CN-121980529-A - Global ionosphere prediction method based on latitude self-adaptive blocking mechanism

CN121980529ACN 121980529 ACN121980529 ACN 121980529ACN-121980529-A

Abstract

The invention relates to a global ionosphere prediction method based on a latitude self-adaptive partitioning mechanism, which comprises the steps of S1 downloading and analyzing GIM products, S2 carrying out normalization processing on VTECs of an ionosphere based on the analyzed GIM products, dividing the processed VTECs into a training set and a verification set, S3 constructing a latitude self-adaptive Vision transformer model and initializing model parameters, S4 designing a latitude partition loss function for the model, S5 training the model by adopting the training set, minimizing the latitude partition loss function to optimize the model parameters, finishing training when the model meets early-stop standards, storing the trained model, S6 carrying out inverse normalization processing on the vertical total electronic content value of the predicted ionosphere output by the model, and generating the predicted GIM products.

Inventors

LI HUI
ZHU CHUANFENG
JIA CHUN
ZANG NAN

Assignees

哈尔滨工程大学

Dates

Publication Date: 20260505
Application Date: 20260121

Claims (7)

1. A global ionosphere prediction method based on a latitude self-adaptive blocking mechanism comprises the following steps: step S1), downloading and analyzing GIM products broadcast by a global satellite navigation service IGS, dividing the analyzed GIM products into a plurality of space grids in a grid mode with equal longitude and latitude, and setting certain time as time resolution; Step S2) based on the GIM product obtained by analysis, carrying out normalization processing on the vertical total electron content of the ionized layer, and dividing the processed vertical total electron content data into a training set and a verification set; step S3), constructing a Vision transformer model with self-adaptive latitude and initializing parameters of the model; Step S4) designing a latitude partition loss function aiming at the self-adaptive Vision transformer model of the latitude; Step S5) training the latitude self-adaptive Vision transformer model by adopting a training set, minimizing the latitude partition loss function to optimize model parameters, ending training when the model meets the early-stop standard, and storing a trained Vision transformer model; step S6) inputting the analyzed GIM product into the trained Vision transformer model, and performing inverse normalization processing on the predicted vertical total electron content value output by the model to finally generate the predicted GIM product conforming to the IONEX format.
2. The global ionosphere prediction method based on the latitude adaptive blocking mechanism according to claim 1, wherein in step S1, GIM products are divided into 71 x 73 spatial grids by equal longitude and latitude grids, time resolution is set to 2 hours, longitude resolution is set to 5 °, and latitude resolution is set to 2.5 °.
3. The global ionosphere prediction method based on the latitude adaptive blocking mechanism according to claim 1, wherein in step S2, outliers and missing values are eliminated based on the labels before normalization processing.
4. The global ionosphere prediction method based on the latitude adaptive blocking mechanism according to claim 1, wherein in step S3, a leachable latitude embedding parameter is a priori added for each latitude corresponding position when the input data is processed through the convolution layer: ; Wherein the method comprises the steps of Representing the features after the position embedding, Representing a learnable latitude embedding parameter, In order to parameterize the convolution operator, , For convolution operations, X is the input data.
5. The global ionosphere prediction method based on the latitude adaptive blocking mechanism according to claim 4, wherein in step S3, the output of multiple attention is added back to the original input through linear mapping, resulting in: ; Wherein, the For the output of a single-layer Vision transformer encoder, The representation layer is normalized and, Is a multi-layer perceptron.
6. The global ionosphere prediction method based on the latitude adaptive blocking mechanism according to claim 1, wherein in step S4, the latitude partition loss function is: ; Wherein the method comprises the steps of As a function of the total loss, As a loss function in the low latitude region, A loss function of a mid-latitude region, As a loss function in the high latitude region, Is the weight coefficient of the low latitude region, Is the weight coefficient of the mid-latitude region, The weight coefficient is the weight coefficient of the high latitude area; The loss function of the low latitude region is: ; The loss function of the mid-latitude region is as follows: ; The loss function of the high latitude region is as follows: ; Wherein, the A set of grid points representing a low latitude region, A set of grid points representing a mid-latitude region, A set of grid points representing a high latitude region, Representing the point of the ith, j grid, Representing the position of the ith, j grid points, Representing the i, j-th grid point, Is the true value of the ionospheric vertical total electron content at the i, j grid point recorded in the dataset, Ionospheric vertical total electron content for the ith, j grid point predicted for Vision transformer model.
7. The global ionosphere prediction method based on the latitude adaptive blocking mechanism according to claim 1, wherein in step S5, model training is performed by adopting an Adam optimizer, a time-along backward transmission method is adopted, and an initial learning rate is set as follows And when the loss value calculated by using the verification set is subjected to continuous rounds and no more decline or the decline amplitude is smaller than a preset threshold value, ending training.

Description

Global ionosphere prediction method based on latitude self-adaptive blocking mechanism Technical Field The invention relates to the technical field of satellite navigation, in particular to a global ionosphere prediction method based on a latitude self-adaptive blocking mechanism. Background The ionosphere is an ionized region of the earth's atmosphere with a height of about 60 to 1000 km, and is filled with a large number of charged particles and free electrons, the spatial distribution of electron density of which significantly affects the propagation characteristics of electromagnetic waves. The total electron content (Total Electron Content, TEC) is taken as a core parameter for representing the ionosphere state, quantitatively describes the integral value of electron density on the propagation path of an electromagnetic signal, and directly determines the propagation delay and phase disturbance of the electromagnetic wave. In order to more clearly describe the spatial distribution characteristics of the total electronic content of the ionosphere, global satellite navigation service (IGS) ionosphere working groups have proposed global ionosphere grid products of the total electronic content of the ionosphere, namely GIM (Global Ionosphere Map) products, inverted the total electronic content of each ionosphere grid point by using the whole-day observables of near 400 GNSS observables on the ground, and formed the products to face global users in the related fields. With the wide application of Global Navigation Satellite Systems (GNSS), the rapid deployment of low-orbit satellite constellations and the increasingly urgent demands of 6G communication on the world-wide integrated network, high-precision ionosphere TEC prediction has become a core problem to be solved in the fields of satellite navigation positioning, spatial weather forecast, radio physics, communication engineering and the like. Traditional ionospheric predictions rely primarily on empirical models and statistical methods. The international reference ionosphere model, the Nequick model and the time sequence extrapolation method based on spherical harmonics have certain prediction capability in geomagnetic calm period, but have obvious limitations in describing complex structures such as rapid changes during geomagnetic disturbance and low latitude ionization abnormality, and are difficult to meet the severe requirements of modern application on prediction precision and timeliness. In recent years, deep learning techniques have provided new technological approaches to ionospheric prediction. Long Short-Term Memory (LSTM), convolutional Long-Term Memory (ConvLSTM), transform, and other architectures exhibit advantages in space-time sequence modeling. However, the existing deep learning method generally adopts equal longitude and latitude grids to carry out spatial discretization on the ionosphere, and ignores the geometric nature of the ionosphere as a spherical physical field. The equal longitude and latitude projection has inherent geometric distortion that the actual earth surface distance difference corresponding to the same longitude and latitude interval at different latitudes is obvious, the distance is maximum near the equator, and the sampling density is gradually reduced to zero towards the two poles, so that the sampling density is seriously uneven on the spherical surface. The geometric distortion not only affects the correct learning of the model on the spatial correlation, but also restricts the further improvement of the prediction precision. Therefore, development of a new deep learning architecture capable of fundamentally solving spherical geometric distortion while conforming to ionospheric physical characteristics is urgently required. Disclosure of Invention In order to solve the problems, the invention provides a global ionosphere prediction method based on a latitude self-adaptive blocking mechanism. The method is specifically applied to the data processing center to obtain global ionosphere grid products, the products provide space-time distribution information of the global ionosphere Vertical Total Electron Content (VTEC) in the IONEX format, and data preprocessing is carried out on the data. Secondly, historical VTEC data are unfolded along a time dimension, and a multi-channel input feature is constructed. The spatial grid data is converted into a latitude band sequence by a convolution embedding layer. Aiming at the difference of the circumferences of longitude circles at different latitude positions, latitude band information is dynamically generated to accurately represent the real space relation on the spherical surface, and the geometric distortion problem existing in the equal longitude and latitude grid representation is effectively solved. And thirdly, capturing long-distance dependence relations among different spatial positions by adopting a multi-layer Vision transformer model, and learning global time