ENSO prediction method, system and device based on space-time fusion neural network

Abstract

The invention discloses an ENSO prediction method, system and device based on a space-time fusion neural network, and belongs to the field of deep learning model prediction. The invention designs a time-space fusion Transform network, the network designs a parallel encoder-decoder structure, and the time-space information can be effectively extracted from the ocean surface temperature data and the # imgabs0 # index at the same time, so that the prediction precision of the # imgabs1 # index is improved. Meanwhile, the network extracts global features by using an attention mechanism in an encoder of the network, and establishes a remote dependency relationship on a target. Compared with a comparison method, the method has better prediction capability in multiple prediction advance periods, the problem that an existing ENSO prediction method underestimates the ENSO event amplitude along with the increase of the prediction advance periods is solved, and the ENSO prediction precision is improved.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning model prediction, and in particular relates to an ENSO prediction method, system and device based on spatiotemporal fusion neural network. Background Art

[0002] El Niño and Southern Oscillation The Earth's Ocean Oscillation (ENSO) is the product of large-scale ocean-atmosphere interactions in the tropical Pacific Ocean and is an important component of the global climate system. ENSO can not only change the state of the Pacific Ocean and atmosphere, but also have an important impact on the global climate, precipitation, agriculture, fisheries and ecosystems. Therefore, in order to recognize and understand the laws of climate change, robust and long-term forward-looking predictions are essential. Traditional numerical model prediction methods consume a lot of computing resources and are difficult to guarantee timeliness. In recent years, with the explosive growth of data, deep learning has rapidly emerged in the field of Earth system science, and researchers have also applied deep learning to ENSO prediction. However, convolutional neural networks and recurrent neural networks have difficulties in learning long-range dependencies and global features, respectively. At the same time, the existing ENSO prediction methods based on deep learning will underestimate the magnitude of ENSO events as the prediction lead time increases, resulting in a decrease in prediction accuracy. Therefore, the current technology used in deep learning for ENSO prediction still has defects such as insufficient accuracy. Summary of the invention

[0003] The purpose of the present invention is to solve the defects of the prior art in ENSO prediction, such as insufficient accuracy and difficulty in long-term prediction, and to provide an ENSO prediction method, system and device based on spatiotemporal fusion neural network. The innovation of the present invention lies in the design of a parallel encoder structure, which can effectively extract sea surface temperature anomalies and The spatiotemporal information of the index can improve The accuracy of index forecast.

[0004] In order to achieve the above-mentioned invention object, the present invention specifically adopts the following technical solutions:

[0005] In a first aspect, the present invention provides an ENSO prediction method based on a spatiotemporal fusion neural network, which comprises the following steps:

[0006] S1. Perform data preprocessing on the acquired original sea surface temperature data to obtain the sea surface temperature anomaly data and the Index data;

[0007] S2. The sea surface temperature anomaly data to be detected is enhanced in time and space, and the sea surface temperature anomaly data after time and space information enhancement is obtained. The exponential data is enhanced with time series information to obtain the time series information enhanced Index data;

[0008] S3. The sea surface temperature anomaly data with enhanced spatiotemporal information and the sea surface temperature anomaly data with enhanced time series information The index data is input into the trained spatiotemporal fusion Transformer network. The first encoder encodes the sea surface temperature anomaly data after spatiotemporal information enhancement to obtain the first encoding output result, and the second encoder encodes the sea surface temperature anomaly data after temporal information enhancement to obtain the first encoding output result. The exponential data is encoded to obtain the second encoding output result, the first encoding output result and the second encoding output result are added together as the final encoding result, and the final encoding result and the sea surface temperature anomaly data after the spatiotemporal information enhancement are input into the decoder to obtain the The forecast results of the index are obtained by using a recursive forecasting strategy for multiple future moments. Index forecast results, completed Index forecast.

[0009] Based on the above scheme, each step can be implemented in the following preferred specific manner.

[0010] As a preferred embodiment of the first aspect, in step S1, the specific method of preprocessing the original sea surface temperature data is as follows:

[0011] S11. For the missing values ​​in the original sea surface temperature data, the missing values ​​belonging to the land area are assigned a value of 0, and the scale of the spatial resolution is set to a preset size, thereby obtaining the initially processed sea surface temperature data;

[0012] S12. taking the difference between the initially processed sea surface temperature data and the historical sea surface temperature climate average as the sea surface temperature anomaly, and obtaining the sea surface temperature anomaly data to be detected;

[0013] S13. Calculation Index, get the Index data.

[0014] As a preferred embodiment of the first aspect, the specific process of step S2 is as follows:

[0015] S21. The sea surface temperature anomaly data to be detected is processed by a time coding function to obtain a first time series coding value, and the sea surface temperature anomaly data to be detected is processed by a space coding function to obtain a first space coding value, and then the first time series coding value, the first space coding value and the sea surface temperature anomaly data to be detected are merged, that is, the sea surface temperature anomaly data to be detected is added with the first time series coding value and the first space coding value as the sea surface temperature anomaly data after the spatiotemporal information is enhanced;

[0016] S22. The time coding function to be detected The index data is processed to obtain a second time series code value, and then the second time series code value and the The index data is integrated to The exponential data is added to the second time series code value as the time series information enhanced Index data.

[0017] As a preferred embodiment of the above-mentioned first aspect, in step S21, the sine function and the cosine function are used as time coding functions to time-code the sea surface temperature anomaly data to be detected, and the time series coding value at the target moment is obtained and used as the first time series coding value; the linear function is used as the space coding function to space-code the sea surface temperature anomaly data to be detected, and the space coding value at the target moment is obtained and used as the first space coding value.

[0018] As a preferred embodiment of the above-mentioned first aspect, in step S3, the first encoder and the second encoder are both formed by cascading multiple layers of encoding blocks in sequence, and the number of encoding block layers in the first encoder and the second encoder is the same, and each layer of encoding block includes a multi-head attention layer and a fully connected feedforward network.

[0019] As a preferred embodiment of the first aspect, the specific processing process in the first layer coding block is as follows:

[0020] S31. Input the abnormal sea surface temperature data after spatiotemporal information enhancement into the first layer encoding block of the first encoder, and obtain the first time series feature after layer normalization and the first multi-head attention layer in sequence, add the abnormal sea surface temperature data after spatiotemporal information enhancement and the first time series feature to obtain the first fused data, and obtain the second time series feature after the first fused data is subjected to layer normalization and the first fully connected feedforward network in sequence, and add the first fused data and the second time series feature as the output of the first layer encoding block in the first encoder;

[0021] S32. After enhancing the timing information The exponential data is input into the first layer encoding block of the second encoder, and then goes through layer normalization and the second multi-head attention layer to obtain the third temporal feature. The exponential data and the third time series feature are added to obtain the second fused data, and the second fused data is sequentially subjected to layer normalization and the second fully connected feedforward network to obtain the fourth time series feature, and the second fused data and the fourth time series feature are added as the output of the first layer encoding block in the second encoder.

[0022] As a preferred embodiment of the above-mentioned first aspect, in step S3, the decoder is composed of multiple layers of decoding blocks cascaded in sequence, and each layer of decoding blocks includes a multi-head attention layer, a fully connected feedforward network and a cross multi-head attention layer.

[0023] As a preferred embodiment of the above-mentioned first aspect, the specific processing process in the first-layer decoding block is as follows: the sea surface temperature anomaly data after spatiotemporal information enhancement is input into the first-layer encoding block of the decoder, and after layer normalization and the third multi-head attention layer in sequence, the fifth time series feature is obtained, the sea surface temperature anomaly data after spatiotemporal information enhancement and the fifth time series feature are added to obtain the third fused data, the third fused data is layer normalized and then passed through the cross multi-head attention layer together with the final encoding result to obtain the sixth time series feature, the third fused data and the sixth time series feature are added to obtain the fourth fused data, the fourth fused data is sequentially passed through layer normalization and the third fully connected feedforward network to obtain the seventh time series feature, and the fourth fused data and the seventh time series feature are added as the output of the first-layer decoding block in the first decoder.

[0024] In a second aspect, the present invention provides an ENSO prediction system based on a spatiotemporal fusion neural network, comprising:

[0025] The preprocessing module is used to preprocess the acquired raw sea surface temperature data to obtain the sea surface temperature anomaly data to be detected and the Index data;

[0026] The spatiotemporal information enhancement module is used to enhance the spatiotemporal information of the sea surface temperature anomaly data to be detected, and obtain the sea surface temperature anomaly data after spatiotemporal information enhancement. The exponential data is enhanced with time series information to obtain the time series information enhanced Index data;

[0027] The prediction module is used to transform the sea surface temperature anomaly data with enhanced spatiotemporal information and the The index data is input into the trained spatiotemporal fusion Transformer network. The first encoder encodes the sea surface temperature anomaly data after spatiotemporal information enhancement to obtain the first encoding output result, and the second encoder encodes the sea surface temperature anomaly data after temporal information enhancement to obtain the first encoding output result. The exponential data is encoded to obtain the second encoding output result, the first encoding output result and the second encoding output result are added together as the final encoding result, and the final encoding result and the sea surface temperature anomaly data after the spatiotemporal information enhancement are input into the decoder to obtain the The forecast results of the index are obtained by using a recursive forecasting strategy for multiple future moments. Index forecast results, completed Index forecast.

[0028] In a third aspect, the present invention provides a computer electronic device comprising a memory and a processor;

[0029] The memory is used to store computer programs;

[0030] The processor is used to implement the ENSO prediction method based on spatiotemporal fusion neural network as described in any one of the solutions of the first aspect when executing the computer program.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] The present invention provides an ENSO prediction method, system and device based on a spatiotemporal fusion neural network. The method enhances the spatiotemporal information of data according to sine and cosine functions and linear functions, and designs a spatiotemporal fusion Transformer network with a parallel encoding-decoding structure based on Transformer to predict The present invention can make up for the deficiencies of the data-driven deep learning model, which is conducive to extracting global features and establishing long-range dependencies on targets, thereby improving the accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Schematic diagram of the steps of the method of the present invention;

[0034] Figure 2 This is a diagram of the spatial-temporal fusion Transformer network structure of the present invention;

[0035] Figure 3 A comparison diagram of the root mean square error results between the method of the present invention and other methods;

[0036] Figure 4 It is a system block diagram of the present invention. DETAILED DESCRIPTION

[0037] In order to make the above-mentioned purpose, features and advantages of the present invention more obvious and easy to understand, the specific implementation mode of the present invention is described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. The technical features in each embodiment of the present invention can be combined accordingly without conflicting with each other.

[0038] In the description of the present invention, it should be understood that the terms "first" and "second" are only used for the purpose of distinguishing descriptions, and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features.

[0039] The present invention studies the existing ENSO prediction and finds that the convolutional neural network and recurrent neural network have difficulties in learning long-range dependencies and global features respectively. At the same time, the existing ENSO prediction method based on deep learning will underestimate the magnitude of ENSO events as the prediction lead time increases, resulting in a decrease in prediction accuracy. Therefore, the present invention proposes an ENSO prediction method based on a spatiotemporal fusion neural network.

[0040] like Figure 1 As shown, in a preferred implementation of the present invention, the above ENSO prediction method based on spatiotemporal fusion neural network includes the following steps S1 to S3. The specific implementation process is described in detail below.

[0041] S1. Perform data preprocessing on the acquired original sea surface temperature data to obtain the sea surface temperature anomaly data and the Index data.

[0042] It should be noted that, in the present invention, the sea surface temperature data is a hybrid data combining ship data, buoy observation data and satellite observation data. The sea surface temperature data can be obtained by real-time measurement of the sea surface, inversion of remote sensing satellite products, or by numerical simulation experiments based on the geodynamic equations to construct an air-sea coupling model. In practical applications, data from different sources and with different spatiotemporal resolutions can be selected according to specific needs.

[0043] In an embodiment of the present invention, the selected original sea surface temperature data include the sixth international coupled model comparison plan (Coupled Model Intercomparison Project Phase 6, CMIP6), the global simple ocean data assimilation analysis system (Simple Ocean Data Assimilation, SODA) and the global ocean data assimilation system (GlobalOcean Data Assimilation System, GODAS) data. Among them, 20 models in the CMIP6 model data set from 1850 to 2000 are selected as training data sets, the SODA reanalysis data set from 1871 to 1972 is selected as a validation data set, and the GODAS reanalysis data set from 1980 to 2022 is selected as a test data set. The spatial range of the sea surface temperature data is selected to be 100°E-255°W and 55°S-60°N, and the time resolution is 1 month.

[0044] It should be noted that, in the present invention, The index data is the three-month sliding average of the sea surface temperature anomaly in the longitude and latitude range of 5°N-5°S and 170°W-120°W, which effectively reflects the changes in ENSO in the form of an index.

[0045] After the original sea surface temperature data is obtained in the above manner, the original sea surface temperature data is preprocessed. As an implementation of the embodiment of the present invention, in the above step S1, the specific method for preprocessing the original sea surface temperature data is as follows:

[0046] S11. For the missing values ​​in the original sea surface temperature data, the missing values ​​belonging to the land area are assigned a value of 0, and the scale of the spatial resolution is set to a preset size to obtain the preliminarily processed sea surface temperature data.

[0047] It should be noted that in the process of obtaining the original sea surface temperature data, data from non-ocean areas such as land and islands will be mixed in. This part of the non-ocean area data will cause missing values ​​in the original sea surface temperature data, so it is necessary to perform preliminary processing according to the above step S11.

[0048] In the embodiment of the present invention, the scale of the spatial resolution of the data is set to 5°×5°.

[0049] S12. The difference between the initially processed sea surface temperature data and the historical sea surface temperature climate average is taken as the sea surface temperature anomaly to obtain the sea surface temperature anomaly data to be detected.

[0050] In the embodiment of the present invention, the historical sea surface temperature climate mean state uses the average value of 30 samples in the most recent 30 years in history as the reference point for the annual value of the corresponding climate element, and selects the statistical value of climate data from 1981 to 2010.

[0051] S13. Calculation Index, get the Index data.

[0052] In an embodiment of the present invention, The ENSO index is an important climate indicator used to monitor the The specific method of the index is as follows: calculate the average value of the sea surface temperature anomaly in the range of 170°W-120°W, 5°S-5°N for three consecutive months to obtain the index to be tested. Index data.

[0053] S2. The sea surface temperature anomaly data to be detected is enhanced in time and space, and the sea surface temperature anomaly data after time and space information enhancement is obtained. The exponential data is enhanced with time series information to obtain the time series information enhanced Index data.

[0054] As an implementation method of the embodiment of the present invention, the specific process of the above step S2 is as follows:

[0055] S21. The sea surface temperature anomaly data to be detected is processed by a time coding function to obtain a first time series coding value, and the sea surface temperature anomaly data to be detected is processed by a space coding function to obtain a first space coding value, and then the first time series coding value, the first space coding value and the sea surface temperature anomaly data to be detected are fused, that is, the sea surface temperature anomaly data to be detected is added with the first time series coding value and the first space coding value as the sea surface temperature anomaly data with enhanced spatiotemporal information.

[0056] It should be noted that in step S21 of the embodiment of the present invention, the time coding function is a sine function and a cosine function, and the space coding function is a linear function. The method for obtaining the coding values ​​in their respective forms using the above-mentioned time coding function and space coding function is as follows: using the sine function and the cosine function as the time coding function, time-coding the sea surface temperature anomaly data to be detected, obtaining the time series coding value at the target moment and using it as the first time series coding value; using the linear function as the space coding function, space-coding the sea surface temperature anomaly data to be detected, obtaining the space coding value at the target moment and using it as the first space coding value.

[0057] S22. The time coding function to be detected The index data is processed to obtain a second time series code value, and then the second time series code value and the The index data is integrated to The exponential data is added to the second time series code value as the time series information enhanced Index data.

[0058] It should be noted that in step S22 of the embodiment of the present invention, the selection of the time coding function is the same as that in step S21, and the specific process of time coding is as described above and will not be repeated here.

[0059] S3. The sea surface temperature anomaly data with enhanced spatiotemporal information and the sea surface temperature anomaly data with enhanced time series information The index data is input into the trained spatiotemporal fusion Transformer network. The first encoder encodes the sea surface temperature anomaly data after spatiotemporal information enhancement to obtain the first encoding output result, and the second encoder encodes the sea surface temperature anomaly data after temporal information enhancement to obtain the first encoding output result. The exponential data is encoded to obtain the second encoding output result, and the first encoding output result and the second encoding output result are added together as the final encoding result X encoder The final encoding result and the sea surface temperature anomaly data with enhanced spatiotemporal information are input into the decoder to obtain the The forecast results of the index are obtained by using a recursive forecasting strategy for multiple future moments. Index forecast results, completed Index forecast.

[0060] It should be noted that the spatiotemporal fusion Transformer network of the present invention adopts a parallel encoder-decoder Transformer structure. Figure 2 The following figure shows the structure of the spatiotemporal fusion Transformer network. The sea surface temperature anomaly data after spatiotemporal information enhancement and the sea surface temperature anomaly data after time series information enhancement The index data will be used as input and encoded by an encoder respectively. The first encoder and the second encoder are composed of multiple layers of encoding blocks cascaded in sequence, and the number of encoding blocks in the first encoder and the second encoder is the same. Each layer of encoding block contains two sub-layers: a multi-head attention layer and a fully connected feedforward network. In addition, layer normalization and residual connection operations are performed on each sub-layer of the encoding block. For the first layer of encoding blocks, its input is the sea surface temperature anomaly data after spatiotemporal information enhancement or the temporal information enhancement. Exponential data; for the remaining layers of coding blocks, the input is the output of the previous layer of coding blocks. Then the results of the output of the last layer of coding blocks of the two encoders are added as the final coding result. Similarly, the decoder of the present invention is also composed of multiple layers of decoding blocks cascaded in sequence. In addition to the two sub-layers contained in the coding block, the decoder also inserts a third sub-layer-a cross-multi-head attention layer, which is used to perform a cross-multi-head attention operation on the output of the multi-head attention layer of the final coding result output by the encoder. Therefore, each layer of the decoding block of the present invention includes a multi-head attention layer, a fully connected feedforward network, and a cross-multi-head attention layer. In addition, layer normalization and residual connection operations are performed on each sub-layer of the decoding block. For the first layer of decoding blocks, its input is the sea surface temperature anomaly data after spatiotemporal information enhancement and the final coding result output by the encoder; for the remaining layers of decoding blocks, its input is the output of the previous layer of decoding blocks and the final coding result output by the encoder.

[0061] The embodiment of the present invention takes the first layer coding block as an example, and the specific processing process is as follows:

[0062] S31. Input the abnormal sea surface temperature data after spatiotemporal information enhancement into the first layer encoding block of the first encoder, and obtain the first time series feature after passing through layer normalization and the first multi-head attention layer in sequence, and add the abnormal sea surface temperature data after spatiotemporal information enhancement and the first time series feature to obtain the first fused data, and pass the first fused data through layer normalization and the first fully connected feedforward network in sequence to obtain the second time series feature, and add the first fused data and the second time series feature as the output of the first layer encoding block in the first encoder.

[0063] In this embodiment, the sea surface temperature anomaly data after the temporal and spatial information enhancement is recorded as X ssta , the process of obtaining the output of the first layer encoding block in the first encoder is:

[0064] X s ′ sta =MHA 1 (LN(X ssta ))+X ssta

[0065]

[0066] Where LN(·) represents layer normalization; MHA 1 represents the first multi-head attention layer; X s ′ sta Represents the first fusion data; FFN 1 represents the first fully connected feed-forward network; Represents the output of the first layer encoding block in the first encoder.

[0067] S32. After enhancing the timing information The exponential data is input into the first layer encoding block of the second encoder, and then goes through layer normalization and the second multi-head attention layer to obtain the third temporal feature. The exponential data and the third time series feature are added to obtain the second fused data, and the second fused data is sequentially subjected to layer normalization and the second fully connected feedforward network to obtain the fourth time series feature, and the second fused data and the fourth time series feature are added as the output of the first layer encoding block in the second encoder.

[0068] In this embodiment, after the timing information is enhanced, Index data is recorded as X index , the process of obtaining the output of the first layer encoding block in the second encoder is:

[0069] X i ′ ndex =MHA 2 (LN(X index ))+X index

[0070]

[0071] Among them, MHA 2 represents the second multi-head attention layer; X i ′ ndex Represents the second fusion data; FFN 2 represents the second fully connected feed-forward network; Represents the output of the first layer encoding block in the second encoder.

[0072] It should also be noted that the implementation method of the first multi-head attention layer and the first multi-head attention layer belongs to the prior art. In the embodiment of the present invention, the specific processing process in each multi-head attention layer is as follows:

[0073] MHA=Concat(attention 1 ,…,attention h )

[0074]

[0075] Among them, Concat represents the concatenation operation; attention 1 ,…,attention hrepresents the output of the 1st, ..., hth multi-head attention; h is the number of multi-head attentions. In the embodiment of the present invention, h = 4 is used; attention represents the output of each single-head attention; Q, K, V represent the query vector, key vector, and value vector respectively; when encoding the sea surface temperature anomaly data, Q, K, V come from the matrix X ssta ; in the When encoding exponential data, Q, K, V come from the matrix X index ; d is the feature dimension of Q, K, V; T represents transposition.

[0076] The embodiment of the present invention takes the first layer decoding block as an example, and the specific processing process is as follows:

[0077] The abnormal sea surface temperature data after spatiotemporal information enhancement is input into the first-layer encoding block of the decoder, and after layer normalization and the third multi-head attention layer, the fifth time series feature is obtained. The abnormal sea surface temperature data after spatiotemporal information enhancement and the fifth time series feature are added to obtain the third fused data. The third fused data is layer normalized and then passed through the cross multi-head attention layer together with the final encoding result to obtain the sixth time series feature. The third fused data and the sixth time series feature are added to obtain the fourth fused data. The fourth fused data is sequentially passed through layer normalization and the third fully connected feedforward network to obtain the seventh time series feature. The fourth fused data and the seventh time series feature are added as the output of the first-layer decoding block in the first decoder.

[0078] In this embodiment, the process of obtaining the output of the first layer decoding block in the first decoder is:

[0079] X s ″ sta =MHA 3 (LN(X ssta ))+X ssta

[0080] X s ″ s ′ ta =CrossAttention(LN(X encoder ),X s ″ sta )+X s ″ sta

[0081]

[0082] Among them, X s ″ staRepresents the third fusion data; MHA 3 represents the third multi-head attention layer; CrossAttention represents the cross multi-head attention layer; X s ″ s ′ ta FFN stands for fourth fusion data; 3 represents the third fully connected feed-forward network; Represents the output of the first layer decoding block in the decoder.

[0083] It should also be noted that in the present invention, the recursive prediction strategy is to first predict the first moment in the future through the trained spatiotemporal fusion Transformer network. The index prediction result is then converted into the first moment of the future prediction. The exponential prediction result is used as input to predict the next moment (time step). This method is used continuously until the entire time range is predicted and multiple future moments are obtained. Index prediction results.

[0084] The present invention will now use a specific example to demonstrate the application effect of the ENSO prediction method based on the spatiotemporal fusion neural network described in S1 to S3 in the above embodiments on a specific data set, so as to facilitate understanding of the essence of the present invention.

[0085] Example

[0086] The specific implementation process of the ENSO prediction method based on the spatiotemporal fusion neural network adopted in this embodiment is as described above and will not be repeated here.

[0087] In this embodiment, a spatiotemporal fusion transformer network (STFTN) is used. The specific network structure is as described above and will not be repeated here. The overall process of index prediction can be divided into four stages: data preprocessing, data spatiotemporal information encoding, model training, and model prediction.

[0088] 1) Data preprocessing: In this embodiment, 20 patterns in the CMIP6 model data set from 1850 to 2000 are selected as training data sets, the SODA reanalysis data set from 1871 to 1972 is selected as a validation data set, the GODAS reanalysis data set from 1980 to 2022 is selected as a test data set, and the spatial range of the sea surface temperature data is selected as 100°E-255°W and 55°S-60°N. According to the above step S1, the original sea surface temperature data is preprocessed.

[0089] 2) Spatial-temporal information encoding: According to the above step S2, the sea surface temperature anomaly data and Exponential data encodes spatiotemporal information as well as temporal information.

[0090] 3) Model training: According to the above step S3, the sea surface temperature anomaly data and Exponential input spatiotemporal fusion Transformer network, and train it.

[0091] 4) Model prediction: The prediction performance of the spatiotemporal fusion Transformer network is verified using test samples. The exponential input is fed into the trained spatiotemporal fusion Transformer network to obtain the next 20 months Index prediction results.

[0092] In this example, in order to test the prediction performance of the spatiotemporal fusion Transformer network (STFTN) in the present invention, deep learning models CNN, ENSTOR, dynamics models CanCM4, GFDL-aer04, GFDL-FLOR-A06 and persistence model are selected as comparisons, and the root mean squared error (RMSE) is used for performance evaluation. The specific results are as follows: Figure 3 As shown in the figure, the spatial-temporal fusion Transformer network (STFTN) proposed in this invention is effective for Index prediction can improve The forecast accuracy of the index.

[0093] It should also be noted that the ENSO prediction method based on the spatiotemporal fusion neural network in the above embodiment can essentially be executed by a computer program or module. Therefore, similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides an ENSO prediction system based on a spatiotemporal fusion neural network corresponding to the ENSO prediction method based on the spatiotemporal fusion neural network provided in the above embodiment, such as Figure 4 As shown, it includes:

[0094] The preprocessing module is used to preprocess the acquired raw sea surface temperature data to obtain the sea surface temperature anomaly data to be detected and the Index data;

[0095] The spatiotemporal information enhancement module is used to enhance the spatiotemporal information of the sea surface temperature anomaly data to be detected, and obtain the sea surface temperature anomaly data after spatiotemporal information enhancement. The exponential data is enhanced with time series information to obtain the time series information enhanced Index data;

[0096] The prediction module is used to transform the sea surface temperature anomaly data with enhanced spatiotemporal information and the The index data is input into the trained spatiotemporal fusion Transformer network. The first encoder encodes the sea surface temperature anomaly data after spatiotemporal information enhancement to obtain the first encoding output result, and the second encoder encodes the sea surface temperature anomaly data after temporal information enhancement to obtain the first encoding output result. The exponential data is encoded to obtain the second encoding output result, the first encoding output result and the second encoding output result are added together as the final encoding result, and the final encoding result and the sea surface temperature anomaly data after the spatiotemporal information enhancement are input into the decoder to obtain the The forecast results of the index are obtained by using a recursive forecasting strategy for multiple future moments. Index forecast results, completed Index forecast.

[0097] Similarly, based on the same inventive concept, another preferred embodiment of the present invention further provides a computer electronic device corresponding to the ENSO prediction method based on spatiotemporal fusion neural network provided in the above embodiment, which includes a memory and a processor;

[0098] The memory is used to store computer programs;

[0099] The processor is used to implement the ENSO prediction method based on spatiotemporal fusion neural network in the above embodiment when executing the computer program.

[0100] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention.

[0101] It is understandable that the above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0102] It should also be noted that those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the system described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the various embodiments provided in this application, the division of steps or modules in the system and method is only a logical function division, and there may be other division methods in actual implementation, such as multiple modules or steps can be combined or integrated together, and a module or step can also be split.

[0103] The above-described embodiment is only a preferred solution of the present invention, but it is not intended to limit the present invention. A person skilled in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.

Claims

1. An ENSO prediction method based on spatiotemporal fusion neural network, characterized in that: The following steps are involved: S1. Perform data preprocessing on the acquired original sea surface temperature data to obtain the sea surface temperature anomaly data and the Index data; S2. The sea surface temperature anomaly data to be detected is enhanced in time and space, and the sea surface temperature anomaly data after time and space information enhancement is obtained. The exponential data is enhanced with time series information to obtain the time series information enhanced Index data; S3. The sea surface temperature anomaly data with enhanced spatiotemporal information and the sea surface temperature anomaly data with enhanced time series information The index data is input into the trained spatiotemporal fusion Transformer network. The first encoder encodes the sea surface temperature anomaly data after spatiotemporal information enhancement to obtain the first encoding output result, and the second encoder encodes the sea surface temperature anomaly data after temporal information enhancement to obtain the first encoding output result. The exponential data is encoded to obtain the second encoding output result, the first encoding output result and the second encoding output result are added together as the final encoding result, and the final encoding result and the sea surface temperature anomaly data after the spatiotemporal information enhancement are input into the decoder to obtain the The forecast results of the index are obtained by using a recursive forecasting strategy for multiple future moments. Index forecast results, completed Index forecast.

2. The ENSO prediction method based on spatiotemporal fusion neural network according to claim 1, characterized in that: In step S1, the specific method for preprocessing the original sea surface temperature data is as follows: S11. For the missing values ​​in the original sea surface temperature data, the missing values ​​belonging to the land area are assigned a value of 0, and the scale of the spatial resolution is set to a preset size, thereby obtaining the initially processed sea surface temperature data; S12. taking the difference between the initially processed sea surface temperature data and the historical sea surface temperature climate average as the sea surface temperature anomaly, and obtaining the sea surface temperature anomaly data to be detected; S13. Calculation Index, get the Index data.

3. The ENSO prediction method based on spatiotemporal fusion neural network according to claim 1, characterized in that: The specific process of step S2 is as follows: S21. The sea surface temperature anomaly data to be detected is processed by a time coding function to obtain a first time series coding value, and the sea surface temperature anomaly data to be detected is processed by a space coding function to obtain a first space coding value, and then the first time series coding value, the first space coding value and the sea surface temperature anomaly data to be detected are merged, that is, the sea surface temperature anomaly data to be detected is added with the first time series coding value and the first space coding value as the sea surface temperature anomaly data after the spatiotemporal information is enhanced; S22. The time coding function to be detected The index data is processed to obtain a second time series code value, and then the second time series code value and the The index data is integrated to The exponential data is added to the second time series code value as the time series information enhanced Index data.

4. The ENSO prediction method based on spatiotemporal fusion neural network according to claim 3, characterized in that: In step S21, the sine function and the cosine function are used as time coding functions to time code the sea surface temperature anomaly data to be detected, and the time series coding value at the target moment is obtained and used as the first time series coding value; the linear function is used as the space coding function to space code the sea surface temperature anomaly data to be detected, and the space coding value at the target moment is obtained and used as the first space coding value.

5. The ENSO prediction method based on spatiotemporal fusion neural network according to claim 1, characterized in that: In step S3, the first encoder and the second encoder are both formed by cascading multiple layers of encoding blocks in sequence, and the number of encoding block layers in the first encoder and the second encoder is the same, and each layer of encoding block includes a multi-head attention layer and a fully connected feedforward network.

6. The ENSO prediction method based on spatiotemporal fusion neural network according to claim 5, characterized in that: The specific processing process in the first layer encoding block is as follows: S31. Input the abnormal sea surface temperature data after spatiotemporal information enhancement into the first layer encoding block of the first encoder, and obtain the first time series feature after layer normalization and the first multi-head attention layer in sequence, add the abnormal sea surface temperature data after spatiotemporal information enhancement and the first time series feature to obtain the first fused data, and obtain the second time series feature after the first fused data is subjected to layer normalization and the first fully connected feedforward network in sequence, and add the first fused data and the second time series feature as the output of the first layer encoding block in the first encoder; S32. After enhancing the timing information The exponential data is input into the first layer encoding block of the second encoder, and then goes through layer normalization and the second multi-head attention layer to obtain the third temporal feature. The exponential data and the third time series feature are added to obtain the second fused data. The second fused data is sequentially subjected to layer normalization and the second fully connected feedforward network to obtain the fourth time series feature. The second fused data and the fourth time series feature are added as the output of the first layer encoding block in the second encoder.

7. The ENSO prediction method based on spatiotemporal fusion neural network according to claim 1, characterized in that: In step S3, the decoder is composed of multiple layers of decoding blocks cascaded in sequence, and each layer of decoding blocks includes a multi-head attention layer, a fully connected feedforward network and a cross multi-head attention layer.

8. The ENSO prediction method based on spatiotemporal fusion neural network according to claim 7, characterized in that: The specific processing process in the first-layer decoding block is as follows: the sea surface temperature anomaly data after spatiotemporal information enhancement is input into the first-layer encoding block of the decoder, and after layer normalization and the third multi-head attention layer, the fifth time series feature is obtained, the sea surface temperature anomaly data after spatiotemporal information enhancement and the fifth time series feature are added to obtain the third fused data, the third fused data is layer normalized and then passed through the cross multi-head attention layer together with the final encoding result to obtain the sixth time series feature, the third fused data and the sixth time series feature are added to obtain the fourth fused data, the fourth fused data is sequentially passed through layer normalization and the third fully connected feedforward network to obtain the seventh time series feature, and the fourth fused data and the seventh time series feature are added as the output of the first-layer decoding block in the first decoder.

9. An ENSO prediction system based on spatiotemporal fusion neural network, characterized in that: include: The preprocessing module is used to preprocess the acquired raw sea surface temperature data to obtain the sea surface temperature anomaly data to be detected and the Index data; The spatiotemporal information enhancement module is used to enhance the spatiotemporal information of the sea surface temperature anomaly data to be detected, and obtain the sea surface temperature anomaly data after spatiotemporal information enhancement. The exponential data is enhanced with time series information to obtain the time series information enhanced Index data; The prediction module is used to transform the sea surface temperature anomaly data with enhanced spatiotemporal information and the The index data is input into the trained spatiotemporal fusion Transformer network. The first encoder encodes the sea surface temperature anomaly data after spatiotemporal information enhancement to obtain the first encoding output result, and the second encoder encodes the sea surface temperature anomaly data after temporal information enhancement to obtain the first encoding output result. The exponential data is encoded to obtain the second encoding output result, the first encoding output result and the second encoding output result are added together as the final encoding result, and the final encoding result and the sea surface temperature anomaly data after the spatiotemporal information enhancement are input into the decoder to obtain the The forecast results of the index are obtained by using a recursive forecasting strategy for multiple future moments. Index forecast results, completed Index forecast.

10. A computer electronic device, characterized in that: including memory and processor; The memory is used to store computer programs; The processor is used to implement the ENSO prediction method based on spatiotemporal fusion neural network as described in any one of claims 1 to 8 when executing the computer program.

Images