Method and system for refining spatial and temporal resolution of ionospheric regions
By using a BP neural network model and multi-source data fusion technology, the problem of insufficient accuracy and resolution in ionospheric modeling in the ocean and low-to-mid latitude regions was solved, enabling rapid acquisition of high-precision ionospheric delay and reducing the demand for computing resources.
Patent Information
- Application Number
- CN202511605144.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-05
- Publication Date
- 2026-03-17
- Estimated Expiration
- 2045-11-05
AI Technical Summary
Existing ionospheric modeling technologies suffer from data gaps in the ocean and some GNSS stations, as well as insufficient accuracy and spatiotemporal resolution during geomagnetic storms and in mid-to-low latitude regions, failing to meet the requirements for high-precision ionospheric modeling.
By employing a BP neural network model and combining multi-source data fusion technology, and utilizing real-time GIM data, solar activity index, and geomagnetic activity index, the spatiotemporal resolution of the ionosphere is refined. High-precision STEC values are calculated through data preprocessing, interpolation, and projection functions to construct a high spatiotemporal resolution ionospheric model.
It enables the rapid generation of high-precision regional ionospheric models without relying on original modeling data sources, reducing computational resource requirements and improving the accuracy and resolution of ionospheric modeling, especially in terms of ionospheric delay acquisition capabilities during geomagnetic storms and in mid-to-low latitude regions.
Smart Images

Figure CN121071493B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method and system for refining the spatiotemporal resolution of the regional ionosphere. Background Technology
[0002] CODE, ESA, NRCan, and WHU use GIMs developed based on 15×15 order spherical harmonic function models, but each has differences in processing details. JPL adopts a linear combination method based on bicubic spline interpolation within 1280 spherical triangular tiles to achieve ionospheric coverage, equivalent to the distribution of 642 vertices. UPC independently calculates TEC, using a multi-layer tomographic scanning model combined with a specific interpolation scheme. CAS's GIMs rely on a combination of spherical harmonic functions and generalized trigonometric series functions to integrate data at global and local scales. IGS's final GIM product is derived by synthesizing GIMs independently calculated by different IAACs and assigning appropriate weights. Traditional ionospheric modeling is mainly based on GPS or GPS / GLONASS measurement data; however, with the development of multiple GNSS systems, especially the introduction of the BeiDou Navigation Satellite System (BDS), more GNSS observation data will promote the improvement of the spatial coverage and accuracy of ionospheric modeling. Studies have shown that using BeiDou CORS network data for ionospheric monitoring, analyzing the impact of BeiDou on ionospheric modeling, and constructing a global ionospheric model based on BDS / GPS dual-system observations all demonstrate the contribution of BDS in improving the accuracy of regional ionospheric models and the accuracy of differential code bias (DCB) determination. However, multi-GNSS measurement data mainly improves the spatial distribution of land areas, while effective GNSS measurements in marine areas are relatively few, which limits the accuracy of GIMs in marine areas and other areas where GNSS stations are unevenly distributed and sparse. To improve the accuracy of GIMs in ocean regions, research based on multi-source data has been developed in recent years. Examples include regional ionospheric models integrating International Reference Ionosphere (IRI), GPS observations, radio occultation data, radar altimeter measurements, and Very Long Baseline Interferometry (VLBI) data; global ionospheric modeling based on GNSS, satellite altimeter, and FORMOSAT-3 / COSMIC (F3 / C) data; and GIMs constructed by integrating GNSS, satellite altimeter, radio occultation, and Doppler Orbitography and Radiopositioning Integrated by Satellite (DORIS) data. All these have significantly improved the accuracy of models in ocean regions, but their accuracy remains insufficient in data-scarce mid- and low-latitude regions.Studies have further confirmed the improvement effect of F3 / C TEC data on GIMs, especially its significance in ocean regions. However, the model's accuracy is low in data-scarce mid- and low-latitude regions and during geomagnetic storms. Furthermore, the model's spatiotemporal resolution cannot meet the requirements of ground-based augmentation systems. Some studies have built global and regional ionospheric monitoring systems using GNSS and Jason satellite data, but these systems cannot accurately represent ionospheric changes in China during geomagnetic storms. Other studies have constructed global ionospheric maps using mobile phones, achieving high spatial resolution in Europe and America. However, because the model data originates from mobile phones, it suffers from uneven spatiotemporal resolution and insufficient accuracy.
[0003] In summary, using only GNSS data for ionospheric modeling suffers from data gaps in the ocean and some sparse areas of GNSS stations. Existing data fusion schemes also lack accuracy and spatiotemporal resolution during geomagnetic storms and in mid-to-low latitude regions, failing to meet the requirements for high-precision ionospheric modeling and ground-based augmentation applications during geomagnetic storms. Therefore, introducing multi-scale data and utilizing multi-source data fusion for complementary advantages is an effective way to achieve high-precision, high-spatiotemporal-resolution ionospheric modeling during geomagnetic storms.
[0004] Overall, current ionospheric products generally suffer from insufficient accuracy and low spatiotemporal resolution, making it difficult to quickly acquire high-precision ionospheric delay data for low- to mid-latitude regions where data is scarce. Existing ionospheric modeling technologies also face limitations such as incomplete coverage of single data sets, negative values in some regions, and insufficient resolution. Summary of the Invention
[0005] To address the shortcomings of current ionospheric products, such as insufficient accuracy and low spatiotemporal resolution, which hinder the rapid acquisition of high-precision ionospheric latency data in low- and mid-latitude regions with scarce data, and the problems of incomplete coverage of single data sources, negative values in some regional models, and insufficient resolution in existing ionospheric modeling techniques, this invention provides a method for refining the spatiotemporal resolution of regional ionospheric data. Once constructed, this method can eliminate the dependence on the original modeling data source and directly generate high-precision regional models quickly using global models without the need for large-scale modeling calculations, thereby significantly reducing the demand for service computing resources.
[0006] According to one aspect of the present invention, a method for refining the spatiotemporal resolution of a regional ionosphere is provided, comprising:
[0007] Obtain real-time GIM data, solar activity index, and geomagnetic activity index;
[0008] The acquired real-time GIM data, solar activity index, and geomagnetic activity index are input into the trained forecast model, and the VTEC mean value of the grid center point is output.
[0009] Based on the output VTEC mean, the STEC value is calculated by combining interpolation and projection functions; wherein, the training of the prediction model includes:
[0010] The raw GIM data and CORS observation data were acquired and preprocessed separately to obtain the GIM-VTEC calculated based on the GIM data and the puncture point VTEC calculated based on the CORS observation data.
[0011] The pre-processed GIM-VTEC and puncture point VTEC are matched according to time and grid range;
[0012] The solar activity index and geomagnetic activity index are obtained and matched with the matched VTEC data by time to construct a training sample set;
[0013] A BP neural network was constructed, with time, longitude, latitude, GIM-VTEC, solar activity index, and geomagnetic activity index as inputs, and the VTEC of the puncture point calculated from CORS observation data as the output.
[0014] The model is trained on the constructed training sample set and output as a trained prediction model.
[0015] As a further technical solution, training on the constructed training sample set also includes:
[0016] A sliding time window is introduced to segment the model training process. After each epoch is completed, the current optimal model parameters are saved, and the optimal model parameters saved in the current epoch are used as the initial weights in the next time window to continue training until all time windows are completed, and then the parameter configuration file is saved.
[0017] As a further technical solution, training on the constructed training sample set also includes:
[0018] Hyperparameter optimization is performed using a combination of grid search followed by random search.
[0019] As a further technical solution, the raw GIM data is acquired and preprocessed, including:
[0020] The original GIM data is subjected to interpolation and encryption processing;
[0021] The interpolated and encrypted data is then subjected to grid division, gross error detection and removal, pure random stationarity verification, and data normalization to obtain GIM-VTEC.
[0022] As a further technical solution, the raw CORS observation data is acquired and preprocessed, including:
[0023] The location of the puncture point is calculated based on the propagation path of the satellite signal;
[0024] STEC was calculated using the dual-frequency pseudorange method in the CORS observation data. Then, by setting the elevation angle threshold and limiting the grid area into which the puncture point falls, puncture points that meet the accuracy requirements were selected.
[0025] The STEC is converted into VTEC using a projection function model. Within each preset time period, multiple puncture point VTECs falling into the same grid are grouped and gross errors are removed. The average value is then used as the puncture point VTEC of the current grid at the current time.
[0026] As a further technical solution, the raw CORS observation data is acquired and preprocessed, including:
[0027] The VTEC corresponding to each puncture point within a preset time period is matched with the GIM-VTEC according to time and grid range.
[0028] According to one aspect of the present invention, a regional ionospheric spatiotemporal resolution refinement system is provided, comprising:
[0029] The first main module is used to acquire real-time GIM data, solar activity index, and geomagnetic activity index;
[0030] The second main module is used to input the acquired real-time GIM data, solar activity index and geomagnetic activity index into the trained forecast model, and output the VTEC mean value of the grid center point;
[0031] The third main module is used to calculate the STEC value based on the output VTEC mean, combined with interpolation and projection functions; wherein, the training of the prediction model includes:
[0032] The raw GIM data and CORS observation data were acquired and preprocessed separately to obtain the GIM-VTEC calculated based on the GIM data and the puncture point VTEC calculated based on the CORS observation data.
[0033] The pre-processed GIM-VTEC and puncture point VTEC are matched according to time and grid range;
[0034] The solar activity index and geomagnetic activity index are obtained and matched with the matched VTEC data by time to construct a training sample set;
[0035] A BP neural network was constructed, with time, longitude, latitude, GIM-VTEC, solar activity index, and geomagnetic activity index as inputs, and the VTEC of the puncture point calculated from CORS observation data as the output.
[0036] The model is trained on the constructed training sample set and output as a trained prediction model.
[0037] As a further technical solution, the training of the prediction model also includes:
[0038] A sliding time window is introduced to segment the model training process. After each epoch is completed, the current optimal model parameters are saved, and the optimal model parameters saved in the current epoch are used as the initial weights in the next time window to continue training until all time windows are completed, and then the parameter configuration file is saved.
[0039] According to one aspect of the present invention, an electronic device is provided, including a memory and a processor, the memory storing program instructions executable by the processor, the processor invoking the program instructions to perform the regional ionospheric spatiotemporal resolution refinement method.
[0040] According to one aspect of the present invention, a non-transitory computer-readable storage medium is provided, the non-transitory computer-readable storage medium storing computer instructions that cause the computer to perform the regional ionospheric spatiotemporal resolution refinement method.
[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0042] This invention integrates multi-source, multi-scale observation data and considers the spatial characteristics and physical mechanisms of ionospheric changes to construct a high-precision, high spatiotemporal resolution real-time ionospheric model. Unlike existing research that focuses on the modeling method itself, this invention's starting point is regional refinement based on a low-precision global model. Once constructed, this method eliminates dependence on the original modeling data source, directly and rapidly generating high-precision regional models from the global model without requiring large-scale modeling calculations, thus significantly reducing the demand for service computing resources. Attached Figure Description
[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0044] Figure 1 This is a flowchart illustrating the regional ionospheric spatiotemporal resolution refinement method provided in an embodiment of the present invention.
[0045] Figure 2 This is a schematic diagram of the modeling process of the BP neural network provided in an embodiment of the present invention.
[0046] Figure 3 This is a schematic diagram of the satellite puncture point coverage area provided in an embodiment of the present invention.
[0047] Figure 4 A schematic diagram of STEC data provided for an embodiment of the present invention.
[0048] Figure 5 A schematic diagram of the ACF of VTEC provided in an embodiment of the present invention.
[0049] Figure 6 This is a schematic diagram of grid encryption and data matching provided in an embodiment of the present invention.
[0050] Figure 7 This is a schematic diagram of data matching and mapping provided for an embodiment of the present invention.
[0051] Figure 8 A schematic diagram of the BP neural network framework provided in an embodiment of the present invention.
[0052] Figure 9 Box plot showing the relationship between hyperparameters and RMSA provided in embodiments of the present invention.
[0053] Figure 10 This is a schematic diagram of a sliding time window provided in an embodiment of the present invention. Detailed Implementation
[0054] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.
[0056] like Figure 1 As shown, this embodiment of the invention provides a method for refining the spatiotemporal resolution of the regional ionosphere. First, real-time GIM data, solar activity index, and geomagnetic activity index are acquired. Then, the acquired real-time GIM data, solar activity index, and geomagnetic activity index are input into the trained forecast model, and the mean VTEC value of the grid center point is output. Then, based on the output mean VTEC value, the STEC value is calculated by combining interpolation and projection functions.
[0057] Backpropagation Neural Network (BP) is a classic feedforward neural network consisting of an input layer, hidden layers, and an output layer. Its basic principle is to continuously adjust the network weights through backpropagation of errors, enabling the model to approximate the nonlinear relationship between the input and output. Despite its relatively simple structure, BP neural networks exhibit stability in handling nonlinear and multivariate coupled problems, possessing good fitting ability and application flexibility.
[0058] In ionospheric modeling, the variation of TEC (thermal activity potential) is influenced by multiple physical factors such as solar activity and geomagnetic disturbances, exhibiting significant temporal evolution characteristics and spatial differences. Using these physical parameters along with temporal and spatial location information as input, a backpropagation (BP) neural network can effectively establish the nonlinear relationship between TEC and TEC, capturing the spatiotemporal variation characteristics of the ionosphere. Compared to more complex deep models, BP neural networks offer lower computational costs and more efficient training while maintaining a certain level of prediction accuracy, making them suitable for high spatiotemporal resolution modeling tasks at a regional scale. Therefore, this invention employs a BP neural network as the core method of the ionospheric prediction model, balancing accuracy and efficiency to model and predict the distribution of TEC in the Kunming ionosphere.
[0059] This invention constructs an ionospheric prediction model, the core objective of which is to address the dependence of ionospheric modeling on real-time CORS station observation data. Because CORS station data involves national security geographic information, its acquisition is often restricted by strict confidentiality regulations and data sharing protocols. When high-precision CORS observation data cannot be obtained in real time, relying on a global ionospheric grid model for ionospheric delay correction will significantly reduce the accuracy of total ionospheric electron content calculation due to the lack of regional constraints from ground-based reference stations, thus affecting the spatial positioning accuracy based on oblique path total electron content.
[0060] Therefore, the forecast model trained in this embodiment of the invention aims to obtain the VTEC value within the grid by inputting only publicly available GIM parameters and space environment indices (including solar activity index and geomagnetic activity index) when CORS data is unavailable. Then, it calculates the STEC using interpolation and projection function models, and outputs a high-precision regional ionospheric delay correction in real time. (See also...) Figure 2 The specific steps for prediction are as follows:
[0061] Step 1, Data Preprocessing: Process and match GIM data and base station data.
[0062] TEC data is a typical type of time series data. Neural networks can only effectively learn patterns and make predictions when it meets the characteristics of stability and non-randomness. Therefore, rigorous preprocessing of TEC data is necessary before modeling to ensure data quality and model reliability. This invention employs a series of key steps in data preprocessing, including: grid partitioning, outlier detection and removal, pure random stationarity testing, and data normalization.
[0063] Figure 3 The image shows the satellite penetration point coverage of GNSS observation stations in Kunming and surrounding areas, with the red rectangle representing the spatial distribution of these stations. In the preliminary experimental phase, observation data with satellite elevation angles ≥30° were initially used for ionospheric model training. However, due to the significant error amplification effect of the projection function at low elevation angles, the model's solution accuracy was unsatisfactory, resulting in a root mean square error (RMS) as high as 17 TECU. Subsequently, through parameter sensitivity analysis, increasing the satellite elevation angle threshold to ≥60° significantly improved model accuracy and reduced the RMS value.
[0064] In the figure, the color of the three grids represents the total number of satellite puncture points within each grid area in a day. Different colors correspond to different density ranges: light yellow represents 0-50 puncture points per day, green represents 50-200 puncture points per day, and blue represents a high-density area with 200-350 puncture points per day. Based on the above analysis results, this embodiment of the invention further screened and determined the research range that could fully cover the Kunming area and had a high density of grid puncture points. The final experimental area was set as follows: latitude range 20°~30°, longitude range 95°~105°.
[0065] In the gross error detection and removal process, identifying gross errors in the data involves two steps:
[0066] 1. Plot a time series graph and analyze the trend of STEC changes, observing for any significant outliers. Some satellites, due to orbital obstruction, signal loss, etc., only have STEC data for certain periods; these data should be directly discarded. STEC represents the total electron density along the oblique path and cannot be negative; outlier data with STEC less than 0 should be removed to ensure the physical validity of the data. For time points with drastic STEC changes, use time-matching analysis to check if the changes are related to geomagnetic storm events. Specifically, this involves checking for significant enhancements in geomagnetic activity before and after the time point, such as a sharp drop in the Dst index or an abnormal increase in the Kp index. If the abrupt change has no obvious geomagnetic storm influence and only occurs in the data of a specific satellite, it is highly likely to be a gross error caused by measurement error or data jumps, and should be removed. After removing obvious outliers, calculate the mean and standard deviation of the overall data, and remove data points exceeding three times the standard deviation. For example... Figure 4 There are obvious gross errors, so they are directly discarded.
[0067] 2. The oblique total electron content (STEC) is converted to the vertical total electron content (VTEC) using a projection factor. All puncture points within the same time period (5-minute temporal resolution) and the same grid cell (5′×5′ spatial resolution) are grouped to facilitate error statistical analysis within local areas. Subsequently, the grouped VTEC values are subjected to a consistency check to determine their degree of deviation. If the VTEC corresponding to a puncture point exceeds a set error threshold, it is identified as a gross error point, and its corresponding observation data is removed to improve the quality and reliability of subsequent modeling data. The determination process is as follows:
[0068] (Formula 1),
[0069] Where: R = 6370km (Earth's radius); h = 350km (assumed height of the ionosphere); Z = 90° (zenith angle).
[0070] In the pure stochastic stationarity test, the autocorrelation function (ACF) is a statistical method that measures the correlation between current values and past values in time series data. The ACF can help determine the time dependence of the data, i.e., whether the data has a predictable structure or is simply random noise.
[0071] Given a time series:
[0072] (Formula 2).
[0073] The correlation of a time series at different time lags k is defined as follows:
[0074] (Formula 3),
[0075] in It is the mean of the time series. Here, k is the variance of the time series, k is the lag order, and the correlation coefficient ranges from 1 to 1. .
[0076] Positive correlation: This indicates that the current value is positively correlated with past values, and this correlation is persistent.
[0077] Negative correlation: This indicates that the current value is negatively correlated with the past value, showing a reversal trend.
[0078] A value close to 0 indicates that there is no significant correlation between the current value and past values, and the data may be random noise.
[0079] Using Python's plot_acf library, you can plot the autocorrelation function, such as... Figure 5 As shown, the initial autocorrelation coefficient (Lag=0) is 1.0, and the correlation between the variable and itself remains at 1. Starting from Lag=1, the autocorrelation coefficient rapidly decreases to approximately 0.8, and then slowly declines thereafter. The fact that the ACF remains high over time indicates that the data is not random noise, but rather exhibits significant persistence and strong autocorrelation. This phenomenon usually signifies a strong trend or seasonal pattern in the data, with values before and after the data point being highly correlated. It can be concluded that the total electron content of the ionosphere exhibits annual, seasonal, and diurnal variations, making it a non-stationary time series with seasonal variability.
[0080] To eliminate the influence of data dimensions, the data is scaled to a fixed range [-1, 1]. Since linear transformation does not change the original data distribution, this method can improve the convergence speed and stability of the model. This method is called data normalization. In this embodiment of the invention, min-max normalization is used. This normalization method is defined as follows:
[0081] (Formula 4),
[0082] X is the original data. and These are the minimum and maximum values of the data, respectively. During the normalization process, extreme values are calculated independently for VTEC, grid latitude and longitude, geomagnetic index, and solar activity index, preserving the statistical independence of each dimension.
[0083] Step 2, Input Feature Construction and Sample Generation: Match by time to construct a complete training sample set.
[0084] To achieve high-precision modeling of regional ionospheric TEC (tectonics), unified spatiotemporal normalization and matching operations are required for the raw observation data. This embodiment of the invention uses GIM data provided by CODE and CORS station observation data as the main data sources, performs independent preprocessing on the two types of data, and fuses them in the spatiotemporal dimensions to construct a training sample set.
[0085] To meet the needs of high spatiotemporal resolution modeling, the original GIM data needs to be interpolated and encrypted, such as... Figure 6 As shown. The original GIM data has a temporal resolution of 1 hour and a spatial resolution of 5°×5°, which cannot meet the requirements for characterizing the detailed changes in the ionosphere at a regional scale. In the spatial dimension, the 5°×2.5° grid is refined to 5′×5′ using bilinear interpolation, that is, each original grid point is refined into several sub-grids, thereby significantly improving the spatial coverage accuracy and better matching the densely distributed areas of CORS sites. In the temporal dimension, the 1-hour temporal resolution is refined to 5-minute intervals using cubic spline interpolation, constructing 12 isochronous high-density slices to capture the short-timescale changes in the ionosphere in greater detail. This results in... Figure 6 The high-density space grid is shown on the right side of the image.
[0086] like Figure 7 As shown, to achieve effective correspondence between GIM data and ground observation data, the puncture point location needs to be calculated based on the satellite signal propagation path. The oblique total electron content is calculated using the dual-frequency pseudorange method in the CORS observation data. Then, by setting an elevation angle threshold and limiting the grid area into which the puncture point falls, puncture points meeting accuracy requirements are selected. A projection function model (thin-layer model) is used to convert STEC into VTEC in the vertical direction, unifying observation values at different elevation angles and eliminating the error caused by ionospheric projection geometry. Within each 5-minute time interval, the VTEC values of multiple puncture points falling into the same grid cell are grouped, and gross errors exceeding a reasonable threshold are removed. The average value is then used as the ionospheric state of that grid at that time. This method effectively reduces the interference of abnormal observations on model training. Finally, the VTEC value corresponding to the puncture point at each 5-minute time point is matched with the interpolated GIM-VTEC.
[0087] Following the steps above, a complete training sample is obtained by matching one year's worth of data. The F10.7, Dst, and SNN external physical driving factors are then matched with the merged puncture point data over time to construct a complete training sample set. For example... Figure 8 As shown, the model inputs are: time, longitude, latitude, VTEC, Dst, SNN, and F10.7 based on GIM data; the model output is: the mean VTEC value of the grid center point calculated from the observation data.
[0088] Step 3, Hyperparameter selection: Optimize the parameter combination using grid search algorithm and random search method.
[0089] The performance of a neural network model largely depends on the settings of its hyperparameters. Common hyperparameters include the learning rate, the number of hidden layer nodes, the activation function, the regularization coefficient, the batch size, and the number of training epochs. These parameters are not automatically obtained during model training but need to be set before training. Appropriate combinations of hyperparameters can significantly improve the model's fitting accuracy and generalization ability, while inappropriate settings can lead to decreased model performance or even serious overfitting or underfitting problems. For example, an excessively high learning rate may prevent the model from converging, while an excessively low learning rate will result in excessively long training times and a tendency to get stuck in local optima; too many hidden layer nodes can easily lead to overfitting, while too few nodes may fail to fully represent the data features, and so on.
[0090] To achieve systematic optimization of model performance, commonly used hyperparameter tuning methods include grid search and random search.
[0091] 1. Grid Search
[0092] Grid search is an exhaustive strategy that constructs all possible combinations within a predefined hyperparameter value space, trains and evaluates the model for each combination, and finally selects the parameter set that performs best. Its main advantage is its comprehensiveness, guaranteeing the finding of a globally optimal or near-optimal solution within the search range. However, its disadvantages are also significant: high computational cost and low search efficiency, especially when the hyperparameters have many dimensions or a wide value range, resulting in enormous computational and time overhead.
[0093] 2. Random Search
[0094] Random search randomly samples a certain number of parameter combinations within a specified parameter distribution range for training and evaluation. Compared to grid search, it has better search efficiency and resource utilization, often achieving better solutions with the same computational budget. Especially in high-dimensional parameter spaces, random search can effectively skip redundant combinations and more quickly locate parameter regions that have a significant impact on the model. However, its results have a certain degree of randomness, and may behave inconsistently in each run, and it lacks the guarantee of a systematic exhaustive search.
[0095] Considering the balance between model stability and computational resources, this embodiment of the invention employs a combined strategy of "grid search followed by random search" for hyperparameter optimization: the first time window of the sample data is selected as the sample data for the pre-experiment, and a grid search method is used; the learning rate is set to a range of 0.01:0.005:0.05; in previous studies, the number of hidden neuron nodes is generally... arrive Between, among and These are the input and output layer nodes, respectively; therefore, in this embodiment of the invention, the number of neurons is set to 6 to 15. Based on the above hyperparameter settings, we can obtain... Figure 9 The box plot shown.
[0096] Figure 9 In the diagram, the x-axis represents the learning rate (lr), and the y-axis represents the root mean square error (RMSE). Curves of different colors represent different numbers of hidden neurons. It can be observed that both low learning rates (<0.01) and high learning rates (>0.04) may lead to high errors and are not suitable for training this model. The pink (13) and dark brown (12) curves are usually in the lower RMSE position, indicating that more hidden neurons help reduce training error, but the possibility of overfitting due to an excessive number of neurons still needs to be considered.
[0097] In summary, in this embodiment of the invention, lr is selected as 0.02 to 0.04 and i is selected as 11 to 15 as the parameter range, and the number of training rounds is 50. In the subsequent formal training, a random search method is used to select parameters within this range for multiple rounds of sampling training.
[0098] Step 4, determine the time window and training strategy: adopt a sliding time window strategy. After each epoch of training, save the current optimal model parameters and use them as the initial model for the next epoch.
[0099] Given the massive amount of data throughout the year, directly performing full-time modeling not only places a heavy burden on computational resources but also easily leads to problems such as convergence difficulties and overfitting during model training. Therefore, this embodiment of the invention introduces a sliding time window method to segment the model training process. Specifically, a typical solar activity cycle (27 days) is used as the time window length, and the training data is progressively slid forward in daily increments, thereby ensuring the model's temporal continuity while capturing subtle changes in ionospheric characteristics across different time periods.
[0100] After each epoch of training, the current optimal model parameters are saved. This optimal model will be used as the initial weights in the next time window to continue training, thereby achieving incremental optimization of the model parameters.
[0101] This strategy, on the one hand, continues the model parameters from the previous window, helping the model maintain a certain continuity across different windows and making the training process more stable; on the other hand, it also provides initialization conditions for data in subsequent time periods that are closer to their distribution characteristics, improving the convergence efficiency of the model during sliding training. Simultaneously, this method ensures that all annual days with available observation data are included in the training, enabling the model to better capture the spatiotemporal variations of the ionosphere. Especially during periods of frequent or drastic ionospheric disturbances, such as during geomagnetic storms, it allows the model to more fully learn local anomaly features. The specific sliding window is as follows... Figure 10 As shown.
[0102] Step 5, divide the training set and test set: divide it into 80% training set and 20% test set.
[0103] Step 6, Model Iteration and Tuning: Visualize and analyze the model's performance metrics, and dynamically adjust parameters such as the number of training epochs and the learning rate based on the training performance.
[0104] Step 7, Error Assessment: Select error indices to evaluate the effectiveness of ionospheric TEC prediction and fitting.
[0105] To evaluate the fitting and prediction level of the constructed ionospheric VTEC prediction model from multiple dimensions, this embodiment of the invention uses root mean square error, mean absolute error, and coefficient of determination (R²) as core evaluation indicators to analyze the model's performance in error control, fluctuation capture, and trend fitting. See Table 1 for details.
[0106] Table 1 Model Performance Parameters
[0107] .
[0108] As shown in Table 1, the model's prediction performance during ionospheric calm periods is significantly better than that during geomagnetic storm disturbances. Under calm conditions, the root mean square errors (RMSEs) for the training and test sets are 1.2659 TECU and 1.3482 TECU, respectively, with the mean absolute error controlled below 0.04 TECU and the standard deviation also being relatively small, indicating low overall error fluctuation. Furthermore, the coefficients of determination reach 0.9257 and 0.9123, respectively, demonstrating a strong ability to fit VTEC changes and exhibiting good stability and accuracy.
[0109] However, during geomagnetic storms, the model showed significant degradation across all four metrics. RMS rose to 3.5986 TECU and 3.8521 TECU, respectively, while MEANA increased to 0.1209 TECU and 0.1583 TECU, with a significantly larger standard deviation and increased volatility. This performance decline reflects the dramatic nonlinear changes in the ionosphere during geomagnetic disturbances, increasing the model's predictive difficulty. Nevertheless, the R² values for the training and test sets remained at 0.8791 and 0.8226, respectively, indicating that the model still possesses some trend recognition and fitting capabilities under geomagnetic storm conditions.
[0110] The implementation of the various embodiments of the present invention is based on programmed processing by a device with processor functionality. Therefore, in practical engineering, the technical solutions and functions of the various embodiments of the present invention are encapsulated into various modules. Based on this reality, and building upon the above embodiments, the embodiments of the present invention provide a regional ionospheric spatiotemporal resolution refinement system, which is used to execute the regional ionospheric spatiotemporal resolution refinement method in the above method embodiments.
[0111] The system includes: a first main module for acquiring real-time GIM data, solar activity index, and geomagnetic activity index; a second main module for inputting the acquired real-time GIM data, solar activity index, and geomagnetic activity index into a trained forecasting model, and outputting the mean VTEC value of the grid center points; and a third main module for calculating the STEC value based on the output mean VTEC value, combined with interpolation and projection functions. The training of the forecasting model includes: acquiring raw GIM data and CORS observation data and preprocessing them separately to obtain the GIM-VTEC and STEC values calculated based on the GIM data. The VTEC at the puncture point is calculated based on CORS observation data; the preprocessed GIM-VTEC and the VTEC at the puncture point are matched according to time and grid range; the solar activity index and geomagnetic activity index are obtained and matched with the matched VTEC data according to time to construct a training sample set; a BP neural network is constructed, with time, longitude, latitude, GIM-VTEC, solar activity index, and geomagnetic activity index as inputs, and the VTEC at the puncture point calculated from CORS observation data as output; the model is trained on the constructed training sample set to output the trained forecast model.
[0112] The regional ionospheric spatiotemporal resolution refinement system provided in this invention addresses the common problems of insufficient accuracy and low spatiotemporal resolution in current ionospheric products, making it difficult to quickly acquire high-precision ionospheric delay data for low- and mid-latitude regions with scarce data. It also addresses the issues of incomplete coverage of single data sources, negative values in some regional models, and insufficient resolution in existing ionospheric modeling technologies. By employing the aforementioned modules and constructing a forecast model based on a BP neural network, the system can eliminate dependence on original modeling data sources and directly generate high-precision regional models quickly using global models without the need for large-scale modeling calculations, thereby significantly reducing the demand for service computing resources.
[0113] It should be noted that the system embodiments provided by the present invention are used not only to implement the methods in the above method embodiments, but also to implement the methods in other method embodiments provided by the present invention. The only difference is that corresponding functional modules are set. The principle is basically the same as that of the above system embodiments provided by the present invention. As long as those skilled in the art can improve the modules in the above system embodiments by referring to the specific technical solutions in other method embodiments and combining technical features to obtain corresponding technical means and technical solutions composed of these technical means, on the basis of the above system embodiments, and on the premise of ensuring the practicality of the technical solutions, they can obtain corresponding system-like embodiments for implementing the methods in other method-like embodiments.
[0114] Based on the same inventive concept as the foregoing embodiments, this embodiment of the invention also provides an electronic device, including a memory and a processor, wherein the memory stores program instructions that are executed by the processor, and the processor invokes the program instructions to execute the regional ionospheric spatiotemporal resolution refinement method.
[0115] In embodiments of the present invention, the memory can be non-volatile memory, such as a hard disk drive (HDD) or a solid-state drive (SSD), or it can be volatile memory, such as random-access memory (RAM). Memory is any other medium capable of carrying or storing desired program code having an instruction or data structure form and accessible by a computer, but is not limited thereto. The memory in embodiments of the present invention can also be a circuit or any other device capable of implementing a storage function for storing program instructions and / or data.
[0116] In this embodiment of the invention, the processor may be a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in this embodiment of the invention. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in this embodiment of the invention can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.
[0117] Based on the same inventive concept as the foregoing embodiments, this embodiment of the invention also provides a non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute the regional ionospheric spatiotemporal resolution refinement method: First, acquire real-time GIM data, solar activity index, and geomagnetic activity index; then, input the acquired real-time GIM data, solar activity index, and geomagnetic activity index into the trained forecast model, and output the mean VTEC value of the grid center point; then, based on the output mean VTEC value, combined with interpolation and projection functions, calculate the STEC value.
[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.
Claims
1. A method for refining the spatial and temporal resolution of the ionospheric layer, characterized in that, The regional refinement based on the global low-precision model comprises: obtaining real-time GIM data, a solar activity index and a geomagnetic activity index; inputting the obtained real-time GIM data, the solar activity index and the geomagnetic activity index into a trained prediction model to output a VTEC average value of a grid center point; based on the output VTEC average value, combining an interpolation function and a projection function to calculate an STEC value; wherein the training of the prediction model comprises: obtaining original GIM data and CORS observation value data and respectively pre-processing to obtain GIM-VTEC calculated based on the GIM data and puncture point VTEC calculated based on the CORS observation value data; wherein the pre-processing of the original GIM data comprises: interpolating and encrypting the original GIM data, and sequentially performing grid division, gross error detection and elimination, pure random stationarity verification and data normalization on the interpolated and encrypted data to obtain GIM-VTEC; the pre-processing of the original CORS observation value data comprises: calculating a puncture point position according to a satellite signal propagation path; calculating STEC by using a double-frequency pseudo-range method in CORS observation value data, and then setting a height angle threshold and limiting a grid region where the puncture point falls to screen puncture points meeting accuracy requirements; converting the STEC into VTEC by using a projection function model, and grouping and eliminating gross errors of the VTEC of multiple puncture points falling into the same grid in each preset time period, and then taking the average value as the puncture point VTEC of the current grid at the current time; the pre-processing of the original CORS observation value data further comprises: matching the VTEC corresponding to the puncture points in each preset time period with the GIM-VTEC according to time and grid range; matching the GIM-VTEC and the puncture point VTEC obtained by pre-processing according to time and grid range; obtaining a solar activity index and a geomagnetic activity index, and matching the VTEC data after matching with the solar activity index and the geomagnetic activity index according to time to construct a training sample set; constructing a BP neural network with time, longitude, latitude, GIM-VTEC, a solar activity index and a geomagnetic activity index as input and the puncture point VTEC calculated from the CORS observation value data as output; training on the constructed training sample set to output a trained prediction model.
2. The method for refining the space-time resolution of the ionospheric region according to claim 1, characterized in that, The training on the constructed training sample set further comprises: introducing a sliding time window to segment the model training process, saving the current optimal model parameters after training each epoch, and taking the optimal model parameters saved by the current epoch as the initial weight to continue training in the next time window until the training of all time windows is completed, and saving the parameter configuration file.
3. The method of claim 2, wherein the ionospheric time-space resolution refinement is performed by a Kalman filter. The training on the constructed training sample set further comprises: adopting a combination strategy of grid search first and random search later to optimize hyperparameters.
4. A system for refining the spatial and temporal resolution of the ionospheric layer, characterized in that, The regional refinement based on the global low-precision model comprises: a first main module for obtaining real-time GIM data, a solar activity index and a geomagnetic activity index; The second main module is configured to input the acquired real-time GIM data, the solar activity index and the geomagnetic activity index into the trained prediction model, and output a VTEC mean value of a grid center point; The third main module is configured to calculate an STEC value based on the output VTEC mean value and in combination with an interpolation function and a projection function; wherein the training of the prediction model comprises: The original GIM data and the CORS observation value data are acquired and preprocessed respectively to obtain a GIM-VTEC calculated based on the GIM data and a piercing point VTEC calculated based on the CORS observation value data; wherein the preprocessing of the original GIM data comprises: interpolation encryption processing on the original GIM data, grid division, gross error detection and elimination, pure random stationarity verification and data normalization on the data after the interpolation encryption processing, to obtain the GIM-VTEC; The preprocessing of the original CORS observation value data comprises: calculating a piercing point position according to a satellite signal propagation path; calculating the STEC by using a dual-frequency pseudo-range method in the CORS observation value data, and then screening piercing points meeting accuracy requirements by setting a height angle threshold and limiting a grid region where the piercing points fall; converting the STEC into a VTEC by using a projection function model, and grouping and eliminating gross errors of the VTEC of multiple piercing points falling into the same grid in each preset time period, and then taking an average value as a piercing point VTEC of the current grid at the current time; The preprocessing of the original CORS observation value data further comprises: matching the VTEC corresponding to the piercing points in each preset time period with the GIM-VTEC according to time and grid range; The GIM-VTEC and the piercing point VTEC obtained after the preprocessing are matched according to time and grid range; The solar activity index and the geomagnetic activity index are acquired, and are matched with the VTEC data after the matching according to time, to construct a training sample set; A BP neural network is constructed, taking time, longitude, latitude, GIM-VTEC, the solar activity index and the geomagnetic activity index as inputs, and taking the piercing point VTEC calculated from the CORS observation value data as output; The training sample set is trained, and a trained prediction model is output.
5. The system for refining the spatial and temporal resolution of the ionospheric region according to claim 4, characterized in that, The training of the prediction model further comprises: A sliding time window is introduced to segment the model training process, the optimal model parameters of the current epoch are saved after training each epoch, and the optimal model parameters saved by the current epoch are taken as initial weights for continuing training in the next time window until the training of all time windows is completed, and the parameter configuration file is saved.
6. An electronic device, comprising: The non-transitory computer readable storage medium stores computer instructions, and the computer instructions cause the computer to execute the regional ionospheric space-time resolution refinement method in any one of claims 1 to 3.
7. A non-transitory computer-readable storage medium, comprising: The non-transitory computer readable storage medium stores computer instructions, and the computer instructions cause the computer to execute the regional ionospheric space-time resolution refinement method in any one of claims 1 to 3.
Citation Information
Patent Citations
Ionized layer electronic forecasting model training method and device and computer equipment
CN116108743A
Method and device for constructing high-resolution ionosphere model
CN119761221A
Method and system for constructing and correcting ionosphere mapping function considering space-time distribution
CN119849567A
