Thermospheric density stratified progressive full-scale prediction system and method for cross-source data fusion

Abstract

The application belongs to the technical field of atmospheric density prediction, and discloses a thermosphere atmospheric density layered progressive full-scale prediction system and method based on cross-source data fusion, which utilizes respective advantages of different gradient data sources, constructs a three-level layered fusion refinement architecture of a benchmark density field, a correction density field and an instantaneous refined density field, and realizes month-year scale long-term prediction, day-week scale medium and short-term prediction and short-term prediction of thermosphere atmospheric density by combining cross-source data calibration and physical constraint modeling. Through the three-level layered refinement architecture, the application fully excavates the long-term coverage advantage of TLE data, the mesoscale variability description advantage of precise orbit data and the high-frequency high-precision advantage of accelerometer data, avoids defects of a single data source, realizes accurate description of full-scale and full-area density, and can improve refinement accuracy and prediction adaptability of thermosphere atmospheric density and reduce data cost.

Description

Technical Field

[0001] This invention belongs to the field of atmospheric density prediction technology, and relates to a hierarchical, progressive, full-scale prediction system and method for thermospheric atmospheric density based on cross-source data fusion. Background Technology

[0002] Thermospheric atmospheric density is a key environmental parameter that regulates the orbital evolution of low-Earth orbit satellites and determines the risk of space debris collisions. Its accurate acquisition and prediction play an irreplaceable role in spacecraft orbital control, collision early warning, and long-term life assessment. Current methods for acquiring and predicting thermospheric atmospheric density suffer from an imbalance between cost and accuracy. Traditional high-precision thermospheric atmospheric density acquisition schemes rely on a large amount of high-precision observation data from spaceborne accelerometers, resulting in extremely high data acquisition and computation costs. Low-cost schemes, on the other hand, lack sufficient accuracy due to the use of a single low-precision data source, making it difficult to support core aerospace applications such as satellite collision early warning.

[0003] Chinese patent CN120163066A discloses an orbit-driven atmospheric density estimation method. This method only considers using historical orbital data from a single data source (TLE) as a benchmark, refining the extrapolated propagation orbit, and then achieving high-precision atmospheric density estimation based on the refined orbital data. Chinese patent CN114861570A discloses a method for predicting the average evolution of the space debris environment and analyzing constellation impacts. This method considers the influence of atmospheric drag during the calculation of space debris environment evolution and constellation impacts, employs an exponential atmospheric density model, and predicts atmospheric density through orbital decay.

[0004] Both of the above methods use a single data source to retrieve thermospheric atmospheric density, which has inherent limitations and cannot simultaneously meet the requirements of coverage and accuracy. While TLE data has the advantages of tracking a large number of targets and long observation duration, its temporal accuracy is low, and the density retrieved based on it is only suitable for climatological statistical analysis and cannot meet the requirements of instantaneous high-precision applications. The density retrieved from precise orbit data has medium accuracy and resolution, but lacks systematic calibration with high-precision data, which limits its absolute accuracy. The density retrieved from spaceborne accelerometer data has the best spatiotemporal resolution and accuracy, but the data coverage is narrow and the acquisition cost is high, making it difficult to achieve global, all-time coverage.

[0005] Chinese patent CN112711736A discloses a calibration method, device, storage medium, and processor for atmospheric density detection data. This method considers determining the calibration coefficients of in-situ atmospheric density detection data through satellite orbit data. Its purpose is to solve the technical problem of poor timeliness in atmospheric density detection data calibration. However, it does not take into account the density of areas that cannot be detected by current thermospheric atmospheric density in-situ data, that is, it does not solve the problem of insufficient density characterization across the entire scale (200-800km altitude).

[0006] Chinese patent CN116136975A discloses a method and device for predicting neutral atmospheric temperature and density based on LSTM neural network. The method considers solar radiation index, geomagnetic activity index, and historical atmospheric density. Based on LSTM neural network, it uses deep learning of historical observation data to obtain data on the changes of neutral atmospheric temperature and density over time.

[0007] Chinese patent CN118468935A discloses a rapid simulation method and system for key parameters of space weather. This method uses a space weather model to simulate and obtain a multi-source key dataset of the thermosphere. Based on an adjusted and optimized Transformer model, a thermosphere mass density simulation and prediction model is constructed to simulate and predict the thermosphere mass density.

[0008] The two methods described above are designed for a single time scale, resulting in poor adaptability of the prediction system, insufficient physical interpretability, and a tendency to produce prediction results that violate the physical laws of the thermosphere. Summary of the Invention

[0009] The technical problem solved by this invention is to provide a layered, progressive, full-scale prediction system and method for thermospheric atmospheric density by cross-source data fusion. By combining multi-source data with physical constraints, the system improves the accuracy and predictive adaptability of thermospheric atmospheric density and reduces data costs.

[0010] This invention is achieved through the following technical solution:

[0011] A cross-source data fusion system for predicting thermospheric atmospheric density in a hierarchical, progressive, full-scale manner includes:

[0012] The data input layer receives multi-source data, including TLE data, precise orbit data, accelerometer data, and space weather index data.

[0013] The data preprocessing layer includes a TLE data preprocessing module, a precise orbit data preprocessing module, an accelerometer data preprocessing module, and a space weather index preprocessing module. These modules perform preprocessing on TLE data, precise orbit data, accelerometer data, and space weather index data, including removing invalid or abnormal data and filling in missing data, to form each dataset and send it to the corresponding module in the refinement layer.

[0014] The refinement layer comprises a three-tiered thermospheric atmospheric density refinement architecture consisting of a bottom-level climatology construction module, a middle-level variability correction module, and a top-level instantaneous refinement module. Each module receives its corresponding dataset, generates it, and outputs the climatology density to the cross-source fusion module. Middle layer corrected density and top-layer refinement density The cross-source fusion module assigns weights to the input data and performs weighted fusion to obtain the refined atmospheric density. , serving as training samples and validation benchmarks for the multi-scale density prediction layer prediction model;

[0015] The multi-scale density prediction layer includes a long-term prediction module, a medium- and short-term prediction module, a short-term prediction module, and a physical constraint module.

[0016] The long-term prediction module is based on climatological state density. Lasso regression models trained with solar activity data are used to predict long-term climatological density from month to year.

[0017] The short-to-medium-term prediction module is based on training during periods of high solar activity and has been... The validated XGBoost model is used for short- to medium-term prediction of thermospheric atmospheric density over days and weeks.

[0018] The short-term prediction module uses the constraints provided by the physical constraint module based on the LSTM model to perform short-term predictions of thermospheric atmospheric density from minutes to hours.

[0019] Furthermore, the long-term prediction module includes an input layer, a feature extraction layer, a Lasso regression model layer, and an output layer; wherein the input layer receives long-term solar activity forecast data, long-term geomagnetic activity statistics, seasonal parameters, and so on. The interannual / seasonal cyclical feature sequences are used to extract features related to solar activity levels, geomagnetic activity levels, and seasonality, which are then sent to the Lasso regression model layer. Constrained by climatological boundary conditions, the Lasso regression model sends the prediction results to the output layer;

[0020] The short-to-medium term prediction module includes an input layer, a feature extraction layer, an XGBoost model layer, and an output layer; the input layer receives 3-day Kp index forecasts, 7-day F10.7 index forecasts, and forecasts for the past 3 months. The time-series data and altitude / latitude / local time parameters are used to extract time-lag features and capture the time delay of the influence of space weather index on atmospheric density, which are then sent to the XGBoost model layer. The XGBoost model then sends the prediction results to the output layer.

[0021] The short-term prediction module includes an input layer, a feature extraction layer, an LSTM model layer, and an output layer. The input layer receives constraints provided by the physical constraint module, real-time space weather index, and instantaneous atmospheric density over the past hour. and The feature extraction layer obtains features including density frequency, density height attenuation constraint, and density increase limit and sends them to the LSTM model layer. The LSTM model then sends the prediction results to the output layer.

[0022] This invention also provides a method for predicting thermospheric atmospheric density in a hierarchical, progressive, full-scale manner by fusing cross-source data, comprising the following operations:

[0023] The data input layer receives multi-source data, including TLE data, precise orbit data, accelerometer data, and space weather index data.

[0024] The data preprocessing layer performs preprocessing on TLE data, precise orbit data, accelerometer data and space weather index data, including removing invalid or abnormal data and filling in missing data, to form each dataset and send it to the corresponding module of the refinement layer.

[0025] The bottom-level climatology construction module, the middle-level variability correction module, and the top-level instantaneous refinement module of the refinement layer receive the TLE dataset, the precise orbit dataset, and the accelerometer dataset, respectively, and perform density refinement at the corresponding levels; and output the climatology density to the cross-source fusion module. Middle layer corrected density and top-layer refinement density The cross-source fusion module assigns weights to the input data and performs weighted fusion to obtain the refined atmospheric density. ;

[0026] The long-term, medium-term, and short-term prediction modules of the multi-scale density prediction layer provide long-term climatological density predictions, medium-term and short-term predictions, and short-term predictions, respectively.

[0027] The long-term forecast is based on climatological density. The Lasso regression model, trained with solar activity data, is used for prediction. The inputs to the model include long-term solar activity forecasts, long-term geomagnetic activity statistics, seasonal parameters, and climatological state density. Interannual / seasonal cyclical characteristic sequences, in terms of climatological density To constrain climatological boundary conditions, we conduct monthly-year long-term climatological state density predictions.

[0028] The aforementioned medium- and short-term forecasts are based on training during periods of high solar activity and refined atmospheric density. The validated XGBoost model was used for prediction, with 3-day Kp exponent forecasts, 7-day F10.7 exponent forecasts, and mid-level corrected densities over the past 3 months as inputs. Using time-series data and altitude / latitude / local time parameters, we can make short- to medium-term predictions of thermospheric atmospheric density over days and weeks.

[0029] The short-term forecast is based on constraints provided by the physical constraint module of the LSTM model. The real-time space weather index and the instantaneous atmospheric density of the past hour are input into the model. and top-layer refining density The parameters are updated in real time using a 1-hour sliding window to predict the thermospheric atmospheric density over 30 minutes to 2 hours.

[0030] Furthermore, the underlying climate state construction module performs underlying density refinement to obtain climate state density. This includes the following refining operations:

[0031] The thermospheric atmospheric density corresponding to TLE data was retrieved using the semi-major axis attenuation method.

[0032] Based on the retrieved thermospheric density, a joint stratification strategy was adopted according to solar activity level, season, and geomagnetic activity level. The global region was divided into grids, and the average density of each grid was calculated to obtain the initial climatological density field. Then, EOF decomposition was performed on the initial climatological density field to extract the first three principal modes and reconstruct the climatological density. ;

[0033] The mid-layer variability correction module performs mid-layer density variability correction and obtains the corrected mid-layer density. This includes the following refining operations:

[0034] The atmospheric density corresponding to the precise orbital data was retrieved using the interpolation differential method, and a linear deviation correction model was established using accelerometer data to calibrate the retrieval results. The following calibration was obtained by fitting the data using the least squares method:

[0035] ;

[0036] in, To achieve precise orbit inversion density after calibration, Atmospheric density derived from precise orbit inversion; This is the proportional correction factor. These are constant correction coefficients;

[0037] Will and Perform spatiotemporal matching and calculate the deviation field. The bias field is extended to a global area, with weighting coefficients set according to altitude. ,pass Obtain the corrected density of the middle layer ;

[0038] The top-level instantaneous refinement module performs instantaneous refinement of the top-level density to obtain the refined top-level density. This includes the following refining operations:

[0039] Instantaneous atmospheric density was retrieved based on accelerometer data and atmospheric drag experienced by the satellite:

[0040] ;

[0041] in, Instantaneous atmospheric density, Due to atmospheric drag on the satellite, For satellite area, For satellite orbital velocity;

[0042] In the area covered by accelerometer data, with instantaneous atmospheric density As the top-level refinement density in this region ;

[0043] In regions with sparse data, according to Obtain the top-level refined density field, where The extrapolation weights are dynamically determined based on the development stage of the geomagnetic storm.

[0044] Furthermore, the cross-source fusion module obtains the fused and refined atmospheric density. for:

[0045] ;

[0046] in, , , They are respectively , , The weights are determined based on data uncertainty and satisfy the following conditions: .

[0047] Furthermore, the constraints provided by the physical constraint module to the LSTM model include:

[0048] Thermospheric atmospheric density decreases exponentially with altitude: ;in, For height, This refers to the elevation of the thermosphere.

[0049] At the same time, the density increase during geostorm periods should be ≤ 5 times the density during calm periods.

[0050] Furthermore, it is also based on the fusion and refinement of atmospheric density The long-term forecasting module, the medium- and short-term forecasting module, and the short-term forecasting module were subjected to tiered verification. If the forecasting model showed prediction bias, the corresponding modules could be adjusted. Weighted adjustment based on the baseline:

[0051] ;

[0052] in To correct the results, These are the original predicted values. For spatiotemporal matching Reference value The weights are dynamic, ranging from 0 to 1.

[0053] The correction process satisfies three constraints to ensure physical consistency: ① The density of the thermospheric atmosphere decreases exponentially with altitude; ② The density increase during geomagnetic storms does not exceed 5 times that during calm periods; ③ The density field is globally spatially continuous, and the gradient between adjacent grids is ≤10% / 5. ° .

[0054] Compared with the prior art, the present invention has the following beneficial technical effects:

[0055] This invention utilizes the advantages of different gradient data sources (bottom-level TLE orbit data, mid-level satellite precise orbit data, and top-level accelerometer data) to construct a hierarchical and progressive refinement architecture ("bottom-level climatological foundation—mid-level variability correction—top-level instantaneous refinement"). The TLE low-precision orbit data provides the basic observations and coarse constraints for atmospheric density inversion, serving as the original data source for density estimation. The mid-level precise orbit data improves orbital position accuracy and enhances the reliability of density inversion, acting as a bridge between the two. The top-level accelerometer data, derived from direct observation of non-gravitational acceleration, provides the most accurate density calibration and serves as a verification benchmark. Furthermore, cross-source data weighted fusion of these three types of data effectively improves the precision of atmospheric density refinement.

[0056] This invention employs a three-tiered, layered refinement architecture design, fully leveraging the advantages of long-term, comprehensive coverage of TLE data, the mesoscale variability characterization of precise orbital data, and the high-frequency, high-precision advantages of accelerometer data. It avoids the limitations of single data sources, achieving accurate characterization of thermospheric atmospheric density across all scales and regions. Simultaneously, it constructs a multi-scale prediction system ("long-term climate prediction—medium- and short-term variability prediction—short-term instantaneous prediction"), combining cross-source data calibration and physical constraint modeling to achieve accurate predictions of thermospheric atmospheric density adaptable to multiple scenarios. This multi-scale prediction system can simultaneously meet the application needs of different time scales, including long-term prediction (monthly-yearly scale), medium- and short-term prediction (day- and weekly scale), and short-term prediction (minute- and hourly scale), improving refinement accuracy and prediction adaptability.

[0057] This invention integrates multi-source data, including TLE data, precise orbit data, and accelerometer data, and combines them with physical constraints. Based on low-cost TLE data, it uses high-precision data only in key stages, significantly reducing data acquisition and computation costs. It achieves high-precision refinement of thermospheric atmospheric density across all scales and regions, and matches the needs of different time scales through a segmented model. It can improve the accuracy and predictive adaptability of thermospheric atmospheric density refinement, reduce data costs, and can be applied to scenarios such as thermospheric climatology, satellite orbit prediction, and space debris modeling, possessing innovation and engineering value.

[0058] The prediction model of this invention incorporates thermospheric atmospheric physical constraints to avoid the physical irrationality of purely data-driven models, thereby improving the reliability and credibility of the prediction results. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the overall architecture of the thermospheric atmospheric density refinement and multi-scale prediction of the present invention;

[0060] Figure 2 This is a schematic diagram of the prediction module architecture of the present invention;

[0061] Figure 3 This is a schematic flowchart of the method of the present invention;

[0062] Figure 4 This is a curve comparing the predicted and actual density data of the thermosphere in this invention; the horizontal axis represents the sample points, and the vertical axis represents the thermosphere density.

[0063] Among them, 100 is the data input layer, 101 is TLE data, 102 is precise orbit data, 103 is accelerometer data, and 104 is space weather index data;

[0064] 200 is the data preprocessing layer, 201 is the TLE data preprocessing module, 202 is the precision orbit data preprocessing module, 203 is the accelerometer data preprocessing module, and 204 is the space weather index data preprocessing module.

[0065] 300 is the three-layer refinement layer, 301 is the bottom-layer climate state construction module, 302 is the middle-layer variability correction module, 303 is the top-layer instantaneous refinement module, and 304 is the cross-source fusion module;

[0066] The system has more than 400 density prediction layers, with 401 being the long-term prediction module, 402 the medium- and short-term prediction module, 403 the short-term prediction module, and 404 the physical constraint module. Detailed Implementation

[0067] The present invention will be further described in detail below with reference to embodiments. These descriptions are for illustrative purposes only and are not intended to limit the scope of the invention.

[0068] This invention constructs a three-level thermospheric atmospheric density refinement framework consisting of "bottom-level climate state foundation, mid-level variability correction, and top-level instantaneous refinement," and proposes a multi-scale collaborative prediction system consisting of "long-term climate prediction, medium- and short-term variability prediction, and short-term instantaneous prediction."

[0069] See Figure 1 A cross-source data fusion system for predicting thermospheric atmospheric density in a hierarchical, progressive, full-scale manner includes:

[0070] The data input layer receives multi-source data, including TLE data, precise orbit data, accelerometer data, and space weather index data.

[0071] The data preprocessing layer includes a TLE data preprocessing module, a precise orbit data preprocessing module, an accelerometer data preprocessing module, and a space weather index preprocessing module. These modules perform preprocessing on TLE data, precise orbit data, accelerometer data, and space weather index data, including removing invalid or abnormal data and filling in missing data, to form each dataset and send it to the corresponding module in the refinement layer.

[0072] The refinement layer comprises a three-tiered thermospheric atmospheric density refinement architecture consisting of a bottom-level climatology construction module, a middle-level variability correction module, and a top-level instantaneous refinement module. Each module receives its corresponding dataset, generates it, and outputs the climatology density to the cross-source fusion module. Middle layer corrected density and top-layer refinement density The cross-source fusion module assigns weights to the input data and performs weighted fusion to obtain the refined atmospheric density. , serving as training samples and validation benchmarks for the multi-scale density prediction layer prediction model;

[0073] The multi-scale density prediction layer includes a long-term prediction module, a medium- and short-term prediction module, a short-term prediction module, and a physical constraint module.

[0074] The long-term prediction module is based on climatological state density. Lasso regression models trained with solar activity data are used to predict long-term climatological density from month to year.

[0075] The short-to-medium-term prediction module is based on training during periods of high solar activity and has been... The validated XGBoost model is used for short- to medium-term prediction of thermospheric atmospheric density over days and weeks.

[0076] The short-term prediction module uses the constraints provided by the physical constraint module based on the LSTM model to perform short-term predictions of thermospheric atmospheric density from minutes to hours.

[0077] Further, see Figure 2 The long-term prediction module includes an input layer, a feature extraction layer, a Lasso regression model layer, and an output layer; wherein the input layer receives long-term solar activity forecast data, long-term geomagnetic activity statistics, seasonal parameters, and so on. The interannual / seasonal cyclical feature sequences are used to extract features related to solar activity levels, geomagnetic activity levels, and seasonality, which are then sent to the Lasso regression model layer. Constrained by climatological boundary conditions, the Lasso regression model sends the prediction results to the output layer;

[0078] The short-to-medium term prediction module includes an input layer, a feature extraction layer, an XGBoost model layer, and an output layer; the input layer receives 3-day Kp index forecasts, 7-day F10.7 index forecasts, and forecasts for the past 3 months. The time-series data and altitude / latitude / local time parameters are used to extract time-lag features and capture the time delay of the influence of space weather index on atmospheric density, which are then sent to the XGBoost model layer. The XGBoost model then sends the prediction results to the output layer.

[0079] The short-term prediction module includes an input layer, a feature extraction layer, an LSTM model layer, and an output layer. The input layer receives constraints provided by the physical constraint module, real-time space weather index, and instantaneous atmospheric density over the past hour. and The feature extraction layer obtains features including density frequency, density height attenuation constraint, and density increase limit and sends them to the LSTM model layer. The LSTM model then sends the prediction results to the output layer.

[0080] Figure 2 Medium time-lag features include extracting features of the F10.7 index with a 3-day lag and features of the Kp index with a 2-hour lag. These features are used to capture the time delay effect of space weather indices on atmospheric density. Physical features include extracting features related to solar activity levels and geomagnetic activity levels. These features are used to introduce prior physical knowledge and enhance the physical interpretability of the model for changes in atmospheric density.

[0081] See Figure 1 , Figure 3 This invention provides a method for hierarchical and progressive full-scale prediction of thermospheric atmospheric density through cross-source data fusion, characterized by the following operations:

[0082] The data input layer receives multi-source data, including TLE data, precise orbit data, accelerometer data, and space weather index data.

[0083] The data preprocessing layer performs preprocessing on TLE data, precise orbit data, accelerometer data and space weather index data, including removing invalid or abnormal data and filling in missing data, to form each dataset and send it to the corresponding module of the refinement layer.

[0084] The bottom-level climatology construction module, the middle-level variability correction module, and the top-level instantaneous refinement module of the refinement layer receive the TLE dataset, the precise orbit dataset, and the accelerometer dataset, respectively, and perform density refinement at the corresponding levels; and output the climatology density to the cross-source fusion module. Middle layer corrected density and top-layer refinement density The cross-source fusion module assigns weights to the input data and performs weighted fusion to obtain the refined atmospheric density. ;

[0085] The long-term, medium-term, and short-term prediction modules of the multi-scale density prediction layer provide long-term climatological density predictions, medium-term and short-term predictions, and short-term predictions, respectively.

[0086] The long-term forecast is based on climatological density. The Lasso regression model, trained with solar activity data, is used for prediction. The inputs to the model include long-term solar activity forecasts, long-term geomagnetic activity statistics, seasonal parameters, and climatological state density. Interannual / seasonal cyclical characteristic sequences, in terms of climatological density To constrain climatological boundary conditions, we conduct monthly-year long-term climatological state density predictions.

[0087] The aforementioned medium- and short-term forecasts are based on training during periods of high solar activity and refined atmospheric density. The validated XGBoost model was used for prediction, with 3-day Kp exponent forecasts, 7-day F10.7 exponent forecasts, and mid-level corrected densities over the past 3 months as inputs. Using time-series data and altitude / latitude / local time parameters, we can make short- to medium-term predictions of thermospheric atmospheric density over days and weeks.

[0088] The short-term forecast is based on constraints provided by the physical constraint module of the LSTM model. The real-time space weather index and the instantaneous atmospheric density of the past hour are input into the model. and top-layer refining density The parameters are updated in real time using a 1-hour sliding window to predict the thermospheric atmospheric density over 30 minutes to 2 hours.

[0089] The following is a detailed explanation of each part.

[0090] 1. Multi-source data preprocessing

[0091] The data preprocessing layer performs standardization preprocessing on the input multi-source data to form a high-quality dataset, including:

[0092] TLE data preprocessing: Collect TLE data covering multiple solar activity cycles globally, and filter valid target data with orbital altitudes of 200-800km and inclinations of 0°-98°; then, combine the solar activity background and use the 3σ criterion to remove outlier data to form a TLE dataset.

[0093] Preprocessing of satellite precise orbit data: Obtain GPS precise orbit data and corresponding laser ranging verification data from multiple satellites, eliminate gross errors through orbit fitting, ensure orbit position accuracy is better than 1m, and form a standardized precise orbit dataset.

[0094] Accelerometer data preprocessing: Obtain raw data from the onboard accelerometer, combine it with satellite attitude data to remove invalid data, and use a low-pass filtering algorithm to filter out sensor noise to obtain a smooth acceleration dataset.

[0095] Space weather index preprocessing: collect solar activity index (F10.7), geomagnetic activity index (Kp, Dst, AE) and solar wind parameters, unify the time base, and use linear interpolation to fill in the missing data to form a space weather index dataset.

[0096] Space weather index datasets are the core driver or input for refining and predicting thermospheric atmospheric density. They quantify the energy injection of solar and geomagnetic activity into the thermosphere, directly determining the long-term trend, short-term fluctuations, and response to extreme events of density. They provide multi-dimensional statistics for the refined density in this invention and provide important input parameters for multi-scale collaborative prediction.

[0097] 2. Layered integration and refinement

[0098] A three-tiered, hierarchical, fusion-refinement architecture, consisting of a bottom-level climatology construction module, a middle-level variability correction module, and a top-level instantaneous refinement module, is constructed. This architecture is then weighted and fused using a cross-source fusion module to achieve full-scale thermospheric atmospheric density refinement.

[0099] (1) Bottom layer: Construction of climatological density of state field driven by TLE data

[0100] The bottom-level climatology construction module uses the semi-major axis attenuation method to invert the thermospheric atmospheric density corresponding to TLE data:

[0101] (1)

[0102] in, Density of the thermosphere (unit: kg / m³). The rate of change of the semi-major axis of the satellite orbit (unit: m / s); The drag coefficient (range 2.0-3.0, preferably 2.5); Satellite surface mass ratio (unit: m² / kg); The semi-major axis of the satellite orbit (unit: m).

[0103] The thermospheric atmospheric density based on inversion adopts a joint stratification strategy according to the solar activity level, season, and geomagnetic activity level. The global region is divided into grids, and the average density of each grid is calculated to obtain the initial climatological density field. Then, the initial climatological density field is decomposed by EOF, and the first three principal modes are extracted and reconstructed to obtain the climatological density. .

[0104] Specifically, according to the solar activity level (low year: F10.7 < 100 sfu; middle year: 100 ≤ F10.7 ≤ 200 sfu; high year: F10.7 > 200 sfu), season (spring: March - May; summer: June - August; autumn: September - November; winter: December - February), and geomagnetic activity level (quiet: Kp ≤ 3; active: 3 < Kp ≤ 5; magnetic storm: Kp > 5) data, a joint stratification strategy is adopted for the solar activity level, season, and geomagnetic activity level, and multi-dimensional classification statistics are carried out on the inverted density data. Among them, the solar activity level, season, and geomagnetic activity level do not simply combine, but jointly serve as stratification conditions. The global region is divided into grids with 5°×5°, and the average density of each grid is calculated to obtain the initial climatological density field. Then, the initial field is decomposed by EOF, and the first three principal modes (cumulative variance interpretation rate ≥ 95%) are extracted and reconstructed to obtain the optimized climatological density. . (2) Middle layer: Variability correction driven by precise orbit data

[0105] The middle layer variability correction module inversely calculates the atmospheric density corresponding to the precise orbit data based on the interpolation differential method, and uses the accelerometer data to establish a linear deviation correction model to calibrate the inversion result. The following calibration is obtained by least squares fitting:

[0106] (2)

[0107] Among them, is the density inverted from the calibrated precise orbit; is the atmospheric density inverted from the precise orbit; is the proportional correction coefficient (value range 0.05 - 0.12); is the constant correction coefficient (value range 10 - ¹ 7 -10 - ¹ 6 kg / m³).

[0108] Furthermore, and are spatiotemporally matched (time matching accuracy ≤ 10 minutes, space matching accuracy ≤ 0.5°), and the deviation field is calculated; the Kriging interpolation method considering the correlation between geomagnetic latitude and local time is used to extend the deviation field to the global 5°×5° grid; the weight coefficient (200-400km: =0.8; 400-600km: =0.6; 600-800km: =0.4), through Obtain the corrected density of the middle layer .

[0109] (3) Top layer: Instantaneous refinement driven by accelerometer data

[0110] The top-level instantaneous refinement module retrieves instantaneous atmospheric density based on preprocessed accelerometer data and atmospheric drag experienced by the satellite:

[0111] (3)

[0112] in, Instantaneous atmospheric density; The atmospheric drag experienced by the satellite (unit: N) is calculated by multiplying the acceleration data by the satellite's mass. Satellite area (unit: m²) 2 ), This refers to the satellite's orbital velocity (unit: m / s).

[0113] Within the accelerometer data coverage area (spatial accuracy ≤1°), instantaneous atmospheric density As the top-level refinement density in this region ; and directly use replace Corresponding grid value;

[0114] In regions with sparse data, according to Obtain the top-level refined density field;

[0115] in, The top-level refined density field of the target mesh; The measured instantaneous density of the neighboring region (distance ≤ 10°) is used to capture instantaneous disturbance characteristics; The climatological density field serves as the benchmark for density refinement. To extrapolate the weights, values ​​are dynamically assigned based on the development stage of the geomagnetic storm, balancing the accuracy of instantaneous disturbances in the measured data with the stability of the climate field; for example, weights are set for the main phase of the geomagnetic storm. Restore phase settings . It is an extrapolation weight of instantaneous disturbance density set based on different development stages of geomagnetic storms (main phase, recovery phase), used to adjust the adjacent area. Contribution to the calculation of instantaneous perturbation density of the target grid.

[0116] (4) Weighted fusion of cross-source data

[0117] Construct a top-down calibration chain for "accelerometer data → precision orbit data → TLE data":

[0118] This invention adopts a top-down calibration chain of "accelerometer data → precise orbit data → TLE data". The uncertainty of the three levels of data is as follows from low to high: accelerometer data (direct density inversion, with the smallest error and the lowest uncertainty) < calibrated precise orbit data (indirect density derivation, with reduced error after calibration) < climatological density data (statistical mean field, with weak timeliness and the highest uncertainty).

[0119] The cross-source fusion module integrates the three-level refinement results by "determining the weight allocation principle based on data uncertainty".

[0120] (4)

[0121] in, To refine atmospheric density; , , The weights for accelerometer data, calibrated precise orbit data, and climatological density of state data are respectively determined to satisfy the following conditions. .

[0122] The weight allocation based on data uncertainty follows an inverse principle: the greater the data uncertainty, the smaller the assigned weight; conversely, the smaller the data uncertainty, the larger the assigned weight. This is not dynamic weighting, but rather determining the optimal fixed weight combination based on the inherent uncertainty levels of the three types of data. This balances fusion accuracy and computational efficiency, meeting the engineering application requirements for thermal density refinement. This invention is preferred. =0.6、 =0.3、 =0.1.

[0123] Fixed weights are a quantitative representation of uncertain ranking: optimal selection. =0.6、 =0.3、 =0.1, essentially transforming the principle of "the lower the uncertainty, the higher the weight" into a fixed value, representing the accelerometer density data with the lowest uncertainty. The corresponding data source) is assigned the highest weight (0.6), and the density data of the precise orbit inversion in the uncertainty period ( Assign a middle weight (0.3) to the climatological density data with the highest uncertainty. Assigning the lowest weight (0.1) satisfies the core logic of uncertainty weighting, simplifies subsequent calculations, and ensures the stability of the fusion result.

[0124] (Fusion and Refinement of Atmospheric Density) is the final output of thermospheric density refinement. Its core function is to integrate the advantages of three levels of data, make up for the shortcomings of a single data source, and obtain high-precision, high-reliability, and high-applicability thermospheric density data, providing the optimal input benchmark for subsequent thermospheric density prediction.

[0125] It possesses high accuracy, particularly through the integration of accelerometer data. ), stability of precise orbital data ( ), the fundamental nature of climate state data ( This avoids the limitations of a single data source (such as partial missing accelerometer data, insufficient response to disturbances in orbital data, and poor timeliness and accuracy of climatological data), and the overall accuracy is higher than the refinement results of any single data source.

[0126] It has high reliability and balances the contribution of data with different uncertainties through weight allocation, reducing the impact of errors from a single data source on the final result. Even if a certain type of data source has a local deviation, it can be offset by the weight adjustment of other data sources, thereby improving the stability and robustness of density data.

[0127] It has high applicability, and the fusion result takes into account both "instantaneous disturbance response" ( ) and "long-term stable background" It can be adapted to different scenarios of thermal density applications (such as low-orbit satellite orbit prediction, collision warning, and on-orbit lifetime estimation), while providing high-quality training and validation data for subsequent prediction models.

[0128] As a refined atmospheric density fusion, it serves as the core input and benchmark data for subsequent multi-scale density prediction of thermospheric atmospheric density. Specific application scenarios and methods are as follows:

[0129] As training samples for the prediction model: By matching the data with corresponding spatial weather indices (F10.7, Kp, Dst, etc.) and spatiotemporal parameters (latitude, longitude, altitude, season), a training dataset is constructed to train machine learning prediction models (such as LSTM). Compared to the refined results from a single data source, Its high precision feature can significantly improve the training effect of the model, enabling the model to learn the correlation between "space weather index and thermosphere density" more accurately.

[0130] As a validation benchmark for the prediction model: the prediction density output by subsequent prediction models needs to be compared with that of the same period. By comparing the models and calculating evaluation indicators such as RMSE, the accuracy of the prediction models can be quantified to determine whether the models meet the needs of practical applications.

[0131] For example, by integrating and refining atmospheric density Or, the instantaneous density was inverted using a spaceborne accelerometer. Using the RMSE core indicator as the benchmark, the long-term, medium-term, and short-term prediction modules were graded and verified. Based on engineering experience and actual usage requirements, the density prediction accuracy was ranked from high to low as follows: short-term density prediction accuracy → medium- and long-term density prediction accuracy → long-term density prediction accuracy. The accuracy requirements were: short-term density prediction accuracy RMSE ≤ 0.25, medium- and long-term density prediction accuracy RMSE ≤ 0.28, and long-term density prediction accuracy RMSE ≤ 0.3.

[0132] As a basis for correcting the prediction results: if subsequent prediction models show prediction deviations under extreme scenarios such as geomagnetic storms, the prediction results can be adjusted accordingly. As a benchmark, through weighted adjustments (such as prediction results and...) By fusing the data in a certain proportion, prediction errors can be reduced and the reliability of the final prediction results can be improved.

[0133] Fusion Refined Atmospheric Density It has the advantages of high precision, full spatiotemporal coverage, and strong physical consistency. It can be used as a correction benchmark for multi-scale prediction results to reduce prediction bias in scenarios such as extreme space weather, data sparsity, and long-term extrapolation.

[0134] 3. Multi-scale density prediction

[0135] A multi-scale collaborative prediction system is constructed by establishing long-term, medium- and short-term, and short-term prediction modules, respectively providing long-term climate density of state prediction, medium- and short-term prediction, and short-term prediction to meet the needs of different time scales:

[0136] Long-term forecasting (month-to-year scale): Constructing a Lasso regression model and using climatology density. The model is trained with solar activity data to achieve climate state-driven operation; parameters such as long-term solar activity forecast data are input into the model, and the climate state density field for the next 1-12 months is predicted with the EOF principal mode coefficient as the prediction target.

[0137] The climatological driving mechanism of the long-term prediction module in this invention is based on the long-term climatological evolution of the thermosphere, and is achieved through climatological density. End-to-end constraints ensure the stability and physical consistency of monthly-to-year scale forecasts:

[0138] ① Driving substrate: Based on covering multiple solar activity cycles The data can effectively and indirectly reflect the three core climatological laws of interannual solar activity variation, seasonal latitudinal differences, and geomagnetic polar disturbances, providing a physical anchor for density prediction.

[0139] ② Model-driven: Lasso regression model Using the principal mode coefficients decomposed by EOF as the prediction target, we can directly model and extrapolate the climatological evolution law, replacing single-grid density fitting, and achieving law-level prediction.

[0140] ③ Input driver: Historical time-series prior features are incorporated into the model input, forming a feature set with solar activity, geomagnetism, and seasonal parameters. The model is constrained by climatological laws to solve the prediction drift problem of traditional single-factor input and improve prediction stability.

[0141] ④ Output driver: based on The predicted density field is reconstructed using the EOF spatial basis functions. To constrain climatological boundary conditions, ensure that predictions conform to the long-term climate characteristics of the thermosphere, and guarantee the physical interpretability and engineering applicability of the results.

[0142] Short- to medium-term forecasts (day-week scale): Construct lightweight XGBoost models with different altitudes and geomagnetic activity levels, and combine them with transfer learning to improve forecast accuracy.

[0143] Short-term forecasting (minute-hour scale): An instantaneous refinement + physically constrained dynamic model is used. Real-time space weather index data is input to construct an LSTM model, and thermospheric physical constraints are introduced.

[0144] Thermospheric atmospheric density decreases exponentially with altitude:

[0145] (5)

[0146] in, Altitude (unit: km); The elevation of the thermosphere (unit: km) is used to ensure that the prediction results conform to the physical law that density decreases exponentially with altitude;

[0147] At the same time, the density increase during the geomagnetic storm period is constrained to be ≤5 times the density during the calm period to avoid physically unreasonable results.

[0148] 4. Results Output

[0149] The system outputs refined and predicted results for different application scenarios, including long-term climate density of state data for climatological research, full-scale refined density fields for thermospheric model calibration, short-term high-frequency prediction data for satellite orbit forecasting, and long-term prediction sequences for space debris modeling.

[0150] Example 1

[0151] The application scenario of this implementation is the refinement and multi-scale prediction of thermospheric atmospheric density in the altitude range of 200-800km. The required data includes: global TLE data from 1990 to the present; GPS precise orbit data (sampling interval of 10 minutes) and laser ranging verification data from CHAMP and Swarm satellites; raw data from onboard accelerometers of CHAMP and Swarm-C satellites (sampling frequency of 1Hz); and F10.7 index, Kp index, Dst index, AE index and solar wind parameters (velocity, density, Bz component) data from 1990 to the present.

[0152] Multi-source data preprocessing

[0153] TLE data preprocessing: TLE data covering more than three solar activity cycles (1990-2023), with orbital altitudes of 200-800 km and inclinations of 0°-98° were selected. Data with abnormal fluctuations in the semi-major axis of the orbit were removed using the 3σ criterion. Valid data segments for low, medium, and high solar activity years were selected using the F10.7 index. Finally, 1.2 million data entries were retained to form a standardized TLE dataset.

[0154] Precise orbit data preprocessing: Obtain GPS precise orbit data (SP3 format) from CHAMP satellite from 2005 to 2008 and Swarm satellite from 2014 to 2023, and simultaneously collect corresponding laser ranging verification data; use the orbit fitting function of the Insight-Spatial Information Analysis Basic Platform software to remove gross errors (data with orbit position deviation > 1m) to ensure that the orbit position accuracy is better than 1m, forming a standardized precise orbit dataset.

[0155] Accelerometer data preprocessing: Obtain raw accelerometer data from CHAMP and Swarm-C satellites, and combine it with satellite attitude data to remove invalid data during attitude flip periods (attitude angle change >10° / s); use a low-pass filter with a cutoff frequency of 0.01Hz to filter out sensor noise, and obtain a smooth acceleration data sequence with a time resolution of 5 minutes.

[0156] Space weather index preprocessing: F10.7 index, Kp index, Dst index, AE index and solar wind parameters from 1990 to 2023 were collected. The time base was unified as UTC. Linear interpolation was used to fill in missing data (missing rate ≤5%) to form a standardized space weather index dataset with a time resolution of 1 hour.

[0157] 1.1. Three-level density hierarchical fusion refinement

[0158] 1.1.1. Construction of the underlying climatological density field

[0159] The thermosphere atmospheric density of TLE data is retrieved using formula (1), where the drag coefficient is... Take 2.5, the satellite area-to-mass ratio Use the publicly available parameters of each satellite; for the problem of low TLE time accuracy, use a 2-hour sliding window average to process the inversion results, and smooth the original inverted density sequence in the time dimension. The steps are as follows:

[0160] Step a: Time series arrangement. Sort all the density values inverted from TLEs according to the epoch time to form a one-dimensional time series ρ(t i ), where t i is the time of the i-th data point;

[0161] Step b: Sliding window definition. Taking the current time t as the center, take a time range of 1 hour before and after (a total of 2 hours) as the sliding window. The window time range is: ;

[0162] Step c: Data screening within the window. Find all the original density inversion values ρ(t j ) whose timestamps fall within this window;

[0163] Step d: Mean calculation. Take the arithmetic mean of all ρ(t j ) within the window as the final density value at the center time t of this window;

[0164] Step e: Window sliding. Slide the window at a fixed time step (such as 10 minutes), and repeat steps 2 - 4. Finally, obtain a continuous and smooth 2-hour sliding average density sequence.

[0165] According to the solar activity level (low year: F10.7 < 100 sfu; middle year: 100 ≤ F10.7 ≤ 200 sfu; high year: F10.7 > 200 sfu), season (spring: March - May; summer: June - August; autumn: September - November; winter: December - February), and geomagnetic activity level (quiet: Kp ≤ 3; active: 3 < Kp ≤ 5; magnetic storm: Kp > 5) data, adopt a joint stratification strategy for the solar activity level, season, and geomagnetic activity level, and conduct multi-dimensional classification statistics on the inverted density data. The three are not simply combined, but jointly used as stratification conditions. Use a 5°×5° grid to divide the global region, calculate the average density of each grid to obtain the initial climatological density field. Then perform EOF decomposition on the initial field, extract the first 3 main modes (for example, the global background density reference mode dominated by solar activity, the latitude difference mode dominated by season, and the polar region perturbation mode dominated by geomagnetic activity), and reconstruct to obtain an optimized climatological density field .

[0166] 1.1.2. Middle atmosphere variability correction

[0167] Inversion of atmospheric density based on interpolation differential method for precise orbit data Density sequences with a temporal resolution of 10 minutes were obtained. The overlapping orbital period of the CHAMP satellite from 2005 to 2008 was selected. Density inversion with accelerometer In comparison, the correction coefficients in formula (2) are obtained by fitting using the least squares method. =0.08、 =5×10 -17 kg / m³, completed The calibration was obtained .

[0168] Will and Perform spatiotemporal matching (time matching accuracy ≤ 10 minutes, spatial matching accuracy ≤ 0.5°) and calculate the deviation field. The bias field was extended to a global 5°×5° grid using Kriging interpolation that considers the correlation between geomagnetic latitude and local time; weighting coefficients were set according to altitude (200-400km). =0.8; 400-600km: =0.6; 600-800km: =0.4), through Obtain the middle layer corrected density field .

[0169] 1.1.3. Top-level instantaneous refinement

[0170] Based on the preprocessed accelerometer data, the instantaneous atmospheric density is inverted using formula (3). ,in Calculated by multiplying acceleration data by satellite mass. The satellite orbital velocity is extracted from precise orbital data. 3σ anomaly detection is used to remove bad data, and data quality is ensured through consistency checks between adjacent orbital segments (interval ≤ 30 minutes) with an error ≤ 5%.

[0171] Within the accelerometer data coverage area (spatial accuracy ≤1°), use directly replace Corresponding grid values; in sparse data regions, construct By weighting the data, the accuracy of instantaneous perturbations in the measured data is balanced with the stability of the climate field, ultimately yielding a high-precision top-level refined density field. Extrapolation weights are set for the main phase of the geomagnetic storm. Restore phase settings .

[0172] 1.1.4. Cross-source weighted fusion

[0173] Set the weights according to formula (4) =0.6、 =0.3、 =0.1, the fused refined atmospheric density is calculated. Complete full-scale density refinement.

[0174] 1.2. Multi-scale collaborative prediction

[0175] 1.2.1. Long-term forecasts (monthly to yearly scale)

[0176] Build a Lasso regression model using the sklearn machine learning library;

[0177] The Lasso regression model was trained using F10.7 index data, geomagnetic Dst data, seasonal parameters, and other data. The interannual / seasonal cyclical feature sequences are used as input features, and the historical climatological density is output using the low-level climatological construction module. Training was performed using sequence and solar activity data, and with To constrain the climatological boundary conditions, the model was trained, and the regularization parameter λ was optimized through cross-validation, with the optimal value determined to be 0.01.

[0178] Input features include: ① long-term solar activity forecast data (1-year forecast of F10.7 index); ② long-term geomagnetic activity statistics (Dst data); ③ seasonal parameters; ④ historical data. The interannual / seasonal cyclical characteristic sequences are classified into four categories.

[0179] Prediction and reconstruction, based on the training phase The EOF basis functions obtained from the decomposition are used to reconstruct the predicted principal mode coefficients, resulting in a long-term climatological density of state prediction field for the next 1-12 months.

[0180] 1.2.2. Medium- and short-term forecasts (daily to weekly scale)

[0181] We constructed a lightweight XGBoost open-source model based on altitude (200-400km, 400-600km, 600-800km) and geomagnetic activity level. We set the number of decision trees to 100 and the maximum depth to 6. The input features include time lag items with F10.7 lag of 3 days and Kp lag of 2 hours (the time lag length is preset based on the time dimension of the input data to extract the corresponding lag features).

[0182] Using transfer learning, the model trained during a period of high solar activity (2013) was transferred to the current forecast scenario, and atmospheric density was refined using fusion. Density verification can be performed by inverting measured accelerometer data from satellites such as CHAMP and Swarm.

[0183] After successful verification, input the 3-day Kp index forecast, the 7-day F10.7 index forecast, and the forecast values ​​for the past 3 months into the model. Using time-series data and altitude / latitude / local time parameters, we can make short- to medium-term forecasts.

[0184] 1.2.3. Short-term forecasting (minute-hour scale)

[0185] A physical constraint + LSTM hybrid model was constructed. The LSTM network was set with 2 hidden layers (64 neurons per layer). The density height decay constraint of formula (5) and the constraint condition that the density increase during the magnetic storm period is ≤ 5 times that during the calm period were introduced.

[0186] Enter the real-time space weather index (AE, Dst index, updated every 1 minute) and the past hour's data. Time series data and The model parameters are updated in real time using a 1-hour sliding window to predict the thermospheric atmospheric density for the next 30 minutes to 2 hours.

[0187] 1.3. Results Output and Effect Verification

[0188] Output results according to application scenarios: output long-term climatological density of state data in NetCDF format to climatological research scenarios; output full-scale refined density fields including uncertainty intervals to thermospheric model calibration scenarios; output short-term high-frequency prediction data with an update frequency of 5 minutes to satellite orbit forecasting and collision warning scenarios; and output long-term density prediction sequences to space debris environment modeling scenarios.

[0189] Density data retrieved from measured accelerometers on satellites such as CHAMP and Swarm, or (Refined atmospheric density) Verification: Calculations were performed at altitudes ranging from 200 to 800 km. The root mean square error (RMSE) of long-term forecasts (12 months), medium- and short-term forecasts (7 days), and short-term forecasts (2 hours). Figure 4 A comparison curve of the predicted density of the thermosphere and the measured density data of the present invention is given, with RMSE = 0.229.

[0190] If the accuracy requirements are met, the result is output; otherwise, the density prediction result is corrected.

[0191] ;

[0192] in To correct the results, These are the original predicted values. For spatiotemporal matching Reference value The dynamic weights, ranging from 0 to 1, are dynamically adjusted considering factors such as forecast duration, intensity of space weather disturbances, and historical model accuracy. The correction process must satisfy three constraints to ensure physical consistency: ① conforming to the exponential decay law of thermospheric atmospheric density with altitude; ② density increase during geomagnetic storms not exceeding 5 times that during calm periods; ③ global spatial continuity of the density field, with gradients between adjacent grids ≤ 10% / 5. ° .

[0193] The embodiments given above are preferred examples for implementing the present invention, and the present invention is not limited to the above embodiments. Any non-essential additions or substitutions made by those skilled in the art based on the technical features of the present invention are within the protection scope of the present invention.

Claims

1. A cross-source data fusion system for predicting thermospheric atmospheric density in a hierarchical, progressive, full-scale manner, characterized in that... include: The data input layer receives multi-source data, including TLE data, precise orbit data, accelerometer data, and space weather index data. The data preprocessing layer includes a TLE data preprocessing module, a precise orbit data preprocessing module, an accelerometer data preprocessing module, and a space weather index preprocessing module. These modules perform preprocessing on TLE data, precise orbit data, accelerometer data, and space weather index data, including removing invalid or abnormal data and filling in missing data, to form each dataset and send it to the corresponding module in the refinement layer. The refinement layer comprises a three-tiered thermospheric atmospheric density refinement architecture consisting of a bottom-level climatology construction module, a middle-level variability correction module, and a top-level instantaneous refinement module. It receives datasets from TLE data, precise orbit data, and accelerometer data to generate and output climatology density to the cross-source fusion module. Middle layer corrected density and top-layer refinement density ; The cross-source fusion module assigns weights to the input data and performs weighted fusion to obtain the refined atmospheric density. , serving as training samples and validation benchmarks for the multi-scale density prediction layer prediction model; The multi-scale density prediction layer includes a long-term prediction module, a medium- and short-term prediction module, a short-term prediction module, and a physical constraint module. The long-term prediction module is based on climatological state density. A Lasso regression model trained with solar activity data from space weather index data, using long-term solar activity forecast data, long-term geomagnetic activity statistics, seasonal parameters, and climatological state density. The interannual / seasonal cyclical feature sequences are used as input features, and the output is the long-term climatological state density prediction results from month to year. The short-to-medium-term prediction module is based on an XGBoost model trained during a period of high solar activity. This model has been further refined. Verification; Inputs are the 3-day Kp index forecast, the 7-day F10.7 index forecast, and the past 3 months' data. Using time-series data and altitude / latitude / local time parameters, output short- to medium-term forecasts of thermospheric atmospheric density over days and weeks; The short-term forecasting module is based on an LSTM model and incorporates constraints provided by a physical constraint module. The inputs are the real-time space weather index and the instantaneous atmospheric density over the past hour. and top-layer refining density Output short-term prediction results of thermospheric atmospheric density in minutes to hours; The underlying climatology construction module is used to refine the underlying density and obtain the climatology density. This includes the following refining operations: The thermospheric atmospheric density corresponding to TLE data was retrieved using the semi-major axis attenuation method. Based on the retrieved thermospheric density, a joint stratification strategy was adopted according to solar activity level, season, and geomagnetic activity level. The global region was divided into grids, and the average density of each grid was calculated to obtain the initial climatological density field. Then, EOF decomposition was performed on the initial climatological density field to extract the first three principal modes and reconstruct the climatological density. ; The mid-layer variability correction module performs mid-layer density variability correction and obtains the corrected mid-layer density. This includes the following refining operations: The atmospheric density corresponding to the precise orbital data was retrieved using the interpolation differential method, and a linear deviation correction model was established using accelerometer data to calibrate the retrieval results. The following calibration was obtained by fitting the data using the least squares method: ； in, To achieve precise orbit inversion density after calibration, Atmospheric density derived from precise orbit inversion; This is the proportional correction factor. These are constant correction coefficients; Will and Perform spatiotemporal matching and calculate the deviation field. The bias field is extended to a global area, with weighting coefficients set according to altitude. ,pass Obtain the corrected density of the middle layer ; The top-level instantaneous refinement module performs instantaneous refinement of the top-level density to obtain the refined top-level density. This includes the following refining operations: Instantaneous atmospheric density was retrieved based on accelerometer data and atmospheric drag experienced by the satellite: ； in, Instantaneous atmospheric density, Due to atmospheric drag on the satellite, For satellite area, For satellite orbital velocity, This is the drag coefficient; In the area covered by accelerometer data, with instantaneous atmospheric density As the top-level refinement density in this region ; In regions with sparse data, according to Obtain the top-level refined density field, where The extrapolation weights are dynamically determined based on the development stage of the geomagnetic storm; The physical constraint module provides the following constraints to the LSTM model: Thermospheric atmospheric density decreases exponentially with altitude: ;in, For height, This refers to the elevation of the thermosphere. At the same time, the density increase during geostorm periods should be ≤ 5 times the density during calm periods.

2. The thermospheric atmospheric density hierarchical progressive full-scale prediction system based on cross-source data fusion as described in claim 1, characterized in that, The long-term prediction module includes an input layer, a feature extraction layer, a Lasso regression model layer, and an output layer; the input layer receives long-term solar activity forecast data, long-term geomagnetic activity statistics, and seasonal parameters. The interannual / seasonal cyclical feature sequences are used to extract features related to solar activity levels, geomagnetic activity levels, and seasonality, which are then sent to the Lasso regression model layer. Constrained by climatological boundary conditions, the Lasso regression model sends the prediction results to the output layer; The short-to-medium term prediction module includes an input layer, a feature extraction layer, an XGBoost model layer, and an output layer; the input layer receives 3-day Kp index forecasts, 7-day F10.7 index forecasts, and forecasts for the past 3 months. The time-series data and altitude / latitude / local time parameters are used to extract time-lag features and capture the time delay of the influence of space weather index on atmospheric density, which are then sent to the XGBoost model layer. The XGBoost model then sends the prediction results to the output layer. The short-term prediction module includes an input layer, a feature extraction layer, an LSTM model layer, and an output layer. The input layer receives constraints provided by the physical constraint module, real-time space weather index, and instantaneous atmospheric density over the past hour. and The feature extraction layer obtains features including density frequency, density height attenuation constraint, and density increase limit and sends them to the LSTM model layer. The LSTM model then sends the prediction results to the output layer.

3. A method for hierarchical, progressive, full-scale prediction of thermospheric atmospheric density through cross-source data fusion, characterized in that... Includes the following operations: The data input layer receives multi-source data, including TLE data, precise orbit data, accelerometer data, and space weather index data. The data preprocessing layer performs preprocessing on TLE data, precise orbit data, accelerometer data and space weather index data, including removing invalid or abnormal data and filling in missing data, to form each dataset and send it to the corresponding module of the refinement layer. The bottom-level climate state construction module, the middle-level variability correction module, and the top-level instantaneous refinement module of the refinement layer receive the TLE dataset, the precise orbit dataset, and the accelerometer dataset, respectively, to carry out density refinement at the corresponding levels. And output climate state density to the cross-source fusion module respectively. Middle layer corrected density and top-layer refinement density ; The cross-source fusion module assigns weights to the input data and performs weighted fusion to obtain the refined atmospheric density. ; The long-term, medium-term, and short-term prediction modules of the multi-scale density prediction layer provide long-term climatological density predictions, medium-term and short-term predictions, and short-term predictions, respectively. The long-term forecast is based on climatological density. The Lasso regression model, trained with solar activity data, is used for prediction. The inputs to the model include long-term solar activity forecasts, long-term geomagnetic activity statistics, seasonal parameters, and climatological state density. Interannual / seasonal cyclical characteristic sequences, in terms of climatological density With climatological boundary conditions as constraints, output the monthly-year long-term climatological density of state prediction results; The aforementioned short-to-medium-term forecasts are based on an XGBoost model trained during a period of high solar activity, and refined using atmospheric density fusion. Validation was performed by inputting 3-day Kp index forecasts, 7-day F10.7 index forecasts, and mid-level correction densities for the past 3 months into the model. Using time-series data and altitude / latitude / local time parameters, output short- to medium-term forecasts of thermospheric atmospheric density over days and weeks; The short-term forecast is based on constraints provided by the physical constraint module of the LSTM model. The real-time space weather index and the instantaneous atmospheric density of the past hour are input into the model. and top-layer refining density The parameters are updated in real time using a 1-hour sliding window, and the thermospheric atmospheric density prediction results are output for 30 minutes to 2 hours. The underlying climatology construction module is used to refine the underlying density and obtain the climatology density. This includes the following refining operations: The thermospheric atmospheric density corresponding to TLE data was retrieved using the semi-major axis attenuation method. Based on the retrieved thermospheric density, a joint stratification strategy was adopted according to solar activity level, season, and geomagnetic activity level. The global region was divided into grids, and the average density of each grid was calculated to obtain the initial climatological density field. Then, EOF decomposition was performed on the initial climatological density field to extract the first three principal modes and reconstruct the climatological density. ; The mid-layer variability correction module performs mid-layer density variability correction and obtains the corrected mid-layer density. This includes the following refining operations: The atmospheric density corresponding to the precise orbital data was retrieved using the interpolation differential method, and a linear deviation correction model was established using accelerometer data to calibrate the retrieval results. The following calibration was obtained by fitting the data using the least squares method: ； in, To achieve precise orbit inversion density after calibration, Atmospheric density derived from precise orbit inversion; This is the proportional correction factor. These are constant correction coefficients; Will and Perform spatiotemporal matching and calculate the deviation field. The bias field is extended to a global area, with weighting coefficients set according to altitude. ,pass Obtain the corrected density of the middle layer ; The top-level instantaneous refinement module performs instantaneous refinement of the top-level density to obtain the refined top-level density. This includes the following refining operations: Instantaneous atmospheric density was retrieved based on accelerometer data and atmospheric drag experienced by the satellite: ； in, Instantaneous atmospheric density, Due to atmospheric drag on the satellite, For satellite area, For satellite orbital velocity, This is the drag coefficient; In the area covered by accelerometer data, with instantaneous atmospheric density As the top-level refinement density in this region ; In regions with sparse data, according to Obtain the top-level refined density field, where The extrapolation weights are dynamically determined based on the development stage of the geomagnetic storm; The physical constraint module provides the following constraints to the LSTM model: Thermospheric atmospheric density decreases exponentially with altitude: ;in, For height, This refers to the elevation of the thermosphere. At the same time, the density increase during geostorm periods should be ≤ 5 times the density during calm periods.

4. The method for hierarchical and progressive full-scale prediction of thermospheric atmospheric density through cross-source data fusion as described in claim 3, characterized in that, The preprocessing includes: TLE data preprocessing: Collect TLE data covering multiple solar activity cycles, filter valid target data, use the 3σ criterion to remove outlier data, and combine the solar activity background to filter valid data segments to form a standardized TLE dataset; Precise orbit data preprocessing: acquire satellite GPS precise orbit data and corresponding laser ranging verification data, perform format standardization conversion, remove gross errors through orbit fitting, and form a standardized precise orbit dataset; Accelerometer data preprocessing: Obtain raw data from the onboard accelerometer, combine it with satellite attitude data to remove invalid data, use a low-pass filtering algorithm to filter out sensor noise, and form a standardized acceleration dataset; Space weather index preprocessing: Collect solar activity index, geomagnetic activity index and solar wind parameters, unify the time base, use linear interpolation to fill in missing data, and form a standardized space weather index dataset.

5. The method for hierarchical, progressive, full-scale prediction of thermospheric atmospheric density through cross-source data fusion as described in claim 3, characterized in that... The cross-source fusion module obtains the fused and refined atmospheric density. for: ； in, , , They are respectively , , The weights are determined based on data uncertainty and satisfy the following conditions: .

6. The method for hierarchical and progressive full-scale prediction of thermospheric atmospheric density through cross-source data fusion as described in claim 3, characterized in that, The Lasso regression model is trained as follows: Using F10.7 index data, geomagnetic Dst data, seasonal parameters, The interannual / seasonal cyclical feature sequences are used as input features, and the historical climatological density is output using the low-level climatological construction module. Training was performed using sequence and solar activity data, and with To constrain the climatological boundary conditions, the model was trained. The aforementioned monthly-year long-term climatological density forecast is a long-term climatological density forecast for the next 1-12 months: Based on the training phase The EOF basis functions obtained from the decomposition are used to reconstruct the predicted principal mode coefficients and output the long-term climatological density of state prediction field for the next 1-12 months.

7. The method for hierarchical, progressive, full-scale prediction of thermospheric atmospheric density through cross-source data fusion as described in claim 3, characterized in that... The XGBoost model is trained as follows: The number and maximum depth of decision trees in the XGBoost model are set, and the input features include time lag terms with F10.7 lag of 3 days and Kp lag of 2 hours. Transfer learning is used to transfer the model trained during periods of high solar activity to the current forecast scenario. Verification is performed, and training is completed once the verification is successful.

8. The method for hierarchical and progressive full-scale prediction of thermospheric atmospheric density through cross-source data fusion as described in claim 3, characterized in that, Also based on the fusion and refinement of atmospheric density The long-term forecasting module, the medium-to-short-term forecasting module, and the short-term forecasting module were subjected to tiered verification. If the output results of the forecasting model did not meet the accuracy requirements, the corresponding module would be re-verified. Weighted adjustment based on the baseline: ； in To correct the results, These are the original predicted values. For spatiotemporal matching Reference value The weights are dynamic, ranging from 0 to 1. The correction process satisfies three constraints to ensure physical consistency: ① The density of the thermospheric atmosphere decreases exponentially with altitude; ② The density increase during geomagnetic storms does not exceed 5 times that during calm periods; ③ The density field is globally spatially continuous, and the gradient between adjacent grids is ≤10% / 5. ° .

Images