Ionized layer irregular body prediction method based on residual compensation

Abstract

The invention relates to the technical field of satellite positioning, in particular to an ionosphere irregular body prediction method based on residual compensation, which comprises the following steps: acquiring relevant ionosphere environment parameters, and constructing a feature data set after preprocessing; and then dividing a data set into a training set and a test set, performing hyper-parameter optimization on the RF model by using a WHO algorithm, and establishing a WHO-RF prediction model to solve the problem of insufficient robustness of a single RF model. After training of a WHO-RF model is completed by using training set data, ROTI is preliminarily predicted through the model, a residual sequence between a predicted value and a measured value is obtained, then a residual compensation model is constructed by using LSTM to improve the problem of low prediction precision of an existing prediction model, and a WHO-RF-LSTM hybrid prediction framework is formed. According to the combined model, global feature extraction of WHO-RF and time sequence residual compensation of LSTM are combined, the ionosphere irregular body intensity prediction precision can be remarkably improved, and therefore the problem that an existing prediction method is low in precision is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of satellite positioning, and particularly to a method for predicting ionospheric irregularities based on residual compensation. Background Art

[0002] The ionosphere is a partially ionized region in the upper atmosphere of the Earth, approximately located in the altitude range of 60 - 1000 km, which contains enough free electrons and significantly affects the propagation of radio waves. The ionosphere contains electron density irregularities of various scales, and generally, the ionosphere behaves as an anisotropic, random, and dispersive medium. These electron density irregular structures have a strong impact on radio wave signals, such as causing rapid and strong perturbations in the amplitude and phase of the radio wave signals, called ionospheric scintillation, which can cause time delay, multipath effect, etc. of the radio wave signals. Ionospheric scintillation caused by ionospheric irregularities has an important impact on radio signals such as satellite communication and satellite navigation, and even leads to the interruption or complete loss of system services. During solar storms, due to a large number of high-energy charged particles released by the sun towards the Earth, the resulting scintillation is more intense. Therefore, the prediction of ionospheric variability is of great significance for space engineering missions, system operation, and scientific research.

[0003] In the past decade, the occurrence intensity of ionospheric irregularities near the low-latitude region has been deeply studied by many scholars. Calculating the ROTI index based on inexpensive geodetic receivers is considered by many scholars to be a sign that can be used to describe and determine ionospheric irregularities and anomalies. If ROTI is used for predicting ionospheric scintillation irregularities, the operating cost will be greatly reduced. Researchers have developed support vector machines, gradient boosting algorithms, and hybrid integration models for predicting irregularities using neural networks.

[0004] However, due to the complex physical process of the formation of ionospheric irregularities, numerous variables, and different non-linear relationships between the variables, the accuracy of the current prediction models still needs to be improved. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for predicting ionospheric irregularities based on residual compensation, aiming to solve the problem of low accuracy of existing prediction methods.

[0006] To achieve the above purpose, the present invention provides a method for predicting ionospheric irregularities based on residual compensation, including the following steps:

[0007] Obtain relevant ionospheric environment parameters, and construct a feature data set after preprocessing;

[0008] Divide the feature data set into a training set and a test set, and use the moving average method to smooth the training set data;

[0009] Automatically optimize the hyperparameters of the RF model using the WHO algorithm to confirm the model parameters;

[0010] Input the training set data into the WHO-RF model for training to construct an ionospheric irregularity intensity prediction model;

[0011] Obtain the residual sequence and construct an LSTM residual compensation model;

[0012] Combine the WHO-RF model and the LSTM residual compensation model to obtain a hybrid prediction framework, and input the test set data into this combined prediction model to obtain the final prediction result of the ionospheric irregularity intensity.

[0013] Among them, in "Construct a feature dataset after preprocessing the relevant ionospheric environment parameters", the feature dataset includes the solar radio flux index F10.7, the geomagnetic index Dst, the interplanetary magnetic field components Bx / By / Bz, the magnetohydrodynamic pressure Flow Pressure, the critical frequency of the F2 layer foF2, the symmetric horizontal component SYM / H of the ring current, and the ionospheric disturbance index ROTI sequence.

[0014] Among them, in "Divide the feature dataset into a training set and a test set, and use the moving average method to smooth the training set data", the moving average method smooths the data with a moving window width of 12 points.

[0015] Among them, in "Automatically optimize the hyperparameters of the RF model using the WHO algorithm to confirm the model parameters", the following steps are included:

[0016] Determine the RF model parameters to be optimized, construct the RF model and use the root mean square error as the fitness function to calculate the parameter fitness value;

[0017] Gradually optimize the parameters by dynamically updating the positions of the stallions and foals. For each group of foals, update their positions according to the positions of the stallions and random factors;

[0018] Recalculate the fitness value for the updated positions and output the model parameters.

[0019] Among them, in "Input the training set data into the WHO-RF model for training to construct an ionospheric irregularity intensity prediction model", the following steps are included:

[0020] Divide the dataset into N subsets, and each subset is used to train a regression tree model;

[0021] Calculate the feature importance through the OOB error rate and analyze the contribution of each feature to the prediction;

[0022] After each regression tree is independently generated, the predicted results of ROTI are output respectively, and the arithmetic mean of the regression results obtained from N decision trees is used to obtain the prediction model for the intensity of ionospheric irregularities.

[0023] Among them, in the "construction of the LSTM residual compensation model", the following steps are included:

[0024] Calculate the difference between the initially output ROTI prediction value of the ionospheric irregularity prediction model and the true value of the training set to obtain the residual sequence;

[0025] Take the residual sequence and relevant ionospheric environment parameters as the input of the LSTM model, and calculate the hidden state and output of the current time step for the data of each time step;

[0026] After each time step is completed, use the BPTT algorithm to calculate the error gradient, update the weight parameters of the LSTM, and repeat the iterative training process until the model converges to obtain the LSTM residual compensation model.

[0027] The method for predicting ionospheric irregularities based on residual compensation of the present invention includes the following steps: obtaining relevant ionospheric environment parameters, constructing a feature data set after preprocessing; dividing the feature data set into a training set and a test set, and using the moving average method to smooth the training set data; automatically optimizing the hyperparameters of the RF model using the WHO algorithm to confirm the model parameters; inputting the training set data into the WHO-RF model for training to construct a prediction model for the intensity of ionospheric irregularities; obtaining the residual sequence and constructing an LSTM residual compensation model; combining the WHO-RF model and the LSTM residual compensation model to obtain a hybrid prediction framework, and inputting the test set data into this combined prediction model to obtain the final prediction result of the intensity of ionospheric irregularities. The present invention constructs a feature data set after obtaining relevant ionospheric environment parameters through preprocessing; then divides the data set into a training set and a test set, and uses the WHO algorithm to optimize the hyperparameters of the RF model to establish a WHO-RF prediction model to solve the problem of insufficient robustness of a single RF model. After completing the training of the WHO-RF model using the training set data, the ROTI is preliminarily predicted through this model, the residual sequence between the predicted value and the measured value is obtained, and then an LSTM is used to construct a residual compensation model to improve the problem of low prediction accuracy of the existing prediction model. Finally, a WHO-RF-LSTM hybrid prediction framework is formed. This combined model combines the global feature extraction of WHO-RF and the temporal residual compensation of LSTM, which can significantly improve the prediction accuracy of the intensity of ionospheric irregularities and provide more reliable technical support for the signal scintillation interruption warning of navigation and electronic communication systems. Thus, the problem of low accuracy of the existing prediction methods is solved. Description of the Drawings

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0029] Figure 1 is a flowchart of the ionospheric irregularity prediction method based on residual compensation provided by the present invention.

[0030] Figure 2 is a specific schematic diagram of the ionospheric irregularity prediction method based on residual compensation of the present invention.

[0031] Figure 3 is a schematic diagram of the WHO algorithm process.

[0032] Figure 4 is a schematic diagram of the principle of the WHO-RF prediction model.

[0033] Figure 5 is a flowchart for automatically optimizing the hyperparameters of the RF model using the WHO algorithm and confirming the model parameters.

[0034] Figure 6 is a flowchart for training the training set data in the WHO-RF model to construct an ionospheric irregularity intensity prediction model.

[0035] Figure 7 is a flowchart for constructing an LSTM residual compensation model. Specific Embodiments

[0036] The following details the embodiments of the present invention. The examples of the embodiments are shown in the drawings, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions throughout. The embodiments described below by referring to the drawings are exemplary and are intended to explain the present invention, and should not be construed as limiting the present invention.

[0037] Please refer to Figures 1 to 7 , the present invention provides an ionospheric irregularity prediction method based on residual compensation, including the following steps:

[0038] S1 Obtain relevant ionospheric environment parameters and construct a feature dataset after preprocessing;

[0039] The feature dataset includes the solar radio flux index F10.7, the geomagnetic index Dst, the interplanetary magnetic field components Bx / By / Bz, the magnetohydrodynamic pressure Flow Pressure, the critical frequency foF2 of the F2 layer, the symmetric horizontal component SYM / H of the ring current, and the ionospheric disturbance index ROTI sequence.

[0040] Specifically, ionospheric background environment parameters related to ionospheric irregularities are extracted from NASA's SPDF (Space Physics Data Facility) products. The eigenvalues used include: the components Bx / By / Bz of the interplanetary magnetic field (IMF), which controls the disturbance changes in the Earth's magnetosphere; the magnetohydrodynamic pressure (Flow Pressure), which is the pressure generated by the interaction between the solar wind and the Earth's magnetic field and characterizes the magnetic field disturbance; the solar radio flux index (F10.7), which measures the level of solar activity and has a strong correlation with sunspots; the geomagnetic activity (Dst) index and the symmetric horizontal component (SYM / H) of the ring current, which are used to judge the degree of magnetic storms;

[0041] Using the scientific data products of the Meridian Project, a national major scientific and technological infrastructure, the critical frequency of the F2 layer (foF2) is extracted. It is the frequency corresponding to the maximum electron density in the ionospheric F2 layer and is closely related to the state of the ionosphere (especially the distribution of electron density);

[0042] The diurnal variation factors HRS and HRC are calculated to reflect the dependence on time (HRS = sin(2π×H / 24), HRC = cos(2π×H / 24), where H represents the time of the hour under consideration, so 0 ≤ H ≤ 23);

[0043] GNSS dual-frequency observation data with a sampling rate of 30 s is obtained. Before calculating the ionospheric TEC rate of change index, i.e., the ionospheric disturbance parameter ROTI, the GNSS data needs to be preprocessed first. The MW combination method is used for cycle slip detection: the combined observable is obtained by subtracting the narrow-lane combination pseudorange from the wide-lane combination phase, and then the difference between epochs is calculated to detect cycle slips and eliminate gross errors. Taking the GPS signal as an example, the wide-lane ambiguity at the i-th epoch is:

[0044]

[0045] where the phase observations at different frequencies are represented by L1 and L2, the frequency observations are represented by f1 and f2, and λ w is the wide-lane combination wavelength. Let be the mean of the wide-lane ambiguities of consecutive epochs; be the variance value of the wide-lane ambiguity at the consecutive i-th epoch, then and The calculation formulas for are:

[0046]

[0047] For judgment is made after taking the difference between epochs, that is, judging Whether it holds. It can be seen that when there is no cycle slip, the measured quantity basically fluctuates stably around zero. If a cycle slip occurs, the wide-lane ambiguity is corrected to to eliminate the influence of the cycle slip. The repair formula is:

[0048]

[0049] where, ΔN w is the jump variable of the wide-lane ambiguity. Then, the ROTI is calculated using the standard deviation of the repaired dual-frequency observation data within 5 minutes by ROT, with the unit of TECU / min. The calculation formula is as follows:

[0050]

[0051] where, the epoch interval is Δt.

[0052] S2 divides the feature data set into a training set and a test set, and uses the moving average method to smooth the training set data;

[0053] The moving average method smooths the data, and the width of the moving window used is 12 points.

[0054] Specifically, the irregular body prediction original data set obtained in step S1 is divided into a training set and a test set according to a certain ratio; the moving average method is used to smooth the data to reduce the influence of random noise. For time series data {x1, x2,..., x T}, the smoothed value is calculated as:

[0055]

[0056] where, w = 12 is the moving window size, x t-i is the i-th data point before the current moment t, and can be calculated only when t ≥ w.

[0057] S3 uses the WHO algorithm to automatically optimize the hyperparameters of the RF model and confirm the model parameters;

[0058] S31 determines the RF model parameters to be optimized, constructs the RF model and uses the root mean square error as the fitness function to calculate the parameter fitness value;

[0059] Specifically, in the process of automatically optimizing the parameters of the RF model using the WHO algorithm, the algorithm first determines that the RF model parameters to be optimized are the number of decision trees n, the number of iterations m, and initializes the population: randomly generate N individuals (parameter combinations), and each individual is represented as x i =(n, m). The population is divided into studs (leaders) and foals (followers). And for each individual xi , construct an RF model and use the root mean square error (RMSE) as the fitness function:

[0060]

[0061] t_sim i is the predicted value of the i-th sample, and t_train i is the true value of the i-th sample, and M is the number of samples in the training set. During the optimization process, the algorithm will search for the parameter combination x that minimizes the RMSE i ;

[0062] S32 gradually optimizes the parameters by dynamically updating the positions of the stallions and foals. For each group of foals, update their positions according to the positions of the stallions and random factors;

[0063] Specifically, gradually optimize the parameters by dynamically updating the positions of the stallions and foals. For each group of foals, update their positions according to the positions of the stallions and random factors:

[0064]

[0065] In the above formula, Stallion is the position of the leader; R is a random number between [-2, 2], mainly used to control the angle between the individual and the leader. The calculation expression of the adaptive mechanism Z is:

[0066] P = R1 < TDR; IDX = (P == 0); Z = R2·IDX + R3·(~IDX)

[0067] In the above formula, P is a vector composed of 0 and 1; R1 and R3 are random vectors uniformly distributed in the [0, 1] space; R2 is a random number between [0, 1]. The random vector R1 that satisfies the condition (P == 0) returns the IDX index; TDR is the adaptive factor, linearly decreasing from 1 to 0, and the expression is as follows:

[0068] TDR = 1 - t×(1 / T max )

[0069] T max is the initial number of iterations. The mating behavior is generated by the mating of adolescent foals assigned to a temporary group and from different ethnic groups. Assume that the foals leaving group i and the foals leaving group j have joined a temporary group respectively. These two foals are male and female. Since these two foals have no family relationship, they will mate after entering adolescence. The expression is:

[0070]

[0071] Crossover = Mean

[0072] In the above formula, X p represents the individual position where an individual p in population k re - enters population k after leaving the group, and its parental positions come from different populations. The leader leads the population to a suitable habitat. If a population dominates a habitat, then this population must leave this place. The calculation formula for the next position of the leader relative to the habitat is:

[0073]

[0074] S33 recalculates the fitness value for the updated position and outputs the model parameters.

[0075] Specifically, recalculate the fitness value for the updated position, and finally output the parameter combination with the highest fitness, which is the global optimal solution WH and is used to configure the RF model; for the specific WHO optimization algorithm process, please refer to Figure 3 as shown.

[0076] S4 inputs the training set data into the WHO - RF model for training to construct an ionospheric irregularity intensity prediction model;

[0077] S41 divides the data set into N subsets, and each subset is used to train a regression tree model;

[0078] Specifically, during the training process, without shuffling the training set data, the data set is divided into N subsets, and each subset is used to train and construct a regression tree. At each node of the tree, m features are randomly selected from M features (m << M), usually According to the principle of minimizing the node impurity, one feature is selected from these m features for branching growth, and the impurity measure used is the Gini impurity, which is expressed as:

[0079]

[0080] P(w j ) is the frequency of the number of samples belonging to w j class at node n accounting for the total number of training samples. This classification tree grows fully to minimize the impurity of each node, and no usual pruning operation is performed. A forest is composed of N regression trees, and new data is predicted according to the generated decision trees, and the final result is obtained by arithmetic averaging according to the voting of each decision tree.

[0081] S42 calculates the feature importance through the OOB error rate and analyzes the contribution of each feature to the prediction;

[0082] Specifically, calculate the feature importance through the OOB error rate and analyze the contribution of each feature to the prediction. VIM j (OOB)Defined as: In each tree of RF, a training bootstrap sample is randomly selected to build a tree, and the out-of-bag (OOB) prediction error rate is calculated. Then, the observations of variable X are randomly permuted, and a tree is built again to calculate the OOB prediction error rate. Finally, the difference between the two OOB error rates is calculated, and the average value among all trees after standardization is the permutation importance VIM of variable X. j After the observations of variable X j are randomly permuted and a tree is built again to calculate the OOB prediction error rate, the permutation importance VIM of variable X j (OOB) is obtained. j The VIM of variable X j (OOB) in the i-th tree is:

[0083]

[0084] where is the number of observed cases in the OOB data of the i-th tree, I(·) is the indicator function, which takes 1 when the two values are equal and 0 when they are not; Y p ∈{0,1} is the true result of the p-th observation, is the predicted result of the i-th tree for the p-th observation in the OOB data before random permutation, is the predicted result of the i-th tree for the p-th observation in the OOB data after random permutation. When variable j does not appear in the i-th tree,

[0085] VIM j (OOB) = 0.

[0086] The permutation importance of variable X j in RF is defined as:

[0087]

[0088] where n is the number of decision trees in RF.

[0089] After each regression tree is independently generated, the predicted results of ROTI are output respectively, and the arithmetic mean of the regression results obtained from N decision trees is used to obtain the ionospheric irregularity intensity prediction model.

[0090] Specifically, after each regression tree is independently generated, the predicted results of ROTI are output respectively, and the arithmetic mean of the regression results obtained from N decision trees is used to obtain the final model prediction value:

[0091]

[0092] The root mean square error of prediction RMSE and the coefficient of determination R 2 are calculated to evaluate the generalization ability:

[0093]

[0094] where: m represents the length of the predicted data; Y pred represents the predicted value of the ionospheric disturbance index prediction model; Y true represents the true value of the ionospheric disturbance index; Cov represents the covariance operation; Var represents the variance operation. The root mean square error RMSE can reflect the deviation degree between the ionospheric disturbance index prediction and the measured value. The smaller the RMSE, the higher the prediction accuracy; the coefficient of determination R 2 can reflect the similarity between the ionospheric disturbance index prediction and the measured value. The closer R 2 is to 1, the more accurate the prediction.

[0095] For the specific training process of the WHO-RF model, please refer to Figure 4 as shown.

[0096] S5 Obtain the residual sequence and construct the LSTM residual compensation model;

[0097] S51 Calculate the difference between the preliminary ROTI prediction value output by the ionospheric irregularity prediction model and the true value of the training set to obtain the residual sequence;

[0098] Specifically, combine the true value data and the prediction value data in step 4.3 to obtain the residual sequence:

[0099]

[0100] S52 Use the residual sequence and related ionospheric environment parameters as the input of the LSTM model, and calculate the hidden state and output of the current time step for each time step of data;

[0101] Specifically, divide the residual sequence data into multiple sequence samples according to the time step length as the input of the LSTM model. During the training process, calculate for each time step of data: determine the memory information to be forgotten through the forget gate, determine the new information to be added through the input gate and the candidate memory unit value, update the memory unit state, and calculate the hidden state and output of the current time step through the output gate.

[0102] S53 After each time step is completed, use the BPTT algorithm to calculate the error gradient, update the weight parameters of the LSTM, and repeat the iterative training process until the model converges to obtain the LSTM residual compensation model.

[0103] Specifically, first initialize the network parameters, including the gating weights and biases; then perform forward propagation step by step in time, calculate the activation values of the forget gate, input gate, and output gate in turn, update the memory unit state and generate the current hidden state, and finally obtain the residual prediction output through linear transformation

[0104]

[0105] W y and b y are output layer parameters, and h t is the current hidden state. After completing all time steps, the mean square error between the predicted value and the true value is calculated as the loss function; then, the gradients are calculated layer by layer starting from the last time step through the BPTT algorithm, and the parameters are updated according to the gradient descent method. The gradient calculation of the gating weight W f is as follows:

[0106]

[0107] T is the total number of time steps, L is the loss function, and C t is the current memory cell state, and f i is the forget gate activation value. The training process is iteratively repeated until the model converges, and finally, the trained LSTM residual model is used for regression prediction. The entire process selectively retains and transmits information through the gating mechanism, effectively capturing long-term dependencies in the time series.

[0108] S6 combines the WHO-RF model and the LSTM residual compensation model to obtain a hybrid prediction framework, and inputs the test set data into this combined prediction model to obtain the final prediction result of the ionospheric irregularity intensity.

[0109] Specifically, the test set is input into the trained WHO-RF hybrid prediction model based on residual compensation for predicting the occurrence and occurrence intensity of ionospheric irregularities.

[0110] Beneficial effects:

[0111] Through deep learning technology, the associations between ionospheric irregularities and various background ionospheric variables are analyzed, and the ionospheric background information strongly related to ionospheric irregularities is screened out. In view of the problem that the ionospheric irregularity samples are few, high-dimensional feature vectors need to be used due to the complex background ionospheric environment, and there are missing values in the data. The RF is good at processing high-dimensional data and its training can be parallelized, with fast speed and still maintaining good accuracy even when there are missing data in the training set. An ionospheric irregularity prediction model is constructed using its excellent performance. At the same time, the wild horse optimization WHO algorithm is combined to optimize the hyperparameters of the RF model, which improves the generalization ability of the ionospheric irregularity prediction model to a certain extent. A residual compensation model is proposed, which combines the global feature extraction of WHO-RF and the temporal residual compensation of LSTM to further improve the accuracy and accuracy of the ionospheric irregularity prediction model.

[0112] The above-disclosed is only the preferred embodiment of the ionospheric irregularity prediction method based on residual compensation of the present invention. Of course, the scope of the rights of the present invention cannot be limited thereby. Those of ordinary skill in the art can understand all or part of the processes of implementing the above embodiments, and the equivalent changes made according to the claims of the present invention still fall within the scope covered by the invention.

Claims

1. An ionospheric irregularity prediction method based on residual compensation, characterized in that It includes the following steps: Obtain relevant ionospheric environment parameters, preprocess them, and construct a feature dataset; Divide the feature dataset into a training set and a test set, and use the moving average method to smooth the training set data; Use the WHO algorithm to automatically optimize the hyperparameters of the RF model and confirm the model parameters; Input the training set data into the WHO-RF model for training to construct an ionospheric irregularity intensity prediction model; Obtain the residual sequence and construct an LSTM residual compensation model; Combine the WHO-RF model and the LSTM residual compensation model to obtain a hybrid prediction framework, input the test set data into this combined prediction model, and obtain the final prediction result of the ionospheric irregularity intensity.

2. The ionospheric irregularity prediction method based on residual compensation according to claim 1, wherein In "Obtain relevant ionospheric environment parameters, preprocess them, and construct a feature dataset", the feature dataset includes the solar radio flux index F10.7, the geomagnetic index Dst, the interplanetary magnetic field components Bx / By / Bz, the magnetohydrodynamic pressure Flow Pressure, the critical frequency foF2 of the F2 layer, the symmetric horizontal component SYM / H of the ring current, and the ionospheric disturbance index ROTI sequence.

3. The ionospheric irregularity prediction method based on residual compensation according to claim 1, wherein In "Divide the feature dataset into a training set and a test set, and use the moving average method to smooth the training set data", the moving average method is used to smooth the data, and the width of the moving window used is 12 points.

4. The ionospheric irregularity prediction method based on residual compensation according to claim 1, wherein In "Use the WHO algorithm to automatically optimize the hyperparameters of the RF model and confirm the model parameters", it includes the following steps: Determine the RF model parameters to be optimized, construct the RF model, and use the root mean square error as the fitness function to calculate the parameter fitness value; Gradually optimize the parameters by dynamically updating the positions of the stallions and foals. For each group of foals, update their positions according to the positions of the stallions and random factors; Recalculate the fitness value for the updated positions and output the model parameters.

5. The ionospheric irregularity prediction method based on residual compensation according to claim 1, wherein In "Input the training set data into the WHO-RF model for training to construct an ionospheric irregularity intensity prediction model", it includes the following steps: Divide the dataset into N subsets, and each subset is used to train a regression tree model; Calculate the feature importance through the OOB error rate and analyze the contribution of each feature to the prediction; After each regression tree is independently generated, respectively output the prediction results of ROTI, and perform an arithmetic average on the regression results obtained from the N decision trees to obtain an ionospheric irregularity intensity prediction model.

6. The ionospheric irregularity prediction method based on residual compensation according to claim 1, characterized in that, In "Construct an LSTM residual compensation model", it includes the following steps: Calculate the difference between the initially output ROTI prediction value of the ionospheric irregularity prediction model and the true value of the training set to obtain a residual sequence; Use the residual sequence and relevant ionospheric environment parameters as the input of the LSTM model, and calculate the hidden state and output of the current time step for the data of each time step; After each time step is completed, use the BPTT algorithm to calculate the error gradient, update the weight parameters of the LSTM, and repeat the iterative training process until the model converges to obtain the LSTM residual compensation model.

Images