Ionized layer modeling method and device based on deep learning enhanced data assimilation
By using a deep learning-based intelligent assimilation network, the problems of high computational complexity and insufficient nonlinearity capture capability in ionospheric modeling are solved, achieving efficient and accurate generation of ionospheric three-dimensional state fields, and improving the practicality and reliability of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-19
- Publication Date
- 2026-04-10
AI Technical Summary
Existing technologies suffer from high computational complexity and poor timeliness in ionospheric modeling, making it difficult to meet the requirements for high-precision real-time assimilation. Furthermore, traditional methods have limited ability to capture strong nonlinear physical processes in the ionosphere, failing to effectively address the spatiotemporal scale mismatch problem and affecting model accuracy and practicality.
A deep learning-based intelligent assimilation network is employed. By constructing a ResNet model and combining multi-source observation data and physical parameters, offline training and online inference are performed to generate a high-precision three-dimensional electron density analysis field of the ionosphere.
It significantly reduces the computational burden, improves the efficiency of ionospheric modeling, enhances the accuracy of characterizing nonlinear physical processes, ensures that the output results conform to physical laws and have adaptability, and improves the practicality and reliability of the model.
Smart Images
Figure CN121835376A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of ionospheric modeling technology, specifically relating to an ionospheric modeling method and apparatus based on deep learning-enhanced data assimilation. Background Technology
[0002] With the increasing application of shortwave communication, satellite navigation, and other radio systems in key areas such as economic, security, and emergency communications, the demand for high-precision, high spatiotemporal resolution ionospheric environmental information has become extremely urgent. As a critical region in the Sun-Earth space environment affecting radio wave propagation and signal quality, accurate modeling and real-time sensing of the ionosphere have become an important foundation for improving the reliability, robustness, and intelligence of related systems.
[0003] Currently, ionospheric modeling mainly relies on numerical models based on physical equations or traditional data assimilation methods. While numerical models such as the International Reference Ionosphere (IRI) and the Thermosphere Ionosphere Electrodynamics General Circulation Model (TIEGCM) have clearly defined physical mechanisms, they inherently contain biases when describing the complex and variable behavior of the ionosphere. Traditional assimilation algorithms, such as the Ensemble Kalman Filter (EnKF), can incorporate observational data to correct the background field, but their implementation has many inherent flaws. These methods generally suffer from high computational complexity, relying on real-time calculation and updating of large-scale covariance matrices, resulting in high computational resource consumption and poor timeliness, making it difficult to meet the requirements of high-precision real-time assimilation. Furthermore, traditional methods are mostly based on linear or weakly nonlinear assumptions, limiting their ability to capture strongly nonlinear physical processes in the ionosphere, such as the equatorial anomaly and the extended-F phenomenon, and failing to fully reflect their true dynamic changes. When integrating observational data such as the sparse and non-uniformly distributed total electron content (TEC) of the global navigation satellite system, the critical frequency of the ordinary wave in the F2 layer (foF2) obtained by the vertical survey instrument, and the peak height of the F2 layer of the ionosphere (hmF2) into the global continuous model field, traditional methods struggle to effectively address the spatiotemporal scale mismatch between the two, affecting the overall accuracy and practicality of the model. Summary of the Invention
[0004] To address the shortcomings of existing technologies, the present invention aims to provide an ionospheric modeling method and apparatus based on deep learning-enhanced data assimilation, thereby solving the problems in the prior art.
[0005] The objective of this invention can be achieved through the following technical solutions: The ionospheric modeling method based on deep learning-enhanced data assimilation includes the following steps: Historical multi-source observation data of the ionosphere, background field data within the corresponding spatiotemporal range, and key physical parameters characterizing external driving conditions were collected. The collected data were preprocessed to construct a sample dataset and divided into training and test sets. Based on ResNet, an intelligent assimilation network is constructed. The network is trained offline using a training set and its performance is tested using a test set to obtain an optimized intelligent assimilation network. The real-time collected observation data, background field data, and current physical parameters are combined and input into the optimized intelligent assimilation network to obtain a full three-dimensional analysis increment field. This field is then superimposed with the background field to generate a three-dimensional electron density analysis field of the ionosphere.
[0006] Furthermore, the historical multi-source ionospheric observation data includes: TEC data retrieved from the Global Navigation Satellite System, foF2 and hmF2 characteristic parameters obtained from ionospheric vertical logging stations, and electron density data obtained from in-situ detection by low-Earth orbit satellites; The key physical parameters characterizing the external driving conditions include: solar radio flux at a wavelength of 10.7 cm, planetary index measuring the degree of geomagnetic disturbance, and three-hour geomagnetic index.
[0007] Furthermore, the preprocessing process includes: removing outliers and data that deviate from the sample mean by more than three standard deviations, and standardizing and normalizing data from different sources and with different dimensions.
[0008] Furthermore, the input layer of the intelligent assimilation network can receive interfaces for various types of data and simultaneously process background field grid data, sparse and non-uniform observation vectors, spatiotemporal coordinate information, and key physical parameters. The bottom layer of the intelligent assimilation network adopts a parameter-shared convolutional layer and residual block structure to extract and learn common physical laws and basic characteristics of the ionosphere from global data. The upper layer of the intelligent assimilation network deploys multiple structurally independent and parameter-adaptive branch subnetworks, which dynamically capture the specific physical processes and evolution patterns of each region through gating mechanisms and attention weights. The output layer of the intelligent assimilation network generates a three-dimensional analysis increment field that is consistent with the background field in spatial dimension through a fully connected layer and dimension reshaping operation.
[0009] Furthermore, when training the intelligent assimilation network, the loss function is: in, R represents the network predictive analysis field. ) represents the network predictive analysis field The residuals, where λ is the regularization weight coefficient; The value of the mean squared error loss function. N The total number of training samples, and The first i The true reference value of the sample and the training network for the first sample i The predicted value for each sample.
[0010] Furthermore, an early stopping strategy is implemented during training. When the loss function on the test set does not show a significant decrease for ten consecutive training epochs, the training process is automatically terminated and the network parameters with the best performance on the test set are saved. The training process employs an adaptive moment estimation optimizer, with an initial learning rate set to 0.1%, and dynamically adjusted using a cosine annealing scheduling strategy. The adaptive moment estimation optimizer is as follows: in, and These represent the parameter values of the model at the current time step and the updated parameter values of the model at the next time step, respectively. To train the network's learning rate, It is the exponential moving average of the squared gradient. It is a very small constant. It is an exponential moving average of the gradient.
[0011] The above-mentioned ionospheric modeling method based on deep learning-enhanced data assimilation is applied to the real-time correction of the ionosphere in radio system forecasts.
[0012] An ionospheric modeling device based on deep learning-enhanced data assimilation performs the modeling method described above, including: Data acquisition and processing module: Collects historical multi-source observation data of the ionosphere, background field data within the corresponding spatiotemporal range, and key physical parameters characterizing external driving conditions. It also preprocesses the collected data, constructs a sample dataset, and divides it into training and test sets. Network training module: Based on ResNet, an intelligent assimilation network is constructed, and the intelligent assimilation network is trained offline using the training set and its performance is tested using the test set to obtain the optimized intelligent assimilation network; Analysis field generation module: Combines real-time collected observation data, background field data and current physical parameters, inputs them into the optimized intelligent assimilation network to obtain a full three-dimensional analysis increment field, and performs superposition calculation with the background field to generate a three-dimensional electron density analysis field of the ionosphere.
[0013] A computer storage medium storing a readable program that, when executed, instructs a computing device to perform the aforementioned ionospheric modeling method based on deep learning-enhanced data assimilation.
[0014] An electronic device includes: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction that causes the processor to perform the operation corresponding to the ionospheric modeling method based on deep learning-enhanced data assimilation described above.
[0015] The beneficial effects of this invention are: 1. This invention significantly reduces the computational burden and improves the efficiency of ionospheric modeling by transforming the complex covariance matrix calculation in traditional assimilation into efficient neural network forward propagation, thus meeting the requirements of real-time high-resolution analysis. Utilizing the powerful nonlinear mapping capabilities of deep neural networks, it effectively overcomes the limitations of linear assumptions in traditional methods, significantly enhancing the accuracy of characterizing complex nonlinear physical processes in the ionosphere. By introducing a physics-guided training mechanism and a shared-private feature decoupling structure, it ensures that the output results not only conform to physical laws but also adaptively fuse multi-source heterogeneous observation data. Employing a framework that separates offline training and online inference, it greatly improves the system's engineering practicality and maintainability while ensuring model generalization performance, providing a complete and reliable solution for constructing a high-precision, high-time-efficiency three-dimensional ionospheric state field.
[0016] 2. This invention establishes the necessary data foundation for high-precision modeling by specifically defining multi-source observation data and external driving parameters. Its function is to explicitly solidify the key observational dimensions and core physical driving factors of the ionospheric state into technical features, ensuring that the intelligent assimilation network can fully learn the complex coupling laws of the ionosphere with spatial, temporal, and external conditions during the training phase. This limitation directly guarantees the model's ability to characterize the characteristics and strongly nonlinear physical processes of different regions globally, thereby significantly improving the physical consistency and spatiotemporal resolution of the final generated electron density analysis field, and enhancing the reliability and practicality of the method in real-time applications.
[0017] 3. This invention, by specifically defining the preprocessing steps as "removing outliers and data deviating from the sample mean by more than three standard deviations" and "performing standardization and normalization," aims to ensure the quality and consistency of input data from the source, laying a reproducible data foundation for the stable training and reliable inference of subsequent deep learning models. This definition clarifies the objective statistical criteria for data cleaning and the standardized operational procedures, effectively eliminating the interference of noise, dimensional differences, and numerical range variations in multi-source heterogeneous data on network training. Its direct technical effect is a significant improvement in the training efficiency, convergence stability, and accuracy and reliability of the final generated ionospheric state field of the intelligent assimilation network, thereby substantially enhancing the robustness and engineering practicality of the entire modeling method.
[0018] 4. This invention, through multi-layered specific constraints on the intelligent assimilation network architecture, transforms the core component "based on ResNet" into an implementable physical structure with a clear functional orientation. First, by defining a unified interface for the input layer to multi-source heterogeneous data, the data foundation for the network to fuse multi-dimensional information to establish precise mapping relationships is clarified. Second, by distinguishing between the underlying shared structure and the high-level adaptive branches, the architecture enforces functional decoupling between learning global physical laws and capturing region-specific phenomena, enabling the network to simultaneously model common ionospheric mechanisms and local anomalies. Finally, by constraining the output layer to generate a three-dimensional incremental field with the same spatial dimension as the background field, it ensures that the network output can be seamlessly synthesized directly with the physical model. These constraints collectively construct a deep neural network sample specifically designed for ionospheric data assimilation. Its technical effect is to endow the model with powerful nonlinear fitting capabilities and physical interpretability, thereby fundamentally solving the bottlenecks of traditional methods in terms of efficiency, nonlinear processing, and multi-scale fusion, providing crucial architectural guarantees for generating high-precision, high-resolution three-dimensional ionospheric state fields.
[0019] 5. This invention establishes a core guarantee mechanism for efficient and stable model convergence by concretizing the optimization algorithm, learning rate strategy, and stopping conditions during the training process into operable parameters and rules. Its function is to transform the abstract process of offline training into a complete technical process with clear mathematical rules, controllable hyperparameters, and automatic decision-making logic. The early stopping strategy dynamically monitors and makes decisions based on test set performance, effectively preventing overfitting and ensuring the generalization ability of the final deployed model. The adaptive moment estimation optimizer and its set initial learning rate, combined with the cosine annealing scheduling strategy, constitute an optimization system that can adaptively adjust the parameter update direction and step size, balancing rapid convergence in the early stages of training with fine-tuning in the later stages. These features work synergistically, and their direct technical effect is to significantly improve the training efficiency and stability of the intelligent assimilation network, enabling the model to reliably learn accurate nonlinear mapping relationships from large-scale historical data. This provides a fully optimized and robust core computing engine for generating high-precision incremental fields in the subsequent online assimilation stage. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 This is a schematic diagram of the offline training process of the deep learning assimilation device of the present invention; Figure 2 This is a flowchart of the online processing procedure for the intelligent assimilation network of the present invention; Figure 3 This is a flowchart of the regional ionospheric reconstruction process based on single-station vertical measurement data of the present invention; Figure 4 This is a comparison chart of the predicted results of the total electron content of the ionosphere in this invention. Detailed Implementation
[0022] 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, and 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.
[0023] Example 1 The ionospheric modeling method based on deep learning-enhanced data assimilation includes the following steps: S1 collects historical multi-source observation data of the ionosphere, background field data within the corresponding spatiotemporal range, and key physical parameters characterizing external driving conditions. The collected data is preprocessed to construct a sample dataset and is divided into training and test sets. The sources of the various data are as follows: The study collected historical multi-source ionospheric observation data covering the entire globe and spanning different solar activity cycles. The main data included: TEC data retrieved from the Global Navigation Satellite System, characteristic parameters such as foF2 and hmF2 obtained from ionospheric vertical logging stations, and electron density data obtained from in-situ detection by low-Earth orbit satellites.
[0024] Simultaneously acquire background field data within the corresponding spatiotemporal range, the source of which can be the output results of mature physical models such as IRI or TIEGCM.
[0025] Key physical parameters characterizing external driving conditions were collected, including: solar radio current at a wavelength of 10.7 cm (F10.7), planetary index (Ap) measuring the degree of geomagnetic disturbance, and three-hour geomagnetic index (Kp).
[0026] The data preprocessing process includes: A rigorous quality control process was implemented for all collected data, removing obvious outliers and data that deviated from the sample mean by more than three standard deviations. Data from different sources and with different dimensions were standardized and normalized to form a spatiotemporally matched sample set of background field, observation data and key physical parameters. The sample set was divided into training set and test set in a 7:3 ratio to ensure that the training data has both spatiotemporal continuity and regional representativeness.
[0027] S2 uses ResNet as its core to build an intelligent assimilation network. The intelligent assimilation network is trained offline using a training set and its performance is tested using a test set to obtain an optimized intelligent assimilation network. 1. Intelligent Assimilation Network The intelligent assimilation network uses a Deep Residual Network (ResNet) that deeply integrates physical mechanisms as its core. The input layer is designed to receive multiple types of data, simultaneously processing background field grid data, sparse and non-uniform observation vectors, finely encoded spatiotemporal coordinate information, and key physical parameters. The network's bottom layer employs parameter-shared convolutional layers and residual block structures to extract and learn common physical laws and fundamental characteristics of the ionosphere from global data. The higher layers deploy multiple structurally independent and parameter-adaptive sub-networks, corresponding to typical ionospheric structures such as the equatorial anomaly region, mid-latitude regions, and polar regions. Gating mechanisms and attention weights dynamically capture the specific physical processes and evolution patterns of each region. The final output layer generates a three-dimensional analytical incremental field that is spatially identical to the background field through fully connected layers and dimensionality reshaping operations. The value of each grid point represents the required physical correction amount for the ionospheric parameters at that location.
[0028] Specifically, intelligent assimilation networks include: The intelligent assimilation network employs a deep residual architecture. Its input layer receives and concatenates multi-source input data, including background field grids, sparse observation vectors, spatiotemporal coordinate encodings, and key physical parameters. The data first enters the bottom-level shared convolutional layers and residual blocks. These layers extract and fuse global spatial features and fundamental physical patterns through convolutional operations and skip connections. Subsequently, the higher layers of the network employ multiple parallel and parameter-adaptive branch subnetworks. Each branch uses gating and attention mechanisms to weight and filter shared features, focusing on capturing the physical processes and evolutionary features of specific regions such as the equatorial anomaly and mid-latitude regions. The processed features are aggregated and dimensionally reshaped by fully connected layers, ultimately outputting a three-dimensional analytical incremental field with the same spatial dimension as the background field, completing the nonlinear mapping from multi-source heterogeneous inputs to global physical corrections.
[0029] 2. Offline training of intelligent assimilation networks like Figure 1 As shown, the training set data is input into the intelligent assimilation network for supervised offline training. As shown in Equation (1), the training process uses the mean square error between the analytical increment field predicted by the network and the high-precision reference analytical field as the main loss function, and introduces a constraint term based on the physical conservation law of the ionosphere as a regularization method to enhance the physical rationality of the output results.
[0030] (1) in, The value of the mean squared error loss function. N The total number of training samples, and The first iThe true reference value of the sample and the training network for the first sample i The predicted value for each sample.
[0031] In the offline training phase, to improve the physical consistency of the output results, a regularization constraint term based on the ionospheric electron density continuity equation is introduced on the basis of the mean square error main loss function.
[0032] Electron density continuity equation: Where, N e Let q be the electron density and q be the electron production rate. For composite coefficients, This is the plasma drift velocity vector.
[0033] This physical constraint term is analyzed by computational network prediction. Constructed by deviations from the laws of physical conservation, specifically, Substituting the background physical parameters at the corresponding time into the electron density continuity equation, calculate its residual R( The total loss function is defined as: in, R represents the network predictive analysis field. ) represents the network predictive analysis field The residuals are given by λ, where λ is the regularization weight coefficient. This constraint term guides the network output to conform to physical laws, effectively improving the model's generalization ability and result reliability under complex space weather conditions.
[0034] As shown in Equation (2), the training process uses an Adaptive Moment Estimation Optimizer (Adam), with the initial learning rate set to 0.1%, and the learning rate value is dynamically adjusted in conjunction with a cosine annealing scheduling strategy.
[0035] Adam: (2) in, and These represent the parameter values of the model at the current time step and the updated parameter values of the model at the next time step, respectively. To train the network's learning rate, It is the exponential moving average of the squared gradient. It is a very small constant (usually 10⁻⁸). It is an exponential moving average of the gradient.
[0036] An early stopping strategy is implemented during training. When the loss function on the test set does not show a significant decrease for ten consecutive training cycles, the training process is automatically terminated and the model parameters with the best performance on the test set are saved. This effectively prevents the model from overfitting and ensures that it has good generalization ability.
[0037] S3 combines real-time collected observation data, background field data, and current physical parameters, inputs them into the optimized intelligent assimilation network to obtain a full three-dimensional analysis increment field, and performs superposition calculations with the background field to generate a three-dimensional electron density analysis field of the ionosphere.
[0038] like Figure 2 As shown, during the actual operational phase, the optimal intelligent assimilation network trained offline is deployed on the assimilation server. After system initialization, it continuously receives real-time observation data streams from global data centers. Within each set assimilation period, such as fifteen minutes, a high-resolution background field is first obtained through short-term forecasting using a numerical model or by employing data persistence methods. Subsequently, the background field at this moment, the real-time arriving sparse observation data and its spatiotemporal coordinates, and the current key physical parameters are combined into a complete input and fed into the pre-trained intelligent assimilation network. Through a single efficient forward propagation calculation, a full three-dimensional analysis increment field covering the entire globe is output within milliseconds. Finally, through a simple field superposition operation, the analysis increment field and the background field are matrix-added to generate a high spatiotemporal resolution ionospheric three-dimensional electron density analysis field that integrates the latest observation information, providing accurate state input for downstream application systems.
[0039] Example 2 In this embodiment, the effectiveness of the proposed method is verified by comparing the prediction results of the total electron content of the ionosphere reconstructed based on deep learning with those of the traditional international reference ionosphere model. The verification process includes: To verify the effectiveness of the method described in this invention, a complete offline training and online assimilation testing environment was constructed. The verification was conducted on a computing server equipped with a high-performance graphics processor, ensuring the computing power required for large-scale neural network model training and real-time assimilation inference. The historical data used for verification came from the Meridian Project, covering different solar activity intensities and geographically encompassing typical ionospheric structures including the equatorial anomaly region, mid-latitude regions, and polar regions, to ensure the spatiotemporal representativeness of the test samples.
[0040] The validation process strictly follows the described two-stage framework. In the offline stage, the intelligent assimilation network is trained using the historical dataset. Key training parameters include batch size, number of iterations, and the hyperparameter settings of the adaptive optimizer. After training, the network weights are fixed and deployed on the assimilation inference server. In the online validation stage, real-time business data streams are simulated. Observational data from the test period, independent of the training set, and a background field are input. The network generates a three-dimensional analytical increment field and synthesizes it with the background field to form the final analytical field.
[0041] The performance evaluation uses root mean square error and correlation coefficient as core quantitative indicators. The total electron content map, electron density profile and key parameters of the ionosphere generated by this invention are systematically compared with the output results of the international reference ionosphere model under the same spatiotemporal conditions.
[0042] The prediction results of the total electron content of the ionosphere are, for example Figure 4 As shown in the results, the predictions obtained by this method are in high agreement with the measured data, with a root mean square error of only 1.3913, while the error of the traditional model reaches 14.4656, representing an improvement in accuracy by more than an order of magnitude. This significant advantage fully demonstrates the superior performance of this method in improving the accuracy of ionospheric parameter estimation, providing more reliable technical support for shortwave frequency planning and communication link quality assessment, and possessing important practical application value.
[0043] Example 3 In this embodiment, an ionospheric modeling device based on deep learning-enhanced data assimilation is proposed, specifically including: Data acquisition and processing module: Collects historical multi-source observation data of the ionosphere, background field data within the corresponding spatiotemporal range, and key physical parameters characterizing external driving conditions. It also preprocesses the collected data, constructs a sample dataset, and divides it into training and test sets. Network training module: Based on ResNet, an intelligent assimilation network is constructed, and the intelligent assimilation network is trained offline using the training set and its performance is tested using the test set to obtain the optimized intelligent assimilation network; Analysis field generation module: Combines real-time collected observation data, background field data and current physical parameters, inputs them into the optimized intelligent assimilation network to obtain a full three-dimensional analysis increment field, and performs superposition calculation with the background field to generate a three-dimensional electron density analysis field of the ionosphere.
[0044] The methods of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code that, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses the code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for performing the methods shown herein.
[0045] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.
Claims
1. A method of ionospheric modeling based on deep learning enhanced data assimilation, characterized in that, The method comprises the following steps: Collecting historical ionospheric multi-source observation data, background field data corresponding to the space-time range, and key physical parameters representing external driving conditions, and pre-processing the collected data to construct a sample data set and divide the training set and the test set; Taking ResNet as the core, constructing an intelligent assimilation network, and using the training set to perform offline training on the intelligent assimilation network, and testing the performance by using the test set to obtain an optimized intelligent assimilation network; Combining the real-time collected observation data, background field data and current physical parameters, inputting the optimized intelligent assimilation network to obtain a full three-dimensional analysis increment field, and performing superposition operation with the background field to generate an ionospheric three-dimensional electron density analysis field.
2. The deep learning enhanced data assimilation based ionospheric modeling method of claim 1, wherein, The historical ionospheric multi-source observation data comprises: TEC data inverted from a global navigation satellite system, foF2 and hmF2 characteristic parameters obtained from an ionospheric vertical sounding station, and electron density data obtained from in-situ detection of a low-orbit satellite; The key physical parameters representing external driving conditions comprise: solar radio flux at 10.7 cm wavelength, planetary index measuring the degree of geomagnetic disturbance, and three-hour geomagnetic index.
3. The deep learning enhanced data assimilation based ionospheric modeling method of claim 1, wherein, The pre-processing process comprises: removing abnormal values and data deviating from the sample mean by more than three standard deviations, and standardizing and normalizing the data of different sources and dimensions.
4. The deep learning enhanced data assimilation based ionospheric modeling method of claim 1, wherein, The input layer of the intelligent assimilation network can receive interfaces of multiple types of data, and simultaneously process background field grid data, sparse and non-uniform observation vectors, space-time coordinate information and key physical parameters; the bottom layer of the intelligent assimilation network adopts a convolution layer and a residual block structure sharing parameters, extracts and learns common physical laws and basic features of the ionosphere from global data; multiple branch sub-networks with independent structure and adaptive parameters are deployed in the upper layer of the intelligent assimilation network, which dynamically captures the special physical processes and evolution modes of each region through a gating mechanism and attention weight; the output layer of the intelligent assimilation network generates a three-dimensional analysis increment field consistent with the background field in the spatial dimension through a fully connected layer and a dimension reshaping operation.
5. The deep learning enhanced data assimilation based ionospheric modeling method of claim 1, wherein, When training the intelligent assimilation network, the loss function is: wherein, represents the network prediction analysis field, R( ) represents the residual of the network prediction analysis field , and λ is the regularization weight coefficient; is the mean square error loss function value, N is the total number of training samples, and are the true reference value of the first i sample and the predicted value of the first i sample by the training network, respectively.
6. The deep learning enhanced data assimilation based ionospheric modeling method according to claim 1 or 5, characterized in that, During the training process, an early stopping strategy is implemented, and when the loss function on the test set does not appear significant decrease for ten consecutive training periods, the training process is automatically terminated and the network parameters with the optimal performance on the test set are saved; The adaptive moment estimation optimizer is used in the training process, the initial learning rate is set to one thousandth, and the cosine annealing scheduling strategy is used to dynamically adjust the learning rate value, and the adaptive moment estimation optimizer is: wherein, with are the parameter values of the model at the current time and the updated parameter values of the model at the next time, respectively, is the learning rate of the training network, is the exponential moving average of the squared gradient, is a small constant, is the exponential moving average of the gradient.
7. The application of the ionospheric modeling method based on deep learning enhanced data assimilation according to any one of claims 1-6 in real-time correction of a predicted ionosphere of a radio system.
8. An ionospheric modeling apparatus based on deep learning enhanced data assimilation, performing the modeling method of any one of claims 1-6, characterized in that, It comprises: A data acquisition and processing module: collecting historical ionospheric multi-source observation data, background field data corresponding to the space-time range, and key physical parameters representing external driving conditions, and pre-processing the collected data to construct a sample data set and divide the training set and the test set; The network training module: taking ResNet as the core, constructing an intelligent assimilation network, and using the training set to train the intelligent assimilation network offline, and using the test set to test the performance, obtaining the optimized intelligent assimilation network; The analysis field generation module: combining the real-time collected observation data, the background field data and the current physical parameters, inputting the optimized intelligent assimilation network, obtaining the full three-dimensional analysis increment field, and superimposing with the background field to generate the ionospheric three-dimensional electron density analysis field.
9. A computer storage medium storing a readable program, characterized in that, When the program is running, the program can instruct the computing device to perform the ionospheric modeling method based on deep learning enhanced data assimilation according to any one of claims 1-6.
10. An electronic device, comprising: Comprise: A processor, a memory, a communication interface and a communication bus, the processor, the memory and the communication interface complete communication with each other through the communication bus; The memory is used to store at least one executable instruction, and the executable instruction makes the processor execute the operation corresponding to the ionospheric modeling method based on deep learning enhanced data assimilation according to any one of claims 1-6.