Ionized layer prediction method and device based on white spectrum method and deep learning

Abstract

The invention discloses an ionosphere prediction method and device based on a white spectrum method and deep learning, and belongs to the technical field of intersection of spatial physics and satellite navigation, and the method comprises the steps: carrying out the preprocessing of multi-source observation data, obtaining time series data, carrying out the white spectrum method analysis of the time series data, and obtaining a disturbance time series; processing the disturbance time sequence to generate a global solar activity level index, a middle and low latitude geomagnetic disturbance index and a polar region geomagnetic disturbance index; the indexes and satellite inversion data serve as input, a deep learning reconstruction model formed by a multi-layer full-connection feedforward neural network is adopted, and ionosphere four-dimensional electron density distribution is output; ionized layer four-dimensional electron density distribution and total electron content observation data serve as input, a space-time prediction model is adopted, and an ionosphere prediction result at the future moment is output. The problems that a traditional spatial index is poor in timeliness and low in resolution and cannot effectively capture the complex time-space characteristics of the ionized layer can be solved.

Description

Technical Field

[0001] This invention relates to the technical field of space physics and satellite navigation, and in particular to a method and apparatus for ionospheric prediction based on white spectrum method and deep learning. Background Technology

[0002] As a crucial component of Earth's upper atmosphere, the ionosphere's internal electron density distribution and dynamic changes in total electron concentration (TEC) directly and significantly impact the positioning accuracy of Global Navigation Satellite Systems (GNSS), the signal quality of High Frequency (HF) communications, and the reliability of various space exploration missions. Therefore, achieving accurate and real-time prediction of the ionospheric state is a cutting-edge research topic in space physics and also possesses significant engineering application value.

[0003] In the field of ionospheric modeling and prediction, a series of traditional solar and geomagnetic activity indices have long been relied upon as key input parameters. For example, the International Reference Ionospheric Model (IRI), as an international standard model, uses the solar 10.7 cm radio current index (F10.7) as one of its key inputs to characterize the impact of solar radiation on ionospheric photochemical processes. Regarding geomagnetic activity, global disturbance levels are typically quantified by the Kp index (or its linearized Ap index); changes in ring current intensity are characterized by the geomagnetic disturbance index Dst; and the intensity of polar energy injection is reflected by the AE index. These indices have been widely applied in empirical models, semi-empirical models, and data-driven models, forming the technical foundation of existing ionospheric prediction methods.

[0004] Currently, in the field of data-driven ionospheric prediction, a common approach is to use deep learning frameworks to directly predict ionospheric parameters using traditional indices and observational data. This type of approach typically involves the following processing flow: First, historical traditional solar and geomagnetic indices (such as F10.7, Dst, Kp, AE, etc.) and ionospheric observational data (such as GNSS-TEC, COSMIC electron density profiles, etc.) are obtained from publicly available data platforms. Then, preprocessing operations such as spatiotemporal alignment, missing value handling, and normalization are performed on the multi-source data. Next, a deep learning architecture is constructed with recurrent neural networks (RNNs), long short-term memory networks (LSTMs), convolutional neural networks (CNNs), or hybrid models (such as ConvLSTMs) at its core. Finally, using historical sequences of traditional indices and ionospheric data as input, the mapping relationship between these indices and future ionospheric states (such as TEC or electron density) is learned, ultimately outputting ionospheric prediction results for a future period.

[0005] These methods, based on traditional indices and deep learning, offer a certain degree of improvement in characterizing nonlinear relationships compared to purely physical or empirical models. However, their performance is fundamentally limited by the inherent limitations of the traditional indices used as input, which can be summarized in the following two aspects:

[0006] 1. Existing space environment driving indices have shortcomings in terms of timeliness and resolution. Currently, solar activity and geomagnetic disturbance indices, such as F10.7, Kp, Dst, and AE indices, which are widely used in space weather and ionospheric models, have significant limitations. Specifically, some indices have low temporal resolution; for example, the F10.7 index is updated daily, and the Kp index is updated every 3 hours, failing to accurately depict short-term, drastic changes in the space environment. More importantly, many key indices suffer from severe release delays; for example, the F10.7 index has a delay of 10-15 days, and the AE index has a delay of approximately 15-30 days. These low resolution and long delays make these traditional indices unsuitable as real-time input sources for ionospheric forecasting and space weather early warning models with high timeliness requirements, thus limiting the accuracy and response speed of the forecasting models.

[0007] 2. Existing ionospheric prediction models have limited ability to learn complex spatiotemporal coupling characteristics. In predicting ionospheric electron density, current techniques often employ models based on Long Short-Term Memory (LSTM) networks and their variants. While these models have certain advantages in processing time-series data and can capture temporal dependencies, the ionosphere, a four-dimensional physical quantity that varies with longitude, latitude, altitude, and time, possesses extremely complex spatial structures and spatiotemporal evolution patterns. Traditional LSTM-based models are insufficient in simultaneously learning and fusing three-dimensional spatial features with temporal dependencies, making it difficult to fully explore and express the spatiotemporal coupling characteristics of the ionosphere. This leads to bottlenecks in prediction accuracy, especially during geomagnetic disturbances.

[0008] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0009] The purpose of this invention is to provide an ionospheric prediction method and device based on white spectrum method and deep learning, which can solve the problems of poor timeliness and low resolution of traditional spatial index, as well as the problem that existing prediction models cannot effectively capture the complex spatiotemporal characteristics of the ionosphere.

[0010] To achieve the above objectives, this invention provides an ionosphere prediction method based on white spectrum analysis and deep learning, comprising the following steps:

[0011] S1: Preprocess the multi-source observation data to obtain time series data, and then perform white spectrum analysis on the time series data to obtain the perturbation time series;

[0012] S2: Process the disturbance time series to generate the global solar activity level index, the mid-to-low latitude geomagnetic disturbance index, and the polar geomagnetic disturbance index;

[0013] S3: Using global solar activity level index, mid-to-low latitude geomagnetic disturbance index, polar geomagnetic disturbance index and satellite inversion data as input, a deep learning reconstruction model composed of multi-layer fully connected feedforward neural network is used to output the four-dimensional electron density distribution of the ionosphere.

[0014] S4: Using observational data on the four-dimensional electron density distribution and total electron content of the ionosphere as input, a spatiotemporal prediction model constructed with a three-dimensional SwinTransformer architecture is used to output the ionosphere prediction results for future times.

[0015] In one embodiment of the present invention, in step S1, the multi-source observation data includes observation data from multi-band coronal images from the Solar Dynamics Observatory and geomagnetic station data; the multi-band coronal images from the Solar Dynamics Observatory include coronal image data in the 94Å, 131Å, 171Å, 193Å, 211Å, 304Å, and 335Å bands; the geomagnetic station data includes observation data from globally distributed mid- and low-latitude geomagnetic stations and three-component observation data from global polar geomagnetic stations.

[0016] In one embodiment of the present invention, the preprocessing of multi-source observation data in step S1 specifically includes: downsampling the resolution of multi-band corona images from the Solar Dynamics Observatory and calculating the horizontal component of geomagnetic station data.

[0017] In one embodiment of the present invention, the white spectrum analysis of time series data specifically includes: first, performing a fast Fourier transform on the time series data to obtain a power spectrum; then, performing local maximum detection on the power spectrum and fitting it through cubic spline interpolation to obtain the upper envelope; after normalization, performing an inverse fast Fourier transform to obtain a perturbation time series after removing non-stationary trends.

[0018] In one embodiment of the present invention, the calculation formula for performing white spectrum analysis on time series data in step S1 is as follows:

[0019]

[0020] in, It is a time series with disturbances after removing non-stationary trends; It is a time series, with time as the axis, formed by the brightness or H component changes of each pixel or geomagnetic station observation point throughout the observation period, where t represents a discrete time point; yes The upper envelope of the power spectrum is obtained by cubic spline interpolation of the local maxima of the power spectrum; It is the value that appears most frequently in the dataset and is used as the normalization benchmark.

[0021] In one embodiment of the present invention, step S2, processing the perturbation time series specifically includes:

[0022] S201: Averaging the perturbation time series;

[0023] S202: The averaged perturbation time series is smoothed by a Savitzky-Golay filter to obtain the perturbation signal; during the filtering process, outliers are detected, and for missing points at the beginning and end of the sequence caused by the white spectrum method, polynomial fitting is used for correction.

[0024] In one embodiment of the present invention, in step S2, the global solar activity level index is obtained by averaging the perturbation time series of multi-band corona images from the Solar Dynamics Observatory to obtain an index characterizing the global solar activity level.

[0025] The mid-to-low latitude geomagnetic disturbance index is obtained by averaging the disturbance signals from mid-to-low latitude geomagnetic stations over time to characterize the intensity of mid-to-low latitude geomagnetic disturbances.

[0026] The polar geomagnetic disturbance index is an index that characterizes the intensity of polar geomagnetic disturbances by calculating the difference between the maximum and minimum values ​​of disturbance signals from polar geomagnetic stations at each time step.

[0027] In one embodiment of the present invention, step S3 specifically includes the following steps:

[0028] S301: Select valid data points with electron density in the altitude range of 60–800 km from satellite inversion data, remove data with electron density values ​​less than or equal to 0, take the natural logarithm of electron density to compress the numerical dynamic range, and calculate the solar zenith angle.

[0029] S302: Perform max-min normalization on the input data;

[0030] S303: Based on timestamps, the global solar activity level index, mid-latitude geomagnetic disturbance index, and polar geomagnetic disturbance index are left-joined with satellite inversion data according to timestamps, so that each satellite inversion data can be matched with the corresponding global solar activity level index, mid-latitude geomagnetic disturbance index, and polar geomagnetic disturbance index and satellite inversion data at the corresponding time.

[0031] S304: Construct a multi-layer fully connected feedforward neural network, which includes an input layer, hidden layers, and an output layer. The input layer has 12-dimensional features, including longitude, latitude, altitude, time, solar zenith angle, global solar activity level index, mid-to-low latitude geomagnetic disturbance index, and polar geomagnetic disturbance index. The hidden layer consists of 5 fully connected layers, each containing batch normalization and ReLU activation functions. The output layer consists of a single linear neuron, which outputs the four-dimensional electron density distribution of the ionosphere in terms of longitude, latitude, altitude, and time.

[0032] In one embodiment of the present invention, step S4 specifically includes the following steps:

[0033] S401: Time-series merging of ionospheric four-dimensional electron density distribution and total electron content observation data;

[0034] S402: Max-min normalization is performed on the four-dimensional electron density data and the total electron content observation data to eliminate the difference in data magnitude, and the data is reconstructed into a five-dimensional tensor form.

[0035] S403: Construct a spatiotemporal prediction model based on a 3D Swing Transformer architecture, which includes an input module, a PatchEmbedding layer, an encoding layer, a downsampling / upsampling module, a decoding layer, a Patch Unembedding layer, and an output recovery module;

[0036] The input module takes into account four-dimensional electron density data in five-dimensional tensor form and total electron content observation data.

[0037] The Patch Embedding layer uses blocks of size [2,4,4] to divide the input data into several local blocks and maps them to 96 dimensions through convolution or linear transformation;

[0038] The coding layer consists of 10 coding network layers and employs a multi-head self-attention mechanism to capture spatiotemporal dependencies at different scales through sliding window operations.

[0039] The downsampling / upsampling module performs spatial resolution downsampling and restoration at different stages of the network to achieve multi-scale feature modeling;

[0040] The decoding layer contains 10 layers of decoding network, forming a symmetrical structure with the encoding layer, and outputs the reconstructed high-dimensional features.

[0041] The Patch Unembedding layer uses convolution operations to restore the high-dimensional feature tensor output by the decoding layer to the original spatiotemporal dimension distribution.

[0042] Output recovery module: Outputs the ionospheric prediction results for future time moments, including the four-dimensional electron density distribution and total electron content prediction results of the ionosphere.

[0043] This invention also provides an ionospheric prediction device based on white spectrum method and deep learning, comprising:

[0044] The data processing module is used to preprocess multi-source observation data to obtain time series data, and then perform white spectrum analysis on the time series data to obtain perturbation time series.

[0045] The index generation module is used to process the disturbance time series and generate the global solar activity level index, the mid-to-low latitude geomagnetic disturbance index, and the polar geomagnetic disturbance index.

[0046] The electron density distribution generation module takes the global solar activity level index, mid-to-low latitude geomagnetic disturbance index, polar geomagnetic disturbance index and satellite inversion data as input, and uses a deep learning reconstruction model composed of a multi-layer fully connected feedforward neural network to output the four-dimensional electron density distribution of the ionosphere.

[0047] The prediction generation module is used to construct a spatiotemporal prediction model using the observation data of the four-dimensional electron density distribution and total electron content of the ionosphere as input, and output the prediction results of the ionosphere at future times.

[0048] Compared with existing technologies, the ionosphere prediction method and apparatus based on white spectrum method and deep learning according to the present invention have the following advantages:

[0049] 1. The proposed index system has achieved a breakthrough in timeliness: the JpG and JpD indices significantly reduce the monitoring delay of geomagnetic disturbances from 15-30 days for the AE index to approximately 0.5 days, and improve the temporal resolution to the minute level. The Jp index also reduces the monitoring delay of solar activity from 10-15 days for F10.7 to approximately 7 days, and improves the resolution to the hour level. This provides an unprecedented high-quality data foundation for the rapid response and near-real-time prediction of ionospheric models.

[0050] 2. Improved reconstruction and prediction accuracy of ionospheric models: Since the driving data (Jp, JpG, JpD) input to the model are "fresher" and more refined, they can more accurately reflect the real changes in the Earth's space environment. Therefore, deep learning models based on these data can more accurately capture the dynamic response of the ionosphere. Especially during violent disturbances such as geomagnetic storms, the reconstruction and prediction accuracy of the model is significantly improved compared with models that rely on traditional indices.

[0051] 3. Enhanced learning ability for complex spatiotemporal characteristics of the ionosphere: The 3D Swin Transformer model adopted has a natural advantage over existing LSTM and its variant models when processing four-dimensional (3D space + 1D time) data. It can learn spatial structure features and temporal evolution laws simultaneously, effectively capturing long-distance and cross-dimensional dependencies, thereby gaining a deeper understanding and simulation of the spatiotemporal evolution of the ionosphere, making the prediction results more physically consistent and more reliable in terms of accuracy. Attached Figure Description

[0052] Figure 1 This is a flowchart of an ionosphere prediction method based on white spectrum method and deep learning according to an embodiment of the present invention;

[0053] Figure 2 This is a flowchart of spatial chain-driven exponent generation according to an embodiment of the present invention.

[0054] Figure 3 This is a flowchart of a deep learning reconstruction model according to an embodiment of the present invention performing white spectrum analysis on time series data;

[0055] Figure 4 This is a flowchart of a spatiotemporal prediction model according to an embodiment of the present invention;

[0056] Figure 5 This is a schematic diagram of an ionospheric prediction device based on white spectrum method and deep learning according to an embodiment of the present invention. Detailed Implementation

[0057] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings, but it should be understood that the scope of protection of the present invention is not limited to the specific embodiments.

[0058] Unless otherwise expressly stated, throughout the specification and claims, the term "comprising" or its variations such as "including" or "comprises" shall be understood to include the stated elements or components without excluding other elements or other components.

[0059] like Figure 1 As shown, a preferred embodiment of the ionosphere prediction method based on white spectrum analysis and deep learning according to the present invention includes the following steps:

[0060] Step S1: Preprocess the multi-source observation data to obtain time series data, and then perform white spectrum analysis on the time series data to obtain the perturbation time series.

[0061] Specifically, such as Figure 2As shown, the multi-source observational data includes observational data from the Solar Dynamics Observatory (SDO) multi-band corona images and geomagnetic observatories.

[0062] The Solar Dynamics Observatory (SDO) multi-band coronal images include coronal image data in the 94 Å, 131 Å, 171 Å, 193 Å, 211 Å, 304 Å and 335 Å bands, which can cover different temperature ranges of solar activity and comprehensively reflect the intensity and distribution of solar surface activity.

[0063] The Solar Dynamics Observatory (SDO) multi-band corona images include wavelengths of 94 Å, 131 Å, 171 Å, 193 Å, 211 Å, 304 Å, 335 Å, 1600 Å, and 1700 Å. It should be noted that the seven bands from 94 to 335 Å primarily represent the upper solar atmosphere, while 1600 and 1700 Å primarily represent the lower solar atmosphere. Therefore, the variations in the 1600 and 1700 Å bands are generally inconsistent with the variations in the other seven bands. Thus, images with wavelengths of 94, 131, 171, 193, 211, 304, and 335 Å are selected in this invention.

[0064] The geomagnetic station data includes observational data from globally distributed mid- and low-latitude geomagnetic stations and three-component (X, Y, Z) observational data from global polar geomagnetic stations.

[0065] Preprocessing of multi-source observation data includes resolution downsampling of SDO multi-band corona images and calculation of the horizontal component (H) of geomagnetic station data.

[0066] Resolution downsampling of SDO multi-band corona images includes: uniformly processing the AIA image of each band into a 100×100 pixel matrix, and extracting the brightness change of each pixel in the matrix over the entire time series as a time series, forming a total of 10,000 key pixel time series.

[0067] Specifically, the bicubic interpolation function of the OpenCV library is used, and the Catmull-Rom spline function is used to perform weighted calculation on the pixels in the 4×4 neighborhood around each pixel, with parameter a being -0.5. The original 4096×4096 image is downsampled into a 100×100 pixel matrix, and the brightness change of each pixel in the matrix is ​​extracted into a time series over the entire time series, forming a total of 10,000 time series.

[0068] The formula for calculating the horizontal component H of geomagnetic station data is as follows:

[0069]

[0070] Where X represents the northward geomagnetic component and Y represents the eastward geomagnetic component. The calculation of the horizontal component H ensures the consistency of the time series.

[0071] Then, time series were generated for the H component of each mid-latitude and polar geomagnetic station as a function of time. At the same time, the time series were time-aligned to ensure that the time intervals of all series were consistent.

[0072] The white spectrum analysis of time series data specifically includes: first, performing a fast Fourier transform on the time series data to obtain the power spectrum; then, detecting local maxima and fitting the power spectrum using cubic spline interpolation to obtain the upper envelope; after normalization, performing an inverse fast Fourier transform to obtain the perturbed time series after removing non-stationary trends. The calculation formula is as follows:

[0073]

[0074] in, It is a time series with disturbances after removing non-stationary trends; It is a time series, with time as the axis, formed by the brightness or H component changes of each pixel or geomagnetic station observation point throughout the observation period, where t represents a discrete time point; yes The upper envelope of the power spectrum is obtained by cubic spline interpolation of the local maxima of the power spectrum; It is the value that appears most frequently in the dataset and is used as the normalization benchmark.

[0075] By applying the white spectrum method to perform time-frequency analysis on the time series of the above 10,000 key pixels and the time series of the H components of all geomagnetic stations, non-stationary features were eliminated, and the corresponding perturbation time series were obtained respectively.

[0076] Step S2: Process the disturbance time series to generate the global solar activity level index, the mid-to-low latitude geomagnetic disturbance index, and the polar geomagnetic disturbance index.

[0077] like Figure 2 As shown, the specific processing of the perturbation time series includes:

[0078] S201: Averaging the perturbation time series:

[0079]

[0080] in, It is a perturbation time series The average value is denoted by n, where n is the number of perturbation time series.

[0081] S202: The Savitzky-Golay filter is used to smooth the fluctuations in the averaged perturbation time series; during the filtering process, outliers are detected, and for missing points at the beginning and end of the sequence caused by the white spectrum method, polynomial fitting is used for correction.

[0082] The Savitzky-Golay filter performs low-order polynomial fitting on continuous subsets of each sequence and denoises the signal using linear least squares. During the filtering process, outliers are detected: if a value exceeds the median within the window by ±3 times the standard deviation (σ), it is marked as an outlier. Outliers are replaced with the previous valid value in their neighborhood. The window ranges from 40 to 100 time points, and the optimal window size that minimizes the number of outliers is selected for processing.

[0083] For missing points at the beginning and end of the sequence caused by the white spectrum method, the specific correction method is: select n+1 valid data points (t) around the missing point. j ,g m (t j Using low-order polynomials Least squares fitting is performed to fill in missing values. Here, parameter n represents the control parameter for the number of neighboring data points used for missing value correction. When a missing value exists at a sequence endpoint, (n+1) valid data points are selected from its neighborhood, and the missing value is estimated using third-order polynomial least squares fitting (based on the best experimental results). The value of parameter n is determined based on signal smoothness and local variation characteristics, and is usually comparable to the Savitzky-Golay filter window length.

[0084] Finally, the corrected perturbation time series The perturbation signal is obtained through Savitzky-Golay filtering. :

[0085]

[0086] in, It is a perturbation time series The standard deviation is used for normalization. The disturbance index is for each band; SG represents the Savitzky-Golay filter operation.

[0087] The Global Solar Activity Index (Jp Index) is an index characterizing the level of global solar activity, derived by averaging perturbation signals from SDO multi-band corona images; its calculation formula is as follows:

[0088]

[0089] in, Let m represent the perturbation signal of the i-th channel in the perturbation time series of the SDO multiband corona image, and m represent the number of channels ultimately retained, i.e., m=7. Originally, the Solar Dynamics Observatory (SDO) multiband corona image had 9 channels. Since the 1600 Å and 1700 Å channels are uncorrelated, the 7 channels 94 Å, 131 Å, 171 Å, 193 Å, 211 Å, 304 Å, and 335 Å are the most effective, hence m=7.

[0090] The mid-to-low latitude geomagnetic disturbance index (JpG index) is an index that characterizes the intensity of geomagnetic disturbances in mid-to-low latitudes by averaging the disturbance signals from mid-to-low latitude geomagnetic stations over time. Its calculation formula is as follows:

[0091]

[0092] Where m represents the number of geomagnetic observatories in the mid- and low-latitude regions. This represents the disturbance signal of the disturbance time series of the i-th mid-to-low latitude geomagnetic station.

[0093] The polar geomagnetic disturbance index (JpD index) is the difference (range) between the maximum and minimum values ​​of disturbance signals from polar geomagnetic stations, calculated hourly. The range reflects the unevenness of disturbance amplitudes among polar stations and provides an index characterizing the intensity of polar geomagnetic disturbances. Its calculation formula is as follows:

[0094]

[0095] in, This represents the disturbance signal of the disturbance time series of the i-th mid-to-low latitude geomagnetic station.

[0096] In this step, by characterizing the multi-source observation data processed by the white spectrum method, indices characterizing the level of solar activity, the intensity of geomagnetic disturbances in mid- and low-latitude regions, and the intensity of geomagnetic disturbances in polar regions are constructed. Together, they constitute the response index system of the solar-geomagnetic-ionospheric system, which can reflect the solar-geomagnetic coupling process with higher timeliness and resolution.

[0097] Step S3: Using the index generated in Step S2 and the satellite inversion data as input, a deep learning reconstruction model composed of a multi-layer fully connected feedforward neural network is used to output the four-dimensional electron density distribution of the ionosphere. The satellite inversion data is the electron density distribution provided by the COSMIC satellite inversion data, including longitude, latitude, altitude, time, solar zenith angle, and observed electron density values.

[0098] Specifically, such as Figure 3 As shown, step S3 specifically includes the following steps:

[0099] S301: Clean the input data; specifically, select valid data points with electron density altitude ranges of 60–800 km from the satellite inversion data, remove data with electron density values ​​less than or equal to 0, take the natural logarithm of the electron density to compress the numerical dynamic range, and calculate the solar zenith angle using the following formula:

[0100]

[0101] in, Indicates the latitude of the observation point. The solar declination angle, For the hour angle, The solar zenith angle represents the intensity of solar radiation incident on the ionosphere at the observation point and is an important parameter affecting electron density.

[0102] S302: Perform max-min normalization on the input data, with the normalization range limited to [0,1], to eliminate dimensional differences between features. The specific calculation method is as follows:

[0103]

[0104] in, These are the original data values. It is the minimum value in the dataset. It is the maximum value in the dataset, ensuring that different features have the same dimensions and range.

[0105] S303: Based on the timestamp, the index generated in step S2 is left-joined with the satellite inversion data (electron density distribution provided by COSMIC satellite inversion data) so that each satellite inversion data can be matched with the index at the corresponding time. Since the time resolution of the index is on the hourly level, matching is only required in the three dimensions of year, day, and hour.

[0106] S304: Construct a multi-layer fully connected feedforward neural network, which includes an input layer, hidden layers, and an output layer.

[0107] The input feature vector of the input layer is 12-dimensional, specifically including longitude, latitude, altitude, time, solar zenith angle, Jp index, JpG index, and JpD index. Latitude is expressed as... and The two components represent the time characteristics, which are further expanded into four dimensions: year, day (year-to-day), hour, and minute.

[0108] The hidden layers consist of five fully connected layers with the following neuron counts: 512, 256, 128, 64, and 16. Each layer includes a linear mapping operation, a batch normalization (BN) operation, and a ReLU activation function. Batch normalization improves training stability and accelerates convergence while effectively suppressing overfitting; the ReLU activation function introduces nonlinear mapping capabilities. Through layer-by-layer linear transformation, normalization, and nonlinear activation, the network of this invention can learn the complex nonlinear relationship between input features and target electron density.

[0109] Output layer: Set up 1 linear output neuron, outputting the four-dimensional electron density distribution of the ionosphere corresponding to the input features in terms of longitude, latitude, altitude and time.

[0110] During model training, mean squared error (MSE) is used as the loss function, and the specific calculation method is as follows:

[0111]

[0112] in, It is the actual output. The model predicts the output, and the mean squared error is used to measure the deviation between the predicted value and the actual electron density value. Simultaneously, a validation set is used to monitor the training process to prevent overfitting and ensure the model's generalization ability. During the continuation process, the Adam optimizer is used for optimization, with 100 training epochs set and an early stopping mechanism employed; training terminates if the loss does not decrease after 10 epochs.

[0113] Through the aforementioned deep neural network modeling process, the four-dimensional electron density distribution of the ionosphere can be reconstructed under the driving force of spatial chain-driven exponents and multi-dimensional spatiotemporal features, achieving a high-precision characterization of the ionospheric state. Applying the trained model to new input data outputs a predicted field of the four-dimensional electron density distribution of the ionosphere, with longitude, latitude, altitude, and time as dimensions; this represents the four-dimensional electron density distribution of the ionosphere, achieving a high-precision reconstruction of the ionospheric state. However, relying solely on the reconstruction results is insufficient to meet the needs of future ionospheric predictions. To achieve forward-looking forecasts of the ionospheric state, it is necessary to further combine the reconstructed four-dimensional electron density with global total electron content (TEC) observation data to construct a prediction network capable of simultaneously modeling spatial and temporal dependencies, thereby achieving a high-precision prediction of the future state of the ionosphere.

[0114] Step S4: Using the four-dimensional electron density distribution and total electron content (TEC) observation data of the ionosphere output in Step S3 as input, the spatiotemporal prediction model constructed using the three-dimensional Swing Transformer architecture outputs the ionosphere prediction results for future times.

[0115] Specifically, such as Figure 4 As shown, step S4 specifically includes the following steps:

[0116] S401: Merge the four-dimensional electron density distribution of the ionosphere output from step S3 with the total electron content observation data in a time series.

[0117] The merged dataset was divided into training and validation sets in a ratio of approximately 30:1. Using the Dst index as the standard, the data from the training years 2011–2018 were divided into storm periods and calm periods on a weekly basis. In each division, one storm period and nine calm periods were selected as the validation set, and the remaining data were used as the training set, thus ensuring that the model has generalization ability in both calm and storm periods.

[0118] S402: Max-min normalize the four-dimensional electron density data and the total electron content observation data to eliminate the difference in data magnitude, and reconstruct the data into a five-dimensional tensor form (batch, time, altitude, longitude, latitude).

[0119] The normalization range is limited to [0,1] to eliminate dimensional differences between features. The specific calculation method is as follows:

[0120]

[0121] in, These are the original data values. It is the minimum value in the dataset. It represents the maximum value in the dataset, ensuring that different features have the same dimensions and range.

[0122] S403: Constructs a spatiotemporal prediction model based on a 3D Swing Transformer architecture, which includes an input module, a PatchEmbedding layer, an encoding layer, a downsampling / upsampling module, a decoding layer, a Patch Unembedding layer, and an output recovery module.

[0123] The input module takes into account four-dimensional electron density data in five-dimensional tensor form and total electron content observation data.

[0124] The Patch Embedding layer uses [2,4,4]-sized patches to divide the input data into several local blocks and maps them to 96 dimensions through convolution or linear transformation;

[0125] The encoder consists of 10 layers (4 basic encoder layers and 6 advanced encoder layers) and employs a multi-head self-attention mechanism (with 4, 6, 12, and 24 heads respectively). It captures spatiotemporal dependencies at different scales through sliding window operations. The advanced encoder layers introduce cross-window attention to enhance the learning ability of long-distance spatiotemporal associations.

[0126] The downsampling / upsampling module performs spatial resolution downsampling and restoration at different stages of the network to achieve multi-scale feature modeling. Specifically, the downsampling module uses a sampling factor of [1,2,2] (no sampling in the time dimension, and halving of the height and spatial dimensions) to achieve feature dimensionality reduction and multi-scale feature extraction. The upsampling module uses transposed convolution to restore the feature dimension with a sampling factor of [1,2,2] to ensure that the spatiotemporal resolution of the output is consistent with that of the input.

[0127] The decoder layer has 10 layers of decoding network (divided into 6 high-level decoding layers and 4 basic decoding layers), forming a symmetrical structure with the encoder layer, gradually restoring the original spatial dimensions and outputting the reconstructed high-dimensional features;

[0128] The Patch Unembedding layer uses convolution operations to restore the high-dimensional feature tensor output by the decoding layer to the original spatiotemporal dimension distribution.

[0129] Output recovery module: Outputs the ionospheric prediction results for future time moments, including the four-dimensional electron density distribution and total electron content prediction results of the ionosphere. The future time moment depends on the duration of the input data; the longer the input, the richer the information, and theoretically, a more stable prediction can be obtained further into the future. However, the error will accumulate with the number of recursion steps.

[0130] For the spatiotemporal prediction model based on the 3D Swing Transformer architecture, the AdamW optimizer (with a learning rate of 1×10⁻) is employed. 4 The weight decay coefficient is set to 1×10⁻ 5 The learning rate scheduling strategy is to decay to 0.5 of the original value every 20 training epochs, the batch size is set to 4, the number of training epochs is set to 100, the loss function is MSE, and the model performance is monitored in real time through the validation set to adjust the training parameters.

[0131] like Figure 5 As shown, an ionospheric prediction device based on white spectrum method and deep learning according to a preferred embodiment of the present invention includes:

[0132] Data processing module 1 is used to preprocess multi-source observation data to obtain time series data, and then perform white spectrum analysis on the time series data to obtain perturbation time series;

[0133] Index generation module 2 is used to process the disturbance time series and generate the global solar activity level index, the mid-to-low latitude geomagnetic disturbance index, and the polar geomagnetic disturbance index;

[0134] The electron density distribution generation module 3 is used to take the global solar activity level index, the mid-to-low latitude geomagnetic disturbance index, the polar geomagnetic disturbance index and satellite inversion data as input, and use a deep learning reconstruction model composed of a multi-layer fully connected feedforward neural network to output the four-dimensional electron density distribution of the ionosphere.

[0135] Prediction generation module 4 is used to take the observation data of the four-dimensional electron density distribution and total electron content of the ionosphere as input, and construct a spatiotemporal prediction model using the three-dimensional Swing Transformer architecture to output the ionosphere prediction results at future times.

[0136] The foregoing description of specific exemplary embodiments of the invention is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it will be apparent that many changes and variations can be made in accordance with the foregoing teachings. The exemplary embodiments were chosen and described in order to explain the specific principles of the invention and its practical application, thereby enabling those skilled in the art to implement and utilize various different exemplary embodiments of the invention, as well as various different choices and variations. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. An ionospheric prediction method based on white spectrum analysis and deep learning, characterized in that, Includes the following steps: S1: Preprocess the multi-source observation data to obtain time series data, and then perform white spectrum analysis on the time series data to obtain the perturbation time series; S2: Process the disturbance time series to generate the global solar activity level index, the mid-to-low latitude geomagnetic disturbance index, and the polar geomagnetic disturbance index; S3: Using global solar activity level index, mid-to-low latitude geomagnetic disturbance index, polar geomagnetic disturbance index and satellite inversion data as input, a deep learning reconstruction model composed of multi-layer fully connected feedforward neural network is used to output the four-dimensional electron density distribution of the ionosphere. S4: Using observational data on the four-dimensional electron density distribution and total electron content of the ionosphere as input, a spatiotemporal prediction model constructed with a three-dimensional SwinTransformer architecture is used to output the ionosphere prediction results for future times.

2. The ionosphere prediction method based on white spectrum method and deep learning as described in claim 1, characterized in that, In step S1, the multi-source observation data includes observation data from the Solar Dynamics Observatory's multi-band coronal images and geomagnetic stations; the Solar Dynamics Observatory's multi-band coronal images include coronal image data in the 94Å, 131Å, 171Å, 193Å, 211Å, 304Å and 335Å bands; the geomagnetic station data includes observation data from globally distributed mid- and low-latitude geomagnetic stations and three-component observation data from global polar geomagnetic stations.

3. The ionosphere prediction method based on white spectrum method and deep learning as described in claim 2, characterized in that, The preprocessing of multi-source observation data in step S1 specifically includes: downsampling the resolution of multi-band corona images from the Solar Dynamics Observatory and calculating the horizontal component of data from geomagnetic stations.

4. The ionosphere prediction method based on white spectrum method and deep learning as described in claim 1, characterized in that, The white spectrum analysis of time series data specifically includes: first, performing a fast Fourier transform on the time series data to obtain the power spectrum; then, performing local maxima detection on the power spectrum and fitting it through cubic spline interpolation to obtain the upper envelope; after normalization, performing an inverse fast Fourier transform to obtain the perturbation time series after removing non-stationary trends.

5. The ionosphere prediction method based on white spectrum method and deep learning as described in claim 1, characterized in that, In step S1, the calculation formula for white spectrum analysis of time series data is as follows: ； in, It is a time series with disturbances after removing non-stationary trends; It is a time series, with time as the axis, formed by the brightness or H component changes of each pixel or geomagnetic station observation point throughout the observation period, where t represents a discrete time point; yes The upper envelope of the power spectrum is obtained by cubic spline interpolation of the local maxima of the power spectrum; It is the value that appears most frequently in the dataset and is used as the normalization benchmark.

6. The ionosphere prediction method based on white spectrum method and deep learning as described in claim 1, characterized in that, Step S2, specifically the processing of the perturbation time series, includes: S201: Averaging the perturbation time series; S202: The averaged perturbation time series is smoothed by a Savitzky-Golay filter to obtain the perturbation signal; during the filtering process, outliers are detected, and for missing points at the beginning and end of the sequence caused by the white spectrum method, polynomial fitting is used for correction.

7. The ionosphere prediction method based on white spectrum method and deep learning as described in claim 5, characterized in that, In step S2, the global solar activity level index is obtained by averaging the perturbation time series of multi-band corona images from the Solar Dynamics Observatory, thus representing the global solar activity level. The mid-to-low latitude geomagnetic disturbance index is obtained by averaging the disturbance signals from mid-to-low latitude geomagnetic stations over time to characterize the intensity of mid-to-low latitude geomagnetic disturbances. The polar geomagnetic disturbance index is an index that characterizes the intensity of polar geomagnetic disturbances by calculating the difference between the maximum and minimum values ​​of disturbance signals from polar geomagnetic stations at each time step.

8. The ionosphere prediction method based on white spectrum method and deep learning as described in claim 1, characterized in that, Step S3 specifically includes the following steps: S301: Select valid data points with electron density in the altitude range of 60–800 km from satellite inversion data, remove data with electron density values ​​less than or equal to 0, take the natural logarithm of electron density to compress the numerical dynamic range, and calculate the solar zenith angle. S302: Perform max-min normalization on the input data; S303: Based on timestamps, the global solar activity level index, mid-latitude geomagnetic disturbance index, and polar geomagnetic disturbance index are left-joined with satellite inversion data according to timestamps, so that each satellite inversion data can be matched with the corresponding global solar activity level index, mid-latitude geomagnetic disturbance index, and polar geomagnetic disturbance index and satellite inversion data at the corresponding time. S304: Construct a multi-layer fully connected feedforward neural network, which includes an input layer, hidden layers, and an output layer. The input layer has 12-dimensional features, including longitude, latitude, altitude, time, solar zenith angle, global solar activity level index, mid-to-low latitude geomagnetic disturbance index, and polar geomagnetic disturbance index. The hidden layer consists of 5 fully connected layers, each containing batch normalization and ReLU activation functions. The output layer consists of a single linear neuron, which outputs the four-dimensional electron density distribution of the ionosphere in terms of longitude, latitude, altitude, and time.

9. The ionosphere prediction method based on white spectrum method and deep learning as described in claim 1, characterized in that, Step S4 specifically includes the following steps: S401: Time-series merging of ionospheric four-dimensional electron density distribution and total electron content observation data; S402: Max-min normalization is performed on the four-dimensional electron density data and the total electron content observation data to eliminate the difference in data magnitude, and the data is reconstructed into a five-dimensional tensor form. S403: Construct a spatiotemporal prediction model based on a 3D Swing Transformer architecture, which includes an input module, a PatchEmbedding layer, an encoding layer, a downsampling / upsampling module, a decoding layer, a Patch Unembedding layer, and an output recovery module; The input module takes into account four-dimensional electron density data in five-dimensional tensor form and total electron content observation data. The Patch Embedding layer uses blocks of size [2,4,4] to divide the input data into several local blocks and maps them to 96 dimensions through convolution or linear transformation; The coding layer consists of 10 coding network layers and employs a multi-head self-attention mechanism to capture spatiotemporal dependencies at different scales through sliding window operations. The downsampling / upsampling module performs spatial resolution downsampling and restoration at different stages of the network to achieve multi-scale feature modeling; The decoding layer contains 10 layers of decoding network, forming a symmetrical structure with the encoding layer, and outputs the reconstructed high-dimensional features. The Patch Unembedding layer uses convolution operations to restore the high-dimensional feature tensor output by the decoding layer to the original spatiotemporal dimension distribution. Output recovery module: Outputs the ionospheric prediction results for future time moments, including the four-dimensional electron density distribution and total electron content prediction results of the ionosphere.

10. An ionospheric prediction device based on white spectrum method and deep learning, characterized in that, include: The data processing module is used to preprocess multi-source observation data to obtain time series data, and then perform white spectrum analysis on the time series data to obtain perturbation time series. The index generation module is used to process the disturbance time series and generate the global solar activity level index, the mid-to-low latitude geomagnetic disturbance index, and the polar geomagnetic disturbance index. The electron density distribution generation module takes the global solar activity level index, mid-to-low latitude geomagnetic disturbance index, polar geomagnetic disturbance index and satellite inversion data as input, and uses a deep learning reconstruction model composed of a multi-layer fully connected feedforward neural network to output the four-dimensional electron density distribution of the ionosphere. The prediction generation module is used to construct a spatiotemporal prediction model using the observation data of the four-dimensional electron density distribution and total electron content of the ionosphere as input, and output the prediction results of the ionosphere at future times.

Images