An ionosphere total electron content prediction method, model training method and system

Abstract

The application discloses an ionosphere total electron content prediction method, a model training method and a system, relates to the field of satellite geodesy and space weather prediction, and the model training method comprises: an ionosphere total electron content prediction model comprising an encoder part of a stacked autoencoder, an Informer network and a decoder part of the stacked autoencoder connected in sequence; wherein the model training comprises stacked autoencoder training and Informer network training. The ionosphere total electron content prediction model is constructed based on the Informer network and the stacked autoencoder, the ionosphere total electron content prediction model is used for ionosphere total electron content prediction, the time complexity is reduced, the memory occupation is reduced, and the ionosphere total electron content prediction efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of satellite geodesy and space weather forecasting, in particular to an ionospheric total electron content prediction method, a model training method and a system. BACKGROUND

[0002] The ionosphere of the earth refers to the part of the high atmosphere of the earth that is ionized, the lower part is called the ionosphere, and the higher part is called the magnetosphere. According to the definition of the Institute of Radio Engineers (IRE), the ionosphere is the entire space from 60 km above the ground to the top of the magnetosphere. The main reason for the ionization of the earth's atmosphere is the interaction between solar high-energy radiation and the upper particles of the earth. The upper atmospheric particles absorb solar radiation energy and ionize, producing a large number of free electrons, thereby producing the ionosphere. The total electron content (TEC) of the ionosphere, also known as the ionospheric electron concentration column content, is the integral of the electron concentration along the height per unit area, and is an important parameter for describing the shape and structure of the ionosphere.

[0003] For the modeling and prediction of ionospheric TEC, a geometric relationship between historical ionospheric TEC data is usually established using fitting techniques to describe the ionospheric data, and the subsequent ionospheric TEC values are extrapolated and predicted on this basis. At present, the global ionospheric total electron content is usually predicted by using the Transformer model, but the time complexity and memory usage of the Transformer model in the process of predicting the global ionospheric total electron content are related to the length of the input sequence, that is, when the length of the input sequence is longer, the time consumed and the memory occupied will be larger, which limits the scalability of the model when a long sequence is input. Moreover, the prediction efficiency is low. SUMMARY

[0004] The purpose of the present application is to provide an ionospheric total electron content prediction method, a model training method and a system.

[0005] To achieve the above-mentioned purpose, the present application provides the following solutions:

[0006] An ionospheric total electron content prediction model training method, comprising:

[0007] obtaining global ionospheric total electron content images of a plurality of historical sample time periods;

[0008] training the ionospheric total electron content prediction model by taking the global ionospheric total electron content images of the historical sample time periods as input and taking the global ionospheric total electron content images of a subsequent historical time period of the historical sample time periods as labels, to obtain a trained ionospheric total electron content prediction model;

[0009] The ionospheric total electron content prediction model comprises an encoder part of a stacked autoencoder, an Informer network and a decoder part of the stacked autoencoder connected in sequence.

[0010] The training comprises stacked autoencoder training and Informer network training.

[0011] The stacked autoencoder training comprises training the stacked autoencoder by taking global ionospheric total electron content images of a historical sample time period as input.

[0012] The Informer network training comprises training the ionospheric total electron content prediction model by taking global ionospheric total electron content images of the historical sample time period as input, taking global ionospheric total electron content images of a historical time period after the historical sample time period as labels, and keeping the weights in the trained stacked autoencoder unchanged, so as to determine the weights in the Informer network.

[0013] Optionally, the Informer network comprises an encoder, a decoder and a fully connected layer.

[0014] The encoder comprises a plurality of pyramid structures connected in sequence, each of the pyramid structures comprises a plurality of self-attention modules connected in sequence, and the self-attention module comprises a multi-head active sparse self-attention mechanism and a distillation operation connected in sequence.

[0015] The decoder is connected with the encoder and comprises a masked multi-head active sparse self-attention mechanism and a multi-head attention mechanism connected in sequence.

[0016] The fully connected layer is connected with the decoder.

[0017] Optionally, before the ionospheric total electron content prediction model is trained by taking global ionospheric total electron content images of a historical sample time period as input and taking global ionospheric total electron content images of a historical time period after the historical sample time period as labels, the method further comprises:

[0018] The global ionospheric total electron content images of the historical sample time period are preprocessed, and the preprocessing comprises linear stretching transformation processing.

[0019] The application further provides an ionospheric total electron content prediction method based on the ionospheric total electron content prediction model.

[0020] Global ionospheric total electron content images of a plurality of historical time periods are acquired.

[0021] inputting the global ionospheric total electron content image of the historical period into the trained ionospheric total electron content prediction model to obtain a global ionospheric total electron content image of a prediction period; the prediction period and the historical period are adjacent periods.

[0022] Optionally, before the global ionospheric total electron content image of the historical period is input into the trained ionospheric total electron content prediction model, the method further comprises:

[0023] preprocessing the global ionospheric total electron content image of the historical period, the preprocessing comprising linear stretching transformation processing.

[0024] The application also provides an ionospheric total electron content prediction model training system, the training system comprising:

[0025] a sample image acquisition module configured to acquire global ionospheric total electron content images of a plurality of historical sample periods;

[0026] a model training module configured to train the ionospheric total electron content prediction model by taking the global ionospheric total electron content images of the historical sample periods as input and taking global ionospheric total electron content images of a next historical period of the historical sample periods as labels, to obtain a trained ionospheric total electron content prediction model.

[0027] The ionospheric total electron content prediction model comprises an encoder part of a stacked autoencoder, an Informer network and a decoder part of the stacked autoencoder connected in sequence.

[0028] The training comprises stacked autoencoder training and Informer network training.

[0029] The stacked autoencoder training comprises training the stacked autoencoder by taking the global ionospheric total electron content images of the historical sample periods as input.

[0030] The Informer network training comprises training the ionospheric total electron content prediction model by keeping the weights in the trained stacked autoencoder unchanged, taking the global ionospheric total electron content images of the historical sample periods as input, taking the global ionospheric total electron content images of the next historical period of the historical sample periods as labels, and determining the weights in the Informer network.

[0031] Optionally, the training system further comprises a preprocessing module.

[0032] The preprocessing module is configured to preprocess the global ionospheric total electron content images of the historical sample periods, the preprocessing comprising linear stretching transformation processing.

[0033] The application further provides an ionospheric total electron content prediction system based on the ionospheric total electron content prediction model.

[0034] a historical image acquisition module, configured to acquire global ionospheric total electron content images of historical time periods;

[0035] a prediction module, configured to input the global ionospheric total electron content images of the historical time periods into the trained ionospheric total electron content prediction model to obtain global ionospheric total electron content images of a prediction time period; the prediction time period is adjacent to the historical time period.

[0036] Optionally, the prediction system further comprises a preprocessing module.

[0037] The preprocessing module is configured to pre-process the global ionospheric total electron content images of the historical time periods, and the pre-processing comprises linear stretch transformation processing.

[0038] According to the embodiments of the application, the following technical effects are achieved: the application provides an ionospheric total electron content prediction method and model training method and system, the method comprising: acquiring global ionospheric total electron content images of historical sample time periods; training an ionospheric total electron content prediction model by taking the global ionospheric total electron content images of the historical sample time periods as input and taking global ionospheric total electron content images of a next historical time period of the historical sample time periods as labels to obtain a trained ionospheric total electron content prediction model; the ionospheric total electron content prediction model comprises an encoder part of a stacked autoencoder, an Informer network and a decoder part of the stacked autoencoder connected in sequence; the training comprises stacked autoencoder training and Informer network training; the stacked autoencoder training comprises taking the global ionospheric total electron content images of the historical sample time periods as input to train the stacked autoencoder; the Informer network training comprises keeping the weights in the trained stacked autoencoder unchanged, taking the global ionospheric total electron content images of the historical sample time periods as input, taking the global ionospheric total electron content images of the next historical time period of the historical sample time periods as labels to train the ionospheric total electron content prediction model to determine the weights in the Informer network. The application constructs an ionospheric total electron content prediction model based on an Informer network and a stacked autoencoder, adopts the ionospheric total electron content prediction model to predict ionospheric total electron content, reduces time complexity, reduces memory occupation and improves ionospheric total electron content prediction efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0040] Figure 1 The ionospheric total electron content prediction model training method flowchart provided for the embodiment 1 of the present application;

[0041] Figure 2 The global ionospheric TEC map prediction model structure schematic diagram provided for the embodiment 1 of the present application;

[0042] Figure 3 The stack autoencoder structure schematic diagram provided for the embodiment 1 of the present application;

[0043] Figure 4 The Informer network structure schematic diagram provided for the embodiment 1 of the present application;

[0044] Figure 5 The global ionospheric TEC map prediction model workflow schematic diagram provided for the embodiment 1 of the present application;

[0045] Figure 6 The ionospheric total electron content prediction method flowchart provided for the embodiment 2 of the present application;

[0046] Figure 7 The root mean square error comparison situation schematic diagram provided for the embodiment 2 of the present application;

[0047] Figure 8 The mean absolute error comparison situation schematic diagram provided for the embodiment 2 of the present application;

[0048] Figure 9 The correlation coefficient comparison situation schematic diagram provided for the embodiment 2 of the present application;

[0049] Figure 10 The ionospheric total electron content prediction model training system block diagram provided for the embodiment 3 of the present application;

[0050] Figure 11 The ionospheric total electron content prediction system block diagram provided for the embodiment 4 of the present application. DETAILED DESCRIPTION

[0051] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of the present application.

[0052] The present application aims to provide an ionospheric total electron content prediction method, a model training method and a system, which constructs an ionospheric total electron content prediction model based on an Informer network (Informer time series prediction model) and a stacked autoencoder (Stacked-AutoEncoder), increases the capacity for long sequence prediction problems, and enables the model to capture long-range dependencies between long time series inputs and outputs; proposes a ProbSparse self-attention mechanism, which has less time complexity and memory usage than the existing standard self-attention mechanism; through a self-attention distilling operation, the input vectors that play a leading role in the stacked self-attention layer can be screened and input into the next layer, greatly reducing the overall spatial complexity; a generative style decoder (Generative Style Decoder) is proposed to obtain a long sequence output, which can obtain the entire decoding sequence at one time, and avoids the propagation of cumulative errors during inference.

[0053] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] Embodiment 1

[0055] As shown in the accompanying drawings, Figure 1 The present application provides an ionospheric total electron content prediction model training method, which comprises:

[0056] S101: Obtain global ionospheric total electron content images of a plurality of historical sample time periods. The global ionospheric total electron content historical images are obtained from the European Space Agency.

[0057] S102: Take the global ionospheric total electron content images of the historical sample time periods as input, and take the global ionospheric total electron content images of the next historical time period after the historical sample time period as label, train the ionospheric total electron content prediction model, and obtain the trained ionospheric total electron content prediction model.

[0058] As shown in the accompanying drawings, Figure 2As shown, the ionospheric total electron content prediction model includes an encoder part of a stacked autoencoder, an informer network and a decoder part of the stacked autoencoder connected in sequence.

[0059] The global ionospheric TEC image prediction model (i.e., the ionospheric total electron content prediction model) proposed in this invention comprises four parts:

[0060] 1. Encoding part: The encoder part of the stacked autoencoder is used to compress and extract features from the global ionospheric TEC image to obtain the feature vector of the global ionospheric TEC image.

[0061] 2. Prediction: Use the Informer network to predict the feature vector of the global ionospheric TEC map.

[0062] 3. Decoding part: The decoder part of the stacked autoencoder is used to reconstruct the global ionospheric TEC image.

[0063] like Figure 5 As shown in the figure, the workflow of the global ionospheric TEC map prediction model based on stacked autoencoders and informer networks is as follows:

[0064] The training includes stacked autoencoder training and Informer network training.

[0065] The stacked autoencoder is trained by taking the global ionospheric total electron content image of the historical sample time period as input to train the stacked autoencoder.

[0066] The informer network training is as follows: keeping the weights in the trained stacked autoencoder unchanged, taking the global ionospheric total electron content image of the historical sample time period as input, and taking the global ionospheric total electron content image of the historical time period after the historical sample time period as a label, the ionospheric total electron content prediction model is trained to determine the weights in the informer network.

[0067] The following is a detailed introduction to stacked autoencoder training and Informer network training:

[0068] Autoencoder is an unsupervised data dimension compression and data feature expression method that can be used for image feature extraction. Autoencoder is a type of neural network. After training, the autoencoder can reduce the dimension of the input data and extract the features of the input data, as well as reconstruct the input data based on the obtained reduced-dimensional features. The autoencoder is mainly divided into two parts, namely the encoder (Encoder) and the decoder (Decoder). Among them, the encoder is responsible for compressing and extracting the feature vector of the input data, and the decoder is responsible for reconstructing the compressed feature vector into the original input data. The stacked autoencoder is an improved network developed on the basis of the autoencoder. Unlike the autoencoder, the encoder and decoder have only one hidden layer. The encoder and decoder of the stacked autoencoder can contain multiple hidden layers. Compared with the autoencoder with a single hidden layer, the stacked autoencoder can better compress and extract the feature vector of the input data, and reconstruct the input data with higher accuracy. The structure of the stacked autoencoder is as follows Figure 3 shown.

[0069] By using the stacked autoencoder to compress the image, the compressed feature vector of the image can be obtained. The steps are as follows:

[0070] 1. In this embodiment, before the global ionospheric total electron content image of the historical sample time period is used as input to train the ionospheric total electron content prediction model, the following steps are also included:

[0071] The global ionospheric total electron content image of the historical sample time period is preprocessed, and the preprocessing includes linear stretch transformation processing.

[0072] The TEC numerical matrix corresponding to the input global ionospheric TEC map is linearly stretched and transformed into a one-dimensional global ionospheric TEC vector. The global ionospheric total electron content image of the historical sample period is the global ionospheric total electron content image released by the European Orbit Determination Center (CODE). The product covers the range of 180° east longitude to 180° west longitude and 87.5° north latitude to 87.5° south latitude, with an accuracy resolution of 5°, a latitude resolution of 2.5°, and a time resolution of 1h. Assume that the original global ionospheric TEC numerical matrix is It is a matrix with 71 rows and 73 columns. The global ionospheric TEC vector after linear stretching is written as n represents the length of the time series. The numerical matrix corresponding to the global ionospheric TEC map is an image describing the distribution and quantity of the total electron content in the ionosphere at different locations around the world. The numerical value is the total electron content in the ionosphere, and the unit is TECU. 1 TECU means that there are 10 12 The elements of each row and column of the numerical matrix correspond to the total electron content of the ionosphere at the latitude and longitude of the coverage area.

[0073] It should be noted that the embodiment obtains global ionospheric total electron content images for 5 days as input data and predicts data for the day after the 5 days as output data for training. For example, global ionospheric total electron content images for days 1-5 of month X are obtained as input data, and a global ionospheric total electron content image for day 6 of month X is obtained as output data.

[0074] 2. The n-length global ionospheric TEC vector is input into the encoder part of the trained stack autoencoder as training data, for compression and extraction of the global ionospheric TEC feature vector, and the obtained global ionospheric TEC feature vector is which is a one-dimensional vector of length 324. The matrix composed of n-length global ionospheric TEC vectors is input into the Informer network as training data, denoted as which is denoted as X tec_en , and the calculation process is as follows:

[0075] The n-length global ionospheric TEC vector used for training is input into the stack autoencoder, the output reconstruction result is compared with the input data and the reconstruction error is calculated, and the network parameters are modified in the reverse direction to achieve the purpose of training the network, as follows:

[0076] 1) The n-length global ionospheric TEC vector is input into the stack autoencoder, that is, into a multi-layer feedforward neural network. In one of the hidden layers of the stack autoencoder, the following three components are included: a batch normalization layer, an activation function layer, and a linear transformation layer.

[0077] Batch normalization layer: Batch Normalization (BN) refers to the standardization of the input of each layer of the neural network, and its purpose is to solve the problem of distribution drift of input data during training, and to prevent the model from overfitting during training. Assuming that the neural network has m neurons in the lth layer and n neurons in the (l-1)th layer, the batch normalization layer input is and the output is The calculation formula of batch normalization is written as:

[0078]

[0079]

[0080] In the formula, γ and β are learning parameters of the batch normalization layer, ∈ is a small value, and its role is to prevent division by 0 when σ 2 is too small to be considered as 0; μ and σ 2 are the mean and variance of the input data, and the calculation formula is as follows:

[0081]

[0082]

[0083] Activation function layer. The Tanh function is selected as the activation function, which is a hyperbolic tangent function, and the output is always in the open interval (-1, 1), and the Input the Tanh activation function layer. The Tanh function of the input value is as follows:

[0084]

[0085] Linear transformation layer. The weight between the lth layer and the (l-1)th hidden layer is denoted as The bias value of the current hidden layer neuron is denoted as The output of this layer of neurons is denoted as Then:

[0086]

[0087] In the above formula The output result of the linear transformation layer is the result input to the next hidden layer.

[0088] 2) Select the mean squared error (MSE) as the index to measure the reconstruction accuracy, and take minimizing the error as the goal. Assuming that the input image one-dimensional vector is x i , and the output reconstructed image one-dimensional vector is The mean squared error of the input and output vectors is calculated, and the calculation formula is:

[0089]

[0090] Then the objective function is written as:

[0091]

[0092] In the formula, n = 1, 2,..., is the total number of pixels of the one-dimensional image vector used for training, i represents the i-th element in the one-dimensional image vector, and are the network weights and bias values, respectively, and l and j, k represent different hidden layers and neurons in the hidden layer of the network.

[0093] 3) Error back propagation. After obtaining the root mean square error of the training set data, the neural network parameters are adjusted by using the error back propagation algorithm (BackPropagation Algorithm, BP). BP algorithm is composed of two processes of forward propagation of signal and backward propagation of error. The weights and bias values in the network are updated according to the error of backward propagation, which is the commonly used neural network learning strategy. In the BP algorithm, the update formula of the weights and bias values of the hidden layer is as follows:

[0094]

[0095]

[0096] In the formula, a is the network learning rate. The parameter update is essentially based on the chain rule and the derivative of the composite function. The derivative of the objective function with respect to the parameters of a certain neuron in a certain hidden layer is obtained. In the process of multiple learning, the network parameters will gradually update in the direction of error reduction.

[0097] After the stack autoencoder is trained, the output of the middle layer of the stack autoencoder is the compressed image feature vector.

[0098] 3. The input data (global ionospheric total electron content image of historical sample time period and its corresponding label) is divided into training set, test set and validation set according to the proportion of 7:2:1, the Informer network is trained, and the random search algorithm is used to optimize the model hyperparameters to find the optimal hyperparameter combination. In the model prediction part, the prediction length is 24, and the predicted future global ionospheric TEC feature matrix is denoted as X tec . The known ionospheric TEC feature vector sequence length of the generative decoder input is 48, and the matrix is denoted as X token . The specific calculation process is as follows:

[0099] As shown in Figure 2 and Figure 4 , the above Informer network includes an encoder, a decoder and a fully connected layer.

[0100] The encoder includes a plurality of sequentially connected pyramid structures, each of the pyramid structures includes a plurality of sequentially connected self-attention modules, and the self-attention module includes a multi-head active sparse self-attention mechanism and a distillation operation connected in sequence.

[0101] The decoder is connected with the encoder and includes a masked multi-head active sparse self-attention mechanism and a multi-head attention mechanism connected in sequence.​​

[0102] The fully connected layer is connected with the decoder.

[0103] The Informer network mainly consists of an encoder with a plurality of multi-head ProbSparse self-attention mechanisms and a distillation operation, a decoder with a plurality of masked multi-head ProbSparse self-attention mechanisms and multi-head attention mechanisms, and a fully connected layer.

[0104] The encoder of the Informer network utilizes the ProbSparse self-attention mechanism layer to perform feature mapping and feature extraction from historical data and calculate attention values, and the calculation formula is as follows:

[0105]

[0106] In the formula, is a sparse matrix with the same size as the Q matrix, and its content only contains the sparsity measure value Q and K are the query vector matrix and the key vector matrix generated by the encoder during the training process, respectively; V represents the value vector matrix; d is the dimension of the input data; the sparsity measure calculation formula of the ProbSparse self-attention mechanism is as follows:

[0107]

[0108] In the above formula, L k The calculation formula of L

[0109]

[0110] The sparsity of the i-th query calculation of the ProbSparse self-attention mechanism is defined as q i , q i and k j are from the query vector matrix Q and the label vector matrix K. The ProbSparse self-attention mechanism selects the result with a higher score in the above sparsity measure calculation, and calculates the attention value for it. For the result with a lower score, the attention value is not calculated.

[0111] The encoder of the Informer network performs distillation operation after ProbSparse self-attention calculation. The distillation operation is designed between the current ProbSparse self-attention layer and the next ProbSparse self-attention layer, and can select the vector that plays a leading role in the operation according to the sparsity calculated by the current ProbSparse self-attention layer, and input the next ProbSparse self-attention layer after one-dimensional convolution, activation function processing and maximum pooling (the middle gray trapezoidal part). Figure 4 Assuming that the j-th ProbSparse self-attention layer is pushed to the j+1-th layer, the formula of the distillation operation between the two is:

[0112]

[0113] In the above formula, [·] AB represents the calculation of the previous ProbSparse self-attention layer before the j-th time, ELU(·) is an activation function, and Conv1d(·) represents one-dimensional convolution calculation. In the above formula, the calculation result of Conv1d(·) is filtered by ELU(·) to select the feature vector that plays a leading role as the input of the j+1-th ProbSparse self-attention layer. The encoder of the Informer network is composed of multiple ProbSparse self-attention layers and distillation operations, and the two are sequentially cycled to form the calculation process of the entire encoder (such as Figure 4 the left half).

[0114] The decoder of the Informer expands the starting token method into a generative structure to replace the decoder structure of the Transformer. The decoder of the Informer is composed of two completely identical multi-head attention mechanism layers, and adds generative inference to alleviate the decrease in prediction speed in the long time sequence prediction process. The Informer network adds a long sequence L token to the input sequence of the decoder, which is composed of a known data sequence of a certain length before the target sequence to be predicted, and the target sequence is denoted as L y . The input data of the decoder is represented as follows:

[0115]

[0116] In the above formula, is a known data sequence of a certain length before the prediction sequence, is a placeholder for the target sequence. As in the above formula, the input of the decoder is a long sequence composed of the known data sequence and the target sequence. To avoid the situation that future unknown data participates in the calculation in the prediction process, the decoder structure adopts a mask multi-head self-attention mechanism layer, which avoids the automatic regression of the multi-head self-attention mechanism layer in the calculation process by setting the mask dot product to -∞. After the calculation of the mask multi-head self-attention mechanism layer, the decoder outputs the prediction result through a forward process to obtain the prediction result of a certain length of time period in the future (for example Figure 4 The right half part.

[0117] The random search algorithm (Random Search) is selected to search for the optimal hyperparameter combination of the Informer network. The random search algorithm randomly samples parameter values within a given hyperparameter range, and then selects the optimal hyperparameter combination from multiple sampling values. The random search algorithm optimization can improve the efficiency of hyperparameter optimization in a high-dimensional hyperparameter space and speed up the speed of finding the optimal hyperparameter combination. In the hyperparameter optimization process of the Informer network, a total of 10 hyperparameters participate in the optimization process, and therefore the random search algorithm is selected for hyperparameter optimization.

[0118] 4. The future global ionospheric TEC feature matrix output by the Informer network is input into the decoder part of the stack autoencoder to reconstruct the global ionospheric TEC image, and the final prediction result is obtained.

[0119] After the above process, the trained ionospheric total electron content prediction model can be obtained.

[0120] Embodiment 2

[0121] As Figure 6 shown, the application also provides an ionospheric total electron content prediction method based on the above ionospheric total electron content prediction model, and the prediction method comprises the following steps:

[0122] S201: Obtain global ionospheric total electron content images of a plurality of historical time periods.

[0123] S202: Input the global ionospheric total electron content images of the historical time periods into the trained ionospheric total electron content prediction model to obtain global ionospheric total electron content images of a prediction time period; the prediction time period and the historical time period are adjacent time periods.

[0124] Specifically, the global ionospheric total electron content images of Y21-25 are obtained, and the global ionospheric total electron content images of Y21-25 are input into the trained ionospheric total electron content prediction model, so that the global ionospheric total electron content images of Y26 can be predicted.

[0125] In the embodiment, before the global ionospheric TEC image of the historical period is input into the trained ionospheric TEC prediction model, the following steps are further included:

[0126] The global ionospheric TEC image of the historical period is preprocessed, and the preprocessing includes linear stretching transformation processing.

[0127] The above method is used for experiments in the embodiment. Global ionospheric TEC data provided by the Center for Orbit Determination in Europe (CODE) is used in the experiments. Data from 2015 to 2017 is selected as a training set, and global ionospheric TEC data in 2018 is predicted. The embodiment only shows the prediction results from 0 o'clock to 8 o'clock on January 1, 2018.

[0128] Figure 7 The root mean square error (RMSE) of the prediction results of the ionospheric TEC prediction model and the prediction products provided by the CODE center from 0 o'clock to 8 o'clock on January 1, 2018 is shown. It can be seen that the model prediction result is more accurate than the CODE product.

[0129] Figure 8 The mean absolute error (MAE) of the prediction results of the ionospheric TEC prediction model and the CODE product is shown. It can be seen that the prediction model result is more accurate than the CODE product.

[0130] Figure 9 The correlation coefficient (CC) of the prediction results of the ionospheric TEC prediction model and the CODE product and the measured global ionospheric TEC data provided by the CODE center is shown. The higher the correlation coefficient value, the stronger the linear relationship between the measured data, and the better the result accuracy. It can be seen that the correlation between the model prediction result and the measured data is stronger than the CODE product.

[0131] The specific values of the above three accuracy evaluation indexes are shown in Table 1:

[0132] Table 1 shows the specific values of the three accuracy indexes

[0133]

[0134]

[0135] The present application firstly uses a stack autoencoder to compress a global ionospheric TEC image and extract a feature vector, then inputs the feature vector into an Informer network for prediction, solves the problems of high computational time complexity and large memory occupancy through a ProSparse self-attention mechanism and a self-attention mechanism distillation operation in the Informer network, solves the problem of low efficiency of output predicted values through an Informer generative decoder, and provides a possibility for ionospheric TEC long-time series prediction. The whole prediction process is more efficient while ensuring accuracy, and finally the output predicted feature vector is reconstructed into a global ionospheric TEC image by the decoder part of the stack autoencoder.

[0136] Embodiment 3

[0137] As Figure 10 shown, the present application also provides an ionospheric total electron content prediction model training system, which comprises:

[0138] A sample image acquisition module T1 is configured to acquire global ionospheric total electron content images of a plurality of historical sample time periods.

[0139] A model training module T2 is configured to train an ionospheric total electron content prediction model by taking the global ionospheric total electron content images of the historical sample time periods as input and taking the global ionospheric total electron content images of a subsequent historical time period of the historical sample time periods as labels, and obtain a trained ionospheric total electron content prediction model.

[0140] The ionospheric total electron content prediction model comprises an encoder part of a stack autoencoder, an Informer network and a decoder part of the stack autoencoder connected in sequence.

[0141] The training comprises stack autoencoder training and Informer network training.

[0142] The stack autoencoder training is configured to train the stack autoencoder by taking the global ionospheric total electron content images of the historical sample time periods as input and taking the global ionospheric total electron content images of a subsequent historical time period of the historical sample time periods as labels.

[0143] The Informer network training is configured to train the ionospheric total electron content prediction model by keeping the weights in the trained stack autoencoder unchanged, taking the global ionospheric total electron content images of the historical sample time periods as input and taking the global ionospheric total electron content images of a subsequent historical time period of the historical sample time periods as labels, and determine the weights in the Informer network.

[0144] In the present embodiment, the training system further comprises a preprocessing module;

[0145] The preprocessing module is configured to preprocess the global ionospheric total electron content images of the historical sample time period, and the preprocessing comprises linear stretch transformation processing.

[0146] Embodiment 4

[0147] As shown in Figure 11 The present application also provides an ionospheric total electron content prediction system based on the ionospheric total electron content prediction model, and the prediction system comprises:

[0148] A historical image acquisition module M1 is configured to acquire global ionospheric total electron content images of a plurality of historical time periods.

[0149] A prediction module M2 is configured to input the global ionospheric total electron content images of the historical time periods into the trained ionospheric total electron content prediction model to obtain global ionospheric total electron content images of a prediction time period; the prediction time period and the historical time period are adjacent time periods.

[0150] In the embodiment, the prediction system further comprises a preprocessing module.

[0151] The preprocessing module is configured to preprocess the global ionospheric total electron content images of the historical time periods, and the preprocessing comprises linear stretch transformation processing.

[0152] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments, and the same or similar parts of each embodiment can be referred to each other. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0153] The principles and implementation manners of the present application are described by using specific examples in the present application, and the above embodiment description is only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation manner and application range will be changed. In conclusion, the content of the specification should not be understood as the limitation of the present application.

Claims

1. A method for training an ionospheric total electron content prediction model, characterized in that: include: Obtain global ionospheric total electron content images for several historical sample time periods; The global ionospheric total electron content image of the historical sample period is used as input, and the global ionospheric total electron content image of the historical period after the historical sample period is used as a label to train the ionospheric total electron content prediction model to obtain a trained ionospheric total electron content prediction model; The ionospheric total electron content prediction model includes an encoder part of a stacked autoencoder, an informer network, and a decoder part of the stacked autoencoder connected in sequence; The training includes stacked autoencoder training and Informer network training; The stacked autoencoder is trained by taking the global ionospheric total electron content image of the historical sample time period as input to train the stacked autoencoder; The informer network training is as follows: keeping the weights in the trained stacked autoencoder unchanged, taking the global ionospheric total electron content image of the historical sample time period as input, and taking the global ionospheric total electron content image of the historical time period after the historical sample time period as a label, the ionospheric total electron content prediction model is trained to determine the weights in the informer network.

2. The ionospheric total electron content prediction model training method according to claim 1, wherein: The Informer network includes an encoder, a decoder and a fully connected layer; The encoder includes a plurality of sequentially connected pyramid structures, each of the pyramid structures includes a plurality of sequentially connected self-attention modules, and the self-attention modules include a sequentially connected multi-head active sparse self-attention mechanism and a distillation operation; The decoder is connected to the encoder and includes a masked multi-head active sparse self-attention mechanism and a multi-head attention mechanism connected in sequence; The fully connected layer is connected to the decoder.

3. The ionospheric total electron content prediction model training method according to claim 1, wherein: Before training the ionospheric total electron content prediction model using the global ionospheric total electron content image of the historical sample time period as input and the global ionospheric total electron content image of the historical time period after the historical sample time period as a label, the following is also included: The global ionospheric total electron content image of the historical sample time period is preprocessed, and the preprocessing includes linear stretch transformation processing.

4. A method for predicting the total electron content of the ionosphere based on the ionosphere total electron content prediction model training method according to claim 1 or 2, characterized in that: The forecasting method comprises: Obtain images of the global ionospheric total electron content for several historical periods; The global ionospheric total electron content image of the historical period is input into a trained ionospheric total electron content prediction model to obtain the global ionospheric total electron content image of the prediction period; the prediction period and the historical period are adjacent periods.

5. The method for predicting the total electron content of the ionosphere according to claim 4, wherein: Before inputting the global ionospheric total electron content image of the historical period into the trained ionospheric total electron content prediction model, the method further includes: The global ionospheric total electron content image of the historical period is preprocessed, and the preprocessing includes linear stretching transformation processing.

6. A system for training a prediction model of total ionospheric electron content, characterized in that: The training system comprises: A sample image acquisition module is used to obtain global ionospheric total electron content images for several historical sample time periods; The model training module is used to train the ionospheric total electron content prediction model using the global ionospheric total electron content image of the historical sample time period as input and the global ionospheric total electron content image of the historical time period after the historical sample time period as a label, thereby obtaining a trained ionospheric total electron content prediction model; The ionospheric total electron content prediction model includes an encoder part of a stacked autoencoder, an informer network, and a decoder part of the stacked autoencoder connected in sequence; The training includes stacked autoencoder training and Informer network training; The stacked autoencoder is trained by taking the global ionospheric total electron content image of the historical sample time period as input to train the stacked autoencoder; The informer network training is as follows: keeping the weights in the trained stacked autoencoder unchanged, taking the global ionospheric total electron content image of the historical sample time period as input, and taking the global ionospheric total electron content image of the historical time period after the historical sample time period as a label, the ionospheric total electron content prediction model is trained to determine the weights in the informer network.

7. The ionospheric total electron content prediction model training system according to claim 6, characterized in that: The Informer network includes an encoder, a decoder and a fully connected layer; The encoder includes a plurality of sequentially connected pyramid structures, each of the pyramid structures includes a plurality of sequentially connected self-attention modules, and the self-attention modules include a sequentially connected multi-head active sparse self-attention mechanism and a distillation operation; The decoder is connected to the encoder and includes a masked multi-head active sparse self-attention mechanism and a multi-head attention mechanism connected in sequence; The fully connected layer is connected to the decoder.

8. The ionospheric total electron content prediction model training system according to claim 6, characterized in that: The system also includes a pre-processing module; The preprocessing module is used to preprocess the global ionospheric total electron content image of the historical sample time period, and the preprocessing includes linear stretching transformation processing.

9. An ionospheric total electron content prediction system based on the ionospheric total electron content prediction model training method according to claim 1 or 2, characterized in that: The forecasting system comprises: A historical image acquisition module is used to obtain global ionospheric total electron content images for several historical periods; The prediction module is used to input the global ionospheric total electron content image of the historical period into the trained ionospheric total electron content prediction model to obtain the global ionospheric total electron content image of the prediction period; the prediction period and the historical period are adjacent periods.

10. The ionospheric total electron content prediction system according to claim 9, characterized in that: The forecast system also includes a preprocessing module; The preprocessing module is used to preprocess the global ionospheric total electron content image of the historical period, and the preprocessing includes linear stretching transformation processing.

Images