El Niño-Southern Oscillation Prediction Methods and Devices

Abstract

This invention provides a method and apparatus for predicting El Niño-Southern Oscillation (ENSO), relating to the field of meteorological forecasting technology. The method includes: inputting a historical meteorological feature sequence of a target sea area into a multi-scale hypergraph module; constructing a multi-scale hypergraph structure to capture the asynchronous dependencies between historical spatiotemporal features at different grid scales, thereby extracting an intermediate feature sequence containing multi-scale spatiotemporal asynchronous dependency information; obtaining the Oceanic Niño Index (ONI) for the historical period based on this intermediate feature sequence; and then inputting it into a frequency domain multi-scale time module to model and reconstruct the multi-scale time series features, outputting a predicted value of the Oceanic Niño Index for the target sea area in the future period. According to this invention, the intensity and evolution trend of future ENSO events can be quantified.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, specifically to El Niño-Southern Oscillation forecasting methods and apparatus. Background Technology

[0002] The El Niño–Southern Oscillation (ENSO) cycle is a fluctuation between abnormally warm (El Niño) and cold (La Niña) conditions in the tropical Pacific Ocean, representing one of the most prominent interannual climate changes on Earth. El Niño refers to a warming of the tropical Pacific that occurs every 2–7 years, while the opposite cold phase is called La Niña. These abnormally warming or cooling conditions are associated with a large-scale east-west sea-level pressure seesaw called the Southern Oscillation (ENSO), which represents the atmospheric manifestation coupled with ENSO. While originating in the tropical Pacific, ENSO's impact extends beyond regional climate, triggering large-scale weather and climate anomalies globally. ENSO forecasts offer humanity an opportunity to mitigate climate anomalies, potentially reducing their social and economic impacts and assisting in the management of natural resources and the environment.

[0003] Current deep learning methods for ENSO prediction mainly include models based on convolutional neural networks (CNNs), recurrent neural networks (RNNs) and their variants, attention mechanisms, and graph neural networks (GNNs). Early work utilized CNNs to extract local spatial features from gridded sea surface temperature (SST) anomaly fields, achieving a 17-month lead time prediction for ENSO events. Subsequently, researchers combined time-series modeling modules such as LSTMs with CNNs to better capture the dynamic evolution of SST sequences, extending the effective prediction capability to 20 months. To further model long-distance spatiotemporal dependencies, attention mechanisms were introduced, including discrete or unified spatiotemporal attention architectures, significantly improving the prediction skill for a 23-month lead time. In recent years, graph neural networks have been used to characterize non-Euclidean connections between global ocean regions, reflecting the characteristics of ocean interactions evolving over time by constructing dynamic or adaptive graph structures. Furthermore, some methods integrate multi-scale feature extraction, spatiotemporal Transformers, or hypergraph modeling strategies to more comprehensively characterize the complex spatiotemporal coupling patterns in the ENSO system, thereby improving the stability and generalization ability of predictions. These methods share a common problem: they generally perform poorly in long-term forecasting scenarios, their forecast accuracy decreases as the forecast duration increases, and the "Spring Prediction Barrier (SPB)" problem causes the forecast accuracy for spring months to be generally lower than that for other months. Summary of the Invention

[0004] This invention provides an El Niño-Southern Oscillation (ENSO) prediction method and apparatus, which allows nodes at different spatial locations to asynchronously receive and integrate historical information from previous moments based on their time differences and correlations. This effectively solves the problem of accuracy degradation caused by neglecting spatiotemporal asynchronous dependence and multi-scale coupling mechanisms in existing ENSO prediction models in long-term predictions.

[0005] Technical solution of the present invention:

[0006] A method for predicting El Niño-Southern Oscillation includes:

[0007] Step S1: Input the meteorological feature sequence of the target sea area within a historical time period into the cross-temporal multi-scale hypergraph module to obtain an intermediate feature sequence. The intermediate feature sequence contains the asynchronous spatiotemporal dependency hidden representation between nodes in the hypergraph structure. The hypergraph structure at each scale is obtained based on the feature sequence of the meteorological feature sequence at the corresponding grid scale. The cross-temporal multi-scale hypergraph module contains multiple hypergraph structures at different grid scales. Each hypergraph structure is constructed based on the spatiotemporal interaction features represented by the meteorological feature sequence at the corresponding scale, and is used to model the asynchronous spatiotemporal dependency relationship between nodes at that scale.

[0008] Step S2: Based on the intermediate feature sequence, obtain the oceanic Niño index for the historical time period;

[0009] Step S3: Input the oceanic Niño index of the historical time period and the prior time series data between the current time and the starting point of the target future time period into the frequency domain multi-scale time module to predict and output the oceanic Niño index of the target sea area in the target future time period.

[0010] An El Niño-Southern Oscillation prediction device for performing the method, comprising:

[0011] The extraction unit is configured to input the meteorological feature sequence of the target sea area in historical time period into the multi-scale hypergraph module to obtain the intermediate feature sequence. The intermediate feature sequence integrates the asynchronous dependency information between nodes in the multi-scale hypergraph structure across time and space. Each multi-scale hypergraph structure is obtained based on the feature sequence at the grid scale corresponding to the meteorological feature sequence.

[0012] The acquisition unit is configured to obtain the oceanic Niño index for the historical time period based on the intermediate feature sequence;

[0013] The prediction unit is configured to input the oceanic Niño index of the historical time period and the target future time period into the frequency domain multi-scale time module to obtain the oceanic Niño index of the target sea area in the target future time period, wherein the oceanic Niño index of the target future time period characterizes the degree of El Niño-Southern Oscillation phenomenon in the target future time period.

[0014] A computer-readable storage medium storing instructions that, when executed by at least one computing device, cause the at least one computing device to perform the El Niño-Southern Oscillation prediction method.

[0015] An electronic device includes: at least one processor and a memory storing program instructions; when the program instructions are read and executed by the processor, the electronic device causes the electronic device to perform the El Niño-Southern Oscillation prediction method.

[0016] A computer program product includes computer instructions that, when executed by a processor, implement the El Niño-Southern Oscillation prediction method.

[0017] Beneficial effects:

[0018] According to the El Niño-Southern Oscillation (ENSO) prediction method and apparatus of this disclosure, a multi-scale hypergraph module is used to obtain an intermediate feature sequence containing asynchronous spatiotemporal dependency information of the target sea area within a historical time period based on the meteorological feature sequence of the target sea area within a historical time period. Specifically, the multi-scale hypergraph module dynamically constructs asynchronous hypergraph structures at multiple grid scales based on the time delay and physical correlation strength between different spatial nodes, and adaptively receives and aggregates historical information from previous time steps. This module not only characterizes the dependency relationships between different spatial locations at the same moment but also captures asynchronous coupling evolution patterns across multiple time windows. Furthermore, a frequency-domain multi-scale time module is used to obtain the oceanic Niño index of the target sea area in the target future time period based on the intermediate feature sequence containing spatial information. In other words, the frequency-domain multi-scale time module performs frequency-domain modeling on the aforementioned intermediate feature sequence, effectively capturing the long-term evolution law of ENSO, and ultimately generating the oceanic Niño index for the target future time period. This allows for the determination of the degree of El Niño-Southern Oscillation phenomenon in the target future time period, improving the efficiency, accuracy, and stability of predicting the evolution trend of ENSO. Therefore, this disclosure can effectively solve the problem that the prediction accuracy of ENSO prediction decreases as the prediction time increases in the prior art. Attached Figure Description

[0019] The above and other objects and features of the embodiments of this disclosure will become clearer from the following description taken in conjunction with the accompanying drawings illustrating the embodiments, wherein:

[0020] Figure 1 This is a flowchart of an El Niño-Southern Oscillation prediction method according to an embodiment of this disclosure;

[0021] Figure 2 This is a system flowchart of the training method and the ENSO prediction method of the cross-temporal multi-scale ENSO prediction model according to an embodiment of this disclosure.

[0022] Figure 3 This is a schematic diagram of the structure of a multi-scale ENSO prediction model across time and space according to an embodiment of this disclosure;

[0023] Figure 4 This is a schematic diagram of the structure of a spatiotemporal multi-scale hypergraph module according to an embodiment of this disclosure;

[0024] Figure 5 This is a schematic diagram of the structure of a frequency domain multi-scale time module according to an embodiment of this disclosure;

[0025] Figure 6 This is a schematic diagram of the overall structure of the spatiotemporal multi-scale ENSO prediction model according to an embodiment of this disclosure;

[0026] Figure 7 This is a block diagram of the El Niño-Southern Oscillation prediction device disclosed herein;

[0027] Figure 8 This is a schematic diagram of the structure of the electronic device disclosed herein. Detailed Implementation

[0028] Exemplary embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings.

[0029] This disclosure addresses the challenges of capturing spatiotemporal information. First, it introduces a multi-scale hypergraph module across spatiotemporal dimensions. Based on the time delay and physical correlation strength between different spatial nodes, it dynamically constructs asynchronous hypergraph structures at multiple grid scales and adaptively receives and aggregates historical information from previous time steps. This module not only characterizes the dependencies between different spatial locations at the same moment but also captures asynchronous coupling evolution patterns across multiple time windows, thereby achieving multi-level, multi-scale modeling of complex air-sea interactions in the ENSO system. Second, in the temporal dimension, a frequency-domain multi-scale time module maps feature sequences from the time domain to the frequency domain. It utilizes the periodic characteristics of different frequency bands in the frequency domain for modeling, effectively alleviating the problem of long-term prediction error accumulation in traditional time-domain modeling. Furthermore, it can recover the time signal through inverse Fourier transform, enabling accurate prediction of the Oceanic Niño Index (ONI) for future time periods. This allows for understanding the extent of El Niño-Southern Oscillation phenomena in the target future time period, improving the efficiency, accuracy, and stability of predicting ENSO evolution trends.

[0030] Figure 1 The flowchart below shows a disclosed embodiment of an El Niño-Southern Oscillation prediction method, which includes the following steps:

[0031] In step S1, the meteorological feature sequence of the target sea area over a historical time period is input into the cross-temporal multi-scale hypergraph module to obtain an intermediate feature sequence. This intermediate feature sequence contains hidden representations of asynchronous spatiotemporal dependencies between nodes in the hypergraph structure. The hypergraph structure at each scale is obtained based on the feature sequence of the meteorological feature sequence at the corresponding grid scale. The cross-temporal multi-scale hypergraph module contains multiple hypergraph structures at different grid scales. Each hypergraph structure is constructed based on the spatiotemporal interaction features represented by the meteorological feature sequence at the corresponding scale, and is used to model the asynchronous spatiotemporal dependencies between nodes at that scale.

[0032] In step S2, the oceanic Niño index for historical time periods is obtained based on the intermediate feature sequence.

[0033] In step S3, the oceanic Niño index of the historical time period and the prior time series data between the current time and the starting point of the target future time period are input into the frequency domain multi-scale time module to obtain the oceanic Niño index of the target sea area in the target future time period. The oceanic Niño index of the target future time period represents the degree of El Niño-Southern Oscillation phenomenon in the target future time period.

[0034] As an example, the El Niño-Southern Oscillation (ENSO) prediction method of this embodiment can be applied to the ENSO prediction model. Specifically, the ENSO prediction model includes a multi-scale hypergraph module, a transformation module, and a frequency-domain multi-scale time module. The transformation module obtains the oceanic Niño index for historical time periods based on intermediate feature sequences. The aforementioned transformation module can be a multilayer perceptron (MLP), and this disclosure does not limit its use.

[0035] As an example, this ENSO prediction model can be trained based on historical meteorological feature sequences and corresponding ONI values ​​for a target sea area. For instance, the historical meteorological feature sequences of the target sea area can be used as training samples, and the ONI values ​​for future time periods of the target sea area can be used as the prediction result labels for the training samples to train the ENSO prediction model. Another example is that if it is necessary to predict the ONI values ​​for the next 12 months based on the meteorological feature sequences of the past 12 months, then during the training of the ENSO prediction model, both the historical and future time periods are set to 12 months. The meteorological feature sequences of the target sea area from 1 to 12 months are used as training samples, and the ONI values ​​for the target sea area from 12 to 24 months are used as the prediction result labels.

[0036] It should be noted that the specific training process will be explained in detail later, and will not be elaborated here.

[0037] As an example, the meteorological characteristic sequence of a target sea area over a historical period can be obtained by preprocessing a global sea surface multi-feature sample database. For instance, the meteorological characteristic sequence of a target sea area over a historical period could be sea surface temperature, heat capacity, etc., from a certain point in the past to the present (e.g., the past 12 months). It should be noted that the global sea surface multi-feature data sample database can be obtained through, but is not limited to, the following addresses: meteorological satellite data download websites, the World Meteorological Organization's climate data storage and query platform, etc.

[0038] The aforementioned multi-scale hypergraph module, based on the characteristics and requirements of the El Niño-Southern Oscillation (ESO) prediction problem, explicitly models a multi-scale coupling mechanism where the state "here and now" influences "there and then" through physical paths. This enhances the ability of nodes at different spatial locations to asynchronously receive and integrate historical information from previous moments based on their time differences and correlations. Simultaneously, a feature sequence construction mechanism from fine-grained to coarse-grained multi-grid scales is introduced to enhance the model's comprehensive extraction capability of local information and global spatial features. By introducing a hypergraph structure, multiple connections are established between distant and nearby nodes in physical space, enabling each spatial node to aggregate information from more related regions. Furthermore, using a month as the basic time step, the module dynamically generates hyperedge structures corresponding to multiple grid scales at each historical time step based on the current and previous meteorological states. Combining the hyperedge evolution trajectory of previous time steps, it adaptively adjusts the node connection weights, hyperedge composition, and topological morphology of the hypergraph at each scale in the current time step. This time-space-scale triple adaptive mechanism enables the model to accurately reproduce the nonlinear, asynchronous, multi-scale dynamic evolution process of the ENSO phenomenon in the real physical world, effectively restoring its complete life cycle from initial disturbance → energy accumulation → phase propagation → event outbreak.

[0039] Based on the characteristics and requirements of the El Niño-Southern Oscillation (ESO) prediction problem, the aforementioned frequency-domain multi-scale time module proposes a frequency-domain-based time series processing strategy. This module maps the historical El Niño index to the frequency domain and decomposes it into time features of multiple different frequency bands, thereby achieving decoupled modeling of short-term fluctuations and long-term trends. In the frequency domain, attention mechanisms are introduced into the time features of each frequency band for feature enhancement and encoding, and then the time features are restored to a time-domain representation through a frequency-domain decoding unit and a frequency-domain fusion unit.

[0040] Optionally, step S1 may include:

[0041] Step S11: Input the meteorological feature sequence of historical time period into the multi-scale data processing unit to obtain the feature sequence of meteorological feature sequence at multiple grid scales; the multi-scale data processing unit is an internal sub-unit of the cross-temporal multi-scale hypergraph module;

[0042] Step S12: Input the feature sequence at the current scale into the corresponding cross-scale asynchronous processing network unit to obtain the hidden features of the asynchronous feature sequence, wherein the hidden features contain the asynchronous spatiotemporal dependencies between nodes in the hypergraph structure at the current scale; wherein the multiple grid scales correspond one-to-one with multiple parallel cross-scale asynchronous processing network units; combine the hidden features of all scales to obtain the intermediate feature sequence.

[0043] As can be seen from the above, the input of the network block in step S1 is the single-scale meteorological feature sequence of the target sea area within a historical time period; the output of each parallel cross-scale asynchronous processing network unit is combined as the overall output of step S1.

[0044] In this embodiment, the meteorological feature sequence of historical time is divided into feature sequences at multiple network scales, which can accurately characterize the evolution of meteorological variables in the target sea area in the time dimension, realize the efficient aggregation of spatial information, and use multiple cross-scale asynchronous processing network units to process the feature sequences at each network scale in parallel, so as to quickly obtain the hidden features of the feature sequences at each network scale, and thus quickly obtain the intermediate feature sequences.

[0045] As an example, the aforementioned multi-scale data processing unit creates feature sequences at different granularities across multiple grid scales from historical meteorological feature sequences. Specifically, based on a preset granularity window size, the multi-scale data processing unit divides the historical meteorological feature sequences into feature sequences at multiple different grid scales, ranging from fine-grained to coarse-grained; for example, the original grid scale of the historical meteorological feature sequences is latitude. ,longitude A total of 24 The global real-world physical space extent is represented by a 48-grid point matrix, with preset granularity window sizes of 24 and 48. Then, a 4-dimensional convolution can be used to obtain the 4-dimensional matrix. 12 and 2 The six feature sequences at two different grid scales represent feature sequence representations from local to global grid scales.

[0046] As an example, the parameters of the multiple parallel cross-scale asynchronous processing network units described above can be shared, and this disclosure does not limit this. Each cross-scale asynchronous processing network unit can employ a cross-spatiotemporal hypergraph learning neural network, and this disclosure also does not limit this.

[0047] Specifically, each cross-scale asynchronous processing network unit is used to output the hidden features of the feature sequences at the corresponding grid scale based on the feature sequences at different grid scales. In other words, the input of each cross-scale asynchronous processing network unit is the feature sequence of the target sea area at a grid scale within a historical time period. The feature sequences at the corresponding grid scale are processed independently to obtain the feature time series with spatial information at the corresponding network scale of the target sea area within a historical time period, i.e. the aforementioned hidden features. Multiple network blocks can derive hidden features at multiple grid scales in parallel and combine multiple hidden features as the output intermediate feature sequence of the cross-temporal multi-scale hypergraph module.

[0048] Optionally, in step S12, the cross-scale asynchronous processing network unit includes a spatial hypergraph learning subunit, a hyperedge asynchronous diffusion subunit, a weight asynchronous optimization subunit, and a hypergraph information aggregation subunit. Step S12 may include:

[0049] Step S121: Input the feature sequence at the current scale into the spatial hypergraph learning subunit to obtain the hypergraph structure of the feature sequence at the current scale, wherein the hypergraph structure of the feature sequence at the current scale includes the hypergraph structure at multiple previous historical moments;

[0050] Step S122: Input the hypergraph structure of the feature sequence at the current scale into the hyperedge asynchronous diffusion subunit in sequence, combine it with the hyperedge structure at the previous historical time, and inject asynchronous interaction information across time windows through a learnable attention mechanism to generate the enhanced current hyperedge features;

[0051] Step S123: Input the enhanced current hyperedge features into the weight asynchronous optimization subunit, calculate the time and space weights according to the time delay and spatial similarity respectively, realize the collaborative calculation of spatiotemporal dynamic weights, and perform adaptive weighted fusion of historical hyperedge features;

[0052] Step S124: Input the historical hyperedge features after adaptive weighted fusion into the hypergraph information aggregation subunit, perform multiple rounds of message passing and node aggregation through graph neural networks or attention mechanisms, and output the hidden feature vector containing asynchronous spatiotemporal dependency information at the current grid scale.

[0053] In this embodiment, the spatial hypergraph learning subunit can construct a hypergraph structure for feature sequences of each preset duration at the corresponding network scale. This allows for adaptive adjustment of the hypergraph structure at each grid scale for each preset duration, taking into account the dynamic evolution characteristics of the time dimension. This enables a more accurate simulation of the spatial dynamic evolution of the ENSO phenomenon in real physical time. Furthermore, based on this hypergraph structure, through the joint mechanism of the hyperedge asynchronous diffusion subunit and the weight asynchronous optimization subunit, the model can dynamically adjust the hyperedge weights and the intensity of historical information injection according to the changes in the network state in previous historical moments (i.e., the differences in feature sequences at each previous historical time step and the relative differences between spatial points). This achieves effective modeling of the nonlinear, multi-scale, asynchronous evolution structure of complex air-sea coupled systems, thereby outputting relatively accurate and physically meaningful hidden features.

[0054] As an example, the aforementioned weight asynchronous optimization subunit can adaptively adjust the hyperedge weights according to the changes in the network state in the previous historical time and space, and selectively aggregate the historical hyperedge information to improve the expressive power and prediction robustness of the hidden features; the aforementioned preset duration can be in months, and this disclosure does not limit it. It should be noted that the preset duration can also be called a "time step".

[0055] As an example, the aforementioned hypergraph information aggregation subunit can adaptively adjust the hyperedge weights according to the changes in network state within each preset duration, aggregate the hypergraph results, and obtain the corresponding accurate hidden features. The preset duration can be in months, and this disclosure does not limit this. It should be noted that the preset duration can also be called a time step. The cross-scale asynchronous processing network unit includes a hypergraph learning subunit, a hyperedge asynchronous diffusion subunit, a weight asynchronous optimization subunit, and a hypergraph information aggregation subunit. The number of channels between each unit is consistent or can be aligned through projection layers to ensure efficient transfer of feature flow between modules. It should be noted that the "change in network state" refers to the change in the value, distribution, or spatial pattern of the input feature sequence within each preset duration (time step), which is the core signal driving the dynamic evolution of the hypergraph structure and the adaptive adjustment of weights.

[0056] The following sections describe the spatial hypergraph learning subunit, the asynchronous hyperedge diffusion subunit, the asynchronous weight optimization subunit, and the hypergraph information aggregation subunit:

[0057] Step S121: The spatial hypergraph learning subunit can construct a hypergraph structure based on the geographical distance between meteorological observation points and the similarity of the meteorological feature sequences of these observation points. Specifically, it can calculate the Euclidean distance between observation points based on the geographical coordinate information of the meteorological observation points in the current feature sequence, and select observation points whose Euclidean distance is less than a first set threshold, incorporating these selected observation points into the same hyperedge. Alternatively, it can construct a hyperedge based on the correlation between the meteorological feature sequences of the meteorological observation points in the current feature sequence. When the correlation between the feature sequences of two meteorological observation points in a historical time period exceeds a second set threshold, the two meteorological observation points are incorporated into the same hyperedge. Furthermore, it can incorporate meteorological observation points that simultaneously meet both of the above conditions into the same hyperedge. This disclosure does not limit this, as long as a set of nodes with high meteorological similarity can be formed. The hypergraph structure output by this spatial graph construction unit is used for subsequent hypergraph information aggregation operations to enhance spatial dependency modeling capabilities.

[0058] It should be noted that, for the s-th grid scale, the hypergraph structure expression constructed based on the geographical distance between meteorological observation points or the similarity of meteorological feature sequences of meteorological observation points can all adopt the following formula:

[0059] ,

[0060] When constructing a hypergraph structure based on the geographical distance between meteorological observation points, This represents the first set threshold; when constructing a hypergraph structure based on the similarity of meteorological feature sequences from meteorological observation points, represents the second set threshold; v represents the node position at the s-th grid scale; and e represents the hyperedge at the s-th grid scale. This represents the hypergraph structure at scale s.

[0061] Step S122: The asynchronous diffusion subunit of the hyperedge outputs asynchronous interaction information across time windows based on the above hypergraph structure. Specifically, within each preset duration, for the hyperedge in the hypergraph structure at the current time, information about it and the historical hyperedge structure at the previous time is collected through a learnable attention mechanism to obtain asynchronous hyperedge interaction information across time windows. That is, historical node information is transmitted through a small number of historical hyperedge-current hyperedge relationships instead of a large number of historical node-current node relationships, thereby realizing a lightweight modeling aggregation path.

[0062] For example, for a certain hyperedge in the current hypergraph, it will receive feature contributions from all the selected hyperedges in the historical hypergraph. The selected hyperedges are those that have spatial overlap or semantic similarity with the current hyperedge. After linear transformation and weighted fusion, the spatiotemporal expressive power of the current hyperedge is enhanced.

[0063] Step S123: The asynchronous weight optimization sub-unit introduces a spatiotemporal joint weight mechanism to adaptively adjust the information injection intensity for different historical moments and different hyperedges. This unit collaboratively designs weights from two dimensions: For temporal dynamic weights, a personalized temporal coding function is designed for each spatial node to capture its unique periodic response pattern (such as 3–6 month oscillations or interannual fluctuations), and the robustness of sparse regions is enhanced through global shared coding. The two are weighted and fused through learnable parameters to form the final temporal dynamic weight; For spatial dynamic weights, the structural similarity between the current hyperedge and historical hyperedges is calculated, and attention weights are allocated based on similarity to prioritize the injection of physically strongly correlated regional information. Finally, the spatial dynamic weight and temporal dynamic weight are weighted and fused to obtain the spatiotemporal dynamic weight, and the hyperedge state at the current moment is updated to adjust the contribution intensity of historical hyperedge features to the current hyperedge.

[0064] It should be noted that the personalized time-domain coding function is for each spatial node. Design its own time-sensitive function This encoding not only introduces absolute time difference The influence of this is also captured through learnable trigonometric function bases to capture the node-specific periodic response patterns:

[0065] ,

[0066] After mapping, we can obtain:

[0067] ,

[0068] in, For nodes Personalized time-domain dynamic encoding vectors; For the current moment With historical moments The time interval between; for The Each component index, value range ; For learnable angular frequency parameters; The dimension of the time-domain encoded vector; For nodes At the present moment Historical moments Personalized time weighting; These are learnable weight vectors; This is the Sigmoid activation function.

[0069] Globally shared encoding functions With time-sensitive function Similar, but parameters Shared by all nodes:

[0070] ,

[0071] After mapping, we can obtain:

[0072] ,

[0073] in, For global time weights; These are learnable weight vectors.

[0074] The final time-domain weights are:

[0075] ,

[0076] in, For the final time weight; For learnable sparsity adjustment parameters; It is an exponential function.

[0077] Step S124: The hypergraph information aggregation subunit receives the hyperedge features after asynchronous diffusion and dynamic weighting, integrates the hyperedges with cross-temporal and spatial information into the hypergraph network, and uses hypergraph convolution operations or hypergraph attention networks to integrate the hyperedges. The propagation path of the "nodes" involves multiple rounds of message passing and aggregation. This allows each node to aggregate multi-mode node information at the current moment while integrating evolutionary information from different spatial locations in previous moments. This significantly enhances the model's ability to spatially model complex ocean-atmosphere coupled systems within a single grid scale.

[0078] Through this embodiment, the above three-stage processing achieves "historical information injection". Weight adaptive adjustment The closed-loop modeling process of "high-order spatial aggregation" allows the model to dynamically adjust the hyperedge structure, diffusion intensity, and aggregation weights based on the changes in network state at each time step (i.e., the difference between the feature sequence at the current time and the previous time), thereby accurately depicting the spatial propagation path, phase delay, and energy transfer process of the ENSO phenomenon in real physical time.

[0079] As an example, in a hyperedge connecting five meteorological observation points, the model first introduces the historical state of the hyperedge over the past three months through an asynchronous diffusion subunit. Then, an asynchronous weighting optimization subunit dynamically adjusts the injection weights of each node based on the current sea surface temperature anomaly distribution. Finally, a weighted fusion of all node features within the hyperedge is performed in the hypergraph information aggregation subunit to generate a comprehensive dynamic representation of the node in the current spatiotemporal context. Furthermore, if the node participates in multiple hyperedges simultaneously (e.g., belonging to the "East-West Pacific Coupling Zone" and the "Equatorial Kelvin Wave Channel Zone"), spatial perception capabilities can be further enhanced through multi-hyperedge aggregation strategies (such as averaging, maximum, or attention-weighted aggregation), outputting more discriminative hidden features.

[0080] According to embodiments of this disclosure, the frequency domain multi-scale time module includes a frequency domain coding unit and a frequency domain decoding unit. The ocean Niño index for a historical time period and prior time series data from the current time to the start point of the target future time period are input into the frequency domain multi-scale time module to obtain the ocean Niño index of the target sea area in the target future time period. This can be achieved by: inputting the ocean Niño index for a historical time period into the frequency domain coding unit to obtain coding features; and inputting the coding features into the frequency domain decoding unit to obtain the ocean Niño index of the target sea area in the target future time period.

[0081] In this embodiment, the ocean Niño index for the input historical time period is mapped to the frequency domain space and decomposed into time features of multiple different frequency segments, thereby achieving decoupled modeling of short-term fluctuations and long-term trends. Moreover, in the frequency domain, the time features of each frequency segment are enhanced and encoded, and then restored to the time domain representation through the frequency domain decoding unit and the frequency domain fusion unit.

[0082] As an example, the frequency domain multi-scale time module can be based on a self-attention mechanism architecture to transform the intermediate feature sequence of the target sea area within a historical time period into the frequency domain and extract the ONI prediction results. It should be noted that the frequency domain multi-scale time module can contain multiple frequency domain operation units, which are described below:

[0083] Step S31: The frequency domain coding unit performs a Discrete Fourier Transform (DFT) on the Ocean Niño Index for historical time periods, mapping the time series of the Ocean Niño Index from the time domain to the frequency domain. This extracts frequency segment features reflecting changes at different time scales, i.e., extracts the temporal features of different frequency segments to capture the periodicity and oscillation patterns in the ENSO phenomenon. Furthermore, it encodes the temporal features of multiple frequency segments by constructing a frequency attention map to highlight the influence of key frequency segments. For example, a frequency domain attention module based on a Transformer structure can be used to model the dependencies between multiple frequency segments and output the corresponding encoded features. Specifically, the frequency domain coding unit constructs a frequency-wise Transformer Encoder in the frequency dimension. Through a multi-head attention mechanism, it captures long-distance dependencies and interactions between different frequencies, thereby outputting the corresponding encoded features. The frequency domain coding unit can stack multiple frequency domain Transformer layers, each containing a frequency attention sublayer and a feedforward network sublayer. Residual connections and normalization operations are used to make the network more stably focus on key frequency segments strongly correlated with ENSO.

[0084] Step S32: The frequency domain decoding unit constructs a prediction structure for future frequency distribution based on the encoded features output by the frequency domain coding unit, i.e., predicting the Oceanic Niño Index (ONI) of the target sea area in the target future time period. For example, the frequency domain decoding unit can adopt a frequency time series modeling network based on a Transformer decoder structure. Specifically, this decoder structure is a frequency-aware Transformer decoder. During the training phase, it can fuse historical frequency representations (i.e., encoded features output by the frequency domain coding unit) and target frequency location information (the real ONI corresponding to the encoded features) to perform sequence prediction in the frequency domain. Each decoding module contains a frequency attention mechanism to progressively predict the evolution characteristics of each future frequency band and output the predicted spectrum results. Then, through inverse Fourier transform (IDFT), the spectrum signal is restored to a time series to obtain the final prediction result of the ONI.

[0085] According to embodiments of this disclosure, the spatiotemporal multi-scale hypergraph module and the frequency domain multi-scale time module are trained as follows: a training sample set is obtained based on the meteorological feature sequence of each meteorological feature among various global sea surface meteorological features and the oceanic Niño index; for each training sample, the following processing is performed: the historical duration meteorological feature sequence in the current training sample is input into the spatiotemporal multi-scale hypergraph module to obtain the estimated intermediate feature sequence; based on the estimated intermediate feature sequence, the historical estimated oceanic Niño index is obtained; the historical estimated oceanic Niño index and the prior oceanic Niño index after the historical time period and before the target future time period are input into the frequency domain multi-scale time module to obtain the target future time period estimated oceanic Niño index; based on the target future time period oceanic Niño index and the actual oceanic Niño index, the parameters of the spatiotemporal multi-scale hypergraph module and the frequency domain multi-scale time module are adjusted.

[0086] The spatiotemporal multi-scale hypergraph module and the frequency domain multi-scale time module can also be trained as follows: A training sample set is obtained based on the meteorological feature sequences and oceanic Niño index for each of various global meteorological features. Each training sample includes a meteorological feature sequence of a first predetermined duration and a true oceanic Niño index of a second predetermined duration following the first predetermined duration. For each training sample, the following processing is performed: The meteorological feature sequence of the first predetermined duration in the current training sample is input into the spatiotemporal multi-scale hypergraph module to obtain a predicted intermediate feature sequence. Based on the predicted intermediate feature sequence, a first predicted oceanic Niño index is obtained, where the first predicted oceanic Niño index is the predicted oceanic Niño index of the first predetermined duration. The first predicted oceanic Niño index and the second predetermined duration following the first predetermined duration are input into the frequency domain multi-scale time module to obtain a second predicted oceanic Niño index, where the second predicted oceanic Niño index is the predicted oceanic Niño index of the second predetermined duration following the first predetermined duration. Based on the second predicted oceanic Niño index and the true oceanic Niño index, the parameters of the adaptive multi-scale spatial hypergraph module and the frequency domain multi-scale time module are adjusted.

[0087] As an example, during the training process, the frequency domain multi-scale time module can also be input with the real ocean Niño index for historical time periods, which is not limited in this disclosure.

[0088] This embodiment allows for convenient and accurate training of multi-scale hypergraph modules across time and space and frequency-domain multi-scale time modules.

[0089] According to embodiments of this disclosure, a training sample set is obtained based on the meteorological feature sequence of each of the various meteorological features of the global sea surface and the oceanic Niño index, including: obtaining various meteorological features of the global sea surface; preprocessing the various meteorological features of the global sea surface to obtain the meteorological feature sequence of each meteorological feature and the oceanic Niño index; performing sequential sliding sampling on the meteorological feature sequence of each meteorological feature and the oceanic Niño index; and obtaining a training sample set based on the sampling results.

[0090] This embodiment uses sampling of preprocessed meteorological feature sequences and oceanic Nino index to increase the sample size and alleviate the problem of insufficient effective data in the marine field.

[0091] As an example, various meteorological features of the global sea surface can be obtained from a global sea surface multi-feature data sample library. This library can be accessed from meteorological satellite data download websites, the World Meteorological Organization's climate data storage and query platform, etc., and this disclosure does not limit its use. The ocean El Niño / Southern Oscillation (OSO) prediction data samples in the global sea surface multi-feature data sample library can be divided into two categories: historical observation data and model simulation data. The model simulation data is observational data simulated based on historical observation data. These two types of data are usually used in combination to improve the accuracy and reliability of predictions. For example, historical observation data can be used to calibrate models and serve as baseline data for model input; model simulation data can help predict the probability and trend of future OSO / Oscillation events.

[0092] As an example, the meteorological feature sequence and oceanic Niño index for each meteorological characteristic can be obtained by preprocessing a global sea surface multi-feature sample database. For instance, the meteorological feature sequence of a target sea area over a historical period can be the sea surface temperature, heat capacity, etc., from a certain point in the past to the present (e.g., the past 12 months), while the ONI over a historical period can be the actual ONI of the target sea area from a certain point in the past to the present (e.g., the past 12 months). It should be noted that the actual ONI over a historical period can be used to characterize the degree of ENSO phenomenon from a certain point in the past to the present (e.g., the past 12 months).

[0093] The aforementioned preprocessing can be performed using data preprocessing software, Python data processing packages, etc., and this disclosure does not limit this approach. The purpose of preprocessing the obtained global sea surface multi-feature sample database is to obtain a sample set in a specific format containing features relevant to the prediction of Ocean El Niño / Southern Oscillation (OSO). The preprocessed features are predictive factors closely related to the timing of OSO, such as sea surface temperature and heat capacity. For ease of understanding, sea surface temperature and heat capacity are described below. Data preprocessing can be performed using data preprocessing software, Python data processing packages, etc., resulting in final data formats for sea surface temperature and heat capacity as meteorological feature sequences and OSO time series data, respectively. The meteorological feature sequence can be represented as... H and W represent latitude and longitude, respectively. ONI time series data can be represented as... .

[0094] As an example, after obtaining the meteorological characteristic sequence and oceanic Niño index of the target sea area over a historical period, the sampling interval can be used as a reference. Sliding sampling was performed on the meteorological feature sequence and ONI. Sampling interval. This can be set for a year, or for the time resolution of the sequence data (usually monthly). The advantage of the latter is that it can greatly increase the sample size, alleviating the problem of insufficient effective data in the marine field. The disadvantage is that data reuse may lead to model overfitting. (Sampling interval for months) Setting it to 12 is equivalent to setting the sampling interval for the year. Set to 1. Therefore, considering insufficient data and the risk of model overfitting, the sampling interval for months is... The range of values ​​can be .

[0095] As an example, the loss can be determined based on the estimated and actual oceanic Niño indices for a target future time period. This loss can then be used to adjust the parameters of the adaptive multi-scale spatial hypergraph module and the frequency-domain multi-scale temporal module. The loss function used to calculate this loss can be the mean squared error loss between the estimated and actual oceanic Niño indices; this disclosure does not limit its application. It should be noted that, regarding training, relevant parameters, methods, and functions can be selected. Meteorological feature sequence samples can be input into the adaptive multi-scale spatiotemporal ENSO prediction model to train the model and obtain a trained adaptive multi-scale spatiotemporal ENSO prediction model.

[0096] To better understand this disclosure, the following is in conjunction with... Figure 2 A systematic explanation will be provided.

[0097] Figure 2The training method for a multi-scale, spatiotemporal ENSO prediction model and the system flow of the ENSO prediction method are demonstrated, such as... Figure 2 As shown, it may include the following steps:

[0098] S210: Obtain a global sea surface multi-feature data sample library;

[0099] S220: Preprocess the data in the global sea surface multi-feature data sample library to extract meteorological feature sequence data and ONI time series data containing features related to the prediction of El Niño-Southern Oscillation.

[0100] S230: Based on the spatiotemporal characteristics and evolution trends of the El Niño-Southern Oscillation (ENSO) in the ocean, a multi-scale ENSO prediction model is constructed and trained using meteorological feature sequence data such as sea surface temperature and heat capacity, as well as ONI time series data.

[0101] S240: Based on a trained spatiotemporal multi-scale ENSO prediction model, predict ONI time series data for future time periods.

[0102] The typical problem addressed by the aforementioned multi-scale, spatiotemporal ENSO prediction model is the extraction of spatiotemporal sequence features and time series prediction. Specifically, it involves extracting intermediate feature sequences containing spatial information from past meteorological feature sequence data, transforming these intermediate feature sequences into past ONI time series data, and then using this past ONI time series data to predict future ONI time series data. The key here lies in capturing the spatial and temporal evolution characteristics of the meteorological feature sequence data.

[0103] Figure 3 The structure of the spatiotemporal multi-scale ENSO prediction model is shown, such as Figure 3 As shown, the input data for the multi-scale ENSO prediction model across time and space is the meteorological characteristic sequence data of the target sea area over the past T time periods. , Figure 3 As shown in the diagram, after the meteorological feature sequence data is extracted by the multi-scale hypergraph module, an intermediate feature sequence containing historical spatiotemporal interaction feature information is obtained. Then, the intermediate feature sequence is transformed into ONI time series data for the past T time periods. Then, time series data The ONI prediction results for the target sea area at future τ times were obtained by frequency domain multi-scale time module. .

[0104] Figure 4 The structure of the multi-scale hypergraph module across time and space is shown, such as... Figure 4As shown, the spatiotemporal multi-scale hypergraph module includes a multi-scale data processing unit and multiple parallel cross-scale asynchronous processing network units. Each cross-scale asynchronous processing network unit includes a spatial hypergraph learning subunit, a hyperedge asynchronous diffusion subunit, a weight asynchronous optimization subunit, and a hypergraph information aggregation subunit.

[0105] The multi-scale hypergraph module targets meteorological feature sequence data from the past. Its purpose is to extract the cross-temporal and spatial interaction dependencies in the meteorological feature sequence data of historical time periods, obtain intermediate feature time series containing historical asynchronous interaction information, and then transform them into the corresponding ONI time series data of the target sea area in the past time period. Figure 4 This illustrates the case where the feature dimension (i.e., number of channels) is 1 at each time step. In reality, the feature dimension at each time step is greater than 1, but the principle remains the same. The functions of each unit in the adaptive multi-scale spatial hypergraph module are described below:

[0106] The aforementioned multi-scale data processing unit can perform scale processing on meteorological feature sequence data from past time periods, such as forming a coarse-grained spatial representation based on two-dimensional convolution. Figure 4 The demonstration focuses on processing time points t-T+1, ..., t. For example, two-dimensional convolution can be used to convolve the spatial grid to obtain a spatiotemporally correlated grid sequence representation of spatial grid points at multiple grid scales (i.e., the sequence features mentioned above), such as... Figure 4 The S shown represents the number of different grid scales. The spatiotemporal correlated grid sequence of each grid scale output by the multi-scale data processing unit is input into a joint unit composed of its corresponding spatial hypergraph learning subunit, hyperedge asynchronous diffusion subunit, weight asynchronous optimization subunit, and hypergraph information aggregation subunit, so as to facilitate subsequent spatial structure modeling and aggregation.

[0107] The aforementioned spatial hypergraph learning subunits construct spatial graph connection structures (i.e., hypergraph structures) at their respective grid scales based on the geographical distance or similarity of meteorological characteristics between marine physical variables. For example, at each time step t in the past (e.g., a month as a time step, which is also the preset duration in the above embodiment), all grid points within a grid scale of the current time step are considered as graph nodes. The Euclidean distance and similarity of meteorological characteristics between each graph node are calculated to form a weighted adjacency matrix, thereby establishing a spatially meaningful graph. Furthermore, in order to capture the potential dynamic correlations between distant regions, a hypergraph structure is introduced, aggregating multiple nodes with similar change patterns in each hyperedge, thereby constructing hypergraph connection relationships that reflect spatial coupling patterns.

[0108] The aforementioned asynchronous diffusion subunit is used to effectively model asynchronous spatial dependencies across time windows without increasing node-level historical storage overhead. Specifically, this unit receives the hypergraph structure of the current time step output by the spatial graph construction unit, and combines it with the historical hyperedge structure of the previous time step, injecting cross-spatial interaction information through a learnable attention mechanism. This achieves "lightweight historical modeling + efficient information injection," avoiding the memory explosion problem caused by storing historical states node by node in traditional methods, while retaining the modeling capability of asynchronous propagation paths between key regions in the ENSO system (such as the delayed response caused by the eastward propagation of Kelvin waves).

[0109] The aforementioned weighted asynchronous optimization subunit is used to assign reasonable and adaptive weights to historical information from different preceding moments, overcoming the limitation of simple time decay functions in characterizing the complex dynamics of the ENSO system. Specifically, this unit collaboratively designs a weighting mechanism from both temporal and spatial dimensions, enabling the model to dynamically adjust the injection intensity of historical information based on the current network state. This achieves "time awareness + spatial similarity + adaptive fusion," allowing the model to accurately identify which historical moments and regions have a critical impact on the current state, thereby improving its ability to predict ENSO event inflection points and phase evolution.

[0110] The aforementioned hypergraph information aggregation subunit is used to perform final information aggregation on the hypergraph structure after asynchronous diffusion and dynamic weighting, outputting hidden features with spatial dependencies and historical asynchronous interaction information. This achieves a three-in-one modeling approach of "multi-scale spatial dependency extraction + historical asynchronous information fusion + dynamic weight adjustment," and the final output hidden features possess strong physical interpretability and high predictive discriminative power.

[0111] It's important to note that in hypergraphs, the node-hyperedge-node aggregation path is a mechanism for information transfer and feature updating. Using this aggregation path, nodes and hyperedges can reinforce each other, thus better representing the complex relationships within the graph structure. This facilitates the modeling and analysis of hypergraph data and has wide applications in various fields such as image segmentation, genetic medicine, and natural language processing. For example, in image segmentation, nodes can be pixels, and hyperedges represent certain relationships between pixels. This aggregation path allows for better utilization of contextual information in the image for pixel classification.

[0112] Figure 5 The structure of the frequency domain multi-scale time module is shown, such as Figure 5 As shown, the frequency domain multi-scale time module includes a frequency domain coding unit and a frequency domain decoding unit.

[0113] The frequency domain multi-scale time module targets ONI time series data of a target sea area over past periods. Its purpose is to capture the temporal evolution patterns within this data and predict future ONI time series data. The functions of each unit in the frequency domain multi-scale time module are described below:

[0114] The aforementioned frequency-domain encoding unit transforms the input ONI time series data from the time domain to the frequency domain to more effectively extract periodic and trend features, i.e., extract the time features of different frequency bands. Simultaneously, deep encoding of the time features of different frequency bands is performed, allowing the construction of a frequency-dimensional attention map through a frequency attention mechanism to identify key frequency bands with significant influence in ENSO evolution.

[0115] The aforementioned frequency-domain decoding unit performs feature transformation on the multi-frequency band time features output by the frequency-domain coding unit, generating prediction results for each frequency band. This is equivalent to transforming the multi-frequency band time features output by the frequency-domain coding unit into a unified feature space representation that facilitates fusion, i.e., the prediction features for future frequency bands. Subsequently, the prediction results for each frequency band are integrated to predict the final ONI—that is, by applying the inverse Fourier transform (IDFT), the fused frequency-domain features are restored to a time-domain signal, predicting the ONI time series data for the target region in future time periods.

[0116] Figure 6 The overall structure of the spatiotemporal multi-scale spatiotemporal ENSO prediction model is shown, such as... Figure 6 As shown, the results of preprocessing the meteorological characteristic sequence of the target sea area over a historical period are presented. The input is a multi-scale hypergraph module that generates a multi-scale, multi-scale, multi-scale, multi-scale interactive meteorological feature time series, i.e., an intermediate feature series. Information from multiple scales of the intermediate feature series is then aggregated using a multilayer perceptron (MLP). The ONI time series data of the target sea area over a historical period were obtained. Then, the ONI time series data [ Input the frequency domain multi-scale time module to predict the final ONI result of the target sea area in the target future time period. ].

[0117] In summary, this disclosure introduces a multi-scale hypergraph module to dynamically model how nodes at different spatial locations asynchronously receive and integrate historical information from previous moments based on their temporal differences and correlations. It accurately characterizes the inherent cross-spatial dynamic coupling mechanism—that is, the state "here and now" influences "there and then" through physical paths—and constructs an ONI time series with cross-spatial interaction characteristics for historical time periods. Subsequently, a frequency domain multi-scale time module is used to perform frequency domain modeling on the ONI time series data of the aforementioned historical time periods, thereby effectively capturing the long-term evolution patterns of ENSO-related variables and ultimately predicting ONI time series data for future time periods. This significantly improves the prediction efficiency, accuracy, and stability of ENSO event evolution trends.

[0118] Figure 7 This is a block diagram illustrating the El Niño-Southern Oscillation prediction device of this disclosure, as shown below. Figure 7 As shown, the device includes an extraction unit 70, an acquisition unit 72, and a prediction unit 74.

[0119] Extraction unit 70 is configured to input the meteorological feature sequence of the target sea area in historical time periods into the multi-scale hypergraph module to obtain intermediate feature sequences, wherein the intermediate feature sequences contain asynchronous spatiotemporal dependencies of the hypergraph structure, and the hypergraph structure at each scale is obtained based on the feature sequence of the meteorological feature sequence at the corresponding grid scale; acquisition unit 72 is configured to obtain the oceanic Niño index in historical time periods based on the intermediate feature sequences; prediction unit 74 is configured to input the oceanic Niño index in historical time periods and the prior time series data between the current time and the starting point of the target future time period into the frequency domain multi-scale time module to obtain the oceanic Niño index of the target sea area in the target future time period.

[0120] According to embodiments of this disclosure, the spatiotemporal multi-scale hypergraph module includes a multi-scale data processing unit and multiple parallel cross-scale asynchronous processing network units. Multiple grid scales correspond one-to-one with multiple parallel cross-scale asynchronous processing network units. The extraction unit 70 is further configured to input meteorological feature sequences from historical time periods into the multi-scale data processing unit to obtain feature sequences of the meteorological feature sequences at multiple grid scales. For each feature sequence at each grid scale, the current feature sequence is input into the corresponding cross-scale asynchronous processing network unit to obtain hidden features of the current feature sequence. These hidden features contain the spatiotemporal relationships between hyperedges in the hypergraph structure of the current feature sequence. The hidden features of the feature sequences at multiple grid scales are then determined as intermediate feature sequences.

[0121] According to embodiments of this disclosure, the cross-scale asynchronous processing network unit includes a spatial graph construction unit, a hyperedge asynchronous diffusion subunit, a weight asynchronous optimization subunit, and a hypergraph information aggregation subunit. The extraction unit 70 is further configured to:

[0122] The feature sequence at the current grid scale is input into the spatial graph construction unit to obtain the dynamic hypergraph structure of the feature sequence within multiple preset time periods (time steps); wherein, each hypergraph structure consists of a set of spatial nodes and hyperedges connecting these nodes, and is used to characterize the physical or statistical correlation between spatial regions at that time.

[0123] The multiple hypergraph structures with preset durations are sequentially input into the hyperedge asynchronous diffusion subunit. Combined with the historical hyperedge structures of the previous time step, asynchronous interaction information across time windows is injected through a learnable attention mechanism to generate enhanced current hyperedge features.

[0124] The enhanced hyperedge features are input into the weighted asynchronous optimization subunit, and the spatiotemporal dynamic weights are calculated collaboratively based on time delay and spatial similarity to adaptively weight and fuse historical hyperedge information.

[0125] The weighted and fused hyperedge features are input into the hypergraph information aggregation subunit. Through hypergraph convolution or attention mechanism, multiple rounds of message passing and node aggregation are performed to output a hidden feature vector with global spatial dependence and historical asynchronous evolution mode at the current grid scale.

[0126] According to embodiments of this disclosure, the frequency domain multi-scale time module includes a frequency domain coding unit and a frequency domain decoding unit. The prediction unit 74 is further configured to input the ocean Niño index of historical time periods into the frequency domain coding unit to obtain coding features; and to input the coding features and the target future time period into the frequency domain decoding unit to obtain the ocean Niño index of the target sea area in the target future time period.

[0127] According to embodiments of this disclosure, the prediction device further includes a training unit configured to train a spatiotemporal multi-scale hypergraph module and a frequency domain multi-scale time module in the following manner: Based on meteorological feature sequences and oceanic Niño indices for each of multiple global meteorological features, a training sample set is obtained, wherein each training sample includes a meteorological feature sequence of a first predetermined duration and a true oceanic Niño index of a second predetermined duration following the first predetermined duration; for each training sample, the following processing is performed: the meteorological feature sequence of the first predetermined duration in the current training sample is input into the spatiotemporal multi-scale hypergraph module to obtain the predicted... Intermediate feature sequence; based on the estimated intermediate feature sequence, a first estimated oceanic Niño index is obtained, wherein the first estimated oceanic Niño index is the estimated oceanic Niño index for a first predetermined duration; the first estimated oceanic Niño index and a second predetermined duration after the first predetermined duration are input into the frequency domain multi-scale time module to obtain a second estimated oceanic Niño index, wherein the second estimated oceanic Niño index is the estimated oceanic Niño index for a second predetermined duration after the first predetermined duration; based on the second estimated oceanic Niño index and the actual oceanic Niño index, the parameters of the adaptive multi-scale spatial hypergraph module and the frequency domain multi-scale time module are adjusted.

[0128] According to embodiments of this disclosure, the training unit is further configured to acquire multiple meteorological features of the global sea surface; preprocess the multiple meteorological features of the global sea surface to obtain a meteorological feature sequence and a marine Niño index for each meteorological feature; perform sequential sliding sampling on the meteorological feature sequence and marine Niño index for each meteorological feature; and acquire a training sample set based on the sampling results.

[0129] The present invention also provides an electronic device, comprising: at least one processor and a memory storing program instructions; when the program instructions are read and executed by the processor, the electronic device performs the above-described method.

[0130] Figure 8 A schematic diagram of the structure of the electronic device provided in this disclosure is shown, such as... Figure 8 As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 may call logical instructions in the memory 830 to execute any of the ENSO prediction methods provided in the above embodiments.

[0131] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0132] It should be noted that the electronic device disclosed herein can implement all the method steps implemented in the above method embodiments and can achieve the same technical effect. Therefore, the parts and beneficial effects that are the same as those in the method embodiments will not be described in detail here.

[0133] According to embodiments of the present disclosure, a computer-readable storage medium is provided that stores instructions, wherein when the instructions are executed by at least one computing device, they cause at least one computing device to perform an El Niño-Southern Oscillation prediction method as described in any of the above embodiments.

[0134] It should be noted that the non-transitory computer-readable storage medium provided in this disclosure can implement all the method steps implemented in the above method embodiments and can achieve the same technical effect. Therefore, the parts and beneficial effects that are the same as those in the method embodiments will not be described in detail here.

[0135] According to embodiments of the present disclosure, a system is provided that includes at least one computing device and at least one storage device for storing instructions, wherein the instructions, when executed by at least one computing device, cause at least one computing device to perform an El Niño-Southern Oscillation prediction method as described in any of the above embodiments.

[0136] According to embodiments of this disclosure, a computer program product is provided, including computer instructions that, when executed by a processor, are used as described in the El Niño-Southern Oscillation prediction method.

[0137] It should be noted that the computer program product provided in this disclosure can implement all the method steps implemented in the above method embodiments and can achieve the same technical effect. Therefore, the parts and beneficial effects that are the same as those in the method embodiments will not be described in detail here.

[0138] While some embodiments of this disclosure have been shown and described, those skilled in the art will understand that modifications may be made to these embodiments without departing from the principles and spirit of this disclosure, which are defined by the claims and their equivalents.

[0139] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0140] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0141] Although the invention has been described with respect to a limited number of embodiments, those skilled in the art will understand from the foregoing description that other embodiments are conceivable within the scope of the invention described herein. Furthermore, it should be noted that the language used in this specification has been chosen primarily for readability and instructional purposes, and not for the purpose of explaining or limiting the subject matter of the invention.

Claims

1. A method for predicting El Niño-Southern Oscillation, characterized in that, include: Step S1: Input the meteorological feature sequence of the target sea area within a historical time period into the cross-temporal multi-scale hypergraph module to obtain an intermediate feature sequence. The intermediate feature sequence contains the asynchronous spatiotemporal dependency hidden representation between nodes in the hypergraph structure. The hypergraph structure at each scale is obtained based on the feature sequence of the meteorological feature sequence at the corresponding grid scale. The cross-temporal multi-scale hypergraph module contains multiple hypergraph structures at different grid scales. Each hypergraph structure is constructed based on the spatiotemporal interaction features represented by the meteorological feature sequence at the corresponding scale, and is used to model the asynchronous spatiotemporal dependency relationship between nodes at that scale. Step S2: Based on the intermediate feature sequence, obtain the oceanic Niño index for the historical time period; Step S3: Input the oceanic Niño index of the historical time period and the prior time series data between the current time and the starting point of the target future time period into the frequency domain multi-scale time module to predict and output the oceanic Niño index of the target sea area in the target future time period.

2. The El Niño-Southern Oscillation prediction method according to claim 1, characterized in that, Step S1 includes: Step S11: Input the meteorological feature sequence of the historical time period into the multi-scale data processing unit to obtain the feature sequence of the meteorological feature sequence at multiple grid scales; the multi-scale data processing unit is an internal sub-unit of the cross-temporal multi-scale hypergraph module; Step S12: Input the feature sequence at the current scale into the corresponding cross-scale asynchronous processing network unit to obtain the hidden features of the asynchronous feature sequence, wherein the hidden features contain the asynchronous spatiotemporal dependencies between nodes in the hypergraph structure at the current scale; wherein the multiple grid scales correspond one-to-one with the multiple parallel cross-scale asynchronous processing network units.

3. The El Niño-Southern Oscillation prediction method according to claim 2, characterized in that, Each of the cross-scale asynchronous processing network units includes a spatial hypergraph learning subunit, a hyperedge asynchronous diffusion subunit, a weight asynchronous optimization subunit, and a hypergraph information aggregation subunit. Step S12 includes: Step S121: Input the feature sequence at the current scale into the spatial hypergraph learning subunit to obtain the hypergraph structure of the feature sequence at the current scale; Step S122: Input the hypergraph structure of the feature sequence at the current scale into the hyperedge asynchronous diffusion subunit, combine it with the hypergraph structure at the previous historical time, and inject asynchronous interaction information across time windows through a learnable attention mechanism to generate the enhanced current hyperedge features; Step S123: Input the enhanced current hyperedge features into the weight asynchronous optimization subunit, calculate the time and space weights according to the time delay and spatial similarity respectively, realize the collaborative calculation of spatiotemporal dynamic weights, and perform adaptive weighted fusion of historical hyperedge features; Step S124: Input the historical hyperedge features after adaptive weighted fusion into the hypergraph information aggregation subunit, perform multiple rounds of message passing and node aggregation through graph neural networks or attention mechanisms, and output the hidden feature vector containing asynchronous spatiotemporal dependency information at the current grid scale.

4. The El Niño-Southern Oscillation prediction method as described in claim 1, characterized in that, The frequency domain multi-scale time module includes a frequency domain coding unit and a frequency domain decoding unit. Step S3 includes: Step S31: The frequency domain coding unit performs discrete Fourier transform on the ocean Niño index for historical time periods, maps the time series of the ocean Niño index from the time domain to the frequency domain, extracts frequency segment features that reflect changes at different time scales, the frequency segment features characterize the time features of different frequency segments, and encodes the time features of multiple frequency segments. By constructing a frequency attention map to highlight the influence of key frequency segments, the corresponding encoded features are output. Step S32: The frequency domain decoding unit constructs a prediction structure of the future frequency distribution based on the coding features output by the frequency domain coding unit. The prediction structure of the future frequency distribution represents the ocean Niño index of the target sea area in the target future time period.

5. The El Niño-Southern Oscillation prediction method according to claim 4, characterized in that, The frequency domain coding unit constructs a self-attention network in the frequency dimension, which captures long-distance dependencies and interactions between different frequencies through a multi-head attention mechanism, and then outputs corresponding coding features. The frequency domain coding unit includes multiple stacked frequency domain Transformer layers, each containing a frequency attention sublayer and a feedforward network sublayer, and uses residual connections and normalization operations to make the network more stably focus on key frequency bands that are strongly correlated with El Niño-Southern Oscillation.

6. The El Niño-Southern Oscillation prediction method according to claim 4, characterized in that, The frequency domain decoding unit adopts a frequency time-series modeling network based on the Transformer decoder structure: the decoder structure is a Transformer decoder, which fuses the encoded features output by the frequency domain coding unit and the corresponding real ocean Niño index during the training phase, and performs sequence prediction in the frequency domain; each layer of the decoding module contains a frequency attention mechanism, which gradually predicts the evolution characteristics of each frequency band in the future and outputs the predicted spectrum results; then, through inverse Fourier transform, the spectrum signal is restored to a time series to obtain the final prediction result of the ocean Niño index.

7. The El Niño-Southern Oscillation prediction method as described in claim 1, characterized in that, The spatiotemporal multi-scale hypergraph module and the frequency domain multi-scale time module are trained in the following manner: A training sample set is obtained based on the meteorological feature sequence of each meteorological feature among various global sea surface meteorological features and the oceanic Niño index; For each training sample, perform the following processing: Input the meteorological feature sequences of historical time periods in the current training samples into the multi-scale hypergraph module to obtain the estimated intermediate feature sequences; Based on the estimated intermediate feature sequence, the estimated oceanic Niño index sequence for historical time periods is obtained; Input the historical time period predicted ocean Niño index sequence and the real ocean Niño index sequence between the current time and the starting point of the target future time period into the frequency domain multi-scale time module to obtain the predicted ocean Niño index for the target future time period. Based on the predicted ocean Niño index for the target future time period and the actual ocean Niño index, the parameters of the spatiotemporal multi-scale hypergraph module and the frequency domain multi-scale time module are adjusted.

8. An El Niño-Southern Oscillation (ELO) prediction device, used to execute the El Niño-Southern Oscillation (ELO) prediction method according to any one of claims 1-7, characterized in that, include: The extraction unit is configured to input the meteorological feature sequence of the target sea area in historical time period into the multi-scale hypergraph module to obtain the intermediate feature sequence. The intermediate feature sequence integrates the asynchronous dependency information between nodes in the multi-scale hypergraph structure across time and space. The hypergraph structure at each scale is obtained based on the feature sequence of the meteorological feature sequence at the corresponding grid scale. The acquisition unit is configured to obtain the oceanic Niño index for the historical time period based on the intermediate feature sequence; The prediction unit is configured to input the oceanic Niño index of the historical time period and the prior time series data between the current time and the starting point of the target future time period into the frequency domain multi-scale time module to obtain the oceanic Niño index of the target sea area in the target future time period.

9. A computer-readable storage medium for storing instructions, characterized in that, When the instruction is executed by at least one computing device, it causes the at least one computing device to perform the El Niño-Southern Oscillation prediction method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: At least one processor and a memory storing program instructions; when the program instructions are read and executed by the processor, the electronic device causes the electronic device to perform the El Niño-Southern Oscillation prediction method as described in any one of claims 1 to 7.

11. A computer program product comprising computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the El Niño-Southern Oscillation prediction method as described in any one of claims 1 to 7.

Images