A space weather disturbance prediction method based on solar eruption event characterization learning

Abstract

The embodiment of the specification discloses a subsequent space weather disturbance prediction method based on solar eruption event representation learning. By taking the observation parameters of solar flare events and coronal mass ejection events, a solar flare feature vector and a coronal mass ejection feature vector are constructed according to the observation parameters, respectively, and an event sample of the flare event and the coronal mass ejection event with a clear corresponding relationship is constructed; the solar flare feature vector is encoded to obtain a corresponding first event representation, and the coronal mass ejection feature vector is encoded to obtain a corresponding second event representation; the first event representation and the second event representation are conditioned and interacted to obtain a fusion representation; the solar flare feature vector and the coronal mass ejection feature vector are spliced to obtain an input feature vector before encoding, and attention factor decomposition machine interaction modeling is performed on the input feature vector before encoding to obtain a feature interaction representation; the fusion representation and the feature interaction representation are jointly enhanced to obtain a solar eruption event embedding representation; and supervised learning is performed according to the solar eruption event embedding representation and the supervised label to generate a target model.

Description

Technical Field

[0001] This specification relates to the field of space weather forecasting, and in particular to a method for predicting subsequent space weather disturbances based on learning the characterization of solar eruption events. Background Technology

[0002] Solar flares and coronal mass ejections (CMEs) are important solar eruption phenomena driving changes in the Sun-Earth space environment. Following intense solar eruptions, subsequent space weather disturbances such as solar proton events and geomagnetic storms often occur, impacting spacecraft operations, satellite communications, navigation and positioning, power systems, and ground-based infrastructure. Therefore, determining whether a single solar eruption will be followed by space weather disturbances is crucial for space weather early warning and risk prevention.

[0003] In existing technologies, prediction studies for solar proton events and geomagnetic storms are mostly conducted separately. Solar proton event predictions typically rely on statistical, empirical, or machine learning models based on characteristics such as flare intensity, active region location, and coronal mass ejection velocity and angular width. Geomagnetic storm predictions, on the other hand, usually focus on assessing their Earth-based validity based on initial observation parameters of coronal mass ejections or near-Earth solar wind conditions. These approaches primarily address single subsequent disturbances and lack a unified technical solution that focuses on analyzing multiple types of subsequent space weather disturbances centered around a single solar eruption.

[0004] Furthermore, most existing methods directly utilize input features for classification, failing to adequately characterize the similarity relationships between different solar eruption events or to form discriminative event representations within the embedding space. For space weather missions with limited sample sizes and uneven distribution of positive and negative samples, relying solely on traditional classification targets for modeling can easily lead to insufficient representational capabilities, thereby affecting prediction performance and model stability. On the other hand, solar flares and coronal mass ejections (CMEs) are physically closely coupled, and modeling either type of feature alone is insufficient to fully reflect the combined impact of solar eruption activity. Simultaneously, second-order interactions between original physical parameters, such as the combination of flare cumulative flux and CME velocity, and source region location and angular width, often significantly influence whether subsequent perturbations occur. Existing technologies still fall short in utilizing these implicit correlations and explicit interactions.

[0005] Therefore, how to construct a technical solution that can combine solar flare events and coronal mass ejection events to more accurately predict the probability of subsequent weather disturbances has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] This specification provides an embodiment of a method for predicting subsequent space weather disturbances based on solar eruption event characterization learning, in order to solve the following technical problem: the need for a more accurate technical solution for predicting the probability of subsequent weather disturbances caused by solar eruption events.

[0007] To solve the above-mentioned technical problems, one or more embodiments of this specification are implemented as follows:

[0008] In a first aspect, embodiments of this specification provide a method for predicting subsequent space weather disturbances based on solar eruption event characterization learning, comprising: acquiring observational parameters of solar flare events and coronal mass ejection events, and constructing event samples of flare events and coronal mass ejection events with a clear correspondence, wherein the supervision label of the event sample is generated according to the subsequent disturbance corresponding to the event sample; constructing a solar flare feature vector based on the observational parameters of the solar flare events, and constructing a coronal mass ejection feature vector based on the observational parameters of the coronal mass ejection events; encoding the solar flare feature vector to obtain the corresponding first event representation, and, for the... The coronal mass ejection feature vector is encoded to obtain the corresponding second event representation; the first event representation and the second event representation are conditionally interacted to obtain a fused representation; the solar flare feature vector and the coronal mass ejection feature vector are concatenated to obtain the input feature vector before encoding, and attention factor decomposition machine interaction modeling is performed on the input feature vector before encoding to obtain the feature interaction representation; the fused representation and the feature interaction representation are jointly enhanced to obtain the solar eruption event embedding representation; supervised learning is performed based on the solar eruption event embedding representation and the supervision label to generate a target model, which is used to predict space weather disturbances.

[0009] In a second aspect, one or more embodiments of this specification provide an electronic device, comprising:

[0010] At least one processor; and,

[0011] A memory communicatively connected to the at least one processor; wherein,

[0012] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method as described in the first aspect.

[0013] In a third aspect, embodiments of this specification provide a non-volatile computer storage medium storing computer-executable instructions, which, when read by a computer from the storage medium, cause one or more processors to perform the method described in the first aspect.

[0014] The above-described at least one technical solution adopted in one or more embodiments of this specification can achieve the following beneficial effects: by acquiring observation parameters of solar flare events and coronal mass ejection events, and constructing event samples of flare events and coronal mass ejection events with a clear correspondence, wherein the supervision label of the event sample is generated according to the subsequent perturbation corresponding to the event sample; constructing a solar flare feature vector based on the observation parameters of the solar flare events, and constructing a coronal mass ejection feature vector based on the observation parameters of the coronal mass ejection events; encoding the solar flare feature vector to obtain the corresponding first event representation, and, for the coronal mass ejection event... The ejection feature vector is encoded to obtain the corresponding second event representation; the first and second event representations are conditionally interacted to obtain a fused representation; the solar flare feature vector and the coronal mass ejection feature vector are concatenated to obtain the input feature vector before encoding, and attention factor decomposition machine interaction modeling is performed on the input feature vector before encoding to obtain the feature interaction representation; the fused representation and the feature interaction representation are jointly enhanced to obtain the solar eruption event embedding representation; supervised learning is performed based on the solar eruption event embedding representation and the supervised label to generate a target model, which is used to predict space weather disturbances. This allows for the construction of event samples using a single solar eruption activity as the basic unit, and the establishment of a unified mapping relationship between it and subsequent solar proton events and geomagnetic storms. This fully explores the coupling relationship between the two types of solar eruption information, improves the discriminative ability of event representations, and trains a more accurate disturbance prediction model, resulting in more accurate predictions of weather disturbances after solar eruption events. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the overall process of a space weather disturbance prediction method based on solar eruption event characterization learning, provided in an embodiment of the present invention.

[0017] Figure 2 This invention provides a schematic diagram of a process for constructing event samples and generating subsequent space weather disturbance labels using a single solar eruption as the basic unit.

[0018] Figure 3 This is a schematic diagram of a bidirectional gating fusion module structure provided in an embodiment of the present invention;

[0019] Figure 4 This is a schematic diagram of the structure of an attention factor decomposition interaction modeling module provided in an embodiment of the present invention;

[0020] Figure 5 A schematic diagram of a supervised contrastive learning and downstream prediction output process provided in an embodiment of the present invention;

[0021] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this specification. Detailed Implementation

[0022] This specification provides an embodiment of a scheme for predicting subsequent space weather disturbances based on solar eruption event characterization learning. The method takes a single solar eruption as the basic object, using solar flare feature vectors constructed from observation parameters of solar flare events and coronal mass ejection (CME) feature vectors constructed from observation parameters of coronal mass ejection (CME) events as joint inputs. It also generates supervisory labels based on whether the eruption is followed by a solar proton event and / or geomagnetic storm, thereby establishing a unified framework for event characterization learning and disturbance prediction.

[0023] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0024] like Figure 1 As shown, Figure 1 This is a flowchart illustrating a method for predicting subsequent space weather disturbances based on solar eruption event characterization learning, provided for one or more embodiments of this specification.

[0025] Figure 1 The process may include the following steps:

[0026] S102: Obtain observation parameters of solar flare events and coronal mass ejection events, and construct event samples of flare events and coronal mass ejection events with clear correspondence, wherein the supervision label of the event sample is generated according to the subsequent perturbation corresponding to the event sample.

[0027] Both solar flare characteristics and coronal mass ejection characteristics are derived from observational parameters. "Parameters" are individual physical observations obtained from the event catalog, while "eigenvectors" are model inputs formed by arranging one or more observational parameters in a predetermined order and then normalizing them. Each observational parameter corresponds to a feature component in the eigenvector.

[0028] In this embodiment, solar flare event data can be obtained from flare catalogs published by authoritative institutions such as SWPC, and coronal mass ejection (CME) event data can be obtained from the event list observed by LASCO. For each sample, its corresponding flare parameters and coronal mass ejection parameters are extracted. The flare parameters may include peak flux, cumulative flux, duration, rise time, and latitude and longitude of the active region; the coronal mass ejection parameters may include angular width, propagation velocity, center position angle, and MPA. The observed parameters of the solar flare events are normalized to obtain solar flare feature vectors; the observed parameters of the coronal mass ejection events are normalized to obtain coronal mass ejection feature vectors. Each observed parameter corresponds to a feature component in the corresponding feature vector.

[0029] Preferably, solar eruptive activity from 1997 to 2023 can be sampled, with each flare-coronary mass ejection event with a clear correlation grouped together as a basic sample unit. It should be understood that the sample time range is not limited to the above-mentioned interval, and those skilled in the art can adjust it according to the data acquisition situation.

[0030] In this approach, event samples are constructed using a single solar eruption as the basic unit, and a unified mapping relationship is established between them and subsequent solar proton events and geomagnetic storms. This allows for the analysis of multiple types of subsequent space weather disturbances from a single solar eruption within the same technical framework.

[0031] like Figure 2 As shown, in this embodiment, each solar flare-coronary mass ejection (CME) event sample can be subsequently perturbed and labeled based on data such as institutional statistics and weekly space environment reports. If the event sample is subsequently accompanied by a solar proton event, it is recorded as a positive category for the solar proton event prediction task; otherwise, it is recorded as a negative category. If the event sample is subsequently accompanied by a geomagnetic storm, it is recorded as a positive category for the geomagnetic storm prediction task; otherwise, it is recorded as a negative category. Furthermore, the consistency between event samples and monitoring labels can be improved through manual verification and cross-validation of multi-source data.

[0032] S104, construct a solar flare feature vector based on the observation parameters of the solar flare event, and construct a coronal mass ejection feature vector based on the observation parameters of the coronal mass ejection event.

[0033] S106, the solar flare feature vector is encoded to obtain the corresponding first event representation, and the coronal mass ejection feature vector is encoded to obtain the corresponding second event representation.

[0034] In this embodiment, considering the differences in physical semantics between solar flare feature vectors and coronal mass ejection feature vectors, two independent multilayer perceptron encoders are used for encoding. Specifically, the flare feature vector is input into the first multilayer perceptron encoder to obtain a first event representation; the coronal mass ejection feature vector is input into the second multilayer perceptron encoder to obtain a second event representation. The two encoders can be composed of multilayer linear transforms and nonlinear activation functions.

[0035] Preferably, in order to reduce the impact of differences in the scales of different input features and stabilize the training process, normalization processing, such as layer normalization processing, can be applied after the output of the first event representation and the second event representation.

[0036] S108, conditional interaction is performed on the first event representation and the second event representation to obtain a fused representation.

[0037] In this embodiment, solar flares and coronal mass ejections are not independent but closely related physically. To fully model the coupling relationship between them, a bidirectional gating fusion mechanism can be adopted. Specifically, a gating vector acting on a first event representation is generated based on a second event representation, and the first event representation is modulated element-wise using the gating vector; simultaneously, a gating vector acting on a second event representation is generated based on the first event representation, and the second event representation is modulated element-wise using the gating vector.

[0038] Subsequently, the modulated first event representation and the modulated second event representation are concatenated and mapped to a unified feature space through a mapping layer to obtain a fused representation. Compared with direct concatenation or direct addition, the above bidirectional gating method can adaptively filter the effective dimensions in the current event representation based on information from another type of event, thereby achieving conditional bidirectional feature interaction.

[0039] like Figure 3 As shown, the bidirectional gating fusion module may include a first gating generation unit, a second gating generation unit, an element-wise modulation unit, and a fusion mapping unit.

[0040] By encoding the feature vectors of solar flares and coronal mass ejections separately, and then using a two-way gated fusion mechanism for conditional interaction, it is beneficial to fully explore the coupling relationship between the two types of solar eruption information and improve the discriminative ability of event representation.

[0041] S110, the solar flare feature vector and the coronal mass ejection feature vector are concatenated to obtain the input feature vector before encoding, and attention factor decomposition machine interactive modeling is performed on the input feature vector before encoding to obtain the feature interactive representation.

[0042] In this embodiment, in addition to the implicit coded spatial interactions described above, the explicit second-order interactions between physical parameters are also of great significance for subsequent perturbation prediction. For example, the combined relationships between solar flare cumulative flux and coronal mass ejection velocity, and between the location of active regions and the width of coronal mass ejection angles, often have a strong discriminative effect on whether solar proton events or geomagnetic storms will occur.

[0043] To address this, an attention factorization machine can be employed. Specifically, for each feature component in the input feature vector before encoding, a corresponding learnable embedding vector is set, and the feature component is multiplied by the corresponding learnable embedding vector to obtain the feature embedding representation of that feature component; element-wise multiplication is performed on the feature embedding representations of any two different feature components to obtain the corresponding second-order interaction vector; importance weights of each second-order interaction vector are calculated using an attention network; and the second-order interaction vectors are weighted and aggregated according to the importance weights to obtain the feature interaction representation.

[0044] Compared to standard factorization machines, attention factorization machines can adaptively identify the most critical feature interactions for the prediction task, thereby improving the model's expressiveness and interpretability. Figure 4 The schematic structure of input feature embedding, second-order interaction construction, attention weighting, and interaction representation output is shown.

[0045] By using attention factor decomposition machines to interactively model and explicitly characterize the important second-order interactions between physical parameters, it is beneficial to simultaneously improve the expressiveness and interpretability of the model.

[0046] The pre-encoding input feature vector is a concatenated vector formed by the solar flare feature vector and the coronal mass ejection (CME) feature vector. Each individual solar flare parameter or CME parameter, after normalization, corresponds to a feature component in the pre-encoding input feature vector. The attention factor decomposition machine constructs a second-order interaction relationship for any two different feature components in the pre-encoding input feature vector.

[0047] Before entering the solar flare encoder and coronal mass ejection encoder, the pre-encoding input feature vector is formed by concatenating the solar flare feature vector and the coronal mass ejection feature vector. Each individual parameter in the flare parameter or coronal mass ejection parameter corresponds to a feature component in the pre-encoding input feature vector after normalization. The attention factor decomposition machine constructs a second-order interaction vector for any two different feature components in this input feature vector, and performs weighted aggregation of each second-order interaction vector through an attention mechanism.

[0048] S112, the fused representation and the feature interaction representation are jointly enhanced to obtain the solar eruption event embedding representation.

[0049] In this embodiment, the feature interaction representation is first projected onto a feature dimension consistent with the fused representation; then, the projected feature interaction representation is superimposed on the fused representation to obtain the solar eruption event embedding representation. This approach preserves the high-level semantic information obtained from gated fusion while injecting explicit interaction priors from the original physical parameter level, thereby simultaneously enhancing the discriminativeness and physical interpretability of the event representation.

[0050] S114, Supervised learning is performed based on the solar eruption event embedding representation and the supervision label to generate a target model, which is used to predict space weather disturbances.

[0051] In this embodiment, supervised contrastive learning is used to constrain the relative positional relationships between samples in the embedding space. For samples with the same supervision label, positive sample pairs are constructed, placing them close to each other in the embedding space; for samples with different supervision labels, negative sample pairs are constructed, placing them far apart from each other in the embedding space. Through optimization of supervised contrastive loss, a more discriminative solar eruption event embedding representation can be learned.

[0052] Preferably, in the case of imbalanced sample classes, a small-batch sampling strategy with a fixed number of positive samples can be adopted to ensure that each batch contains a sufficient number of similar sample pairs, thereby enhancing the training signal for supervised contrastive learning.

[0053] After obtaining the trained embedding representation of solar eruption events, the target model can perform at least two types of downstream tasks.

[0054] The first type of downstream task is K-nearest neighbor prediction based on similar historical case retrieval. Specifically, a sample library is constructed using the event embeddings of the training samples; for the event to be predicted, the cosine similarity between its event embedding and the embeddings of each historical event in the sample library is calculated; the K nearest neighbor samples with the highest similarity are selected; and a weighted voting strategy with temperature scaling is used to calculate the probability that the event to be predicted belongs to the positive class. In this way, not only can the probability of the occurrence of solar proton events or geomagnetic storms be output, but the historical samples most similar to the current event can also be returned, providing case references for business personnel.

[0055] The second type of downstream task is direct probability prediction based on a classifier. Specifically, a classifier can be fed into the embedded representation of the solar eruption event, and a method of joint optimization using supervised contrastive loss and cross-entropy loss can be adopted to simultaneously learn the discriminative event representation and the classification decision boundary. The classifier can output the probability of solar proton events and the probability of geomagnetic storms, respectively.

[0056] like Figure 5 As shown, the supervised contrastive learning and downstream prediction output process can include supervised label-driven contrastive training, K-nearest neighbor retrieval voting, and classifier probability output.

[0057] By utilizing supervised contrastive learning to construct a discriminative event embedding space, we can support both the retrieval of similar historical events and the prediction of the probability of subsequent disturbances, thereby enhancing the practical value of operational space weather forecasts.

[0058] Preferably, during the model training phase, a small batch sampling method with a fixed number of positive samples can be used, and parameters such as hidden layer dimension, projection dimension, attention dimension, temperature parameter, learning rate, and loss weight can be set according to the task type.

[0059] Optionally, the performance of the method can be evaluated using evaluation metrics such as accuracy, recall, precision, false alarm rate, F1 score, and AUC score.

[0060] By acquiring observational parameters of solar flare events and coronal mass ejection (CME) events, and constructing event samples with clear correspondences between them, wherein the supervision labels of the event samples are generated based on the subsequent perturbations corresponding to the event samples; solar flare feature vectors are constructed based on the observational parameters of the solar flare events, and CME feature vectors are constructed based on the observational parameters of the CME events; the solar flare feature vectors are encoded to obtain the corresponding first event representation, and the CME feature vectors are encoded to obtain the corresponding second event representation; the first event representation and the second event representation are conditionally interacted to obtain a fused representation; the solar flare feature vector and the CME feature vector are concatenated to obtain the input feature vector before encoding, and attention factor decomposition machine interactive modeling is performed on the input feature vector before encoding to obtain the feature interaction representation; the fused representation and the feature interaction representation are jointly enhanced to obtain the solar eruption event embedding representation; supervised learning is performed based on the solar eruption event embedding representation and the supervision labels to generate a target model, which is used to predict space weather perturbations. This allows for the construction of event samples using single solar eruption events as the basic unit, and the establishment of a unified mapping relationship between these events and subsequent solar proton events and geomagnetic storms. By fully exploring the coupling relationship between the two types of solar eruption information, the ability to distinguish event representations is improved, thereby training a more accurate disturbance prediction model that can more accurately predict weather disturbances after solar eruption events.

[0061] In the second aspect, such as Figure 6 As shown, Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this specification. The device includes:

[0062] At least one processor; and,

[0063] A memory communicatively connected to the at least one processor; wherein,

[0064] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method as described in the first aspect.

[0065] In a third aspect, based on the same idea, embodiments of this specification also provide a non-volatile computer storage medium corresponding to the above method, storing computer-executable instructions, which, when read by a computer from the storage medium, cause one or more processors to execute the method described in the first aspect.

[0066] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0067] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0068] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0069] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A method for predicting subsequent space weather disturbances based on solar eruption event representation learning, comprising: Observational parameters of solar flare events and coronal mass ejection events are obtained, and event samples of flare events and coronal mass ejection events with clear correspondence are constructed. The supervision labels of the event samples are generated based on the subsequent perturbations corresponding to the event samples. A solar flare feature vector is constructed based on the observation parameters of the solar flare event, and a coronal mass ejection feature vector is constructed based on the observation parameters of the coronal mass ejection event. The solar flare feature vector is encoded to obtain the corresponding first event representation, and the coronal mass ejection feature vector is encoded to obtain the corresponding second event representation; By conditionally interacting with the first event representation and the second event representation, a fused representation is obtained. The solar flare feature vector and the coronal mass ejection feature vector are concatenated to obtain the input feature vector before encoding. Then, attention factor decomposition machine interactive modeling is performed on the input feature vector before encoding to obtain the feature interactive representation. The fused representation and the feature interaction representation are jointly enhanced to obtain the solar eruption event embedding representation; Supervised learning is performed based on the solar eruption event embedding representation and the supervision label to generate a target model, which is used to predict space weather disturbances.

2. The method as described in claim 1, wherein, The solar flare feature vector includes feature components corresponding to one or more of the observed parameters, namely, peak flux of the flare, cumulative flux of the flare, duration of the flare, rise time of the flare, longitude of the active region and latitude of the active region; the coronal mass ejection feature vector includes feature components corresponding to one or more of the observed parameters, namely, angular width, propagation velocity, center position angle and MPA.

3. The method as described in claim 1, wherein, The supervision label for the event sample is generated based on the subsequent perturbation corresponding to the event sample, including: Based on the event catalog, the space environment weekly report, and the list of effective coronal mass ejection events on Earth, determine whether the event samples are subsequently accompanied by solar proton events or geomagnetic storms; If the event sample is subsequently followed by a solar proton event, the solar proton event prediction task is recorded as a positive class; otherwise, it is recorded as a negative class. If the event sample is subsequently accompanied by a geomagnetic storm, it is recorded as a positive class for the geomagnetic storm prediction task; otherwise, it is recorded as a negative class.

4. The method of claim 3, wherein, Supervised learning is performed based on the solar eruption event embedding representation and the supervised label, including: Construct positive sample pairs using supervised labeled samples of the same class, and construct negative sample pairs using supervised labeled samples of different classes; Supervised contrastive loss constrains similar samples to be close to each other in the embedding space and dissimilar samples to be far apart.

5. The method of claim 1, wherein, The method further includes: predicting the forecast results of subsequent space weather disturbances based on the target model, specifically including: Obtain samples of solar eruption events to be predicted, generate corresponding embedding representations of these events, and use the target model to make predictions on these embedding representations as follows: The probability of the target event occurring is output based on K-nearest neighbor retrieval; The probability of a solar proton event is directly output based on a classifier. The probability of geomagnetic storms is directly output based on the classifier. Return a sample of historical solar eruption events similar to the target event.

6. The method of claim 5, wherein, The probability of occurrence of the target event is output based on K-nearest neighbor retrieval, including: Calculate the cosine similarity between the embedding representation of the solar eruption event to be predicted and the known positive class sample embedding representation; Select the K nearest neighbor samples with the highest similarity; A weighted voting strategy with temperature scaling is used to calculate the probability that the target event belongs to the positive class.

7. The method of claim 1, wherein, Conditional interaction is performed on the first event representation and the second event representation to obtain a fused representation, including: A gating vector is generated based on the second event representation and applied to the first event representation, and the first event representation is modulated element-wise using the gating vector; Simultaneously, a gating vector is generated based on the first event representation and applied to the second event representation, and the second event representation is modulated element-wise using the gating vector; The modulated first event representation and the modulated second event representation are concatenated and mapped to a unified feature space through a mapping layer to obtain a fused representation.

8. The method of claim 1, wherein, The input feature vector before encoding is subjected to attention factorization machine interaction modeling to obtain feature interaction representation, including: The solar flare feature vector and the coronal mass ejection feature vector are concatenated to obtain the input feature vector before encoding; For each feature component in the input feature vector before encoding, the feature component is multiplied by the corresponding learnable embedding vector to obtain the corresponding feature embedding representation; Element-wise multiplication of the feature embedding representations of any two different feature components yields the corresponding second-order interaction vector. The importance weights of each second-order interaction vector are calculated using an attention network. The second-order interaction vectors are weighted and aggregated according to the importance weights to obtain the feature interaction representation.

9. The method of claim 1, wherein, The fused representation and the feature interaction representation are jointly enhanced to obtain a solar eruption event embedding representation, including: Project the feature interaction representation onto a feature dimension consistent with the fused representation; The projected feature interaction representation is superimposed with the fused representation to obtain the solar eruption event embedding representation.

10. An electronic device, comprising: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, such that the at least one processor can perform the method as described in any one of claims 1 to 9.

Images