A method and device for predicting the duration of substorm expansion phase based on machine learning algorithm

Abstract

The application discloses a substorm expansion phase duration prediction method based on a machine learning algorithm, receives substorm event data to be detected, inputs a substorm expansion phase duration prediction model based on a machine learning algorithm which is established and trained in advance, and obtains a substorm expansion phase duration prediction result. The substorm expansion phase duration prediction model based on the machine learning algorithm comprises an LSTM layer, the output result of the LSTM layer is subjected to a convolution layer, the output result of the convolution layer is input into a full connection layer, and the substorm expansion phase duration prediction result is obtained through the full connection layer. The application also discloses a substorm expansion phase duration prediction device, which comprises a memory for receiving and storing data and programs and a processor for executing programs to realize the substorm expansion phase duration prediction method. After the model is trained, parameters do not need to be adjusted, and only corresponding time period data needs to be input into the model, so that the prediction result of the corresponding event time period can be output. The application is the first modeling and prediction of a key parameter of the entire substorm.

Description

Technical Field

[0001] This application relates to the field of research on the application of machine learning technology in space physics, specifically to a method and device for predicting the duration of substorm expansion based on machine learning algorithms. Background Technology

[0002] A geomagnetic substorm (hereinafter referred to as a substorm) is a geomagnetic disturbance event in the Earth's space environment characterized by a violent burst of energy. It primarily occurs in the tail of the magnetosphere (night side) near Earth and is a transient response of the Earth's magnetic field to the energy input from the solar wind. Substorms occur approximately 3 to 4 times per day. Besides causing drastic changes in the Earth's polar space environment, they also induce strong disturbances in the Earth's ionosphere, affecting GPS navigation signal reception and communication between satellites and Earth. Substorms are often accompanied by enhanced auroral phenomena. The region with the strongest and most varied auroral intensity on Earth is an elliptical band between 65° and 70° north and south latitude, which is connected to the region with the most intense magnetospheric substorm activity.

[0003] Substorms typically last 2-3 hours and are generally divided into an initiation phase, an expansion phase, and a recovery phase. The initiation phase is the initial state, characterized by a rapid intensification of localized auroral regions, generally lasting about 5-10 minutes. During the expansion phase, the area of ​​the main auroral oval region expands significantly, with both its equatorial and polar boundaries extending outwards, accompanied by an increase in auroral intensity. The duration of the expansion phase is usually variable, ranging from 10 to 100 minutes. During the recovery phase, as geomagnetic activity gradually subsides, the auroral intensity and area gradually return to normal.

[0004] Generally, the initiation and evolution of substorms are considered to be highly correlated with magnetopause reconnection activity, often occurring when the interplanetary magnetic field is southward. Solar wind energy is injected into Earth's magnetosphere relatively efficiently, easily triggering magnetospheric substorms. Other physical parameters, such as solar dynamic pressure, velocity, and interplanetary magnetic field strength, also influence the evolution of substorms. Substorms are an important process in Earth's magnetospheric activity, and observing and modeling them can help us understand the entire magnetospheric activity process.

[0005] In the prior art, for example, Chinese Patent Application Document 1 (CN117910656A), published on April 19, 2024, discloses "A method for short-term prediction of geomagnetic storms based on a hybrid deep learning model. First, Dst data information is acquired; then, linear interpolation is used to process outlier and missing data; next, by reviewing historical data, Dst and external input feature parameters are transformed into supervised sequence pairs; then, the dataset is partitioned and preprocessed with zero-mean standardization; then, a hybrid deep learning Dst prediction model consisting of a pre-trained network and a prediction network is built; the model is trained to obtain the optimal model structure; finally, the trained Dst prediction model is used to predict the output results of the test set, evaluate the model performance, and complete the short-term prediction of geomagnetic storms." The method of this invention utilizes historical Dst data information to construct a hybrid deep learning model composed of a pre-trained network and a prediction network, improving the accuracy of prediction. It employs Monte Carlo dropout technique to estimate the uncertainty of the model, providing a prediction interval, and achieving relatively accurate short-term prediction of geomagnetic storms, suitable for Dst prediction 1-4 hours in advance.

[0006] For example, Chinese patent application document 2, CN119919989A, published on May 2, 2025, discloses "a method and device for identifying auroral substorms based on visual eye-tracking patterns and deep learning, mainly addressing the problems of low accuracy and lack of basis for model design in existing auroral substorm identification methods. Its implementation scheme includes: collecting eye-tracking data from multiple space physicists to generate expert eye-tracking gaze pattern maps for various types of auroral substorms; establishing an auroral substorm eye-tracking dataset; further designing a visual eye-tracking pattern prediction module to learn the experts' visual gaze patterns; and setting up a substorm sequence identification module embedded with visual patterns, which incorporates space physics knowledge and expert visual gaze patterns. Finally, a deep learning model suitable for auroral substorm sequence identification is constructed and trained, and the trained model is used to identify auroral sequences. This invention not only enhances the representation of auroral substorm features but also improves the accuracy and rationality of auroral substorm image sequence identification, and can be used for automated space physics event identification."

[0007] The aforementioned patent document 1 primarily addresses the prediction of geomagnetic storms. In the field of space weather, geomagnetic storms and substorms have significantly different definitions. Geomagnetic storms are mainly detected by geomagnetic stations at the equator, detecting magnetic field fluctuations. Their primary cause is the loop current in the magnetic field, and they are mainly defined by the Dst index. The Dst index is typically defined as the average disturbance value over one hour; for prediction purposes, its timescale is on the order of hours, which aligns well with the evolution and timescale of geomagnetic storm events. Substorms, however, have a more complex mechanism, and there is considerable debate in domestic and international research. From a physical perspective, the energy released by a substorm is usually less than that of a geomagnetic storm. Some believe that multiple consecutive substorms are a factor in the generation of geomagnetic storms, but many others believe that substorms and geomagnetic storms are not directly related. The definition of a substorm also typically differs from that of a geomagnetic storm; it is mainly predicted by magnetic field fluctuations detected by geomagnetic stations at high latitudes, and it has similar indices such as AP, AE, and AL. These indices, similar to the Dst index, are also detected and defined on an hourly basis. However, substorms tend to be shorter in duration than geomagnetic storms, typically lasting only 30-120 minutes, while the entire cycle of a geomagnetic storm can range from two days to a week. Furthermore, substorms occur much more frequently than geomagnetic storms, with an average of over 20 events per week, while the frequency of geomagnetic storms is generally calculated on a monthly basis. Therefore, from a predictive perspective, substorm prediction is a short-term, near-term prediction, and its requirements are quite different from those for geomagnetic storms. Another characteristic of substorms is the change in their auroral region. As substorms evolve, the auroral region shows a significant increase in auroral intensity, and the auroral ovum expands towards its equatorial and polar boundaries. Therefore, auroral intensity is also a good way to describe the evolution of substorms.

[0008] Patent document 2 primarily relies on expert experience to identify and label aurora images during substorms, determining whether an aurora observation image represents an aurora during a substorm. It mainly targets the field of image recognition, providing a foundation for scientists to extract and screen aurora images related to substorms. Previous research indicates that predictions related to geomagnetic storms and substorms mainly focus on predicting geomagnetic indices, such as Dst, AE, AP, and AL, which are defined by the average value of one-hour geomagnetic disturbances. Patent document 2 can only predict whether an event will occur, not the duration of the event itself.

[0009] Currently, there is no existing technology to predict the duration of substorms. Summary of the Invention

[0010] The purpose of this application is to solve the important problem of predicting the duration of the expansion phase in substorm modeling, and to construct a time-series prediction method based on deep learning technology to predict the duration of the expansion phase of a specific substorm event from observational data of physical parameters of the solar and interplanetary magnetic fields, thereby meeting the modeling needs of researchers for substorm event research.

[0011] To achieve the above objectives, the technical solution of this application provides a method for predicting the duration of a substorm expansion period based on a machine learning algorithm, comprising: receiving data of a substorm event to be detected, inputting a pre-established and trained substorm expansion period duration prediction model based on a machine learning algorithm, and obtaining a substorm expansion period duration prediction result; The substorm expansion duration prediction model based on machine learning algorithms includes: an LSTM layer, the output of which is passed through a convolutional layer, the output of which is input to a fully connected layer, and the predicted substorm expansion duration is obtained through the fully connected layer.

[0012] As an improvement to the above method, receiving the substorm event data to be detected includes: using solar wind and interplanetary magnetic field data as input data for the prediction model to establish a correspondence between the input data and the duration of the substorm expansion period; and constructing label data for the duration of the substorm expansion period.

[0013] As another improvement to the above method, the construction of the labeled data includes: using extreme ultraviolet observation data, screening out manually labeled substorm events and corresponding aurora image sequences during the expansion period, and then, based on the total intensity of the aurora, finding the time period during which the total intensity of the aurora continues to increase as the expansion period time of the substorm, and labeling the duration of the expansion period.

[0014] As another improvement to the above method, it also includes preprocessing the received substorm event data to be detected, including outlier handling, missing value filling, and total aurora intensity calculation.

[0015] As an improvement to the above method, training a substorm expansion duration prediction model based on machine learning algorithms includes the following steps: Construct a dataset by extracting solar wind and interplanetary magnetic field observation data that begin S minutes before the start time of the list of observed substorm events and last for T minutes, and constructing it as T. A matrix tensor of K, containing T minutes of time-series data and K physical parameters of observational data of solar wind and interplanetary magnetic fields; the K physical parameters include the average magnetic field strength, the components of the magnetic field in the x, y, and z directions in the Earth's geocentric inertial frame and the Earth's geocentric solar magnetosphere, the root mean square standard deviation of the magnetic field strength scalar, the root mean square standard deviation of the magnetic field vector, the velocity of the solar wind, the velocity components of the solar wind in the x, y, and z directions, the number density of protons in the solar wind, the temperature of protons in the solar wind, the dynamic pressure of the solar wind, the electric field strength in the solar wind, plasma parameters, Alfvén Mach number, and magnetoacoustic Mach number; the corresponding observational data are mapped to the interval [0,1] to preserve the original distribution characteristics of the data; The matrix tensors of the dataset are fed into the LSTM layer for computation; The output of the LSTM layer is extracted and then convolutional operations are performed using a convolutional layer. The output of the convolutional layer is fed into the fully connected layer for linear transformation, which is then mapped to the classification result to obtain the predicted duration of the sub-burst inflation period.

[0016] As a further improvement to the above method, outliers and missing values ​​in the data are handled by linear interpolation.

[0017] As a further improvement to the above method, an evaluation step for the prediction model is also included: using the labeled dataset sample data as input to the substorm expansion period duration prediction model, adjusting the parameters to train the duration prediction model, obtaining the model extraction results and evaluating them; the evaluation is based on the error between the predicted expansion period duration and the actual expansion period duration, as well as the average error. The evaluation step of the prediction model further includes: dividing the dataset into N equal parts with the same amount of data; using one different part as the test set for each training and model evaluation, and using the other N-1 parts as the training set; repeating the training and testing evaluation N times to ensure that each part of the dataset has been tested; and using the average result of the N tests as the result of cross-validation.

[0018] To achieve the above objectives, the technical solution of this application also provides a substorm expansion period duration prediction device based on machine learning algorithm, comprising: a dataset unit for receiving and storing substorm event data to be detected; A substorm expansion duration prediction model based on machine learning algorithms is used to receive substorm event data to be detected after training, and obtain the substorm expansion duration prediction results. The substorm expansion duration prediction model based on machine learning algorithm includes: an LSTM layer, the output of which is passed through a convolutional layer, the output of which is input to a fully connected layer, and the predicted substorm expansion duration is obtained through the fully connected layer; and the substorm expansion duration prediction model based on machine learning algorithm is trained according to the above training method.

[0019] To achieve the above objectives, this application also provides a device for predicting the duration of a substorm expansion phase based on a machine learning algorithm, characterized in that it includes: Memory, used to receive and store data and programs; A processor is used to execute the program to implement the above-described method for predicting the duration of subburst expansion based on machine learning algorithms.

[0020] For the purposes described above, this application also provides a computer-readable storage medium, characterized in that it includes a computer program that can be executed by a processor to implement the substorm expansion period duration prediction method described above.

[0021] The technical advantages and effects of this application are as follows: Based on the typical characteristics of magnetospheric substorm events, this invention proposes a substorm expansion period duration prediction model using a Long Short-Term Memory (LSTM) deep learning network model for time-series prediction. This model implements a bottom-to-top-level inference process and ultimately predicts the duration of the substorm expansion period. The proposed solution utilizes an end-to-end approach; after the model is trained, no further parameter adjustments are required. Simply inputting the corresponding time period data into the model directly outputs the prediction results for the corresponding event time period. This is the first time that key parameters of the entire substorm have been modeled and predicted. Attached Figure Description

[0022] Figure 1 A flowchart illustrating the substorm expansion period duration prediction method in a specific implementation; Figure 2 A list of substorm events for specific implementation methods; Figure 3 Examples of slices for OMNI's solar wind and interplanetary magnetic field data; Figure 4 This is a schematic diagram of an image sequence during the substorm expansion phase, as described in a specific implementation method. Figure 5 This is a schematic diagram of a substorm expansion period duration prediction model network based on machine learning algorithms, illustrating a specific implementation method. Detailed Implementation

[0023] The technical solutions provided in this application are further illustrated below with reference to the embodiments.

[0024] This invention aims to predict the duration of a substorm event. While substorms and geomagnetic storms can be predicted by forecasting the geomagnetic index, the time precision of the geomagnetic index is limited. Furthermore, the occurrence of geomagnetic storms and substorms is significantly influenced by the solar wind and interplanetary magnetic field, involving magnetopause reconnection and instability. Therefore, parameters such as the solar wind and interplanetary magnetic field are also applicable to substorm prediction. Higher-precision observational data for solar wind and other physical parameters exist; satellites detecting the solar wind and interplanetary magnetic field upstream of the magnetopause can achieve a time precision of one minute. Therefore, this invention utilizes solar wind parameters to predict the duration of the substorm expansion phase, using the total auroral intensity as the labeling factor. As mentioned earlier, auroral images are also an important descriptive feature of substorms; therefore, this invention focuses on predicting and modeling the duration of substorms (expansion phase duration).

[0025] Variations in auroral intensity are a typical characteristic of substorms; therefore, this invention uses auroral extreme ultraviolet (EUVI) observations from the Polar satellite as a description of the substorm state. (Polar satellite is one of the four spacecraft in the Global Geospace Science (GGS) program and one of the six spacecraft in the International Solar Geophysics (ISTP) program. POLAR provides multi-wavelength imaging of the aurora, measures plasma entry into the polar magnetosphere and geomagnetic tail, plasma flow into and out of the ionosphere, and particle energy deposition in the ionosphere and upper atmosphere.) The UVI optical sensor operates in four filter bandpasses: two atomic oxygen lines centered at 130.4 and 135.6 nm, and two molecular nitrogen Lyman-Birge-Hopfield bands centered at approximately 150 nm (LBHS, Lyman-Birge-Hopfield Short, i.e., around the first 150 nm band) and 170 nm (LBHL, Lyman-Birge-Hopfield Long, referring to the first 170 nm band). With an 8° circular field of view, Polar / UVI provides a typical spatial resolution of approximately 30–40 km, and the observed aurora radiation is assumed to be at an altitude of 120 km.

[0026] In "Liou, K. (2010). Polar ultraviolet imager observation of auroralbreakup. Journal of Geophysical Research, 115(A12), A12219. https: / / doi.org / 10.1029 / 2010ja015578" and "Frey, HU, SB Mende, V. Angelopoulos, and EF Donovan (2004), Substorm onset observations by IMAGE-FUV, J. Geophys. Res. The article ", 109, A10304, doi:10.1029 / 2004JA010607" lists over 2000 substorm initiation events observed using Polar / UVI and determined manually. Therefore, this invention utilizes the substorm initiation events calibrated by Liou and Frey, and further calibrates the duration of the expansion phase for each substorm event by observing changes in UVI aurora intensity. Combined with observational parameters of the solar and interplanetary magnetic fields, a time-series prediction algorithm based on machine learning is used, trained with labels, to construct a substorm expansion phase duration prediction model. This algorithm is mainly used to predict the duration of the expansion phase of specific substorm events, improving the ability to model substorm observation data and promoting statistical analysis research on big data related to substorm events.

[0027] In order to achieve the above-mentioned objectives of the invention, such as Figure 1 As shown, the technical solution of this embodiment mainly includes the following steps: 1. Raw data collection: Select raw observational data of the solar wind and interplanetary magnetic field from the OMNI website (https: / / omniweb.gsfc.nasa.gov / form / omni_min.html) or other sources, along with Polar / UVI observational images, selecting image sequences during the substorm according to the list of substorm initiation events.

[0028] 2. Data preprocessing (data cleaning): This includes data preprocessing such as outlier handling, missing value imputation, and total aurora intensity calculation; 3. Dataset Construction: By manually selecting samples containing substorm events, and labeling the duration of their expansion phase as data based on aurora observation image sequences during the period, we constructed the labeled dataset required for training the substorm expansion phase duration prediction model by further labeling solar wind and interplanetary magnetic field data slices prior to the substorm onset time. 4. Model Construction: like Figure 5 As shown, the sub-burst inflation period duration prediction model structure includes: an LSTM (Long Short-Term Memory) prediction network model layer, followed by a convolutional layer for feature extraction, and then a fully connected layer for duration prediction.

[0029] To verify the effectiveness of the above model in predicting the duration of the sub-burst expansion phase, a combination of LSTM layers, convolutional layers, and fully connected layers was chosen as the superior network model structure. After trying different network structures such as CNN, this model structure performed best on the test set.

[0030] 5. Model Evaluation: Cross-validation is used to evaluate model performance. Specifically, the dataset of substorm events is divided into N equal parts (e.g., 5 parts). For each training and model evaluation, a different part is used as the test set, and the other N-1 parts (e.g., 4 parts) are used as the training set. This training and testing process is repeated N times, ensuring that each part of the dataset has been used as the test set. The average of these N iterations is used as the cross-validation result. This cross-validation method minimizes random errors introduced by selecting the test and training sets.

[0031] The sample data from the labeled solar wind and interplanetary magnetic field datasets were used as input to the selected substorm expansion duration prediction model. The parameters were adjusted to train the prediction model, and the model prediction results were evaluated by comparing them with the previously labeled substorm expansion duration labels.

[0032] The following further elaborates on the specific steps of the substorm expansion period duration prediction method based on machine learning algorithm provided in the specific embodiments of the present invention. The method includes: Step 1 is data collection; First, based on literature and the required Polar satellite extreme ultraviolet auroral (UVI) image data, manually labeled substorm events from 1996 to 2000, along with corresponding substorm auroral image sequence files, were selected. Auroral image sequences are typically stored in PNG format with a time resolution of 3 minutes. Next, based on the substorm event list, solar wind and interplanetary magnetic field observation data during the substorm events were selected. Using magnetopause location data provided by the OMNI website, observation files with a time resolution of 1 minute were chosen. The file header mainly records the corresponding observation auxiliary parameters and coordinate system characteristics, and details the physical meaning and storage format of each observed parameter.

[0033] Step 2 is data preprocessing; This includes handling outliers and filling in missing values. Because the raw data may contain outliers at certain times or be missing for short periods during instrument operation, it is necessary to process these outliers and missing values ​​using methods such as linear interpolation. Temporal resampling ensures the validity of solar wind, interplanetary magnetic field, and aurora image data.

[0034] Step 3 is dataset construction; Using solar wind and interplanetary magnetic field data as inputs to the prediction model, a correspondence between the inputs and the duration of the substorm expansion period is established. The label data for the duration of the substorm expansion period is labeled according to the following method: using extreme ultraviolet observation data, manually labeled substorm events and corresponding aurora image sequences during the expansion period are selected. Then, based on the total aurora intensity, the time period during which the total aurora intensity continues to increase is identified as the expansion period of the substorm, and the duration of the expansion period is labeled.

[0035] Specifically, by collecting publicly published papers and based on the annotation of substorm events by Liou et al. and manual correction, this invention initially screened over 2000 substorm events from 1996 to 2000, and annotated the characteristics required for this study for each event accordingly. The start and end times, aurora intensity changes, and duration of each substorm were mainly recorded as event tags. Based on the corresponding substorm start time, data on the solar and interplanetary magnetic fields provided by the OMNI website were used to slice the corresponding time segments, obtaining the corresponding substorm time data segments and corresponding tags. Here, based on the correlations in the papers, this invention specifically selected data from a time period of S (e.g., 7) minutes before the substorm start to a total of T (e.g., 30) minutes after the substorm start as the data slice for a single substorm event.

[0036] The process of constructing the training dataset is as follows: Based on the start time (onset) of the list of observed substorm events, solar wind and interplanetary magnetic field observation data starting 7 minutes before (e.g., 7 minutes before) and lasting for 30 minutes are extracted, including K (e.g., 19) physical parameters. This data is constructed as a 30*19 matrix tensor, which contains 30 minutes of time-series data and 19 solar wind and interplanetary magnetic field parameters. Specifically, the 19 physical parameters are: average magnetic field strength, components of the magnetic field in the x, y, and z directions in the Earth's geocentric inertial frame (GSE) and the Earth's geocentric heliomagnetic system (GSM), root mean square standard deviation of the magnetic field strength scalar, root mean square standard deviation of the magnetic field vector, solar wind velocity, velocity components of the solar wind in the x, y, and z directions, proton number density in the solar wind, temperature of protons in the solar wind, dynamic pressure of the solar wind, electric field strength in the solar wind, plasma parameters, Alfvén Mach number, and magnetoacoustic Mach number. The corresponding observation data are mapped to the interval [0,1] to preserve the original distribution characteristics of the data.

[0037] Step 4 is model construction and testing; Considering the characteristics of solar wind and interplanetary magnetic field time-series observation data and the mission's objective of predicting the duration of the substorm expansion phase, the LSTM network, which excels in multi-scale feature fusion in commonly used prediction tasks, is adopted as the network model for predicting the duration of the substorm expansion phase. This model improves the performance of the deep neural network by incorporating long-term features from each layer and integrating short-term memory spans. Finally, the output layer of the LSTM network is passed through a convolutional layer to extract its feature information, and then passed through a fully connected network to predict the output result. Simultaneously, this network model is experimentally compared with other commonly used one-dimensional feature network models to verify the advantages of the LSTM model using the training dataset and labeled data of this invention in predicting the duration of the substorm expansion phase.

[0038] Step 5 is model evaluation. The labeled dataset samples were used as input to the selected substorm inflation duration prediction model. Parameters were adjusted (e.g., trying different batch sizes and learning rates for each training iteration to evaluate the model's training performance under different parameters) to train the duration prediction model. The model's extraction results were then evaluated. Here, real-world results were used for evaluation. For each substorm event, the inflation duration labeled using aurora intensity changes was compared to the predicted inflation duration, and the error and average error were calculated. For example, if the inflation duration of a substorm event was 30 minutes, and the predicted inflation duration was 33 minutes, then the error for a single event was 3 minutes, or 10%. Finally, the average error of the test set data was calculated.

[0039] To better illustrate the advantages and implementation of this technical solution, the specific execution flow of this particular embodiment is described below.

[0040] First, based on literature and the required Polar satellite extreme ultraviolet (UVI) observation data, manually labeled substorm events from 1996 to 2000, along with corresponding aurora image sequences during their expansion periods, were selected. To construct a duration prediction method, based on the physical laws and data characteristics of substorm phenomena, OMNI solar wind and interplanetary magnetic field data were used as inputs to the prediction model, establishing a correspondence between the inputs and the duration of substorm expansion periods. Further data preprocessing is required, primarily including missing data point completion and outlier correction.

[0041] Figure 2 It shows the start time, start and end locations, latitude and longitude, etc. of each substorm event.

[0042] Where Date_UTC represents: Date represents the date, UTC represents Coordinated Universal Time, MLT represents Magnetic Local Time, MLAT represents Magnetic Latitude (all are geomagnetic coordinate systems), GLON represents Geographic Longitude, and GLAT represents Geographic Latitude.

[0043] Based on the corresponding substorm event time, this invention uses data from solar wind and interplanetary magnetic fields to slice the corresponding time segments to obtain the corresponding substorm event data segments.

[0044] Figure 3 This is a sample of solar wind and interplanetary magnetic field data slices created based on the substorm onset time. This data serves as the input to the model. The time precision of this data is 1 minute. Nineteen physical parameters from this data are used as input to the model, and a 30*19 matrix tensor is created for each sample. The specific physical meaning and format of each column are as follows: 1 Field magnitude average, nT F8.2 2 BX, nT (GSE, GSM) F8.2 3 BY, nT (GSE) F8.2 4 BZ, nT (GSE) F8.2 5 BY, nT (GSM) F8.2 6 BZ, nT (GSM) F8.2 7 RMS SD B scalar, nT F8.2 8 RMS SD B field vector, nT F8.2 9 Speed, km / s F8.1 10 Vx Velocity,km / s F8.1 11 Vy Velocity, km / s F8.1 12 Vz Velocity, km / s F8.1 13 Proton Density, n / cc F8.2 14 Proton Temperature, K F8.2 15 Flow pressure, nPa F8.2 16 Electric field, mV / m F8.2 17 Plasma beta F5.2 18 Alfven mach number F5.2 19 Magnetosonic Mach number F5.2 The English meanings are as follows: (Field magnitude average, nT) refers to the average magnetic field strength; BX, nT (GSE, GSM) refers to the x-component magnetic field strength; BY, nT (GSE) refers to the y-component magnetic field strength in GSE coordinates; BZ, nT (GSE) refers to the z-component magnetic field strength in GSE coordinates; BY, nT (GSM) refers to the y-component magnetic field strength in GSM coordinates; BZ, nT (GSM) refers to the z-component magnetic field strength in GSM coordinates; RMS SD B scalar, nT refers to the root mean square standard deviation of the magnetic field strength scalar; RMS SD B field vector, nT refers to the root mean square standard deviation of the magnetic field strength vector. All magnetic field units are nT. Speed, km / s refers to the average velocity of the solar wind, in km / s; Vx Velocity, km / s refers to the x-component of the solar wind velocity, in km / s; Vy Velocity, km / s refers to the y-component of the solar wind velocity, in km / s; Vz Velocity, km / s refers to the z-component of the solar wind velocity, in km / s; Proton Density, n / cc refers to the solar proton density, in protons per cubic centimeter; Proton Temperature refers to the proton temperature, in K; Flow pressure refers to the solar wind dynamic pressure, in nPa; Electric field refers to the solar wind electric field intensity, in mV / m; Plasma beta refers to the plasma parameter, in units of 1; Alfven Mach number refers to the Alfven Mach number (the ratio of solar wind velocity to Alfven velocity), in units of 1; Magnetosonic Mach number refers to the magnetoacoustic Mach number (the ratio of solar wind velocity to magnetosonic velocity), in units of 1. Format refers to the storage format of each parameter. Finally, the data is mapped to the [0,1] interval, preserving the original distribution characteristics of the data.

[0045] After constructing the dataset of input parameters, corresponding labels also need to be built. For a substorm event, based on the aurora observation image sequence of Polar / UVI, the time period during which the total aurora intensity continuously increases is identified as the inflation period of the substorm, and the duration of the inflation period is further labeled.

[0046] like Figure 4 The image shown is a typical sequence of aurora images observed during a substorm's expansion. If the sequence consists of seven images, the expansion period of this event would be approximately 21 minutes. The duration of each event is used as a label to prepare for the construction of a subsequent duration prediction model. Figure 4This is a sequence of aurora observation images for a substorm event, where the numbers represent time, such as 19961215-025811a, where 19961215 represents December 15, 1996, and 025811 represents 02:58:11 UTC. like Figure 5 The diagram shows the network model structure of the substorm expansion period duration prediction method based on machine learning algorithms, as described in a specific embodiment of this invention. The model is then trained, using solar wind and interplanetary magnetic fields as input conditions for prediction. Based on the start time (onset) of the observed substorm event list, solar wind and interplanetary magnetic field observation data from 42 minutes to 7 minutes prior to that onset time are extracted, including a total of 19 corresponding physical parameters. This data is constructed as a 30*19 matrix tensor (30 minutes of time-series data, 19 solar wind and interplanetary magnetic field parameters). After constructing the sample to fit the model input format, the input data first passes through an LSTM layer, then the output of the LSTM layer passes through a convolutional layer, the convolutional layer is flattened to input the fully connected layer, and finally, the fully connected layer yields the predicted output. This invention creates the input dataset and labels based on the model structure, thereby achieving the technical objective of predicting the duration of the substorm expansion period.

[0047] Finally, after training the model, its metrics were validated. Based on the prepared dataset, one-third of the data from each year of events between 1996 and 2000, totaling 721 substorm events, were selected as the test set, with the remaining data serving as the training set. Testing showed that the average error duration of the substorm duration prediction model was approximately 5.6 minutes, with an error ratio of approximately 18.3%. For the task of predicting the duration of the substorm expansion period, the model's output results are within a very reliable range. After the model is trained, no further parameter adjustments are needed; simply inputting the corresponding time period data into the model directly outputs the prediction results for the corresponding event time period. This invention is the first to model and predict the key parameters of the entire substorm.

[0048] From the above description of specific embodiments of the present invention, it is evident that the advantages of the present invention are as follows: First, the technical solution of this invention introduces deep learning technology for the first time to mine information from solar wind and interplanetary magnetic field data and the evolution of substorms, realizes a semantic reasoning process from the bottom to the top, and predicts and models the duration of the substorm expansion period.

[0049] Finally, the model is robust. After training, no parameters need to be adjusted. Simply input the data for the corresponding time period into the model, and it can directly output the duration prediction results of the corresponding substorm event, which plays a key role in the modeling of the entire substorm.

[0050] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the duration of substorm expansion phases based on machine learning algorithms, characterized in that, include: Receive the data of the substorm event to be detected, input it into the pre-established and trained substorm expansion period duration prediction model based on machine learning algorithm, and obtain the substorm expansion period duration prediction result; The substorm inflation duration prediction model based on machine learning algorithm includes: an LSTM layer, the output of the LSTM layer is passed through a convolutional layer, the output of the convolutional layer is input to a fully connected layer, and the predicted substorm inflation duration is obtained through the fully connected layer. The process of receiving substorm event data to be detected includes: using solar wind and interplanetary magnetic field data as input data for the prediction model, establishing a correspondence between the input data and the duration of the substorm expansion period; and constructing label data for the duration of the substorm expansion period. The construction of the labeled data includes: using extreme ultraviolet observation data, screening out manually labeled substorm events and corresponding aurora image sequences during the expansion period, and then, based on the total intensity of the aurora, finding the time period during which the total intensity of the aurora continues to increase as the expansion period of the substorm, and labeling the duration of the expansion period; The training of a substorm expansion duration prediction model based on machine learning algorithms includes the following steps: The dataset was constructed by extracting solar wind and interplanetary magnetic field observation data for 30 minutes, starting 7 minutes before the start time of the list of observed substorm events, based on the start time of that start time. A matrix tensor of K is constructed, containing 30 minutes of time-series data and K physical parameters of observational data on the solar wind and interplanetary magnetic field. These K physical parameters include: the average magnetic field strength, the components of the magnetic field in the x, y, and z directions in the Earth's geocentric inertial frame and the Earth's geocentric solar magnetosphere, the root mean square standard deviation of the magnetic field strength scalar, the root mean square standard deviation of the magnetic field vector, the velocity of the solar wind, the velocity components of the solar wind in the x, y, and z directions, the number density of protons in the solar wind, the temperature of protons in the solar wind, the dynamic pressure of the solar wind, the electric field strength in the solar wind, plasma parameters, the Alfvén Mach number, and the magnetoacoustic Mach number. The corresponding observational data are mapped to the interval [0,1] to preserve the original distribution characteristics of the data. The matrix tensors of the dataset are fed into the LSTM layer for computation; The output of the LSTM layer is extracted and then convolutional operations are performed using a convolutional layer. The output of the convolutional layer is fed into the fully connected layer for linear transformation, which is then mapped to the classification result to obtain the predicted duration of the sub-burst inflation period.

2. The method for predicting the duration of substorm expansion based on machine learning algorithms according to claim 1, characterized in that, It also includes preprocessing the received substorm event data to be detected, including outlier handling, missing value filling, and total aurora intensity calculation.

3. The method for predicting the duration of substorm expansion phase based on machine learning algorithms according to claim 2, characterized in that, Outliers and missing values ​​in the data are handled using linear interpolation.

4. The method for predicting the duration of substorm expansion phase based on machine learning algorithms according to claim 1, characterized in that, It also includes a prediction model evaluation step: using labeled dataset sample data as input to the substorm expansion period duration prediction model, adjusting parameters to train the duration prediction model, obtaining the model extraction results and evaluating them; the evaluation is based on the error between the predicted expansion period duration and the actual expansion period duration, as well as the average error. The evaluation step of the prediction model further includes: dividing the dataset into N equal parts with the same amount of data; using one different part as the test set for each training and model evaluation, and using the other N-1 parts as the training set; repeating the training and testing evaluation N times to ensure that each part of the dataset has been tested; and using the average result of the N tests as the result of cross-validation.

5. A device for predicting the duration of a substorm expansion phase based on a machine learning algorithm, characterized in that, include: Data set unit, used to receive and store data on substorm events to be detected; A substorm expansion duration prediction model based on machine learning algorithms is used to receive substorm event data to be detected after training, and obtain the substorm expansion duration prediction results. The substorm inflation duration prediction model based on machine learning algorithm includes: an LSTM layer, the output of the LSTM layer is passed through a convolutional layer, the output of the convolutional layer is input to a fully connected layer, and the predicted substorm inflation duration is obtained through the fully connected layer. The substorm expansion duration prediction model based on machine learning algorithm is trained according to the method of claim 1.

6. A device for predicting the duration of a substorm expansion phase based on a machine learning algorithm, characterized in that, include: Memory, used to receive and store data and computer programs; A processor is configured to execute the computer program to implement the substorm expansion duration prediction method based on machine learning algorithm as described in claim 1.

7. A computer-readable storage medium, characterized in that, The method includes a computer program that can be executed by a processor to implement the substorm expansion period duration prediction method as described in claim 1.

Images