Modeling method and device for global ionospheric foF2, and detection equipment

Abstract

This application relates to the field of ionospheric environment prediction technology, and discloses a modeling method for global ionospheric foF2, comprising: constructing a sample dataset based on acquired raw satellite occultation observation data, raw space environment data, and raw space environment observation data; wherein the sample dataset includes a training set, a test set, and a validation set; performing five-fold cross-validation on the XG regression framework using a Bayesian optimized stochastic hyperparameter optimization method to select the optimal hyperparameter combination to complete the construction and training of the XGBoost prediction model; constructing a lightweight XGBoost model based on the XGBoost prediction model and machine learning model interpretation methods; validating the performance of the XGBoost model using the validation set or test set; and constructing a global ionospheric foF2 prediction system based on the performance validation results. This method can achieve high-precision global prediction of foF2, balancing model prediction accuracy and interpretability. This application also discloses a modeling device and detection equipment for global ionospheric foF2.

Description

Technical Field

[0001] This application relates to the field of ionospheric environment forecasting technology, for example to a modeling method and apparatus, and detection equipment for global ionospheric foF2. Background Technology

[0002] Currently, the critical frequency of the F2 layer of the ionosphere (foF2) is a key parameter determining the maximum usable frequency for shortwave communication, and its accurate prediction is of great significance to both defense and civilian applications. Existing foF2 prediction methods are mainly divided into three categories:

[0003] 1. Empirical / climatological models: such as the International Reference Ionosphere (IRI) model and the Klobuchar model. These models are computationally efficient, but their accuracy is insufficient under extreme space weather conditions, and they are difficult to capture detailed regional features such as equatorial anomalies and winter anomalies.

[0004] 2. Physical models: such as TIE-GCM. These models are based on first principles, have high computational complexity, are sensitive to initial conditions, and are difficult to meet the needs of real-time applications.

[0005] 3. Artificial intelligence models: such as recurrent neural networks (RNN) and long short-term memory networks (LSTM). Although these models improve prediction accuracy, their "black box" nature makes the decision-making process difficult to trace and the contribution of each driving factor impossible to quantify, becoming a bottleneck restricting their application in critical tasks.

[0006] The XGBoost algorithm has shown potential in ionospheric parameter prediction due to its efficiency, built-in regularization, and good feature selection capabilities. Meanwhile, the game-theory-based SHAP (SHapley Additive exPlanations) method can provide a unified interpretive framework for complex models. However, there is currently a lack of a foF2 modeling method that combines high-precision prediction with physical interpretability and has been rigorously validated on a global scale using multi-source data. Summary of the Invention

[0007] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments, but rather as a prelude to the detailed description that follows.

[0008] This disclosure provides a modeling method, apparatus, and detection device for global ionospheric foF2, to achieve high-precision global prediction of foF2, while balancing model prediction accuracy and interpretability.

[0009] In some embodiments, the modeling method includes: constructing a sample dataset based on the acquired raw satellite occultation observation data, raw space environment data, and raw space environment observation data; wherein the sample dataset includes a training set, a test set, and a validation set; performing five-fold cross-validation on the XG regression framework using a Bayesian-optimized stochastic hyperparameter optimization method to select the optimal hyperparameter combination to complete the construction and training of the XGBoost prediction model; constructing a lightweight XGBoost model based on the XGBoost prediction model and machine learning model interpretation methods; validating the performance of the XGBoost model using the validation set or test set, and constructing a global ionospheric foF2 prediction system based on the performance validation results; wherein the space environment observation data is obtained by cleaning the raw space environment observation data.

[0010] In some embodiments, the apparatus includes a processor and a memory storing program instructions, the processor being configured to, when executing the program instructions, perform the modeling method for global ionosphere foF2 as described above.

[0011] In some embodiments, the detection device includes: a detection device body; and a modeling apparatus for global ionosphere foF2 as described above, installed on the detection device body.

[0012] The modeling method, apparatus, and detection device for global ionosphere foF2 provided in this disclosure can achieve the following technical effects:

[0013] This disclosure integrates machine learning model interpretation methods with XGBoost prediction models and applies them to global ionospheric foF2 modeling to achieve a high degree of consistency between the interpretation results and ionospheric physical mechanisms, providing a quantitative tool for ionospheric physical mechanism research. Simultaneously, feature lightweight optimization improves the model's real-time performance and engineering adaptability. Furthermore, this disclosure verifies the XGBoost model's performance using validation or test sets and constructs a global ionospheric foF2 prediction system based on the performance verification results to support end-to-end prediction and interpretable output. Based on this, high-precision global foF2 prediction is achieved, quantifying the driving role of each feature in the prediction results, balancing model prediction accuracy and interpretability.

[0014] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description

[0015] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations and drawings do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are shown as similar elements. The drawings are not to be scaled. And wherein:

[0016] Figure 1 This is a schematic diagram of a modeling method for the global ionosphere foF2 provided in an embodiment of this disclosure;

[0017] Figure 2 This is a schematic diagram of another modeling method for the global ionosphere foF2 provided in this disclosure embodiment;

[0018] Figure 3 This is a schematic diagram of another modeling method for the global ionosphere foF2 provided in this disclosure embodiment;

[0019] Figure 4 This is a schematic diagram of another modeling method for the global ionosphere foF2 provided in this disclosure embodiment;

[0020] Figure 5 This is a schematic diagram of another modeling method for the global ionosphere foF2 provided in this disclosure embodiment;

[0021] Figure 6 This is a schematic diagram illustrating the relationship between solar activity and ionospheric parameters;

[0022] Figure 7 This is a schematic diagram of the framework of the lightweight XGBoost model;

[0023] Figure 8 This is a schematic diagram of the feature contribution analysis results obtained by ranking the feature contributions of the XGBoost prediction model according to the feature contribution index.

[0024] Figure 9 This is a scatter plot of the prediction density of the XGBoost prediction model;

[0025] Figure 10 This is a schematic diagram showing the median absolute error of the lightweight XGBoost model prediction and the RF model prediction at different local times and in different seasons.

[0026] Figure 11 This is a scatter plot of the prediction density of the lightweight XGBoost model, the IRI-2020 model, and the NPDM model.

[0027] Figure 12 This is a schematic diagram of a modeling apparatus for the global ionosphere foF2 provided in an embodiment of this disclosure. Detailed Implementation

[0028] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0029] Unless otherwise stated, the term "multiple" means two or more.

[0030] In this embodiment of the disclosure, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B. The term "correspondence" can refer to an association or binding relationship; A corresponding to B means that there is an association or binding relationship between A and B.

[0031] Combination Figure 1 As shown, this disclosure provides a modeling method for global ionosphere foF2, including:

[0032] S01, the detection equipment constructs a sample dataset based on the acquired raw satellite occultation observation data, raw space environment data, and raw space environment observation data. The sample dataset includes a training set, a test set, and a validation set.

[0033] S02, the detection device performs five-fold cross-validation on the XG regression framework using the Bayesian-optimized stochastic hyperparameter optimization method to select the optimal hyperparameter combination to complete the construction and training of the XGBoost prediction model.

[0034] In this step, five-fold cross-validation is used for model performance evaluation. Five-fold cross-validation involves dividing the entire sample dataset into five equal subsets, selecting one subset as the test set, and using the remaining four subsets as the training set to perform five iterations of training and testing. The average score is then calculated based on the results of the five experiments. This objectively assesses the model's generalization ability.

[0035] S03, the detection device constructs a lightweight XGBoost model based on the XGBoost prediction model and the machine learning model interpretation method.

[0036] S04, the detection equipment verifies the performance of the XGBoost model using a validation or test set, and constructs a global ionospheric foF2 prediction system based on the performance verification results. The space environment observation data is obtained by cleaning the raw space environment observation data.

[0037] The modeling method for global ionospheric foF2 provided in this disclosure first constructs a sample dataset based on the acquired raw satellite occultation observation data and raw space environment data. Then, a five-fold cross-validation of the XGBoost regression framework is performed using a Bayesian-optimized stochastic hyperparameter optimization method to select the optimal hyperparameter combination for constructing and training the XGBoost prediction model. Next, a lightweight XGBoost model is constructed based on the XGBoost prediction model and machine learning model interpretation methods. Finally, the XGBoost model's performance is validated using a validation or test set, and a global ionospheric foF2 prediction model is constructed based on the performance validation results. This disclosure integrates machine learning model interpretation methods with the XGBoost prediction model and applies them to global ionospheric foF2 modeling to achieve a high degree of consistency between the interpretation results and the ionospheric physical mechanisms, providing a quantitative tool for ionospheric physical mechanism research. Simultaneously, lightweight feature optimization improves the model's real-time performance and engineering adaptability. Furthermore, embodiments of this disclosure also perform performance verification of the XGBoost model based on a validation set or test set, and construct a global ionospheric foF2 prediction system based on the performance verification results to support end-to-end prediction and interpretable output. Based on this, high-precision global prediction of foF2 is achieved, the driving effect of each feature on the prediction results is quantified, and both model prediction accuracy and interpretability are considered.

[0038] Optionally, combined Figure 2 As shown, the detection device performs five-fold cross-validation on the XG regression framework using a Bayesian-optimized stochastic hyperparameter optimization method to select the optimal hyperparameter combination for constructing and training the XGBoost prediction model, including:

[0039] S11, the detection device relies on the XG regression framework and performs five-fold cross-validation through a Bayesian-optimized stochastic hyperparameter optimization method, with root mean square error as the evaluation index.

[0040] In this step, by comparing the prediction performance of the RF (Random Forest) model and the XGBoost model, it was found that the XGBoost model had an RMSE (Root Mean Square Error) of 0.928MHz on the test set, and R... 2 The coefficient of determination (RMSE) is 0.868, which is significantly better than the RF model (RMSE is 1.549MHz, R...). 2 Since R0.769, the XGBoost regression framework was chosen to construct the foF2 prediction model. Where R0.769... 2R is used to characterize how well a model fits the observed data and reflects the ability of the model's independent variables to explain changes in the dependent variable. 2 The value range is [0,1]. The closer the value is to 1, the higher the simulation fitting accuracy and the better the consistency between the prediction result and the real data.

[0041] S12, the detection device randomly selects a set number of parameter combinations and configures an early stop mechanism, and selects the optimal hyperparameter combination of the validation set for model training to build an XGBoost prediction model.

[0042] The hyperparameter combination includes hyperparameters, which may include some or all of the following: tree depth, subsample ratio, column sampling ratio, L1 penalty term, L2 penalty term, minimum split gain, learning rate, and number of iterations.

[0043] In this step, the probe device selects the optimal hyperparameter combination for the validation set, including: the probe device selects the optimal hyperparameter combination for the validation set according to the following hyperparameter search conditions. These hyperparameter search conditions include, for example: tree depth greater than or equal to 7 and less than or equal to 15; subsample ratio greater than or equal to 0.3 and less than or equal to 1; column sampling ratio greater than or equal to 0.6 and less than or equal to 1; L1 penalty term greater than or equal to 0 and less than or equal to 5; L2 penalty term greater than or equal to 0 and less than or equal to 15; minimum split gain greater than or equal to 0.001 and less than or equal to 1; learning rate greater than or equal to 0.001 and less than or equal to 0.3; and number of iterations greater than or equal to 800 and less than or equal to 2000. It should be noted that the hyperparameter search conditions can be flexibly set according to actual modeling needs.

[0044] The early stopping mechanism includes terminating model training and saving the terminated model if the validation set loss does not decrease within S consecutive validation cycles. Here, S is, for example, 5, 7, or 9. The specific value of S can be flexibly set according to actual modeling needs.

[0045] Thus, based on the XG regression framework, five-fold cross-validation is performed using a Bayesian-optimized stochastic hyperparameter optimization method, with root mean square error as the evaluation metric. Simultaneously, a set number of parameter combinations are randomly selected, and an early stopping mechanism is configured to screen the optimal hyperparameter combination on the validation set for model training to construct the XGBoost prediction model. This achieves the construction and training of the XGBoost prediction model.

[0046] Optionally, the detection device performs five-fold cross-validation on the XG regression framework using a Bayesian-optimized stochastic hyperparameter optimization method to select the optimal hyperparameter combination for constructing and training the XGBoost prediction model, and also includes:

[0047] The objective function of the XG regression framework configured for the detection equipment is the squared error loss.

[0048] The detection device constructs multiple decision trees and trains each decision tree based on the prediction residuals of the preceding trees to obtain the decision output of each decision tree.

[0049] The detection device performs weighted processing on the decision output of each decision tree to obtain the foF2 prediction result.

[0050] This enables the construction and training of the XGBoost prediction model, providing a reliable prediction model for subsequent performance verification of the XGBoost model.

[0051] In some alternative embodiments, the detection device constructs multiple decision trees and trains each decision tree based on the prediction residuals of the preceding trees to obtain the decision output of each decision tree. This includes: the detection device inputs a training set and foF2 values ​​to the generated XGBoost prediction model, iteratively constructs a third preset number of decision trees, and trains each decision tree based on the prediction residuals of the preceding trees to obtain the output of each decision tree.

[0052] In this embodiment of the disclosure, the machine learning model interpretation method includes the SHAP method.

[0053] Optionally, combined Figure 3 As shown, the detection device constructs a lightweight XGBoost model based on the XGBoost prediction model and machine learning model interpretation method, including:

[0054] S21, the detection device randomly selects a first preset number of samples from the training set to construct a sample baseline distribution.

[0055] In this step, the sample baseline distribution is used to estimate the expected predicted value when features are missing. The first preset number is, for example, 10,000.

[0056] S22, the detection device randomly selects a second preset number of samples from the test set and calculates the feature contribution index.

[0057] In this step, the feature contribution metric includes the SHAP value. Calculating the feature contribution metric involves using TreeSHAP (Tree-based SHapley Additive exPlanations) to calculate the SHAP value for each feature data point. TreeSHAP is a highly efficient feature attribution algorithm designed for tree-based ensemble prediction models. It leverages the hierarchical branching structure of a decision tree, quantifies and allocates the marginal contribution of features based on the distribution weights of node samples, and accumulates the weights of each input feature on the output result layer by layer. This allows for rapid and accurate local attribution and global feature importance analysis of the model's prediction results. Here, the second preset quantity is, for example, 3000.

[0058] S23, the detection equipment ranks the XGBoost prediction model according to the feature contribution index to obtain the dominant features. Among them, the dominant features include local time and latitude, solar radiation index, annual day, and year.

[0059] In this step, the detection device sorts the feature contributions of the XGBoost prediction model according to the feature contribution index to obtain the dominant features. This includes: the detection device quantifies the contribution of each feature data to the foF2 prediction using the SHAP value to obtain the prediction contribution of each feature data; the detection device sorts the prediction contributions of all feature data in descending order and determines the feature data corresponding to the top N prediction contributions as the dominant features. The prediction contribution of each feature data is consistent with the ionospheric physical mechanism. Here, the feature data includes 14 basic feature data and 2 combined feature data. The basic feature data includes spatiotemporal feature data (8 in total), solar activity feature data (3 in total), and electromagnetic disturbance feature data (3 in total). The combined feature data includes spatiotemporal coupling feature data and geomagnetic coupling feature data. The spatiotemporal feature data can be referred to in step S31, and the combined feature data can be referred to in step S32. N is greater than or equal to 3. Preferably, N is 5.

[0060] In some optional embodiments, the detection device quantifies the contribution of each feature data to the foF2 prediction using SHAP values, obtaining the predicted contribution of each feature data. The detection device sorts the predicted contributions of all feature data in descending order, determining the feature data corresponding to the top five predicted contributions as dominant features. The dominant features include local time (LT), latitude (Lat), solar radiation index (F10.7), day of the year (Doy), and year (Year). The contribution of local time (LT) reaches 31.1%. Understandably, after obtaining the predicted contribution of each feature data, the detection device can use a SHAP beehive diagram to visualize and analyze the predicted contribution, thus revealing the positive or negative contribution of a single feature data to the foF2 prediction under different values.

[0061] S24, the detection device performs feature ablation processing on the XGBoost prediction model based on the dominant features to construct a lightweight XGBoost model.

[0062] In some optional embodiments, the dominant features include local time (LT), latitude (Lat), solar radiation index (F10.7), day of the year (Doy), and year (Year). The detection device performs feature ablation processing on the XGBoost prediction model based on the dominant features to construct a lightweight XGBoost model. This includes comparing the model performance of the XGBoost model constructed using 16 feature data and the aforementioned dominant features to obtain model performance results.

[0063] This achieves lightweight optimization of the XGBoost prediction model.

[0064] Optionally, space environment data is obtained by cleaning raw space environment data, which includes solar activity parameters and geomagnetic activity parameters.

[0065] Combination Figure 4 As shown, the detection equipment constructs a sample dataset based on the acquired raw satellite occultation observation data, raw space environment data, and raw space environment observation data, including:

[0066] S31, the detection equipment obtains spatiotemporal characteristic data based on satellite occultation observation data, solar activity characteristic data based on solar activity parameters, and electromagnetic disturbance characteristic data based on geomagnetic activity parameters. Specifically, the satellite occultation observation data is obtained by cleaning the raw satellite occultation observation data, and the space environment observation data is obtained by cleaning the raw space environment observation data.

[0067] In this step, spatiotemporal characteristic data includes year, month, day, UTC (UT), solar day, local time (LT), latitude (Lat), and longitude (Lon). Solar activity characteristic data includes solar radio flux (F10.7), sunspot number (SSN), and solar wind speed (Vsw). Electromagnetic disturbance characteristic data includes global geomagnetic activity indices and equatorial geomagnetic disturbance indices (Dst). The global geomagnetic activity indices include the Planetary K-index (KP) and the Planetary A-index (AP).

[0068] The raw satellite occultation observation data includes occultation observation data from COSMIC-2 and occultation observation data from COSMIC-1 for a specified period, such as 2007–2023. The raw space environment observation data includes GRACE (Gravity Recovery and Climate Experiment) satellite data and GIRO (Global Ionosphere Radio Observatory) vertical probe data.

[0069] In some alternative embodiments, the raw space environment data includes raw solar activity parameters and raw geomagnetic activity parameters. Raw satellite occultation observation data and GRACE satellite data are obtained from the CDAAC website. GIRO vertical probe data are obtained from the Lowell GIRO Data Center website. Raw solar activity parameters are obtained from the OMNI Data Network website. Specifically, the raw solar activity parameters are solar activity data from 2007 to 2023.

[0070] S32, the detection equipment generates spatiotemporal coupling characteristic data based on UTC and longitude Lon, and generates geomagnetic coupling characteristic data based on the equatorial geomagnetic disturbance index Dst and latitude Lat, in order to construct combined characteristic data.

[0071] In this step, the detection device generates spatiotemporal coupling characteristic data based on UTC and longitude Lon, and generates geomagnetic coupling characteristic data based on the equatorial geomagnetic disturbance index Dst and latitude Lat, in order to construct combined characteristic data. This includes: the detection device obtains spatiotemporal coupling characteristic data based on the product of UTC and longitude Lon, and obtains geomagnetic coupling characteristic data based on the product of the equatorial geomagnetic disturbance index Dst and latitude Lat.

[0072] S33, the detection equipment inputs spatiotemporal characteristic data, solar activity characteristic data, electromagnetic disturbance characteristic data, combined characteristic data, and space environment observation data to a feature normalizer for normalization processing to obtain effective spatiotemporal characteristic data, effective activity characteristic data, effective electromagnetic disturbance characteristic data, effective combined characteristic data, and effective space environment observation data. The variable foF2 is then input to a target variable normalizer for normalization processing to obtain the target variable. The space environment observation data is obtained by data cleaning of the original space environment observation data.

[0073] S34. The detection equipment selects effective spatiotemporal characteristic data from years of high solar activity and years of low solar activity as the test set, and selects effective spatiotemporal characteristic data from other years as the training set.

[0074] In this step, a high solar activity year refers to a year within the solar cycle when the intensity of activities such as sunspot number, solar radio flux, and solar wind is at its peak, while a low solar activity year refers to a year within the solar cycle when the intensity of various solar activities falls to its lowest point. For example, 2014 was a high solar activity year, and 2019 was a low solar activity year.

[0075] S35, the detection equipment constructs a validation set based on effective space environment observation data.

[0076] S36, the detection device constructs and generates a sample dataset based on the test set, training set, and validation set.

[0077] In this way, the constructed sample dataset remains stable in both years of high and low solar activity, and can accurately reproduce typical ionospheric morphologies such as equatorial anomalies, Weddell Sea anomalies, and winter anomalies, providing reliable data samples for the construction and training of XGBoost prediction models.

[0078] Optionally, the detection equipment performs data cleaning in the following manner:

[0079] The detection equipment is configured with the peak height of the F2 layer, the effective range of foF2, and the effective range of the electron density gradient, and the electron density profile is configured to meet the horizontal angle condition. The electron density gradient is the change in electron density between adjacent heights. The height difference with the adjacent height The ratio of electron density gradient to electron density gradient is expressed as... .

[0080] The detection equipment is configured with an invalid value of 9999, which is NAN. Here, NAN represents a null value.

[0081] The detection equipment uses the isolated forest algorithm to remove extreme anomalies and marks them as NAN.

[0082] The detection equipment interpolates NAN using linear interpolation.

[0083] In this way, by configuring the peak height of the F2 layer, the effective range of foF2, and the effective range of the electron density gradient, and by configuring the horizontal angle condition of the electron density profile, the validity of the data can be ensured and the influence of horizontal differences on the results can be eliminated. By configuring the invalid value 9999 as NAN, using the isolated forest algorithm to remove extreme outliers and marking them as NAN, and using linear interpolation to complete the missing data, the continuity of the data in time can be guaranteed.

[0084] In some optional embodiments, the detection device removes extreme anomalous data using the Isolation Forest algorithm, including: the detection device uses the Isolation Forest algorithm to detect and obtain anomalous data; the detection device determines whether the anomalous data is extreme anomalous data based on the outlier ratio; wherein the outlier ratio is set to 0.0001; if the anomalous data is extreme anomalous data, the detection device removes the extreme anomalous data; if the anomalous data is not extreme anomalous data, the detection device retains the anomalous data.

[0085] Optionally, the detection device is configured with the F2 layer peak height and the effective range of foF2, and the effective range of the electron density gradient, including:

[0086] The detection equipment is configured with a peak height of F2 layer greater than or equal to 100km and less than or equal to 500km.

[0087] The detection equipment is configured with foF2 greater than or equal to 0Hz and less than or equal to 20Hz.

[0088] The detection equipment is configured to have an electron density profile that meets the horizontal angle requirements, including:

[0089] The detection equipment is configured with an electron density gradient greater than or equal to -0.01el / (m 3 •km) and less than or equal to 0.01el / (m 3 •km), the variation of the electron density profile in the horizontal 5˚ direction is less than or equal to 0.001el / m 4 Among them, el / (m 3 ∙km) represents the electron number density gradient along a kilometer path.

[0090] This ensures the validity of the data and keeps the horizontal difference in the electron density profile within 5˚ to eliminate the influence of horizontal differences on the results.

[0091] In some alternative embodiments, the detection device is configured to vary the electron density profile in the horizontal 5˚ direction. Less than or equal to 0.001el / m 4 ,include:

[0092] .in, .

[0093] Optionally, the detection device is configured with the peak height of the F2 layer and the effective range of foF2, the effective range of the electron density gradient, and also includes: the deviation between the original electron density and the electron density after the 9-point window moving average is less than or equal to 1.5.

[0094] The electron density after the 9-point window moving average refers to the arithmetic mean of the sampling data of a total of 9 consecutive time series centered on the current sampling point.

[0095] This allows for a second round of data screening and cleaning.

[0096] Optionally, space environment observation data includes GRACE satellite data and GIRO vertical survey instrument data. Combined Figure 5 As shown, the detection equipment verifies the performance of the XGBoost model using a validation set or test set, and constructs a global ionospheric foF2 prediction system based on the performance verification results, including:

[0097] S41, the detection equipment performs performance verification on the lightweight XGBoost model based on the validation set, compares the prediction accuracy of the lightweight XGBoost model with the space environment parameter prediction model, and verifies the ability of the lightweight XGBoost model to reproduce typical ionospheric morphology based on the validation set.

[0098] In this step, the space environment parameter prediction models include IRI-2020 (International Reference Ionosphere 2020 model), NPDM (Neustrelitz Peak Density Model), and the Random Forest (RF) model. Typical ionospheric morphologies include the magnetic equatorial anomaly, the Weddell Sea anomaly, and the winter anomaly.

[0099] S42, the detection device stores a lightweight XGBoost model, feature normalizer, and target variable normalizer to encapsulate and generate a global ionospheric foF2 prediction system.

[0100] In this way, the XGBoost model is validated using a validation or test set, and a global ionospheric foF2 prediction system is built based on the validation results to support end-to-end prediction and interpretable output. This achieves high-precision global foF2 prediction while balancing model prediction accuracy and interpretability.

[0101] In some alternative embodiments, the detection device uses a validation set to validate the performance of the lightweight XGBoost model, obtaining that the XGBoost model maintains R² during both the high solar activity year of 2014 and the low solar activity year of 2019. 2 ≥0.77, the prediction error for mid-to-high latitudes is reduced by more than 50% compared to the IRI-2020 and NPDM models.

[0102] The detection equipment also used a validation set to verify the ability of the lightweight XGBoost model to reproduce the ionospheric magnetic equatorial anomaly, the Weddell Sea anomaly, and the winter anomaly. The XGBoost model can accurately present the bimodal structure of the equatorial anomaly at ±15° latitude on both sides of the magnetic equator, and successfully reproduce the anomalous characteristics of the Weddell Sea where fof2 is higher at night in summer than during the day and fof2 is higher during the winter day in mid-latitudes than during the summer day.

[0103] It should be noted that the detection equipment performs performance verification on the XGBoost model using a validation set or test set, and constructs a global ionospheric foF2 prediction system based on the performance verification results. This also includes:

[0104] The detection equipment performs basic performance verification of the lightweight XGBoost model based on the test set and compares the prediction accuracy of the lightweight XGBoost model with that of the space environment parameter prediction model.

[0105] In some alternative embodiments, the probing device uses a test set to verify the performance of the lightweight XGBoost model and obtain the R-value of the XGBoost model. 2 The frequency response was 0.886, and the RMSE was 0.964MHz.

[0106] Optionally, the detection device stores a lightweight XGBoost model, feature normalizer, and target variable normalizer to encapsulate and generate a global ionospheric foF2 prediction system, including:

[0107] The detection equipment generates an input module, a preprocessing module, a prediction module, an inverse normalization module, and an output module.

[0108] The detection equipment encapsulates an input module, a preprocessing module, a prediction module, an inverse normalization module, and an output module to generate a global ionospheric foF2 prediction system.

[0109] The system comprises the following modules: Input module receives raw satellite occultation observation data and raw space environment data. Preprocessor module cleans the raw satellite occultation observation data and raw space environment data, and performs feature normalization on spatiotemporal feature data, solar activity feature data, electromagnetic disturbance feature data, combined feature data, and space environment observation data. Prediction module loads a lightweight XGBoost model and outputs normalized foF2 prediction values. Denormalization module converts the normalized foF2 prediction values ​​into actual foF2 prediction values ​​in MHz units. Output module generates the prediction contribution of all feature data.

[0110] Optionally, the detection device stores a lightweight XGBoost model, feature normalizer, and target variable normalizer to encapsulate and generate a global ionospheric foF2 prediction system, and also includes:

[0111] After the detection equipment encapsulates an input module, a preprocessing module, a prediction module, an inverse normalization module, and an output module to generate a global ionospheric foF2 prediction system, it stores the global ionospheric foF2 prediction system as a serialized model file.

[0112] The detection device loads the serialized model file and outputs the actual foF2 prediction value and an interpretability report of the prediction contribution of the feature data.

[0113] The predictive contribution interpretability report for feature data includes a ranking of the feature contributions of all feature data and / or a SHAP beehive diagram.

[0114] This enables the local deployment of the global ionospheric foF2 prediction system. After loading the serialized model file, the response time is no more than 1 second. It supports end-to-end prediction and interpretability output, and synchronously provides interpretability reports on the prediction contribution of feature data, providing decision support for engineering applications such as shortwave communication frequency planning, satellite navigation ionospheric correction, and space weather anomaly early warning.

[0115] Optionally, serialize the model file, such as a .pkl model file.

[0116] In some optional embodiments, the detection device is configured with Python 2023.2.1, pandas (2.2.3), xgboost (2.1.3), and shap (0.48.0) libraries. This enables the loading of serialized model files and the local deployment and application of the global ionospheric foF2 prediction system. In specific applications, users input raw satellite occultation observation data, raw space environment data, and raw space environment observation data into the serialized model file, and after loading the model file, they can obtain the actual foF2 prediction value and the prediction contribution of the feature data.

[0117] In a practical application, Figure 6 (a) shows the correlation of solar activity parameters. Figure 6 (b) shows the effect of solar flux on the ionosphere. Figure 6 (c) is the probability density distribution of foF2 values ​​obtained by using existing foF2 prediction methods, and the probability density distribution of foF2 values ​​obtained by cleaning foF2 using the method of this application. Figure 6 (d) is a schematic diagram of the monthly average change trends of F10.7, SSN, and foF2.

[0118] Figure 7 This is a schematic diagram of the framework of the lightweight XGBoost model.

[0119] Figure 8This is a schematic diagram illustrating the feature contribution analysis results obtained by ranking the feature contributions of the XGBoost prediction model based on the feature contribution index. Among them, Figure 8 (a) is the ranking of feature importance. Figure 8 (b) is the key feature LT influencing the trend. Figure 8 (c) shows the thermal distribution of SHAP values. Figure 8 (d) shows the distribution of feature SHAP values. Here, the key feature LT is... Figure 8 (a) The feature data with the highest feature importance ranking. Combined with... Figure 8 (a) It can be seen that the dominant features include five data points: local time LT, latitude Lat, solar radiation index F10.7, day of the year Doy, and year Year.

[0120] Figure 9 This is a scatter plot of the predicted density of the XGBoost model. Figure 9 In this context, XGmodel represents the XGBoost model. The test set includes 218,018 feature data points. The probing device uses this test set to validate the performance of the lightweight XGBoost model, obtaining its R-value. 2 The frequency response was 0.886, and the RMSE was 0.964MHz.

[0121] Figure 10 This is a schematic diagram showing the median absolute error of the foF2 predictions of the lightweight XGBoost model and the RF model at different local times and in different seasons.

[0122] Figure 11 This is a scatter plot of the predicted density for the lightweight XGBoost model, IRI-2020, and NPDM model. Combined with... Figure 11 As shown, the R of the XGBoost model 2 The XGBoost model outperforms both the IRI-2020 and NPDM models in terms of RMSE.

[0123] The detection device constructs a lightweight XGBoost model based on the XGBoost prediction model and machine learning model interpretation method, specifically including:

[0124] First, the detection device randomly selects 10,000 samples from the training set to construct a baseline distribution, and randomly selects 3,000 samples from the test set to calculate the SHAP value. The calculation process employs a dynamic programming strategy to recursively traverse the decision tree path.

[0125] Then, the detection equipment ranked the features of the XGBoost prediction model according to the SHAP value to obtain the dominant features. These dominant features included local time and latitude, solar radiation index, day of the year, and year, and were ranked from highest to lowest contribution as follows: local time, latitude, solar radiation index, day of the year, and year. The total contribution of these five features exceeded 75%.

[0126] In another practical application, the dominant features include local time (LT), latitude (Lat), solar radiation index (F10.7), day of the year (Doy), and year (Year).

[0127] The detection device performs feature ablation processing on the XGBoost prediction model based on the dominant features to construct a lightweight XGBoost model, specifically including:

[0128] The detection device compares the model performance of the XGBoost model constructed based on the 16 feature data and the aforementioned dominant features on the same test set to obtain the model performance results. The model performance results are shown in Table 1. The core feature subset model represents the XGBoost model constructed based on the aforementioned dominant features, and the full feature baseline model represents the XGBoost prediction model (i.e., the XGBoost model constructed based on the 16 feature data).

[0129] Table 1. Comparison of model performance results between the core feature subset and the full feature model.

[0130] Core Feature Subset Model 0.848 1.066 0.784 Full-feature benchmark model 0.886 0.964 0.724 Change -0.038 +0.102 +0.060

[0131] As shown in Table 1, with a 62.5% reduction in feature dimensions, the R-value of the lightweight XGBoost model is... 2 It decreased by only 0.038, RMSE increased by only 0.102MHz, MAE increased by only 0.060MHz, and remained at the same level. 2 The prediction accuracy is ≥0.848. This verifies that the machine learning model interpretation method can achieve lightweight optimization of the XGBoost prediction model, and that the interpretation method has good reliability.

[0132] In another practical application, the raw space observation data includes GRACE satellite data from 2014 (15,425 entries) and vertical log data from 7 GIRO stations from 2019 (54,030 entries). The locations of the 7 stations can be found in Table 2.

[0133] The detection equipment validates the XGBoost model using a validation or test set, and constructs a global ionospheric foF2 prediction system based on the validation results, including:

[0134] The detection equipment validated the performance of the lightweight XGBoost model based on the validation set, comparing the prediction accuracy of the lightweight XGBoost model with that of the space environment parameter prediction model, and learned that in K... P When ≤3, the R-value of this XGBoost model is 2 It is 0.80, and in K P When the value is ALL (i.e., the range of values ​​is not limited), the R-value of this XGBoost model is... 2 The value is 0.77. Meanwhile, this XGBoost model performs well in the equatorial region, the low latitude region of the Southern Hemisphere, and the high latitude region of the Northern Hemisphere. 2 All satisfy R 2 ≤0.74, mid-to-high latitude RMSE≤0.59MHz, representing an error reduction of over 50% compared to space environment parameter prediction models. Specifically, in K... P For values ​​≤3, the performance verification results of the lightweight XGBoost model are shown in Table 2.

[0135] The detection device performed basic performance verification on the lightweight XGBoost model based on a test set (218,018 samples), and obtained the R-value of the XGBoost model. 2 With a high foF2 value of 0.886 and an RMSE of 0.964MHz, the high-density region is concentrated in the typical range of [2,8]Hz, and the model prediction stability is outstanding.

[0136] The XGBoost model performance was compared using two test sets: one where the original satellite occultation data only included COSMIC-1 occultation observations from 2007 to 2023 (Case 1), and another where the original satellite occultation data included both COSMIC-1 and COSMIC-2 occultation observations from 2007 to 2023 (Case 2). Performance verification showed that in low-latitude regions, the MAE decreased from 0.792MHz to 0.773MHz, and the RMSE decreased from 1.042MHz to 1.015MHz. In low-latitude regions, the R... 2 The R² value reached 0.89, significantly outperforming the IRI-2020 model and the NPDM model. Specifically, the IRI-2020 model's R² value... 2 The R-value of the NPDM model is 0.80. 2 It is 0.72.

[0137] Table 2 In K P Performance validation results of the lightweight XGBoost model when ≤3

[0138] Equatorial region -7.95° -14.40° 0.88 1.44 1.13 Low latitudes of the Northern Hemisphere 21.43° -158.15° 0.72 1.76 1.29 Southern Hemisphere low latitudes -29.73° -53.71° 0.81 1.78 1.34 Mid-latitudes of the Northern Hemisphere 40.80° 0.50° 0.82 0.59 0.40 Mid-latitudes of the Southern Hemisphere -34.42° 19.22° 0.84 0.71 0.53 High latitudes of the Northern Hemisphere 69.60° 19.20° 0.74 0.37 0.29 Southern Hemisphere High Latitudes -74.62° 164.24° 0.52 0.74 0.52

[0139] Combination Figure 12As shown, this disclosure provides a modeling apparatus 70 for the global ionosphere foF2, including a processor 700 and a memory 701. Optionally, the apparatus 70 may further include a communication interface 702 and a bus 703. The processor 700, communication interface 702, and memory 701 can communicate with each other via the bus 703. The communication interface 702 can be used for information transmission. The processor 700 can call logical instructions in the memory 701 to execute the modeling method for the global ionosphere foF2 described in the above embodiment.

[0140] Furthermore, the logic instructions in the aforementioned memory 701 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0141] The memory 701, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as program instructions / modules corresponding to the methods in the embodiments of this disclosure. The processor 700 executes functional applications and data processing by running the program instructions / modules stored in the memory 701, thereby implementing the modeling method for global ionosphere foF2 described in the above embodiments.

[0142] The memory 701 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 701 may include high-speed random access memory and may also include non-volatile memory.

[0143] This disclosure provides a detection device, including: a detection device body and the aforementioned modeling device 70 for global ionosphere foF2. The modeling device 70 for global ionosphere foF2 is mounted on the detection device body. The mounting relationship described herein is not limited to placement inside the detection device body, but also includes mounting connections with other components of the detection device, including but not limited to physical connections, electrical connections, or signal transmission connections. Those skilled in the art will understand that the modeling device 70 for global ionosphere foF2 can be adapted to feasible detection device bodies to achieve other feasible embodiments.

[0144] The foregoing description and accompanying drawings fully illustrate embodiments of this disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Moreover, the terminology used in this application is for describing embodiments only and is not intended to limit the claims. As used in the description of embodiments and claims, the singular forms “a,” “an,” and “the” are intended to equally include the plural forms unless the context clearly indicates otherwise. Similarly, the term “and / or” as used in this application means including one or more of the associated listed items and all possible combinations thereof. Additionally, when used in this application, the term "comprise" and its variations "comprises" and / or "comprising" refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Without further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element. In this document, each embodiment may focus on the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, the relevant parts can be referred to the description of the method section.

Claims

1. A modeling method for the global ionosphere foF2, characterized in that, include: Based on the acquired raw satellite occultation observation data and raw space environment data, a sample dataset is constructed; the sample dataset includes a training set, a test set, and a validation set. Five-fold cross-validation was performed on the XG regression framework using the Bayesian optimization stochastic hyperparameter optimization method to select the optimal hyperparameter combination for the construction and training of the XGBoost prediction model. A lightweight XGBoost model is constructed based on the XGBoost prediction model and the machine learning model interpretation method. The XGBoost model is validated using a validation set or test set, and a global ionospheric foF2 prediction system is constructed based on the performance validation results. The space environment observation data is obtained by cleaning the original space environment observation data.

2. The method according to claim 1, characterized in that, Five-fold cross-validation was performed on the XG regression framework using a Bayesian-optimized stochastic hyperparameter optimization method to select the optimal hyperparameter combination for constructing and training the XGBoost prediction model, including: Based on the XG regression framework, five-fold cross-validation was performed using a Bayesian-optimized stochastic hyperparameter optimization method, with root mean square error as the evaluation index. Randomly select a set number of parameter combinations and configure an early stopping mechanism to screen the optimal hyperparameter combination on the validation set for model training to build an XGBoost prediction model; The hyperparameter combination includes hyperparameters, which may include some or all of the following: tree depth, subsample ratio, column sampling ratio, L1 penalty term, L2 penalty term, minimum split gain, learning rate, and number of iterations.

3. The method according to claim 2, characterized in that, Five-fold cross-validation was performed on the XG regression framework using a Bayesian-optimized stochastic hyperparameter optimization method to select the optimal hyperparameter combination for constructing and training the XGBoost prediction model. This also included: The objective function of the XG regression framework is configured as the squared error loss. Construct multiple decision trees and train each decision tree based on the prediction residuals of the preceding trees to obtain the decision output of each decision tree; The decision outputs of each decision tree are weighted to obtain the foF2 prediction results.

4. The method according to claim 1, characterized in that, Based on the XGBoost prediction model and machine learning model interpretation methods, a lightweight XGBoost model is constructed, including: A first preset number of samples are randomly drawn from the training set to construct a sample baseline distribution; A second preset number of samples are randomly selected from the test set, and the feature contribution index is calculated. The XGBoost prediction model is ranked according to its feature contribution index to obtain the dominant features; among them, the dominant features include local time and latitude, solar radiation index, annual day, and year; Based on the dominant features, feature ablation processing is performed on the XGBoost prediction model to construct a lightweight XGBoost model.

5. The method according to claim 1, characterized in that, Space environment data is obtained by cleaning raw space environment data, which includes solar activity parameters and geomagnetic activity parameters. Based on the acquired raw satellite occultation observation data and raw space environment data, a sample dataset is constructed, including: Spatiotemporal characteristic data were obtained from satellite occultation observation data, solar activity characteristic data were obtained from solar activity parameters, and electromagnetic disturbance characteristic data were obtained from geomagnetic activity parameters. Among them, satellite occultation observation data was obtained by data cleaning of raw satellite occultation observation data, and space environment observation data was obtained by data cleaning of raw space environment observation data. Spatiotemporal characteristic data includes year, month, day, UTC, solar day, local time, latitude (Lat), and longitude (Lon). Solar activity characteristic data includes solar radio current (F10.7), sunspot number (SSN), and solar wind speed (Vsw). Electromagnetic disturbance characteristic data includes global geomagnetic activity index and equatorial geomagnetic disturbance index (Dst). Spatiotemporal coupling characteristic data are generated based on UTC and longitude Lon, and geomagnetic coupling characteristic data are generated based on equatorial geomagnetic disturbance index Dst and latitude Lat, in order to construct combined characteristic data; The spatiotemporal feature data, solar activity feature data, electromagnetic disturbance feature data, combined feature data, and space environment observation data are input into a feature normalizer for normalization processing to obtain effective spatiotemporal feature data, effective activity feature data, effective electromagnetic disturbance feature data, effective combined feature data, and effective space environment observation data. The variable foF2 is then input into a target variable normalizer for normalization processing to obtain the target variable. The space environment observation data is obtained by cleaning the original space environment observation data. Effective spatiotemporal feature data from years of high solar activity and years of low solar activity were selected as the test set, and effective spatiotemporal feature data from other years were selected as the training set. A validation set is constructed based on valid space environment observation data; Based on the test set, training set, and validation set, a sample dataset is constructed.

6. The method according to claim 5, characterized in that, Data cleaning should be performed in the following manner: Configure the peak height of the F2 layer and the effective range of foF2, as well as the effective range of the electron density gradient; The electron density profile is configured to meet the horizontal angle condition; Configure the invalid value 9999 as NAN; Extreme outliers are removed using the Isolation Forest algorithm and then labeled as NAN. NAN is interpolated using linear interpolation.

7. The method according to claim 6, characterized in that, Configure the peak height of the F2 layer and the effective range of foF2, as well as the effective range of the electron density gradient, including: The peak height of the F2 layer is configured to be greater than or equal to 100km and less than or equal to 500km; Configure foF2 to be greater than or equal to 0Hz and less than or equal to 20Hz; Configure the electron density profile to meet the horizontal angle condition, including: Configure the electron density gradient to be greater than or equal to -0.01el / (m 3 •km) and less than or equal to 0.01el / (m 3 •km), the variation of the electron density profile in the horizontal 5˚ direction is less than or equal to 0.001el / m 4 .

8. The method according to claim 5, characterized in that, The XGBoost model is validated using a validation or test set. Based on the validation results, a global ionospheric foF2 prediction system is constructed, including: The performance of the lightweight XGBoost model was verified based on the validation set. The prediction accuracy of the lightweight XGBoost model was compared with that of the space environment parameter prediction model. The ability of the lightweight XGBoost model to reproduce typical ionospheric morphology was also verified based on the validation set. Save the lightweight XGBoost model, feature normalizer, and target variable normalizer to encapsulate and generate a global ionospheric foF2 prediction system.

9. A modeling apparatus for the global ionosphere foF2, comprising a processor and a memory storing program instructions, characterized in that, The processor is configured to, when running the program instructions, execute the modeling method for global ionosphere foF2 as described in any one of claims 1 to 8.

10. A detection device, characterized in that, include: The detection equipment itself; The modeling apparatus for global ionosphere foF2 as described in claim 9 is installed on the main body of the detection device.

Images