prediction is performed using a non-iterative prediction method.

[0105] Preferably, according to the data preprocessing in step B2), the embodiment of the present application uses the following formula to perform the inverse process to obtain an understandable prediction result:

[0106]

[0107] in, This is the prediction result.

[0108] C5) Calculate various ENSO-related indices and evaluation indicators based on the prediction results of the segmented deep neural network ENSO intelligent prediction model.

[0109] Preferably, the embodiment of the present application utilizes and The index identifies the occurrence of an ENSO event, as well as its intrinsic properties such as type, intensity, and duration. It is calculated by taking the average of the sea surface temperature anomalies within a fixed region. The index covers the range (5°N-5°S, 150°W-90°W). The index covers the range (5°N-5°S, 160°E-150°W). The index covers the range of (5°N-5°S, 170°W-120°W).

[0110] When constructing the ENSO intelligent prediction model, the embodiment of the present application explicitly incorporates the prior ENSO physical mechanism. In particular, in the process of constructing a segmented deep neural network based on the spatiotemporal series prediction model, the different characteristics of the dominant physical mechanisms in the occurrence and disappearance of ENSO are taken into account. Compared with the general and naive ENSO deep learning prediction model (such as using only a single model or using multiple unrelated models for prediction), it has better pertinence and advancement, can effectively improve the accuracy of ENSO prediction, and improve ENSO prediction skills. At the same time, from the perspective of model operation, the deep learning model designed in the embodiment of the present application can save a lot of computing resources compared to the traditional dynamic model.

[0111] On the other hand, the embodiment of the present invention also provides an ENSO intelligent prediction device based on a segmented deep neural network, such as Figure 3 As shown, the device includes:

[0112] D.1. Initial data acquisition module, used to obtain the initial forecast time, initial data and forecast duration, etc.

[0113] Preferably, the embodiment of the present application verifies that when the input sequence length of each physical variable is 9 months, the performance and resource usage of the model reach an optimal balance.

[0114] D.2. Data preprocessing module, used to convert the initial data into the input format specified by the model;

[0115] D.3. Prediction module, used to transmit the converted input data to the ENSO intelligent prediction model and run the model to obtain prediction results;

[0116] D.4. Post-processing module, used to post-process the model output and convert the output data into understandable meteorological grid data and various ENSO indices and quantitative evaluation indicators.

[0117] Preferably, the embodiment of the present application adopts a non-iterative prediction strategy for long time series prediction, and its behavior is as follows Figure 4 As shown in Figure 2, the forecast results for the following month are not added to the input, but the initial input data are used to obtain the ENSO multivariate forecast at one time.

[0118] The ENSO intelligent prediction device provided in the embodiment of the present application has the same technical features as the ENSO intelligent prediction method based on the segmented deep neural network provided in the above embodiment, so it can also solve the same technical problems and achieve the same technical effects.

[0119] Furthermore, the ENSO intelligent prediction device based on a segmented deep neural network provided in the embodiment of the present application can be specific hardware on the device or software or firmware installed on the device. The device provided in the embodiment of the present application has the same implementation principle and technical effects as the aforementioned method embodiment. For the sake of brief description, for matters not mentioned in the device embodiment, reference can be made to the corresponding content in the aforementioned method embodiment. Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can all refer to the corresponding processes in the aforementioned method embodiment, and will not be repeated here.

[0120] On the other hand, an embodiment of the present invention further provides a computer device E, Figure 5 A computer device E provided in an embodiment of the present application can be used to execute an ENSO intelligent prediction method based on a segmented deep neural network. The computer device E comprises a memory E.1, a processor E.2, a graphics card E.3, and a bus. The memory E.1 stores machine-readable instructions executable by the processor E.2 and the graphics card E.3. When the computer device E is running, the processor E.2, the graphics card E.3, and the memory E.1 communicate via the bus, and the processor E.2 and the graphics card E.3 jointly execute the machine-readable instructions to perform the steps of the aforementioned ENSO intelligent prediction method based on a segmented deep neural network.

[0121] Specifically, the aforementioned memory E.1, processor E.2, and graphics card E.3 can be general-purpose memory, processor, and graphics card, without specific limitations. When processor E.2 and graphics card E.3 execute the computer program stored in memory E.1, they can implement the aforementioned ENSO intelligent prediction method based on a segmented deep neural network. The graphics card E.3 executes instructions related to the ENSO intelligent prediction model, while the processor E.2 executes other control instructions, such as those for controlling input and output streams.

[0122] On the other hand, corresponding to the above-mentioned ENSO intelligent prediction method based on a segmented deep neural network, embodiments of the present application further provide a computer-readable storage medium for storing the ENSO intelligent prediction method based on a segmented deep neural network. The computer-readable storage medium stores machine-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions cause the processor to execute the steps of the above-mentioned ENSO prediction method based on a segmented deep neural network.

[0123] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. An ENSO intelligent prediction method based on a segmented deep neural network, characterized in that: The following steps are involved: 1) Based on the unbalanced characteristics of the ENSO evolution process, an ENSO intelligent prediction model based on a segmented deep neural network is constructed; 2) Training the constructed ENSO intelligent prediction model based on segmented deep neural network; 3) Determine the model input data and preprocess the input data, and use the preprocessed data to predict ENSO through the ENSO intelligent prediction model based on the segmented deep neural network; The specific steps to build an ENSO intelligent prediction model based on a piecewise deep neural network include: 11) Select the required ENSO forecast length and divide the ENSO forecast into two segments: the front and the back according to the segmentation positioning model; 12) Based on the different evolution processes and ocean-atmosphere characteristics of the first and second phases of ENSO, dedicated encoders and decoders for the first and second phases are designed based on a deep learning model for spatiotemporal series prediction. 13) Based on the different evolution processes and ocean-atmosphere characteristics of the first and second phases of ENSO, a weighted feature fusion model is constructed for the first and second phases respectively, based on the deep learning model of spatiotemporal series prediction; 14) Based on the structure constructed in steps 12) and 13), an ENSO intelligent prediction model based on a piecewise deep neural network is created; The expression of the segmentation and positioning model is: Where x t is the ENSO prediction dataset, t is the step size of ENSO prediction, t seg The dividing position of the front and back sections, is the floor rounding function, seg(·) is the model used to calculate the front and back segment segmentation positions, f seg is the slope of seg(·).

2. The ENSO intelligent prediction method based on a segmented deep neural network according to claim 1 is characterized in that: The expression of the ENSO intelligent prediction model based on the segmented deep neural network is: In the formula, concat(·) means concatenating the features of different segments together, encoder th (·) and decoder th (·) are respectively the encoder and decoder designed based on the spatiotemporal sequence prediction deep learning model in step 12), th (·) is the weighted feature fusion device designed in step 13), is the output of this segment decoder, is the weighted fusion feature obtained by the weighted feature fusion device, and They are the encoder outputs of this segment and another segment respectively.

3. The ENSO intelligent prediction method based on a segmented deep neural network according to claim 1 is characterized in that: The specific steps of step 2) include: 21) Based on the ocean-air characteristics of ENSO, select variables that play a key role in ENSO prediction and construct an ENSO prediction dataset; 22) Perform quality control and data preprocessing on different physical variables in the ENSO prediction dataset; 23) The preprocessed dataset is divided into training set, validation set and test set, and the training set is used to train the constructed ENSO intelligent prediction model.

4. The ENSO intelligent prediction method based on a segmented deep neural network according to claim 1 is characterized in that: The specific steps of step 3) include: 31) Select the reporting start time based on actual business needs, obtain the initial time data, and perform quality control and data preprocessing. The processed results serve as the input of the ENSO intelligent prediction model based on the segmented deep neural network; 32) Input the processed initial moment data into the front-end and back-end encoders of the segmented deep neural network ENSO intelligent prediction model for feature encoding to determine the segmental features of ESNO; 33) The features after the front and rear segment coding are spliced ​​and input into their respective weighted feature fusion devices for feature fusion to obtain weighted fusion features; 34) The weighted fusion features of the front and back segments are respectively input into the front and back segment decoders of the segmented deep neural network ENSO intelligent prediction model for decoding, and the data is post-processed to obtain the ENSO grid prediction results, and long-term ENSO prediction is performed using a non-iterative prediction method; 35) Based on the prediction results of the ENSO intelligent prediction model based on a segmented deep neural network, various ENSO-related indices and evaluation indicators are calculated.

5. The ENSO intelligent prediction method based on a segmented deep neural network according to claim 1 is characterized in that: The calculation formula of the weighted feature fusion is: Where, is the weighted fusion feature, and They are the encoder outputs of this segment and another segment respectively, and each segment constructs its own weighted feature fusion merge th (·), α is the weighting matrix.

6. An ENSO intelligent prediction device based on a segmented deep neural network, characterized in that: include: Initial data acquisition module, which obtains the initial forecast time, initial data and forecast duration of ENSO; The data preprocessing module converts the acquired initial moment data into the input format specified by the model, and the converted data is used as input data; The prediction module transmits input data to the ENSO intelligent prediction model based on the segmented deep neural network, runs the model, and obtains the prediction results; The result acquisition module processes the model output and converts the output data into meteorological grid data and various ENSO indices and quantitative evaluation indicators; The specific steps to build an ENSO intelligent prediction model based on a piecewise deep neural network include: 11) Select the required ENSO forecast length and divide the ENSO forecast into two segments: the front and the back according to the segmentation positioning model; 12) Based on the different evolution processes and ocean-atmosphere characteristics of the first and second phases of ENSO, dedicated encoders and decoders for the first and second phases are designed based on a deep learning model for spatiotemporal series prediction. 13) Based on the different evolution processes and ocean-atmosphere characteristics of the first and second phases of ENSO, a weighted feature fusion model is constructed for the first and second phases respectively, based on the deep learning model of spatiotemporal series prediction; 14) Based on the structure constructed in steps 12) and 13), an ENSO intelligent prediction model based on a piecewise deep neural network is created; The expression of the segmentation and positioning model is: Where x t is the ENSO prediction dataset, t is the step size of ENSO prediction, and t seg The dividing position of the front and back sections, is the floor rounding function, seg(·) is the model used to calculate the front and back segment segmentation positions, f seg is the slope of seg(·).

7. A computer device, characterized in that: include: A processor, a memory, and a computer program stored on the memory and capable of running on the processor, wherein when the computer program is executed by the processor, the steps of the ENSO intelligent prediction method based on a segmented deep neural network are implemented as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the ENSO intelligent prediction method based on a segmented deep neural network are implemented as described in any one of claims 1 to 5.

Images