Solar radio burst prediction method and system based on full-disk solar magnetogram

Abstract

This invention discloses a method and system for predicting solar radio bursts based on a complete solar magnetograph, belonging to the field of solar radio technology. It includes: acquiring a complete solar magnetograph; inputting the complete solar magnetograph into a pre-trained improved ResNet model to obtain image feature vectors; dividing the image feature vectors into pseudo-time series and inputting them into an improved LSTM network to obtain solar radio burst prediction results; the improved ResNet model introduces a custom residual layer to extract small-scale features from the complete solar magnetograph, and the improved LSTM network introduces a gating mechanism to process the image feature vectors at different time scales. This method can improve the accuracy of solar radio burst prediction without utilizing solar radio spectrum data; it solves the problem that existing predictions using solar radio spectra are easily interfered with by the accuracy of the input data observations.

Description

Technical Field

[0001] This invention relates to the field of solar radio technology, and in particular to a method and system for predicting solar radio bursts based on a complete solar magnetogram. Background Technology

[0002] The statements in this section merely refer to the background art related to this invention and do not necessarily constitute prior art.

[0003] Solar radio bursts can be seen as a marker for predicting space weather disasters. Solar radio data analysis is an important tool for predicting solar radio bursts based on the whole-disk solar magnetograph. Existing research has applied deep learning and machine learning techniques to analyze solar radio data. For example, a solar radio database was developed using deep belief networks to extract features from solar radio spectra; a recurrent neural network specifically designed for time series analysis was used to classify solar radio spectrum data; a pre-trained visual geometric swarm network (VGG-16) model was used to classify solar radio spectra, improving the true positive rate (TPR) by 12.2% compared to existing methods; a univariate time series prediction deep residual neural architecture (N-BEATS) for predicting solar radio flux was established, achieving better performance than multi-flux neural network methods in prediction; and a radio burst detection system was designed using AlexNet, designed to adapt to different solar spectra to provide high-accuracy predictions of solar radio bursts based on the whole-disk solar magnetograph.

[0004] In summary, previous studies have mostly focused on solar radio spectrum. Undoubtedly, solar radio bursts have very obvious characteristic information on solar radio spectrum. However, the observation of solar radio spectrum has extremely high requirements for instruments and is therefore easily affected by sudden events.

[0005] A full-disk solar magnetograph is a "snapshot" of the Sun at a specific moment, preserving information about solar activity at that time, such as the relative positions of active regions. However, solar activity is continuous, meaning a full-disk solar magnetograph can be viewed as sequential data. But this characteristic has been relatively rarely utilized in previous solar radio studies, and the features present in sequential data cannot be well utilized. Existing algorithms applied to the processing of full-disk solar magnetographs still have the following problems:

[0006] (1) The whole disk of the solar magnetogram contains features at different scales, but existing deep learning models (such as VGG-16 and AlexNet) are not good at capturing small-scale features, resulting in limited ability to process fine structures.

[0007] (2) The whole solar magnetogram has a complex spatial structure and a large amount of data. Existing models may not have sufficient generalization ability when processing magnetograms at different time points and in different regions, resulting in a decrease in prediction accuracy on new data.

[0008] (3) Although some models (such as recurrent neural networks) are used to process time series data, they do not fully consider the combination of spatial and temporal information and cannot fully capture the dynamic changes of solar activity. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a method, system, electronic device, computer-readable storage medium, and computer program product for predicting solar radio bursts based on a complete solar magnetograph. It uses a complete solar magnetograph as research data, and analyzes the complete solar magnetograph using an improved ResNet network and an improved LSTM network to extract features hidden in the sequence for predicting solar radio bursts.

[0010] In a first aspect, the present invention provides a method for predicting solar radio bursts based on a complete solar magnetogram;

[0011] A method for predicting solar radio bursts based on a complete solar magnetogram includes:

[0012] A complete solar magnetograph is obtained, and the complete solar magnetograph is input into a pre-trained improved ResNet model to obtain image feature vectors;

[0013] The image feature vector is divided into a pseudo-time series and input into an improved LSTM network to obtain the solar radio burst prediction results.

[0014] The improved ResNet model introduces a custom residual layer to extract small-scale features from the whole-disk solar magnetogram, and the improved LSTM network introduces a gating mechanism to process the image feature vectors at different time scales.

[0015] In some implementations, the custom residual layer includes multiple custom residual modules connected in sequence, wherein the custom residual modules introduce a multilayer perceptron to perform a nonlinear transformation on the features of each channel.

[0016] In some implementations, the custom residual module is subsequently connected to a first convolutional unit, a second convolutional unit, a global average pooling layer, a multilayer perceptron, and a ReLU activation function.

[0017] In some implementations, the improved ResNet model includes sequentially connected convolutional layers, batch normalization layers, ReLU activation functions, max pooling layers, multiple custom residual layers, global pooling layers, and fully connected layers.

[0018] In some implementations, the improved LSTM network introduces a dynamic attention mechanism to dynamically adjust the degree of attention the improved LSTM network pays to different time scales, and outputs solar radio burst prediction results based on the whole-disk solar magnetogram.

[0019] In some implementations, the gating mechanism is represented as:

[0020] G t =σ(ω) G [h t-1 ,x t ]+b G );

[0021] In the formula, σ is the activation function of the LSTM network, and h t-1 Given the hidden state of the previous time step t-1, x t For the input at the current time t, w G b is the cyclic weight for the additional gating state. G The bias for the additional gating state.

[0022] Secondly, the present invention provides a solar radio burst prediction system based on a complete solar magnetogram;

[0023] A solar radio burst prediction system based on a complete solar magnetogram includes:

[0024] The image feature extraction module is configured to: acquire a full-disk solar magnetograph, input the full-disk solar magnetograph into a pre-trained improved ResNet model, and acquire image feature vectors;

[0025] The solar radio burst prediction module is configured to: divide the image feature vector into a pseudo-time series and input it into an improved LSTM network to obtain the solar radio burst prediction result;

[0026] The improved ResNet model introduces a custom residual module to extract small-scale features from the full-disk solar magnetogram, and the improved LSTM network introduces a gating mechanism to process the image feature vectors at different time scales.

[0027] Thirdly, the present invention provides an electronic device;

[0028] An electronic device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described method for predicting solar radio bursts based on a full-disk solar magnetogram.

[0029] Fourthly, the present invention provides a computer-readable storage medium;

[0030] A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the above-described method for predicting solar radio bursts based on a full-disk solar magnetogram.

[0031] Fifthly, the present invention provides a computer program product;

[0032] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the above-described method for predicting solar radio bursts based on a full-disk solar magnetogram.

[0033] Compared with the prior art, the beneficial effects of the present invention are:

[0034] 1. The technical solution provided by this invention utilizes an improved ResNet to extract spatial features from a full-disk solar magnetogram. By introducing a multilayer perceptron (MLP) through a custom residual layer, the features of each channel are further processed, enhancing the expressive power of the model and enabling the network to better capture small-scale features, thus achieving full extraction of spatial information.

[0035] 2. The technical solution provided by this invention introduces an additional gating mechanism to improve the LSTM network, giving the model more degrees of freedom to control which information should be retained, updated or discarded. This can help the model better capture complex temporal dependencies and provides the model with additional sublinear transformation paths, enabling it to better analyze complex sequence patterns and effectively capture information from different time spans in the full-disk solar magnetogram at a certain moment.

[0036] 3. The technical solution provided by this invention introduces a dynamic attention mechanism, which enables the LSTM network to dynamically adjust the degree of attention to different time steps through context, so that the model can more focused on those time steps that have a greater impact on the final decision. Attached Figure Description

[0037] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0038] Figure 1 This is a flowchart provided for an embodiment of the present invention;

[0039] Figure 2 A schematic diagram of the network architecture of the improved ResNet model provided in an embodiment of the present invention;

[0040] Figure 3 This is a schematic diagram of the network architecture of the custom residual module provided in an embodiment of the present invention;

[0041] Figure 4 A schematic diagram of the image feature vector output by the improved ResNet model provided in an embodiment of the present invention;

[0042] Figure 5 A schematic diagram of the network architecture of the improved LSTM network provided in an embodiment of the present invention;

[0043] Figure 6 A flowchart illustrating the dynamic attention mechanism provided in an embodiment of the present invention;

[0044] Figure 7 This is an example image showing the visualization result after image feature vector segmentation provided in an embodiment of the present invention;

[0045] Figure 8 This is an example diagram of a complete solar magnetograph provided in an embodiment of the present invention;

[0046] Figure 9 An example diagram of the evaluation results provided in an embodiment of the present invention. Detailed Implementation

[0047] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0048] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0049] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0050] Example 1

[0051] Existing methods for predicting solar radio bursts rely on solar radio spectrum data. However, the acquisition of solar spectrum data depends on high-precision instruments and is highly susceptible to unforeseen events, resulting in poor accuracy of the basic data and affecting prediction performance. Therefore, this invention provides a method for predicting solar radio bursts based on a complete solar magnetogram. Considering the continuity of solar activity, the complete solar magnetogram is treated as sequential data, and spatial feature analysis and extraction are performed using an improved ResNet model and an improved LSTM network to predict solar radio bursts.

[0052] Next, combined Figures 1-9 This embodiment provides a detailed description of a solar radio burst prediction method based on a complete solar magnetic map. The method includes the following steps:

[0053] S1. Obtain the complete solar magnetic map.

[0054] A full-disk solar magnetograph visually displays the distribution of the magnetic field on the Sun's surface. The image is displayed in grayscale, with black and white areas representing opposite magnetic fields. As a single-frame image over time, compared to a spectrogram, a full-disk solar magnetograph clearly represents the state of solar activity at a specific moment and has higher resolution. This means the image contains rich features, providing a clear view of the magnetic field distribution on the Sun's surface. Compared to a spectrogram, a full-disk solar magnetograph can be likened to a "snapshot" of solar activity at a particular moment; each image contains all the characteristics of the currently observed solar activity, effectively representing the state of active solar regions.

[0055] In this embodiment, the full-disk solar magnetograms are obtained from the Michelson-Doppler Imager (MDI) on the Solar and Heliospheric Observatory (SOHO) satellite. The MDI performs full-disk observations of the light field every 96 minutes. These full-disk magnetograms are recorded by its 1024×1024 CCD detector with a field of view of 34.

[0056] S2. Input the full solar magnetogram into the pre-trained improved ResNet model to obtain image feature vectors.

[0057] A complete solar magnetogram is a complex image dataset containing a wealth of spatial information; it may include features at different scales, such as local magnetic field activity and overall solar activity.

[0058] Compared to shallow convolutional neural networks, the ResNet model has stronger feature extraction capabilities, effectively extracting useful spatial features from the entire solar magnetogram. However, as the number of network layers increases, deep networks often face the problems of vanishing or exploding gradients, making the model difficult to train. Furthermore, although the ResNet network can capture features of different scales through its multi-layer convolution and pooling operations, small-scale features may be gradually ignored during the convolution and pooling processes in deep network structures. Therefore, in this embodiment, the ResNet model is improved by introducing a custom residual layer to comprehensively extract small-scale features from the entire solar magnetogram.

[0059] As one implementation, the improved ResNet model includes a convolutional layer, a batch normalization layer, a ReLU activation function, a max pooling layer, four custom residual layers, a global pooling layer, and a fully connected layer connected in sequence. The custom residual layers include two custom residual modules connected in sequence.

[0060] Specifically, taking a complete solar magnetogram as input, the input undergoes a convolutional layer with a 7x7 kernel and a stride of 2. This is followed by batch normalization for stable training and the introduction of non-linearity using the ReLU activation function, resulting in an output feature map. This feature map is then processed by a max-pooling layer to reduce its size before being input into a custom residual layer. The first custom residual layer has 64 input channels, 64 output channels, and a stride of 1; the second has 64 input channels, 128 output channels, and a stride of 2; the third has 128 input channels, 256 output channels, and a stride of 2; and the fourth has 256 input channels, 512 output channels, and a stride of 2.

[0061] The output of the fourth defined residual layer is used as the input of the global pooling layer, outputting a 512-channel 1x1 feature map; the 1x1 feature map is expanded and passed through a fully connected layer to output a 512-dimensional image feature vector.

[0062] Furthermore, each custom residual layer contains two custom residual modules. Each custom residual module includes a first convolutional unit, a second convolutional unit, a global average pooling layer, a multilayer perceptron, and a ReLU activation function connected in sequence. The input of the first convolutional unit and the output of the multilayer perceptron are residually connected. The first and second convolutional units have the same structure, including a convolutional layer, a BN layer, and a ReLU activation function connected in sequence. The multilayer perceptron includes an input layer, a linear layer, a ReLU activation function, another linear layer, a ReLU activation function, a Reshape layer, and an output layer connected in sequence.

[0063] The specific implementation of the custom residual module is as follows: After processing the output feature image using two sets of convolutional units with 3x3 convolutional kernels, batch normalization (BN), and ReLU activation function, global average pooling is performed to compress the features of each channel into a single value, forming a two-dimensional tensor. The pooled features are then transformed nonlinearly through an MLP while keeping the output size unchanged. The output of the MLP is then adjusted to a size equal to that of the original feature map, and residual connections are performed to finally output the processed feature image.

[0064] In this embodiment, by introducing an MLP module, the features extracted by the residual module are nonlinearly transformed, and the input image process residuals are connected to ensure that small-scale features are not ignored during convolution training.

[0065] In this embodiment, the custom residual module receives inputs in_chinnels, out_chinnels, blocks, and stride, representing the input channels, output channels, number of modules, and stride, respectively. In this embodiment, the number of custom residual modules in the custom residual layer is fixed at 2.

[0066] S3. Divide the image feature vector into a pseudo-time series and input it into an improved LSTM network to obtain the solar radio burst prediction results based on the whole-disk solar magnetogram.

[0067] Among them, the improved LSTM network introduces a gating mechanism to process image feature vectors at different time scales, and introduces a dynamic attention mechanism to dynamically adjust the attention of the improved LSTM network to different time steps, outputting solar radio burst prediction results based on the whole-disk solar magnetogram.

[0068] As one implementation method, the specific process of S3 is as follows:

[0069] S301. Divide the input 512-dimensional image feature vector into a pseudo-time series.

[0070] Specifically, the 512-dimensional image feature vector is divided into four sub-vectors, each corresponding to a time step, and each group of sub-vectors is input into the LSTM as sequence data.

[0071] Since the improved LSTM network only accepts time series data as input, the image feature vector is processed into a continuous pseudo-time series. The time step is the unit of number of the processed pseudo-time series. In the current case, the value corresponding to the time step is 4, which means that the data input to the improved LSTM network is a set of pseudo-time series data containing four feature sub-vectors.

[0072] S302. Input the sub-vectors into the improved LSTM network for processing, and use the dynamic attention mechanism to perform weighted calculations on the processing results to output the solar radio burst prediction results.

[0073] In this embodiment, an improved LSTM network is used to process sub-vectors that are continuously distributed in the spatial dimension. This enables the full learning of the spatial information contained in the continuous spatial sequence data, thereby realizing the extraction of information from different time spans in the full-disk solar magnetogram.

[0074] LSTM (Long Short-Term Memory) is a special type of recurrent neural network (RNN) specifically designed for processing and predicting time series data. Traditional RNNs are prone to gradient vanishing or exploding problems when processing long sequences, making it difficult for the model to capture long-term dependencies. LSTM effectively solves these problems by introducing memory units and gate mechanisms (input gate, forget gate, and output gate). In an LSTM network, memory units store the network's state and can remember information over long periods. The input gate controls the extent to which information is added to the memory unit, the forget gate determines which information in the memory unit should be forgotten, and the output gate determines the extent to which information in the memory unit is output. Through the combination of these gates, LSTM can selectively retain and discard information, thus effectively learning and remembering dependencies over long time spans.

[0075] A complete solar magnetogram at a given moment contains information spanning different time periods, which existing LSTM models cannot effectively capture. Therefore, in this embodiment, an additional gating mechanism G is introduced, giving the model more degrees of freedom to control which information should be retained, updated, or discarded. This helps the model better capture complex temporal dependencies. The new gating mechanism essentially provides the model with an additional semilinear transformation path, enabling it to better analyze complex sequence patterns.

[0076] Furthermore, in the improved LSTM network, the forget gate f t Represented as:

[0077] f t =σ(ω) f [h t-1 ,x t ]+b f );

[0078] Input gate i t Represented as:

[0079] i t =σ(ω) i [h t-1 ,x t ]+bi );

[0080] Output gate o i Represented as:

[0081] o t =σ(ω) o [h t-1 ,x t ]+b o );

[0082] The gating mechanism calculation formula is expressed as follows:

[0083] G t =σ(ω) G [h t-1 ,x t ]+b G );

[0084] Based on this gating mechanism, the LSTM network state unit c at time t t Represented as:

[0085]

[0086] Candidate state information at time t Represented as:

[0087]

[0088] Candidate state information h at time t t Represented as:

[0089] h t =o t ·tanh(c t );

[0090] In the formula, x t For the current input, i t ω is the input gate value. f ω t ω o ω c These are the loop weights for each state, b f b t b o b c The biases h for each state are respectively. t-1 c t-1 Let be the hidden state and candidate state of the previous time step, and σ and tanh be the activation functions of the improved LSTM network, namely the sigmoid function and the hyperbolic tangent function, respectively.

[0091] f t Control the state unit c of the previous moment t-1Whether the information in the previous time step is transmitted to the current time step determines the proportion of information output from the previous time step that is retained. t Control the input x at the current moment t and hidden state h t-1 Thus, selectively acquiring the current input information, o t Then, based on the new input information and the output information from the previous time step, the output at the current time step can be calculated, c. t Used to store and transmit time-series information.

[0092] G t Essentially, this adds an extra layer of control between the forgetting and input mechanisms of the classic LSTM, serving as a supplementary adjustment to the memory units and allowing for better control over C. t To make additional additions or reductions, compared to traditional LSTM based on f t i t The method to update the memory state, G t An additional adjustment channel is provided, enhancing the model's ability to adapt to features at different time scales of the input data.

[0093] While improved LSTM networks can capture the temporal dependencies of long sequences, they focus on a uniform amount of information at each time step, failing to differentiate the importance of information at each step. Analysis of the entire solar magnetograph reveals that each image contains both active and inactive solar regions. Clearly, the features of active solar regions deserve greater attention due to their close connection to solar radio bursts. Therefore, this embodiment introduces a dynamic attention mechanism, allowing the model to dynamically adjust its focus on different time steps based on context, enabling it to concentrate more effectively on those time steps that have a greater impact on the final decision. The specific process is as follows:

[0094] (1) The dynamic attention mechanism receives the output of the improved LSTM as input data LSTM_out and passes in the attention head. The default number of attention heads is 4 (because the previous time step division number is 4).

[0095] (2) Each attention head independently calculates its attention weights and generates a feature context vector according to the following steps:

[0096] (201) Use linear transformation to process the input data LSTM_out and calculate the attention weights of the current attention head.

[0097] (202) Use the softmax function to normalize the attention weights to ensure that the sum of the attention weights is 1.

[0098] (203) Multiply the normalized attention weights with the elements of LSTM_out one by one, and sum them along the time step dimension to generate the feature vector of the current attention head.

[0099] (204) The outputs of all attention heads are concatenated into a single context feature vector, and the dimension is compressed through a linear layer to generate the final context feature vector.

[0100] (3) Feature output and classification:

[0101] Specifically, the generated context feature vector is passed to a fully connected layer, where it undergoes feature transformation through the first fully connected layer fc1. Then, it passes through a Dropuu layer to prevent overfitting by randomly deactivating some neurons to enhance the model's generalization ability. Next, it passes through a second fully connected layer fc2 to compress the feature dimension from 512 to 128, and then through a fully connected layer fc3 to compress the feature dimension from 128 to 2. The output serves as the prediction result for solar radio bursts.

[0102] As one implementation method, before executing S2 and S3, the improved LSTM network and the improved ResNet model need to be trained. The specific process is as follows:

[0103] (1) A solar radio burst log file provided on an official website was obtained, which systematically records solar radio burst events from January 1967 to January 2011. In the log file, the seventh column describes the type of solar radio burst event that occurred within a specified time interval. This log file systematically compiles observational data from various observatories.

[0104] Based on information recorded in the log files of solar radio burst observations, the data were divided into two categories: "outbursts" and "non-outbursts." For each set of records, if any number of solar radio bursts occurred in the corresponding daily data in the log file, it was classified as an "outburst"; otherwise, it was classified as a "non-outburst." In the obtained full-disk solar magnetogram dataset, there are 2100 records with solar bursts and 936 records without solar bursts.

[0105] In order to verify the accuracy of the method in predicting solar radio bursts, the common ten-fold cross-validation method was used to divide the dataset, that is, to divide the dataset into ten equal-sized subsets, with the proportion of data in each subset being the same as that in the original dataset.

[0106] In addition, considering the possibility that too many image features may cause small pixel features to be ignored, the input dataset was subjected to feature engineering and normalization processing, including random horizontal flipping, random vertical flipping, and random rotation, to obtain the processed image data.

[0107] (2) The improved LSTM network and the improved ResNet model are trained using the processed data as the training set, and the prediction performance of the model is evaluated using evaluation metrics.

[0108] For example, solar radio burst prediction is considered a binary classification task, where burst samples are classified as positive samples, and non-burst samples are classified as negative samples. The evaluation parameters are: TP (TruePositive), FP (False Positive), TN (True Negative), and FN (False Negative), which are defined in detail in Table 1.

[0109] Table 1 Definitions of TP, FP, TN, and FN

[0110]

[0111] Based on the evaluation parameters, four evaluation metrics were used to assess the performance of the prediction model: accuracy, precision, recall, and false positive rate (FAR). Accuracy, precision, recall, and false positive rate are all important metrics for evaluating predictions. Accuracy measures the proportion of correctly predicted samples, considering the accuracy of predictions for the true positive category. Recall assesses the ability to successfully capture positive samples, and false positive rate quantifies the proportion of incorrectly predicted positive samples in the negative category. Detailed definitions are given in Table 2.

[0112] Table 2 Definitions of Accuracy, Precision, Recall, and False Alarm Rate

[0113]

[0114] Table 3 shows the results of ten confusion matrices obtained based on the ten-fold cross-validation method. As can be seen from the table, the number of samples predicted as true positives and true negatives is relatively high, which indicates that the model has good performance in practical applications.

[0115] Table 3 shows the results of ten datasets and their evaluation parameters obtained based on the ten-fold cross-validation method.

[0116]

[0117]

[0118] Table 4 presents the calculation results of the evaluation indicators for these ten sets of data. As can be seen from the table, the four evaluation indicators fluctuate within a very small range. It is also worth noting that the accuracy rate has always remained above 0.87 and the precision has always remained above 0.89, which indicates that the model proposed in this embodiment has a very good predictive effect.

[0119] Table 4 shows the ten datasets obtained based on the ten-fold cross-validation method and their evaluation metrics.

[0120]

[0121] Combination Figure 9 This provides a more intuitive view of the fluctuation range of the evaluation indicators, showing that the fluctuation range of the results is relatively small.

[0122] Example 2

[0123] This embodiment discloses a solar radio burst prediction system based on a complete solar magnetogram, including:

[0124] The image feature extraction module is configured to: acquire a full-disk solar magnetograph, input the full-disk solar magnetograph into a pre-trained improved ResNet model, and acquire image feature vectors;

[0125] The solar radio burst prediction module is configured to: divide the image feature vector into a pseudo-time series and input it into an improved LSTM network to obtain solar radio burst prediction results based on the whole disk solar magnetogram;

[0126] The improved ResNet model introduces a custom residual module to extract small-scale features from the full-disk solar magnetogram, and the improved LSTM network introduces a gating mechanism to process the image feature vectors at different time scales.

[0127] It should be noted that the image feature extraction module and the solar radio burst prediction module mentioned above correspond to the steps in Embodiment 1. The examples and application scenarios implemented by these modules and their corresponding steps are the same, but they are not limited to the content disclosed in Embodiment 1. It should also be noted that these modules, as part of the system, can be executed in a computer system, such as a set of computer-executable instructions.

[0128] Example 3

[0129] Embodiment 3 of the present invention provides an electronic device, including a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, they complete the steps of the above-mentioned method for predicting solar radio bursts based on a full-disk solar magnetogram.

[0130] Example 4

[0131] Embodiment 4 of the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the above-described method for predicting solar radio bursts based on a full-disk solar magnetogram.

[0132] Example 5

[0133] Embodiment 5 of the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the above-described method for predicting solar radio bursts based on a full-disk solar magnetogram.

[0134] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0135] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0136] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0137] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0138] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting solar radio bursts based on a complete solar magnetogram, characterized in that, include: A complete solar magnetograph is obtained, and the complete solar magnetograph is input into a pre-trained improved ResNet model to obtain image feature vectors; The image feature vector is divided into a pseudo-time series and input into an improved LSTM network to obtain the solar radio burst prediction results. The improved ResNet model introduces a custom residual layer to extract small-scale features from the whole-disk solar magnetogram, and the improved LSTM network introduces a gating mechanism to process the image feature vectors at different time scales. The improved ResNet model consists of sequentially connected convolutional layers, batch normalization layers, ReLU activation functions, max pooling layers, four custom residual layers, a global pooling layer, and a fully connected layer. Each custom residual layer comprises two sequentially connected custom residual modules. Each custom residual module consists of sequentially connected first convolutional unit, second convolutional unit, global average pooling layer, multilayer perceptron, and ReLU activation function. The input of the first convolutional unit and the output of the multilayer perceptron are residually connected. The first and second convolutional units have the same structure, consisting of sequentially connected convolutional layers, batch normalization layers, and ReLU activation functions. The multilayer perceptron consists of sequentially connected input layers, linear layers, ReLU activation functions, reshape layers, and output layers. The specific implementation of the custom residual module is as follows: After processing the output feature image by using two sets of convolutional units with 3x3 convolutional kernels, batch normalization (BN), and ReLU activation function, global average pooling is performed to compress the features of each channel into a single value, forming a two-dimensional tensor. The pooled features are then transformed nonlinearly through MLP while keeping the output size unchanged. The output of MLP is then adjusted to a size equal to that of the original feature map, and residual connections are performed to finally output the processed feature image. An improved LSTM network is introduced by a gating mechanism to process image feature vectors at different time scales. A dynamic attention mechanism is also introduced to dynamically adjust the attention level of the improved LSTM network to different time steps, and output solar radio burst prediction results based on the whole-disk solar magnetogram. The input 512-dimensional image feature vector is divided into pseudo-time series, and the sub-vectors are input into an improved LSTM network for processing. The processing results are weighted and calculated through a dynamic attention mechanism to output the solar radio burst prediction results.

2. The solar radio burst prediction method based on a complete solar magnetogram as described in claim 1, characterized in that, The custom residual module introduces a multilayer perceptron to perform nonlinear transformations on the features of each channel.

3. The solar radio burst prediction method based on a complete solar magnetogram as described in claim 1, characterized in that, The gating mechanism is represented as follows: ； In the formula, For the activation function of the LSTM network, This is the hidden state at the previous time step t-1. The input is the current time t. w G For the cyclic weights of the additional gating states, b G The bias for the additional gating state.

4. A solar radio burst prediction system based on a complete solar magnetogram, characterized in that, include: The image feature extraction module is configured to: acquire a full-disk solar magnetograph, input the full-disk solar magnetograph into a pre-trained improved ResNet model, and acquire image feature vectors; The solar radio burst prediction module is configured to: divide the image feature vector into a pseudo-time series and input it into an improved LSTM network to obtain the solar radio burst prediction result; The improved ResNet model introduces a custom residual module to extract small-scale features of the whole-disk solar magnetogram, and the improved LSTM network introduces a gating mechanism to process the image feature vectors at different time scales. The improved ResNet model consists of sequentially connected convolutional layers, batch normalization layers, ReLU activation functions, max pooling layers, four custom residual layers, a global pooling layer, and a fully connected layer. Each custom residual layer comprises two sequentially connected custom residual modules. Each custom residual module consists of sequentially connected first convolutional unit, second convolutional unit, global average pooling layer, multilayer perceptron, and ReLU activation function. The input of the first convolutional unit and the output of the multilayer perceptron are residually connected. The first and second convolutional units have the same structure, consisting of sequentially connected convolutional layers, batch normalization layers, and ReLU activation functions. The multilayer perceptron consists of sequentially connected input layers, linear layers, ReLU activation functions, reshape layers, and output layers. The specific implementation of the custom residual module is as follows: After processing the output feature image by using two sets of convolutional units with 3x3 convolutional kernels, batch normalization (BN), and ReLU activation function, global average pooling is performed to compress the features of each channel into a single value, forming a two-dimensional tensor. The pooled features are then transformed nonlinearly through MLP while keeping the output size unchanged. The output of MLP is then adjusted to a size equal to that of the original feature map, and residual connections are performed to finally output the processed feature image. An improved LSTM network is introduced by a gating mechanism to process image feature vectors at different time scales. A dynamic attention mechanism is also introduced to dynamically adjust the attention level of the improved LSTM network to different time steps, and output solar radio burst prediction results based on the whole-disk solar magnetogram. The input 512-dimensional image feature vector is divided into pseudo-time series, and the sub-vectors are input into an improved LSTM network for processing. The processing results are weighted and calculated through a dynamic attention mechanism to output the solar radio burst prediction results.

5. An electronic device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the solar radio burst prediction method based on a full-disk solar magnetogram as described in any one of claims 1-3.

6. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the solar radio burst prediction method based on the whole-disk solar magnetogram as described in any one of claims 1-3.

7. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the solar radio burst prediction method based on the whole-disk solar magnetogram as described in any one of claims 1-3.

Images