Disturbance storm time index prediction method and device

Abstract

The invention discloses a disturbance storm time index prediction method and device, and the method comprises the steps: obtaining a second coding season component; decomposing the first superposition data to obtain a first decoding season component and a first decoding trend component; performing frequency domain learning on the second coding seasonal component and the first decoding seasonal component, and fusing the second coding seasonal component and the first decoding seasonal component with the first decoding seasonal component to obtain a first fused component; decomposing the first fusion component to obtain a second decoding season component and a second decoding trend component; decomposing the second fusion component to obtain a third decoding season component and a third decoding trend component; fusing the third decoding season component, the historical trend component sequence, the first decoding trend component, the second decoding trend component and the third decoding trend component to obtain a third fused component; and determining a disturbance storm time index prediction value based on the third fusion component. According to the invention, the precision of disturbance storm time index prediction can be improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method and device for predicting a disturbance storm time index. Background Art

[0002] The ring current is a particularly important current system from east to west in the Earth's magnetosphere. It is mainly composed of high-energy particles of 10–200 keV and induces a southward magnetic field on the surface that is opposite to the direction of the Earth's magnetic field. When the solar wind strengthens, more particles are injected into the inner magnetosphere, causing the ring current to strengthen, which significantly weakens the horizontal component of the surface magnetic field and triggers the geomagnetic storm (abbreviated as geomagnetic storm). Geomagnetic storms can seriously affect the Earth's electromagnetic environment and may cause a wide range of major impacts, including communication problems, satellite failures, induced currents in power grids and oil pipelines, etc. The violent disturbance of the horizontal component of the surface magnetic field in the middle and low latitudes caused by the ring current of the Earth's magnetosphere can be characterized by the disturbance storm time (Dst) index. As an important indicator for measuring geomagnetic activity, the Dst index can clearly reflect the occurrence time and intensity of the geomagnetic storm. Generally speaking, during the development of a geomagnetic storm, the enhancement of the ring current leads to a decrease in the intensity of the horizontal component of the geomagnetic field in low-latitude areas on the surface. Accordingly, the Dst index begins to decrease from about zero before the geomagnetic storm to a negative value. The decline of the Dst index can be used as a sign of the beginning of a magnetic storm. The Dst index also serves as a representation of the first strengthening and then weakening of the ring current, reflecting the occurrence and recovery of the magnetic storm. In daily life, the Dst index needs to be predicted to warn of geomagnetic storms. However, the inherent complex nonlinear relationship between the Dst index and the solar wind parameters leads to inaccurate predictions of the disturbance storm time index. Traditional prediction models often cannot skillfully capture the subtle changes and correlations between these factors, resulting in low accuracy in the prediction of the disturbance storm time index. Summary of the invention

[0003] The embodiments of the present application provide a disturbance storm time index prediction method and device, which can improve the accuracy of disturbance storm time index prediction.

[0004] In a first aspect, the present application provides a method for predicting disturbance storm time index, comprising:

[0005] Obtain solar wind combined parameter series data and target index prediction model;

[0006] Inputting the solar wind combined parameter sequence data into the encoder of the target index prediction model to obtain the second encoded seasonal component;

[0007] Inputting the historical seasonal component sequence into the adaptive frequency enhancement block of the decoder of the target index prediction model to perform frequency domain enhancement to obtain first enhanced data;

[0008] Superimposing the first enhanced data and the historical seasonal component sequence to obtain first superimposed data;

[0009] decomposing the first stacked data using a hybrid expert decomposition block in a decoder of the target index prediction model to obtain a first decoded seasonal component and a first decoded trend component;

[0010] The second encoded seasonal component and the first decoded seasonal component are input into the adaptive non-stationary frequency enhanced attention module of the decoder of the target index prediction model for frequency domain learning, and then fused with the first decoded seasonal component to obtain a first fused component;

[0011] decomposing the first fused component using a hybrid expert decomposition block in a decoder of the target index prediction model to obtain a second decoded seasonal component and a second decoded trend component;

[0012] After inputting the second decoded seasonal component into the feedforward neural network of the decoder of the target index prediction model, the second decoded seasonal component is fused with the second decoded seasonal component to obtain a second fused component;

[0013] decomposing the second fused component using a hybrid expert decomposition block in a decoder of the target index prediction model to obtain a third decoded seasonal component and a third decoded trend component;

[0014] Fusing the third decoded seasonal component, the historical trend component sequence, the first decoded trend component, the second decoded trend component, and the third decoded trend component to obtain a third fused component;

[0015] The disturbance storm time index prediction value is determined based on the third fusion component.

[0016] In a second aspect, the disturbance storm time index prediction device provided by the present application includes:

[0017] An acquisition module is used to obtain solar wind combined parameter sequence data and target index prediction model;

[0018] An input module, used for inputting solar wind combined parameter sequence data into an encoder of a target index prediction model to obtain a second encoded seasonal component;

[0019] An enhancement module, used for inputting the historical seasonal component sequence into an adaptive frequency enhancement block of a decoder of a target index prediction model for frequency domain enhancement to obtain first enhanced data;

[0020] A superposition module, used for superimposing the first enhanced data and the historical seasonal component sequence to obtain first superimposed data;

[0021] A first decomposition module, configured to decompose the first superimposed data using a hybrid expert decomposition block in a decoder of a target index prediction model to obtain a first decoded seasonal component and a first decoded trend component;

[0022] A first fusion module is used to input the second encoded seasonal component and the first decoded seasonal component into an adaptive non-stationary frequency enhanced attention module of a decoder of a target index prediction model for frequency domain learning, and then fuse them with the first decoded seasonal component to obtain a first fused component;

[0023] A second decomposition module, configured to decompose the first fusion component using a hybrid expert decomposition block in a decoder of the target index prediction model to obtain a second decoded seasonal component and a second decoded trend component;

[0024] A second fusion module is used to input the second decoded seasonal component into a feedforward neural network of a decoder of a target index prediction model, and then fuse it with the second decoded seasonal component to obtain a second fused component;

[0025] A third decomposition module, configured to decompose the second fusion component using a hybrid expert decomposition block in a decoder of the target index prediction model to obtain a third decoded seasonal component and a third decoded trend component;

[0026] A third fusion module is used to fuse the third decoded seasonal component, the historical trend component sequence, the first decoded trend component, the second decoded trend component, and the third decoded trend component to obtain a third fusion component;

[0027] A determination module is used to determine a disturbance storm time index prediction value based on the third fusion component.

[0028] In a third aspect, the electronic device provided in the present application includes a memory and a processor, the memory stores a computer program, and the processor is used to run the computer program in the memory to implement the steps in the disturbance storm time index prediction method provided in the present application.

[0029] In a fourth aspect, the computer-readable storage medium provided in the present application stores a plurality of instructions, which are suitable for a processor to load to implement the steps in the disturbance storm time index prediction method provided in the present application.

[0030] In a fifth aspect, the computer program product provided in the present application includes a computer program or instructions, which, when executed by a processor, implements the steps in the disturbance storm time index prediction method provided in the present application.

[0031] In the present application, compared with the related art, the disturbance storm time index prediction method includes: obtaining solar wind combination parameter sequence data and a target index prediction model; inputting the solar wind combination parameter sequence data into the encoder of the target index prediction model to obtain a second encoded seasonal component; inputting the historical seasonal component sequence into the adaptive frequency enhancement block of the decoder of the target index prediction model for frequency domain enhancement to obtain first enhanced data; superimposing the first enhanced data and the historical seasonal component sequence to obtain first superimposed data; using the mixed expert decomposition block in the decoder of the target index prediction model to decompose the first superimposed data to obtain a first decoded seasonal component and a first decoded trend component; inputting the second encoded seasonal component and the first decoded seasonal component into the adaptive non-stationary frequency enhancement attention block of the decoder of the target index prediction model After the module performs frequency domain learning, it is fused with the first decoded seasonal component to obtain the first fused component; the first fused component is decomposed using the hybrid expert decomposition block in the decoder of the target index prediction model to obtain the second decoded seasonal component and the second decoded trend component; after the second decoded seasonal component is input into the feedforward neural network of the decoder of the target index prediction model, it is fused with the second decoded seasonal component to obtain the second fused component; the second fused component is decomposed using the hybrid expert decomposition block in the decoder of the target index prediction model to obtain the third decoded seasonal component and the third decoded trend component; the third decoded seasonal component, the historical trend component sequence, the first decoded trend component, the second decoded trend component, and the third decoded trend component are fused to obtain the third fused component; the disturbance storm time index prediction value is determined based on the third fused component. The present application cleverly combines the adaptive frequency optimization technology and the non-stationary frequency enhancement strategy to improve the accuracy of the disturbance storm time index prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0033] Figure 1 It is a scene schematic diagram of a disturbance storm time index prediction system provided by an embodiment of the present application;

[0034] Figure 2 It is a schematic diagram of an embodiment of a disturbance storm time index prediction method provided in an embodiment of the present application;

[0035] Figure 3 It is a structural schematic diagram of a target index prediction model in a disturbance storm time index prediction method provided in an embodiment of the present application;

[0036] Figure 4It is a schematic diagram of observation data and prediction results of Dst index predicted 1, 3 and 6 hours in advance in 6 geomagnetic storms in the disturbance storm time index prediction method provided in an embodiment of the present application;

[0037] Figure 5 It is a schematic diagram of the comparison results between the predicted values ​​and observed values ​​of the Dst index in six geomagnetic storms in the disturbance storm time index prediction method provided in an embodiment of the present application;

[0038] Figure 6 It is a structural schematic diagram of an embodiment of a disturbance storm time index prediction device provided in an embodiment of the present application;

[0039] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0040] It should be noted that the principles of the present application are illustrated by implementing them in an appropriate computing environment. The following description is based on the illustrated specific embodiments of the present application and should not be considered as limiting other specific embodiments of the present application that are not described in detail herein.

[0041] In the following description of the present application, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it can be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0042] In the following description of the present application, the terms "first\second\third" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application described here can be implemented in an order other than that illustrated or described here.

[0043] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0044] The traditional Dst (disturbance storm time) index prediction methods mainly include the following methods:

[0045] (1) Differential equation prediction method based on traditional experience.

[0046] It is generally believed that the ring current injection rate Q is linearly related to the solar wind-magnetosphere coupling function, which is expressed as a function of the interplanetary magnetic field and interplanetary plasma parameters. In the early days, it was believed that the ring current decay time τ was a constant, but recent studies have shown that τ varies with interplanetary conditions. The energy coupling function ε' has a good correlation with the intensity of the magnetic storm, especially for large and extra-large magnetic storms, ε' has a good correlation with the intensity of the magnetic storm. Therefore, using the coupling function in energy form, the calculation equations for Q and τ are as follows:

[0047]

[0048] Using solar wind observation data as input parameters and the quicklookDst index provided by the International Geomagnetic Data Center as the initial value of the iteration, the above formula (1.1) is used for iteration to realize the prediction of the Dst index.

[0049] (2) Dst index prediction method based on support vector machine (SVM) model.

[0050] Support vector machine (SVM) is based on the VC dimension theory in statistical learning (VC dimension is mainly used to study the speed and generalizability of the learning process to achieve consistent convergence, and uses relevant statistical theories to determine important indicators of the learning ability of related function sets. It can usually be understood as the complexity of the problem. The higher the VC dimension, the more complex the problem). It is a machine learning method that aims to minimize structural risk. This method is similar to multi-layer perceptron networks and radial basis function networks. It was originally used to solve pattern classification problems, and then it was developed to solve nonlinear regression problems. The main theoretical idea of ​​support vector machine is to transform the linear inseparable problem of low-dimensional input space into high-dimensional feature space through mapping, making it a linear separable problem, and establish a hyperplane as a decision surface, so as to achieve the purpose of maximizing the isolation edge between positive examples and negative examples.

[0051] This method uses solar wind observation data (three components of the interplanetary magnetic field, three components of the solar wind speed, proton number density, temperature, solar wind dynamic pressure, plasma beta value, electric field y component, Alfvén speed, clock angle, cone angle and pitch angle) as model input, and uses K-fold cross validation to improve the reliability of the model prediction results. Then, the data set is divided into a training set and a test set. Usually, the training set accounts for a larger proportion and is used for model training. The test set is used for model verification and evaluation. Then, a suitable SVM kernel function is selected, such as a linear kernel, a polynomial kernel or a Gaussian kernel (RBF kernel). The kernel function formula is as follows:

[0052] 1. Linear kernel function:

[0053] K(x,x i )=x T ,xi (1.2)

[0054] 2. Polynomial kernel function:

[0055] K(x,x i )=(γx T x i +r) P ,γ>0 (1.3)

[0056] 3.RBF kernel function:

[0057] K(x,x i )=exp(-γ||xx i || 2 ),γ>0 (1.4)

[0058] 4. Two-layer perceptron kernel function:

[0059] K(x,x i ) = tanh(γx T x i +r) P (1.5)

[0060] Map the data to a high-dimensional space to make the data linearly separable or easier to process. Use the training set data to train the SVM model and find the best hyperplane that maximizes the margin between the predicted value and the actual value. Use the trained SVM model to predict the new input data and get the predicted value of the Dst index.

[0061] (3) Dst index prediction method based on long short-term memory (LSTM) model.

[0062] LSTM is a type of recurrent neural network that is very suitable for predicting time series data. The LSTM model has complex hidden units that can selectively add or reduce information, so it can accurately model data with short-term or long-term dependencies. LSTM has two transmission states, the cell state (C) and the hidden state (H). First, it decides to update the state from the previous cell state (C). t-1 ) is discarded, and the previous output (H t-1 ) and the current input (X t ) through the forget gate (F t ) to make a decision. Then it decides what information to store in the cell state (C). First, the input gate (I t ) decides which values ​​to update and passes a tanh layer to create candidate vectors (C t ), and then update the cell state (C t Finally, through the output gate (Ot) and the cell state (Ct ) to get the final output (H t ).

[0063] The differential equation prediction method based on traditional experience has the problem of inaccurate parameter estimation. Different scenarios have different parameters. The linear correlation has limited representation capabilities, and the correlation between the current injection rate Q and the solar wind-magnetosphere coupling function is linear or nonlinear. At present, there is no absolute theoretical support. If the ring current injection rate Q and the solar wind-magnetosphere coupling function are not linear, the prediction method based on this relationship will definitely be inaccurate. It can be seen from formula (1.1) that the differential equation prediction method based on traditional experience requires a large number of estimated parameter values, such as 7.26, -3.83, 8.70, etc. However, these parameters are estimated by various researchers based on experience, and there are certain errors. This leads to inaccurate prediction models. Moreover, in different scenarios, these parameters have different values, and the prediction range of the prediction model is very limited.

[0064] The Dst index prediction method based on the support vector machine (SVM) model can only predict events, not hourly predictions. The Dst index prediction method using the support vector machine (SVM) model only focuses on the prediction superiority of the details of the magnetic storm event, and does not involve the accuracy of predictions several hours in advance. In addition, when the magnetic storm data set reaches a certain scale, the training of the model will become very time-consuming and even exceed the limitations of computing resources.

[0065] The Dst index prediction method based on the long short-term memory (LSTM) model has high computational complexity and limited parallel processing capabilities. Due to its complex internal structure and large number of parameters, the LSTM model requires more computing resources. This computing demand is particularly significant when processing large-scale data sets or long sequences, which may cause the training process to take a long time. In addition, since the LSTM structure contains multiple fully connected layers (MLP) and needs to store cell states and hidden states, it will occupy a lot of memory resources during the training process.

[0066] In order to improve the effect of disturbance storm time index prediction, the embodiments of the present application provide a disturbance storm time index prediction method, a disturbance storm time index prediction device, an electronic device, a computer readable storage medium, and a computer program product. Among them, the disturbance storm time index prediction method can be executed by the disturbance storm time index prediction device, or by an electronic device integrated with the disturbance storm time index prediction device.

[0067] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0068] Please refer to Figure 1 , the present application also provides a disturbance storm time index prediction system, such as Figure 1 As shown, the disturbance storm time index prediction system includes an electronic device 100, and the disturbance storm time index prediction device provided by the present application is integrated in the electronic device 100.

[0069] Among them, the electronic device 100 can be any device equipped with a processor and has processing capabilities, such as mobile electronic devices with processors such as smart phones, tablet computers, PDAs, laptops, smart speakers, or fixed electronic devices with processors such as desktop computers, televisions, servers, industrial equipment, etc.

[0070] In addition, if Figure 1 As shown, the disturbance storm time index prediction system may further include a memory 200 for storing original data, intermediate data, and result data.

[0071] In the embodiment of the present application, the memory 200 can be a cloud storage. Cloud storage is a new concept extended and developed from the concept of cloud computing. A distributed cloud storage system (hereinafter referred to as the storage system) refers to a storage system that uses cluster applications, grid technology, and distributed storage file systems to bring together a large number of different types of storage devices (storage devices are also called storage nodes) in the network through application software or application interfaces to work together and provide external data storage and business access functions.

[0072] At present, the storage method of the storage system is: create a logical volume, and when creating a logical volume, allocate physical storage space for each logical volume. The physical storage space may be composed of disks of a storage device or several storage devices. The client stores data on a logical volume, that is, stores the data on the file system. The file system divides the data into many parts, each of which is an object. The object contains not only data but also additional information such as data identification (ID, ID entity). The file system writes each object into the physical storage space of the logical volume, and the file system records the storage location information of each object, so that when the client requests to access the data, the file system can allow the client to access the data according to the storage location information of each object.

[0073] The process of the storage system allocating physical storage space to a logical volume is as follows: based on the estimated capacity of the objects stored in the logical volume (this estimate often has a large margin relative to the actual capacity of the objects to be stored) and the grouping of independent redundant disk arrays (RAID, Redundant Array of Independent Disk), the physical storage space is divided into stripes in advance. A logical volume can be understood as a stripe, thereby allocating physical storage space to the logical volume.

[0074] It should be noted that Figure 1 The scenario diagram of the disturbance storm time index prediction system shown is only an example. The disturbance storm time index prediction system and scenario described in the embodiment of the present application are intended to more clearly illustrate the technical solution of the embodiment of the present application, and do not constitute a limitation on the technical solution provided in the embodiment of the present application. A person of ordinary skill in the art can know that with the evolution of the disturbance storm time index prediction system and the emergence of new business scenarios, the technical solution provided in the embodiment of the present application is also applicable to similar technical problems.

[0075] It should be noted that the serial numbers of the following embodiments are not intended to limit the preferred order of the embodiments.

[0076] Please refer to Figure 2 , Figure 2 FIG. 1 is a flow chart of an embodiment of a disturbance storm time index prediction method provided in an embodiment of the present application. Figure 2 As shown, the process of the disturbance storm time index prediction method provided by this application is as follows:

[0077] 201. Obtain solar wind combined parameter sequence data and target index prediction model.

[0078] In an embodiment of the present application, the solar wind combined parameter sequence data includes the total interplanetary magnetic field, the north-south component of the interplanetary magnetic field, the proton temperature, the proton density, the solar wind speed, the solar wind dynamic pressure and the electric field.

[0079] In an embodiment of the present application, a disturbance storm time index prediction method includes:

[0080] (1) Acquire a historical data set, wherein the historical data set includes a plurality of historical sequences of solar wind combination parameters and corresponding disturbance storm time index observation values ​​in a preset time period.

[0081] In the embodiment of the present application, the preset time period can be set according to the specific situation. For example, the preset time period is from January 1, 2008 to December 31, 2020. The historical data set mainly consists of multiple solar wind combination parameter historical sequences and corresponding disturbance storm time index observations. This data set comes from the OMNI database in NASA's National Space Science Data Center (NSSDC). The time resolution of the solar wind combination parameter historical sequence and the corresponding disturbance storm time index observations remains at one hour throughout the data collection period. The solar wind combination parameters encapsulated in this data set include the total interplanetary magnetic field (B IMF ), the z component (or north-south component) of the interplanetary magnetic field (B z ) (expressed in the Geocentric Solar Magnetosphere (GSM) coordinate system), proton temperature (T), proton density (D), solar wind speed (V sw ), solar wind dynamic pressure (P), and electric field (E). Considering the time span and hourly resolution of the dataset, it consists of 43,824 data points. The specific choice of this time range ensures reliable coverage of the relevant phases of the solar cycle, providing an overall view of the underlying patterns and variations in the data.

[0082] (2) Perform preprocessing operations on the historical data set to obtain a preprocessed data set.

[0083] In the embodiment of the present application, the historical data set is preprocessed to obtain the preprocessed data set, including: configuring the URL of the source data, so that the historical data set can be quickly downloaded from the network. The downloaded historical data set is preprocessed, such as aligning the data and deleting useless attribute values.

[0084] (3) Split the preprocessed dataset into training set, validation set, and test set according to the preset ratio.

[0085] In the embodiment of the present application, the preprocessed data set is divided into a training set, a validation set, and a test set in a ratio of 6:2:2, and Z-Score normalization is performed.

[0086] (4) Obtain the target index prediction model based on the training set.

[0087] In the model setting, the dimension of the model is configured to 512, and 8 attention heads are used to process the data. The target index prediction model includes 2 encoders and 1 decoder. In training, a batch size of 32 is used, and the model is run for 10 epochs with an initial learning rate of 0.0001. To prevent overfitting, the dropout rate is set to 0.05, and the GELU activation function is used.

[0088] (5) Test the target index prediction model based on multiple key indicators and test sets.

[0089] Among them, multiple key indicators are MAE, RMSE and correlation coefficient (R), which reflect different aspects of the accuracy and reliability of the model respectively.

[0090] Mean Absolute Error (MAE):

[0091]

[0092] Root Mean Square Error (RMSE):

[0093]

[0094] Correlation coefficient (R):

[0095]

[0096] The prediction model is trained cyclically using the data of the training set and the validation set, and the prediction of the model is evaluated using the test set data.

[0097] Forward propagation: The process by which data passes through a neural network, starting at the input layer and calculating the output of each layer, layer by layer, until the final prediction result (or activation value of the output layer) is obtained. In this step, the model uses the current parameter values ​​to calculate the prediction for the given input. For each layer, the output of the previous layer is taken as input, and the output of the current layer is obtained by weighted summing (plus bias term) and then applying activation function.

[0098] Calculate loss: The loss function measures the difference between the model prediction and the true label. It is the goal of model optimization, and the training process is to try to find a set of parameters that minimize the loss function. Use loss functions (such as mean absolute error, mean square error, and correlation coefficient) to compare the model prediction with the true label and calculate the loss value.

[0099] Backpropagation: The process of computing the gradients (or derivatives) of the loss function with respect to the model parameters. These gradients indicate how the parameters should be adjusted to reduce the loss. Starting from the output layer, the gradients of the parameters of each layer are computed backwards, layer by layer, until the input layer. This usually involves the application of the chain rule to calculate the contribution of each parameter to the loss.

[0100] Parameter update: Based on the gradient obtained by backpropagation, the optimization algorithm (such as stochastic gradient descent, Adam, etc.) is used to update the parameters of the model. The goal is to reduce the value of the loss function. Usually, the parameters are updated with a small step size (learning rate) in the opposite direction of the gradient. For example, for each parameter, the new value is equal to the old value minus the learning rate multiplied by the gradient of the parameter.

[0101] Validation model: Validation model is the process of evaluating the performance of the model on a dataset independent of the training data. This helps monitor whether the model is overfitting and adjust the training strategy. At some point in the training process (such as at the end of each epoch), the loss and performance metrics of the model are calculated using the validation dataset.

[0102] Adjust the learning rate: The learning rate is a hyperparameter that determines the step size of the parameter update. An appropriate learning rate can speed up training and find better parameter values. Dynamically adjust the learning rate based on information such as validation performance and loss curves. Common methods include learning rate decay (such as exponential decay, cosine decay), learning rate warm-up, etc.

[0103] Save model parameters: Save model parameters (including weights and biases) to a file (such as HDF5, PyTorch's .pth file).

[0104] 202. Input the solar wind combined parameter sequence data into an encoder of a target index prediction model to obtain a second encoded seasonal component.

[0105] like Figure 3 As shown, in an embodiment of the present application, the target index prediction model includes N encoders and M decoders. In an embodiment of the present application, N is 2 and M is 1. The encoder includes a reversible instance normalization module (RevIN), an adaptive frequency enhancement block (AFEB), a feedforward neural network (FeedForward), and two hybrid expert decomposition blocks (MOEDecomp). The adaptive frequency enhancement block (AFEB) is used to adjust the frequency properties of the input sequence; the hybrid expert decomposition block (MOEDecomp) separates the sequence into two components, seasonal (S) and trend (T) components, and the feedforward neural network (FeedForward) refines the seasonal signal for prediction. It is the embedded historical sequence data, namely the solar wind combination parameter sequence data.

[0106] The encoder adopts a multi-layer structure, the structure is as follows:

[0107]

[0108] in, represents the output of the l-layer encoder layer, represents the input of the (l-1)th encoder layer. l is the number of encoder layers, l∈[1,N]. is the embedded historical sequence data, i.e., the solar wind combined parameter sequence data. Encoder(·) is formalized as follows:

[0109]

[0110]

[0111] in, is the seasonal component after the first MOEDecomp decomposition block in layer l. AFEB(·) is the adaptive frequency enhancement block. represents the input of the (l-1)th encoder layer. is the seasonal component decomposed by the second MOEDecomp decomposition block in layer l.

[0112] In the embodiment of the present application, the solar wind combined parameter sequence data is input into the encoder of the target index prediction model to obtain the second encoded seasonal component, including:

[0113] (1) The solar wind combined parameter sequence data is input into the reversible instance normalization module of the encoder of the target index prediction model for reversible instance normalization to obtain first standardized data.

[0114] The solar wind combined parameter sequence data is input into the reversible instance normalization module (RevIN) of the encoder of the target index prediction model for reversible instance normalization to obtain the first standardized data.

[0115] In the field of time series forecasting, accurately predicting future values ​​is often a complicated matter due to the problem of distribution shift. This problem often occurs when the statistical properties of time series data (such as mean and variance) change over time, resulting in distribution differences between the training dataset and the test dataset. This difference can greatly reduce the performance of the forecasting model because they can only scale well to known data. Therefore, solving the distribution shift problem is very important to improve the accuracy and reliability of time series forecasting models. To solve the distribution shift problem, this application proposes a reversible instance normalization module (RevIN). The method consists of a symmetrically structured normalization and denormalization layer, which normalizes the time series data to a standard distribution and restores the original output distribution. RevIN uses learnable affine transformation parameters to adaptively scale and shift the normalized data, so that the model maintains the statistical properties of the input data throughout the forecasting process. The RevIN process consists of two key stages: normalization and denormalization.

[0116] Normalization: In the multivariate time series forecasting task, given a set of input instances X = {x 1 ,x 2 ,…,x K}, K is the number of input instances. For each input instance T x is the length of the input sequence and D is the number of feature variables. The mean and variance of this input instance are calculated as follows:

[0117]

[0118] in, Indicates that the d features of the i instance are in all T x The average value of the time step, T x is the length of the input sequence. Represents the values ​​of d features of instance i at time step t. Indicates that the d features of the i instance are in all T x We then use these statistics to estimate the variance of the input data To normalize:

[0119]

[0120] in, represents the normalized value of d features of instance i at time step t. γ d and β d is the learnable affine parameter of each feature variable and ò is a small constant to ensure numerical stability.

[0121] Denormalization: Based on transformed input instances The predicted value of Denormalize, T y is the length of the prediction sequence. We can get the final prediction of the model by the following method

[0122]

[0123] in, is the final prediction output of RevIN denormalization, i.e., the first standardized data, which is the predicted value of d features in i instances at time step t. is the predicted output of the model. The parameter γ d and β d is a learnable affine parameter corresponding to d features, allowing the model to adjust the normalization value according to the statistical properties of the data. In addition, ò is a small constant used to prevent division by zero during normalization to ensure numerical stability. Through this two-stage process, RevIN can effectively alleviate the impact of distribution drift, enabling time series forecasting models to better adapt to changes in data distribution and improve forecasting accuracy.

[0124] (2) The first standardized data is input into the adaptive frequency enhancement block of the encoder of the target index prediction model for frequency domain enhancement to obtain second enhanced data.

[0125] The first standardized data is input into an adaptive frequency enhancement block (AFEB) of an encoder of a target index prediction model for frequency domain enhancement to obtain second enhanced data.

[0126] In the embodiment of the present application, the adaptive frequency enhancement block (AFEB) processes the input X through a multi-layer perceptron (MLP) en / de ∈R L×D The potential representation Q∈R is generated from L×D .

[0127] Q = MLP(X en / de ),

[0128] Where, the Fourier transform of Q is expressed as The frequency domain is obtained by fast Fourier transform (FFT). In the frequency domain, only the α% (0<α<100) component is selected as follows:

[0129]

[0130] in, For the selected component, Select α (·) is the random selection operation of the frequency components in Q′(M<<N), F(·) is the fast Fourier transform (FFT), Transformation via BlockMLP:

[0131]

[0132] in, for The output after BlockMLP processing. This BlockMLP has two operations: 1. The ReLU activation function is applied to 2. Linear transformation W 2 Applied to the above results. Among them, W 1 and W 2 is the weight of BlockMLP, and ReLU is the rectified linear unit activation function. Frequency domain signal Thinning is performed using a soft thresholding technique via the SoftShrink function:

[0133]

[0134] in, is to apply the SoftShrink function to The feedback signal after λ is a regularization parameter used to adjust the sparsity level of the result representation. The SoftShrink function effectively The small-amplitude elements of are set to zero while retaining the large-amplitude elements, thereby improving the sparsity of the signal representation. After soft thresholding, is padded with zeros and converted back to the time domain using an inverse fast Fourier transform (IFFT):

[0135]

[0136] Among them, Y represents the frequency domain signal filled by Apply the inverse Fourier transform F -1 (·) The final output is obtained. This operation ensures that the output data is consistent with the original data structure, thus producing Reflects the processed data, that is, the second enhanced data.

[0137] (3) Superimposing the second enhanced data and the first standardized data to obtain second superimposed data.

[0138] (4) Decomposing the second superimposed data using a hybrid expert decomposition block in an encoder of the target index prediction model to obtain a first encoded seasonal component.

[0139] The second superimposed data is decomposed using a mixed expert decomposition block (MOEDecomp) in the encoder of the target index prediction model to obtain a first encoded seasonal component and a trend component.

[0140] The Mixed Expert Decomposition block (MOEDecomp) solves the problem of extracting trend components from complex periodic data. It combines average pooling filters of different sizes to extract different trends and uses a Softmax function to dynamically weight these trends for synthesis. This process is encapsulated as:

[0141] X trend =Softmax(L(x))×F(x),

[0142] X seasonal =XX trend ,

[0143] Among them, X trend represents the trend component of the input sequence data X, which is obtained by applying the Softmax function to the output L(x) and then multiplying it with the pooling filter F(x). seasonal is the seasonal component of the data X, which is obtained by subtracting the trend component X from the original data X trend The calculated value is the original data X and the trend component X trend The process definition of MOEDecomp is as follows:

[0144] X seasonal ,X trend =MOEDecomp(X).

[0145] That is, the first coded seasonal component.

[0146] (5) After the first coded seasonal component is input into the feedforward neural network of the encoder of the target index prediction model, it is fused with the first coded seasonal component to obtain the first coded fusion component.

[0147] After the first coded seasonal component is input into the feedforward neural network (FeedForward) of the encoder of the target index prediction model, it is fused with the first coded seasonal component to obtain the first coded fusion component.

[0148] (6) Using the hybrid expert decomposition block in the decoder of the target index prediction model, the first coded fusion component is decomposed to obtain the second coded seasonal component.

[0149] The mixed expert decomposition block (MOEDecomp) in the decoder using the target index prediction model decomposes the first coded fusion component into a second coded seasonal component. That is, the second encoded seasonal component.

[0150] 203. Input the historical seasonal component sequence into the adaptive frequency enhancement block of the decoder of the target index prediction model for frequency domain enhancement to obtain first enhanced data.

[0151] like Figure 3 As shown, in the embodiment of the present application, the historical seasonal component sequence An adaptive frequency enhancement block (AFEB) of a decoder inputting the target index prediction model performs frequency domain enhancement to obtain first enhanced data.

[0152] like Figure 3 As shown in the figure, in the embodiment of the present application, the decoder includes a reversible instance normalization module (RevIN), an adaptive frequency enhancement block (AFEB), an adaptive non-stationary frequency enhancement attention module (ANFEA), a feedforward neural network (FeedForward) and three mixed expert decomposition blocks (MOEDecomp). The decoder also adopts a multi-layer structure. The input of each layer and is the output of the previous layer. In the decoder, the input After the adaptive frequency enhancement block (AFEB) and residual connection, it is decomposed into seasonal components through the MOEDecomp module and trend component Then, and encoder output It is input into the adaptive non-stationary frequency enhanced attention module (ANFEA) and another MOEDecomp block to obtain and Then, After the feedforward network and the final MOEDecomp block, we generate and Seasonal Section The output of this layer The trend part T of this layer l de It is obtained by weighted summing the trend part of this layer and the trend part of the previous layer. The multi-layer structure of the decoder is as follows:

[0153]

[0154] Among them, is the seasonal component output of the l-th layer decoder, T l de is the trend component output of the l-th layer decoder. l represents the number of the decoder, l∈[1,M]. and is the embedded historical sequence data, is the historical seasonal component series, is the historical trend component sequence. The Decoder(·) structure can be formalized as:

[0155]

[0156] in, and Respectively represent the seasonal component and trend component of the i-th decomposition block in the l-th layer. l,i (i∈{1,2,3}) represents the projection of extracting i trends ANFEA(·) is the adaptive non-stationary frequency enhanced attention. The seasonal component of the last layer and trend component Add them together to get the final prediction result. is the seasonal component after depth transformation, achieved by projecting it into the target dimension.

[0157] 204. Superimpose the first enhanced data and the historical seasonal component sequence to obtain first superimposed data.

[0158] 205. Decompose the first superimposed data using a hybrid expert decomposition block in a decoder of the target index prediction model to obtain a first decoded seasonal component and a first decoded trend component.

[0159] 206. After inputting the second encoded seasonal component and the first decoded seasonal component into the adaptive non-stationary frequency enhanced attention module of the decoder of the target index prediction model for frequency domain learning, the components are fused with the first decoded seasonal component to obtain a first fused component.

[0160] In the embodiment of the present application, the second encoded seasonal component and the first decoded seasonal component are input into the adaptive non-stationary frequency enhanced attention module of the decoder of the target index prediction model for frequency domain learning, and then fused with the first decoded seasonal component to obtain a first fused component, including:

[0161] The two second encoded seasonal components are taken as K vector and V vector respectively, and the first decoded seasonal component is taken as Q vector. After input into the adaptive non-stationary frequency enhanced attention module of the decoder of the target index prediction model for frequency domain learning, they are fused with the first decoded seasonal component to obtain the first fused component.

[0162] In this application, the Adaptive Non-Stationary Frequency Enhanced Attention (ANFEA) mechanism can analyze the paradigm of sequence data in the frequency domain and process complex non-stationary signals.

[0163] This module redesigns the input components to query the Q vector, K vector and V vector. The criss-cross attention is obtained by deriving a query from the decoder, denoted as q = x en ·w q ,in At the same time, through k = x de ·w k and v = x de ·w v Computes the keys and values ​​from the encoder.

[0164] The ANFEA module performs Fourier transform on the encoded query, key and value to obtain the frequency domain and In the selection process of frequency domain representation, the components of α% (0<α<100) are randomly selected as follows:

[0165]

[0166] Among them, Select α (·) is a mechanism that randomly selects a fraction α% of frequency components from each Fourier transformed input, allowing the model to adaptively manage the dynamic range of frequencies present in non-stationary signals. The core of this prediction model is the frequency domain attention mechanism, which is encapsulated as ANFEA(q,k,v). This mechanism is crucial for capturing the temporal dynamics of data in spectral format. It focuses on the frequency domain in the following way:

[0167]

[0168] Where ANFEA(q,k,v) represents an attention-based nonlinear feature extraction operation for query q, key k, and value v matrices. The operation consists of using the dot product of the query matrix and the key matrix Calculate the attention score, followed by the sigmoid activation σ. These scores have an effect on the value matrix in the attention mechanism Perform weighting. Perform soft thresholding on the weighted value SoftShrink λ (·) Introduce sparsity, then perform padding and inverse Fourier transform F -1 (·). Finally, the result is scaled by τ and shifted by Δ. After the attention calculation, a soft thresholding operation SoftShrink is applied λ (·), guides sparsity and enhances frequency domain signals The robustness is as follows:

[0169]

[0170] in, Represents the output tensor resulting from applying the SoftShrink function to the product of the attention score and the value. λ (·) performs soft thresholding on the input tensor, keeping values ​​above a threshold λ and setting others to zero. σ(·) represents the sigmoid activation function, introducing nonlinearity into the computation. and are the query matrix, key matrix, and value matrix, respectively. The expression sign(·) returns the sign of each element in the tensor. The expression max(|·|-λ,0) computes the maximum of the elements between the absolute value of the tensor subtracted by λ and zero, effectively performing soft thresholding. To ensure the fidelity of non-stationary characteristics in the transformed signal, adaptive scaling τ and shifting ΔΔ are rigorously applied to the frequency-to-time domain conversion as follows:

[0171]

[0172] Where Y represents the tensor filled by applying The inverse Fourier transform (F -1 ) is obtained, followed by element-wise multiplication of the scaling factor τ and addition of the shift factor Δ. The Padding(·) operation ensures that The structure of the original data is kept consistent, so that the final output Y of the frequency domain learning that reflects the processed data is obtained. By focusing on important frequency components and reducing noise, the ANFEA module enhances the prediction accuracy of non-stationary sequence data. The final output Y of the frequency domain learning is fused with the first decoded seasonal component to obtain the first fused component.

[0173] 207. Decompose the first fusion component using the mixed expert decomposition block in the decoder of the target index prediction model to obtain a second decoded seasonal component and a second decoded trend component.

[0174] 208. After the second decoded seasonal component is input into the feedforward neural network of the decoder of the target index prediction model, it is fused with the second decoded seasonal component to obtain a second fused component.

[0175] 209. Decompose the second fusion component using the mixed expert decomposition block in the decoder of the target index prediction model to obtain a third decoded seasonal component and a third decoded trend component.

[0176] 210. Fuse the third decoded seasonal component, the historical trend component sequence, the first decoded trend component, the second decoded trend component, and the third decoded trend component to obtain a third fused component.

[0177] 211. Determine a predicted value of a disturbance storm time index based on the third fusion component.

[0178] In an embodiment of the present application, a disturbance storm time index prediction value is determined based on the third fusion component, including: inputting the third fusion component into a reversible instance normalization module of a decoder of a target index prediction model for normalization and denormalization in sequence to obtain a disturbance storm time index prediction value.

[0179] In order to evaluate the reliability of the target index prediction model ANFEformer under severe geomagnetic disturbances, its performance under strong storm conditions was rigorously analyzed. Six magnetic storm events were selected for prediction comparison, including two very large storms with a minimum Dst less than -200nT (March and June 2015), two with a minimum Dst less than -175nT (December 2015 and August 2018), and two with a minimum Dst less than -125nT (October 2016 and May 2017). Table 4 shows the evaluation index values ​​of the prediction results of the six magnetic storm events. Figure 4 The observation data and forecast results of the Dst index predicted 1, 3 and 6 hours in advance in 6 geomagnetic storms are shown. Figure 5 The comparison between the predicted and observed values ​​of the Dst index is shown in detail in Figure 2. From these six events, it can be seen that the one-hour forecast results accurately reflect the three stages of the magnetic storm: the initial stage, the main stage, and the recovery stage. The forecast of the sudden start of the magnetic storm (SC) is also good, which is consistent with the observations. The three-hour forecast also includes all four parts, but the consistency with the observations is reduced. The six-hour forecast also does a better job in predicting the three stages of the storm, but is weaker in predicting the SC.

[0180] In the empirical evaluation, the disturbance storm time (Dst) values ​​predicted by the forecast model were compared with the corresponding actual recorded data. A series of scatter plots for different forecast intervals show the degree of agreement between the forecast and observation. The R values ​​of the forecast horizon are all high, especially for the 1-3h forecast, with R values ​​ranging from 0.827 to 0.986, which has good linear robustness. At the same time, the MAE values ​​represented on the secondary vertical axis are always contained in a narrow range, which proves the accuracy of the model. The data points clustered near the best fit line of the scatter plot indicate that the forecast model performs well.

[0181] Table 1: Prediction results of 6 geomagnetic storms

[0182]

[0183] Compared with the related art, the disturbance storm time index prediction method includes: obtaining solar wind combination parameter sequence data and a target index prediction model; inputting the solar wind combination parameter sequence data into the encoder of the target index prediction model to obtain a second encoded seasonal component; inputting the historical seasonal component sequence into the adaptive frequency enhancement block of the decoder of the target index prediction model for frequency domain enhancement to obtain first enhanced data; superimposing the first enhanced data and the historical seasonal component sequence to obtain first superimposed data; using the mixed expert decomposition block in the decoder of the target index prediction model to decompose the first superimposed data to obtain a first decoded seasonal component and a first decoded trend component; inputting the second encoded seasonal component and the first decoded seasonal component into the adaptive non-stationary frequency enhancement attention module of the decoder of the target index prediction model After frequency domain learning, it is fused with the first decoded seasonal component to obtain the first fused component; the first fused component is decomposed using the hybrid expert decomposition block in the decoder of the target index prediction model to obtain the second decoded seasonal component and the second decoded trend component; after the second decoded seasonal component is input into the feedforward neural network of the decoder of the target index prediction model, it is fused with the second decoded seasonal component to obtain the second fused component; the second fused component is decomposed using the hybrid expert decomposition block in the decoder of the target index prediction model to obtain the third decoded seasonal component and the third decoded trend component; the third decoded seasonal component, the historical trend component sequence, the first decoded trend component, the second decoded trend component, and the third decoded trend component are fused to obtain the third fused component; the disturbance storm time index prediction value is determined based on the third fused component. The present application cleverly combines the adaptive frequency optimization technology and the non-stationary frequency enhancement strategy to improve the accuracy of the disturbance storm time index prediction.

[0184] In order to facilitate better implementation of the disturbance storm time index prediction method provided in the embodiment of the present application, the embodiment of the present application also provides a disturbance storm time index prediction device based on the above disturbance storm time index prediction method. The meanings of the terms are the same as those in the above disturbance storm time index prediction method. For specific implementation details, please refer to the description in the above method embodiment.

[0185] Please refer to Figure 6 , Figure 6 : is a schematic diagram of the structure of an embodiment of a disturbance storm time index prediction device provided in an embodiment of the present application, and the disturbance storm time index prediction device may include:

[0186] An acquisition module 701 is used to acquire solar wind combined parameter sequence data and a target index prediction model;

[0187] An input module 702, used to input the solar wind combined parameter sequence data into an encoder of the target index prediction model to obtain a second encoded seasonal component;

[0188] An enhancement module 703 is used to input the historical seasonal component sequence into an adaptive frequency enhancement block of a decoder of a target index prediction model for frequency domain enhancement to obtain first enhanced data;

[0189] A superposition module 704 is used to superimpose the first enhanced data and the historical seasonal component sequence to obtain first superimposed data;

[0190] A first decomposition module 705 is used to decompose the first superposition data using a hybrid expert decomposition block in a decoder of a target index prediction model to obtain a first decoded seasonal component and a first decoded trend component;

[0191] A first fusion module 706 is used to input the second encoded seasonal component and the first decoded seasonal component into the adaptive non-stationary frequency enhanced attention module of the decoder of the target index prediction model for frequency domain learning, and then fuse them with the first decoded seasonal component to obtain a first fused component;

[0192] A second decomposition module 707, configured to decompose the first fusion component using a hybrid expert decomposition block in a decoder of the target index prediction model to obtain a second decoded seasonal component and a second decoded trend component;

[0193] A second fusion module 708 is used to input the second decoded seasonal component into the feedforward neural network of the decoder of the target index prediction model, and then fuse it with the second decoded seasonal component to obtain a second fused component;

[0194] A third decomposition module 709 is used to decompose the second fusion component using the hybrid expert decomposition block in the decoder of the target index prediction model to obtain a third decoded seasonal component and a third decoded trend component;

[0195] A third fusion module 710 is used to fuse the third decoded seasonal component, the historical trend component sequence, the first decoded trend component, the second decoded trend component, and the third decoded trend component to obtain a third fused component;

[0196] The determination module 711 is used to determine the disturbance storm time index prediction value based on the third fusion component.

[0197] The specific implementation of each of the above modules can be found in the previous embodiments and will not be described in detail here.

[0198] Please refer to Figure 7 , Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application.

[0199] The electronic device may include components such as a processor 101 with one or more processing cores, a memory 102 with one or more computer-readable storage media, a power supply 103, and an input unit 104. Those skilled in the art will appreciate that the electronic device structure shown in the figure does not constitute a limitation on the electronic device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently. Among them:

[0200] The processor 101 is the control center of the electronic device, which uses various interfaces and lines to connect various parts of the entire electronic device, and executes various functions of the electronic device and processes data by running or executing software programs and / or modules stored in the memory 102, and calling data stored in the memory 102. Optionally, the processor 101 may include one or more processing cores; optionally, the processor 101 may integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface, and application programs, and the modem processor mainly processes wireless communications. It is understandable that the above-mentioned modem processor may not be integrated into the processor 101.

[0201] The memory 102 can be used to store software programs and modules. The processor 101 executes various functional applications and disturbance storm time index prediction by running the software programs and modules stored in the memory 102. The memory 102 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device, etc. In addition, the memory 102 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage devices. Accordingly, the memory 102 may also include a memory controller to provide the processor 101 with access to the memory 102.

[0202] The electronic device also includes a power supply 103 for supplying power to each component. Optionally, the power supply 103 can be logically connected to the processor 101 through a power management system, so as to manage charging, discharging, power consumption and other functions through the power management system. The power supply 103 can also include one or more DC or AC power supplies, recharging systems, power failure detection circuits, power converters or inverters, power status indicators and other arbitrary components.

[0203] The electronic device may further include an input unit 104, which may be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal input related to user settings and function control.

[0204] Although not shown, the electronic device may also include a display unit, an image acquisition element, etc., which will not be described in detail here. Specifically in this embodiment, the processor 101 in the electronic device will load the executable code corresponding to one or more computer programs into the memory 102 according to the following instructions, and the processor 101 will execute the steps in the disturbance storm time index prediction method provided in this application, such as:

[0205] The solar wind combined parameter sequence data and the target index prediction model are obtained; the solar wind combined parameter sequence data is input into the encoder of the target index prediction model to obtain the second encoded seasonal component; the historical seasonal component sequence is input into the adaptive frequency enhancement block of the decoder of the target index prediction model for frequency domain enhancement to obtain the first enhanced data; the first enhanced data and the historical seasonal component sequence are superimposed to obtain the first superimposed data; the first superimposed data is decomposed using the mixed expert decomposition block in the decoder of the target index prediction model to obtain the first decoded seasonal component and the first decoded trend component; the second encoded seasonal component and the first decoded seasonal component are input into the adaptive non-stationary frequency enhancement attention module of the decoder of the target index prediction model for frequency domain learning, and then the first decoded seasonal component and the first decoded seasonal component are combined with each other. The decoded seasonal components are fused to obtain a first fused component; the first fused component is decomposed using a hybrid expert decomposition block in the decoder of the target index prediction model to obtain a second decoded seasonal component and a second decoded trend component; the second decoded seasonal component is input into the feedforward neural network of the decoder of the target index prediction model, and then fused with the second decoded seasonal component to obtain a second fused component; the second fused component is decomposed using a hybrid expert decomposition block in the decoder of the target index prediction model to obtain a third decoded seasonal component and a third decoded trend component; the third decoded seasonal component, the historical trend component sequence, the first decoded trend component, the second decoded trend component, and the third decoded trend component are fused to obtain a third fused component; the disturbance storm time index prediction value is determined based on the third fused component.

[0206] It should be noted that the electronic device provided in the embodiment of the present application and the disturbance storm time index prediction method in the above embodiment belong to the same concept, and its specific implementation process is detailed in the above related embodiments and will not be repeated here.

[0207] The present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program stored therein is executed on a processor of an electronic device provided in an embodiment of the present application, the processor of the electronic device executes the steps in the disturbance storm time index prediction method provided in the present application. The storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0208] The present application also provides a computer program product or a computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes various optional implementations of the above-mentioned disturbance storm time index prediction method.

[0209] The above is a detailed introduction to a disturbance storm time index prediction method and device provided by the present application. This article uses specific examples to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

[0210] It should be noted that when the above embodiments of the present application are applied to specific products or technologies, the relevant data of the user is involved, and the user's permission or consent is required, and the collection, use and processing of the relevant data need to comply with the relevant laws, regulations and standards of the relevant region.

Claims

1. A disturbance storm time index prediction method, characterized in that: The disturbance storm time index prediction method comprises: Obtain solar wind combined parameter series data and target index prediction model; Inputting the solar wind combined parameter sequence data into the encoder of the target index prediction model to obtain the second encoded seasonal component; Inputting the historical seasonal component sequence into the adaptive frequency enhancement block of the decoder of the target index prediction model to perform frequency domain enhancement to obtain first enhanced data; Superimposing the first enhanced data and the historical seasonal component sequence to obtain first superimposed data; decomposing the first stacked data using a hybrid expert decomposition block in a decoder of the target index prediction model to obtain a first decoded seasonal component and a first decoded trend component; The second encoded seasonal component and the first decoded seasonal component are input into the adaptive non-stationary frequency enhanced attention module of the decoder of the target index prediction model for frequency domain learning, and then fused with the first decoded seasonal component to obtain a first fused component; decomposing the first fused component using a hybrid expert decomposition block in a decoder of the target index prediction model to obtain a second decoded seasonal component and a second decoded trend component; After inputting the second decoded seasonal component into the feedforward neural network of the decoder of the target index prediction model, the second decoded seasonal component is fused with the second decoded seasonal component to obtain a second fused component; decomposing the second fused component using a hybrid expert decomposition block in a decoder of the target index prediction model to obtain a third decoded seasonal component and a third decoded trend component; Fusing the third decoded seasonal component, the historical trend component sequence, the first decoded trend component, the second decoded trend component, and the third decoded trend component to obtain a third fused component; The disturbance storm time index prediction value is determined based on the third fusion component.

2. The disturbance storm time index prediction method according to claim 1, characterized in that: The method of determining the disturbance storm time index prediction value based on the third fusion component includes: The third fusion component is input into the reversible instance normalization module of the decoder of the target index prediction model for normalization and denormalization in sequence to obtain the predicted value of the disturbance storm time index.

3. The disturbance storm time index prediction method according to claim 2, characterized in that: The solar wind combined parameter sequence data includes the total interplanetary magnetic field, the north-south component of the interplanetary magnetic field, the proton temperature, the proton density, the solar wind speed, the solar wind dynamic pressure and the electric field.

4. The disturbance storm time index prediction method according to claim 3, characterized in that: The step of inputting the solar wind combined parameter sequence data into an encoder of a target index prediction model to obtain a second encoded seasonal component comprises: Inputting the solar wind combined parameter sequence data into the reversible instance normalization module of the encoder of the target index prediction model for reversible instance normalization to obtain first standardized data; Inputting the first standardized data into an adaptive frequency enhancement block of an encoder of a target index prediction model for frequency domain enhancement to obtain second enhanced data; Superimposing the second enhanced data and the first standardized data to obtain second superimposed data; decomposing the second stacked data using a hybrid expert decomposition block in an encoder of the target index prediction model to obtain a first encoded seasonal component; After inputting the first coded seasonal component into the feedforward neural network of the encoder of the target index prediction model, the first coded seasonal component is fused with the first coded seasonal component to obtain a first coded fused component; The hybrid expert decomposition block in the decoder using the target index prediction model decomposes the first coded fusion component to obtain a second coded seasonal component.

5. The disturbance storm time index prediction method according to claim 4, characterized in that: The second encoded seasonal component and the first decoded seasonal component are input into the adaptive non-stationary frequency enhanced attention module of the decoder of the target index prediction model for frequency domain learning, and then fused with the first decoded seasonal component to obtain a first fused component, including: The two second encoded seasonal components are taken as K vector and V vector respectively, and the first decoded seasonal component is taken as Q vector. After input into the adaptive non-stationary frequency enhanced attention module of the decoder of the target index prediction model for frequency domain learning, they are fused with the first decoded seasonal component to obtain the first fused component.

6. The disturbance storm time index prediction method according to claim 5, characterized in that: The disturbance storm time index prediction method comprises: Acquire a historical data set, wherein the historical data set includes a plurality of solar wind combination parameter historical sequences and corresponding disturbance storm time index observation values ​​in a preset time period; Performing preprocessing operations on the historical data set to obtain a preprocessed data set; Split the preprocessed dataset into training set, validation set and test set according to the preset ratio; A target index prediction model is obtained based on training of the training set; Test the target index prediction model based on multiple key indicators and test sets.

7. A disturbance storm time index prediction device, characterized in that: The disturbance storm time index prediction device comprises: An acquisition module is used to obtain solar wind combined parameter sequence data and target index prediction model; An input module, used for inputting solar wind combined parameter sequence data into an encoder of a target index prediction model to obtain a second encoded seasonal component; An enhancement module, used for inputting the historical seasonal component sequence into an adaptive frequency enhancement block of a decoder of a target index prediction model for frequency domain enhancement to obtain first enhanced data; A superposition module, used for superimposing the first enhanced data and the historical seasonal component sequence to obtain first superimposed data; A first decomposition module, configured to decompose the first superimposed data using a hybrid expert decomposition block in a decoder of a target index prediction model to obtain a first decoded seasonal component and a first decoded trend component; A first fusion module is used to input the second encoded seasonal component and the first decoded seasonal component into an adaptive non-stationary frequency enhanced attention module of a decoder of a target index prediction model for frequency domain learning, and then fuse them with the first decoded seasonal component to obtain a first fused component; A second decomposition module, configured to decompose the first fusion component using a hybrid expert decomposition block in a decoder of the target index prediction model to obtain a second decoded seasonal component and a second decoded trend component; A second fusion module is used to input the second decoded seasonal component into a feedforward neural network of a decoder of a target index prediction model, and then fuse it with the second decoded seasonal component to obtain a second fused component; A third decomposition module, configured to decompose the second fusion component using a hybrid expert decomposition block in a decoder of the target index prediction model to obtain a third decoded seasonal component and a third decoded trend component; A third fusion module is used to fuse the third decoded seasonal component, the historical trend component sequence, the first decoded trend component, the second decoded trend component, and the third decoded trend component to obtain a third fusion component; A determination module is used to determine a disturbance storm time index prediction value based on the third fusion component.

8. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the steps in the disturbance storm time index prediction method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor to execute the steps in the disturbance storm time index prediction method according to any one of claims 1 to 6.

10. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, the steps in the disturbance storm time index prediction method according to any one of claims 1 to 6 are implemented.

Images