Adaptive low earth orbit satellite clock error real-time prediction system based on multi-method fusion

Abstract

The application provides a kind of based on the adaptive low-orbit satellite clock error real-time prediction system of multi-method fusion, comprising: clock error prediction model subsystem, for providing the first prediction method based on polynomial periodic term model, the second prediction method based on pre-training plus fine-tuning Autoformer model and the third prediction method based on wavelet analysis LSTM model and the corresponding available information to adaptive selection subsystem;Adaptive selection subsystem is used to select the prediction method that meets the calculation resource condition, data available condition and prediction accuracy requirement from the first prediction method to the third prediction method according to available information and then feedback to clock error prediction model subsystem, to complete clock error prediction by it.The application selects the prediction method adaptively for different application scenarios, combined with user demand and resource allocation and other factors, can improve the robustness of clock error prediction system, and provides high-precision, high-reliable low-orbit satellite clock error prediction product for users.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of satellite navigation technology, and particularly relates to an adaptive low-orbit satellite clock error real-time prediction system based on multi-method fusion. BACKGROUND

[0002] In recent years, with the rapid development of the low-orbit satellite field, the number of low-orbit satellites is increasing, and the use of low-orbit satellites to enhance the GNSS system has also attracted attention from all walks of life. In the low-orbit navigation enhancement system, based on the real-time clock error and the orbit product of the low-orbit satellite, the ground user can obtain enhanced PNT (navigation, positioning and timing) service. However, in the real-time determination of the low-orbit satellite clock error and the orbit product, whether it is data transmission or product solving process, a certain time will be consumed, resulting in a delay of several seconds to several hours of the low-orbit satellite clock error and orbit product. Therefore, the near real-time / after-determined low-orbit satellite clock error and orbit product need to be predicted to meet the real-time application requirements. However, compared with the GNSS satellite clock error, the low-orbit satellite clock error prediction faces greater challenges due to the influence of more systematic factors, such as the relativistic effect, the influence of temperature on the hardware delay, etc.

[0003] At present, the low-orbit satellite clock error prediction algorithm is basically based on the polynomial combined periodic term model, but this model is difficult to accurately describe the relatively complex systematic phenomena contained in the low-orbit satellite clock error, especially the high-frequency periodic term, resulting in limited low-orbit satellite clock error prediction accuracy. SUMMARY

[0004] In order to solve the above problems existing in the prior art, the application provides an adaptive low-orbit satellite clock error real-time prediction system based on multi-method fusion. The technical problems to be solved by the application are realized through the following technical scheme:

[0005] An adaptive low-orbit satellite clock error real-time prediction system based on multi-method fusion comprises a clock error prediction model subsystem and an adaptive selection subsystem.

[0006] The clock error prediction model subsystem is configured to provide a first prediction method based on a polynomial periodic term model, a second prediction method based on a pre-trained and fine-tuned Autoformer model, and a third prediction method based on a wavelet analysis LSTM model, and corresponding available information to the adaptive selection subsystem; wherein the available information includes required computing resources and historical clock error data for input.

[0007] The adaptive selection subsystem is configured to select a prediction method that meets the computing resource condition, the data availability condition and the prediction accuracy from the first prediction method to the third prediction method according to the available information to obtain a selection result, and feed back the selection result to the clock error prediction model subsystem.

[0008] The clock difference prediction model subsystem is configured to utilize the corresponding prediction method of the selection result to perform clock difference prediction.

[0009] Beneficial effects:

[0010] The application provides a self-adaptive low-orbit satellite clock difference real-time prediction system based on multi-method fusion, which comprises a clock difference prediction model subsystem, a self-adaptive selection subsystem and a clock difference prediction model subsystem. The clock difference prediction model subsystem is configured to provide a first prediction method based on a polynomial periodic term model, a second prediction method based on a pre-training and fine-tuning Autoformer model and a third prediction method based on a wavelet analysis LSTM model and corresponding available information to the self-adaptive selection subsystem. The self-adaptive selection subsystem is configured to select a prediction method meeting the calculation resource condition, the data availability condition and the prediction accuracy from the first prediction method to the third prediction method according to the available information to obtain a selection result, and feed back the selection result to the clock difference prediction model subsystem. The clock difference prediction model subsystem is configured to utilize the corresponding prediction method of the selection result to perform clock difference prediction. The application solves the problem of poor long-term prediction accuracy of low-orbit satellites, and can adaptively select a prediction method to perform clock difference prediction according to different application scenarios, user needs and resource allocation and other factors, thereby improving the robustness of the low-orbit satellite clock difference prediction system and providing high-precision and high-reliability low-orbit satellite predicted clock difference products for users.

[0011] The application will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 is a schematic diagram of a self-adaptive low-orbit satellite clock difference real-time prediction system based on multi-method fusion provided by the application;

[0013] Figure 2 is a schematic diagram of the implementation process of a self-adaptive low-orbit satellite clock difference real-time prediction system based on multi-method fusion provided by the application;

[0014] Figure 3 is a schematic diagram of the pre-training and model fine-tuning process in the second prediction method provided by the application;

[0015] Figure 4 is a schematic diagram of the prediction process of the third prediction method provided by the application. DETAILED DESCRIPTION

[0016] The application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0017] The application is aimed at different application scenarios, and under the premise of comprehensively evaluating user demand and existing resources, a long short-term memory (LSTM) low-orbit clock difference prediction method based on wavelet analysis and a low-orbit clock difference prediction method based on pre-training fine-tuning Autoformer are proposed, and an adaptive low-orbit satellite clock difference real-time prediction system based on multi-method fusion is designed, so as to obtain high-precision low-orbit satellite real-time clock difference by adaptively selecting the best clock difference prediction method and improve the robustness of the clock difference prediction system.

[0018] Firstly, the system proposes a LSTM low-orbit clock difference prediction method based on wavelet analysis and an Autoformer low-orbit clock difference prediction method based on pre-training fine-tuning, to solve the problem of poor long-term prediction accuracy of low-orbit satellites; secondly, an adaptive clock difference prediction system is constructed for different application scenarios, and according to user demand, resource configuration and other factors, a polynomial periodic term model method, a low-orbit clock difference prediction method based on wavelet analysis LSTM model and a clock difference prediction method based on pre-training fine-tuning Autoformer model are adaptively selected to improve the robustness of the low-orbit satellite clock difference prediction system and provide high-precision and high-reliable low-orbit satellite predicted clock difference products for users.

[0019] As shown in Figure 1 The application provides an adaptive low-orbit satellite clock difference real-time prediction system based on multi-method fusion, which comprises a clock difference prediction model subsystem and an adaptive selection subsystem.

[0020] The clock difference prediction model subsystem is used to provide a first prediction method based on a polynomial periodic term model, a second prediction method based on a pre-training fine-tuning Autoformer model and a third prediction method based on a wavelet analysis LSTM model and corresponding available information to the adaptive selection subsystem; wherein the available information comprises required computing resources and historical clock difference data for input.

[0021] The adaptive selection subsystem is used to select a prediction method meeting the computing resource condition, data availability condition and prediction accuracy from the first prediction method to the third prediction method according to the available information, and feed back the selection result to the clock difference prediction model subsystem.

[0022] The clock difference prediction model subsystem is used to perform clock difference prediction by using the prediction method corresponding to the selection result.

[0023] It is worth noting that the clock error prediction model subsystem of the present application includes three clock error prediction methods, which are a prediction method based on a polynomial period term model, a prediction method based on a pre-training plus fine-tuning Autoformer model and a prediction method based on a wavelet analysis LSTM model. In order to facilitate analysis and description, the three prediction methods are marked as the first prediction method, the second prediction method and the third prediction method.

[0024] The first prediction method has the advantages of fast operation speed and less resource occupation, and the disadvantage of poor medium and long-term prediction results.

[0025] The second prediction method involves the pre-training and fine-tuning of the Autoformer model. The pre-training part helps the model master the general pattern and periodic structure of the clock error change through the autocorrelation mechanism and series decomposition structure of the model. Model fine-tuning can learn from the latest clock error sequence. In the face of the difference of low-orbit satellite clock error cross-day data, through incremental training and parameter fine-tuning, it can quickly adapt to sudden changes, dynamically update the model, improve the clock error prediction accuracy, and enhance the generality and robustness of the model. The advantage of this model is that it can capture the long-term change rule in the clock error sequence, which is beneficial to improve the medium and long-term clock error prediction accuracy.

[0026] The third prediction method uses wavelet transform to decompose different frequency components in the low-orbit satellite clock error, and trains the decomposed sequences through the long-short period memory network model, which can analyze the multi-scale information of the sequence and improve the prediction accuracy of the model. Its advantage is that it combines deep learning network and wavelet analysis, and shows higher accuracy in medium and long-term clock error prediction. The disadvantage is that it relies on more historical clock error data and requires more resources.

[0027] In combination with Figure 1 and Figure 2 , the adaptive selection subsystem of the present application includes an availability judgment module and a prediction accuracy comparison module.

[0028] The availability judgment module is used to select a preliminary prediction method that meets the calculation resource condition and the data availability condition from the first prediction method based on the polynomial period term model, the second prediction method based on the pre-training plus fine-tuning Autoformer model and the third prediction method based on the wavelet analysis LSTM model according to the available information.

[0029] The clock difference prediction system of the application faces various application scenarios and has wide applicability. The available clock difference prediction model can be selected according to different hardware platforms and resource configurations. The three prediction methods used by the system have different requirements for the use of computing resources. The deep learning model training requires a large amount of computing resources, while the polynomial periodic term model has a significantly lower resource requirement. Therefore, the availability information of the method is determined according to the computing resources. The specific implementation of this step is to determine whether the computing resources meet the requirements for each system when the module is started. The prediction methods that do not meet the requirements are set to an unavailable state. For the prediction methods that meet the resource requirements, the availability is checked and marked as available, and the system waits for the next step.

[0030] The clock difference prediction needs existing clock difference data as the data input of the model. The length and quality of the existing clock difference data in different scenarios are different, which will affect the use of the clock difference prediction model. For the deep learning model, including the LSTM model based on wavelet analysis and the Autoformer model with pre-training and fine-tuning, a large number of existing clock difference sequences are needed as training data when predicting the clock difference. If there is a lack of clock difference data, the model cannot be trained, so the clock difference data of the satellite needs to be judged. The specific implementation of this step is that the module judges the data availability for different prediction models: for the LSTM model and the Autoformer model, the most basic historical clock difference data is used as the model training input, and the model is marked as available. If the data is insufficient, the LSTM and Autoformer models are not available, and the model is marked as unavailable. For the polynomial periodic term model, the data requirement is the lowest, but the data availability still needs to be judged.

[0031] The prediction accuracy comparison module is configured to select the preliminary prediction method with the highest prediction accuracy from all the preliminary prediction methods that meet the data availability conditions as the available prediction method to obtain a selection result, and feed back the selection result to the clock difference prediction model subsystem.

[0032] The function of the module is to determine priority information according to the empirical accuracy performance of different clock difference prediction lengths of each prediction method. The three clock difference prediction methods contained in the clock difference prediction method subsystem perform differently at different clock difference prediction lengths. Generally, the polynomial period term model performs better in short-term prediction, the Autoformer model and the LSTM model are better in medium and long-term prediction accuracy, but the performances of the two models in specific prediction lengths are still different. The module specifically includes two steps. The first step is initial priority setting. During the model training process, the prediction accuracy of the existing data is tested for different clock difference prediction lengths. According to the accuracy performance of different models on the test data, the initial priority of model selection is set according to the historical data-based empirical accuracy for different prediction lengths. The second step is the priority updating strategy based on empirical accuracy. For available clock difference prediction models, the prediction accuracy is evaluated every day, the clock difference prediction accuracy of each model within the last three days for different prediction lengths is compared, the latest prediction accuracy result is obtained, the priority information is updated according to the accuracy ranking, and the optimal clock difference prediction model is selected to ensure higher prediction accuracy.

[0033] In a specific embodiment of the present application, in combination with Figure 1 and Figure 2 , the availability judgment module is specifically used for:

[0034] judging whether there is a prediction method whose required computing resources do not exceed the computing resource threshold from the first prediction method based on the polynomial period term model, the second prediction method based on the pre-trained and fine-tuned Autoformer model, and the third prediction method based on the wavelet analysis-based LSTM model, and if so, determining the prediction method as a preliminary prediction method;

[0035] wherein the computing resources of each prediction method can be known in advance according to existing data, and the computing resource threshold is a value set in advance according to industry experience.

[0036] For each preliminary prediction method, if the preliminary prediction method has historical clock difference data for input, and the number of historical clock difference data exceeds a predetermined number threshold, it is determined that the preliminary prediction method meets the data availability condition; otherwise, the preliminary prediction method is marked as unavailable.

[0037] In a specific embodiment of the present application, in combination with Figure 1 and Figure 2 , the prediction accuracy comparison module is specifically used for:

[0038] S1, if the preliminary prediction method meeting the data availability condition is the first prediction method, it is determined that the first prediction method is determined as a to-be-screened prediction method;

[0039] S2, if the preliminary prediction method meeting the data available condition is a second prediction method, determining whether the second prediction method exists a corresponding fine-tuned Autoformer model; if the second prediction method does not exist a corresponding fine-tuned Autoformer model, generating a fine-tuned Autoformer model; checking whether the fine-tuned Autoformer model needs to be updated, if yes, updating the fine-tuned Autoformer model to obtain a latest fine-tuned Autoformer model, and then determining the second prediction method as a prediction method to be screened; if no, directly determining the second prediction method as the prediction method to be screened;

[0040] Specifically, S2 includes:

[0041] If the preliminary prediction method meeting the data available condition is a second prediction method, determining whether the second prediction method exists a corresponding fine-tuned Autoformer model;

[0042] If the second prediction method does not exist a corresponding fine-tuned Autoformer model, determining whether a pre-trained Autoformer model corresponding to the second prediction method exists, if no, training a pre-trained Autoformer model by using long-term clock difference data; if yes, retraining the existing pre-trained Autoformer model by using the latest clock difference data to obtain a fine-tuned Autoformer model, and then determining the second prediction method as the prediction method to be screened after the generation of the fine-tuned Autoformer model;

[0043] If the second prediction method exists a fine-tuned Autoformer model, checking whether the fine-tuned Autoformer model needs to be updated, if yes, updating the fine-tuned Autoformer model, and then determining the second prediction method as the prediction method to be screened after the updating; wherein, the condition of whether to update is whether the generation time of the fine-tuned Autoformer model is within a predetermined valid period.

[0044] In combination with Figure 2 and Figure 3, the pre-training Autoformer model is pre-trained using a large amount of data, and the fine-tuning Autoformer model is fine-tuned using the latest clock difference data. The pre-training part helps the model master the general pattern and periodic structure of the clock difference change through the autocorrelation mechanism and series decomposition structure of the Autoformer model. The model fine-tuning can learn from the latest clock difference sequence, and through incremental training and parameter fine-tuning, it can quickly adapt to sudden changes in the face of low-orbit satellite clock difference cross-day data differences, dynamically update the model, improve the clock difference prediction accuracy, and enhance the generality and robustness of the model. The advantage of the Autoformer model is that it can capture the long-term change rule in the clock difference sequence, which is beneficial to improve the medium and long-term clock difference prediction accuracy.

[0045] For the pre-training part of the Autoformer model, first, the data is pre-processed to obtain long-term large-scale satellite clock difference data for pre-training . Secondly, the Autoformer architecture is used to decompose the sequence trend item and periodic item information.

[0046]

[0047] wherein, represents the trend item part, represents the seasonal, i.e. periodic item part. The model aims to predict the clock difference value at future time:

[0048]

[0049] The loss function is represented as:

[0050]

[0051] The pre-trained Autoformer model is generated, and the trained Autoformer model parameters are saved:

[0052]

[0053] The fine-tuning part of the Autoformer model first selects the latest small amount of clock difference data as input, ; secondly, the parameters of the pre-trained Autoformer model are loaded as initial parameters, and the parameters of the Autoformer model are adjusted on small-scale data for new clock difference data features:

[0054]

[0055] Get new Autoformer model parameters:

[0056]

[0057] wherein is the learning rate, is the gradient of the loss function with respect to the parameters, indicating the direction and size of the influence of the current parameters on the error.

[0058] S3, if the preliminary prediction method meeting the data available condition is a third prediction method, it is judged whether the third prediction method has a corresponding LSTM model; if not, an LSTM model is generated; it is checked whether the LSTM model needs to be updated; if yes, the LSTM model is updated to obtain the latest LSTM model, and the third prediction method is determined as a to-be-screened prediction method; if not, the third prediction method is directly determined as a to-be-screened prediction method.

[0059] S4, for each to-be-screened prediction method, the prediction accuracy of the to-be-screened prediction method is calculated according to the difference between the historical prediction clock error data and the actual clock error data of the to-be-screened prediction method.

[0060] S5, for each to-be-screened prediction method, the priority is set according to the prediction accuracy of the to-be-screened prediction method, and the to-be-screened prediction method corresponding to the highest priority is selected as an available prediction method.

[0061] In a specific embodiment of the present application, the prediction accuracy comparison module, after obtaining the available prediction method, uses the clock error data predicted by the available prediction method and the future actual clock error data to calculate a new prediction accuracy, thereby updating the priority.

[0062] In a specific embodiment of the present application, the clock error prediction model subsystem is specifically used for:

[0063] receiving a selection result;

[0064] if the selection result is that the first prediction method is an available prediction method, the clock error is predicted by using the prediction method based on the polynomial period term prediction model;

[0065] if the selection result is that the second prediction method is an available prediction method, the clock error is predicted by using the prediction method based on the pre-training fine-tuning Autoformer model.

[0066] if the selection result is that the third prediction method is an available prediction method, the clock error is predicted by using the prediction method based on the wavelet analysis LSTM model.

[0067] In a specific embodiment of the present application, the process of clock error prediction using the prediction method based on the polynomial periodic term prediction model comprises:

[0068] Polynomial fitting is performed on the historical clock error data of the low-orbit satellite to obtain fitting residuals;

[0069] Least square estimation is performed on the periodic term coefficients using the fitting residuals to obtain periodic term information of the clock error; wherein the periodic term information comprises amplitude, phase, frequency and other information of the periodic term;

[0070] A polynomial periodic term prediction model of the clock error is established using the periodic term information of the clock error;

[0071] The clock error is predicted using the polynomial periodic term prediction model to obtain predicted clock error data.

[0072] In a specific embodiment of the present application, referring to Figure 4 , the process of clock error prediction using the prediction method based on the LSTM model of wavelet analysis comprises:

[0073] The historical clock error data of the low-orbit satellite is obtained and preprocessed to obtain clock error training data;

[0074] In this step, the clock error data of the low-orbit satellite is obtained Data preprocessing is performed to remove trend items and outliers, and low-orbit satellite clock error data for model training is obtained.

[0075] The clock error training data is composed into a clock error sequence, and wavelet transform decomposition is performed on the clock error sequence to obtain wavelet components of each layer;

[0076] The input low-orbit satellite clock error sequence is decomposed by wavelet transform. The maximum overlap discrete wavelet transform has the advantages of no down-sampling, shift invariance, etc., and is suitable for signals of any length. Its formula is represented as:

[0077]

[0078]

[0079]

[0080] wherein is the detail function of the j th layer, is the approximation function of the th layer, is the filter length, is the sequence length, and all indices in the formula are taken modulo to ensure the processing of the loop boundary. and represent the first layer filter, which is expressed as:

[0081]

[0082]

[0083] wherein, , are high-pass and low-pass filters, respectively. For the input clock difference sequence data, the decomposition results of the first layer, i.e., wavelet components, are obtained by maximum overlap discrete wavelet transform:

[0084] ；

[0085] For each wavelet component of each layer, an independent LSTM model is constructed, and the prediction result of the wavelet component is generated by using the LSTM model;

[0086] The decomposed wavelet components are independently predicted as variables, i.e., for each wavelet component , an independent LSTM model is constructed to generate a multi-step prediction result:

[0087]

[0088] wherein is the input time step, is the output prediction sequence step.

[0089] The prediction results of all wavelet components are wavelet reconstructed to obtain the predicted clock difference data.

[0090] The wavelet reconstruction of the component prediction results is an inverse wavelet transform process, which is expressed as:

[0091]

[0092] Finally, the clock difference prediction result in the future time period is obtained

[0093]

[0094] In a specific embodiment of the present application, the process of predicting the clock difference by using the prediction method based on the pre-training fine-tuning Autoformer model includes:

[0095] The clock difference prediction method based on the pre-training fine-tuning Autoformer model is used to perform clock difference prediction on the fine-tuned Autoformer model to obtain predicted clock difference data.

[0096] According to the new prediction model, i.e., the fine-tuned Autoformer model, the corresponding predicted clock difference sequence of the future time is obtained.

[0097]

[0098] In addition, the clock difference prediction model has an updating function. The low-orbit satellite data is constantly updated, and the model needs to be retrained periodically. New clock difference data is added every interval to retrain once, so as to prevent the model from deviating after long-term operation. For fine-tuning, the training set is smaller, and the model training is faster, so the model can be quickly updated and adjusted.

[0099] It should be noted that the terms "first", "second" in the present application are only used for description purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined as "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise specifically limited.

[0100] The above is a further detailed description of the present application in combination with specific preferred embodiments, and the specific implementation of the present application cannot be limited to these descriptions. For ordinary skilled persons in the technical field to which the present application belongs, a number of simple deductions or replacements can be made without departing from the concept of the present application, and all of them should be regarded as falling within the protection scope of the present application.

Claims

1. An adaptive low-Earth orbit satellite clock bias real-time prediction system based on multi-method fusion, characterized in that, include: Clock error prediction model subsystem and adaptive selection subsystem; The clock error prediction model subsystem is used to provide the adaptive selection subsystem with a first prediction method based on a polynomial periodic term model, a second prediction method based on a pre-trained and fine-tuned Autoformer model, and a third prediction method based on a wavelet analysis LSTM model; wherein, the available information includes the required computing resources and historical clock error data for input. The adaptive selection subsystem is used to select a forecasting method that meets the computational resource conditions, data availability conditions, and forecast accuracy from the first forecasting method to the third forecasting method based on the available information, and to feed the selection result back to the clock difference forecasting model subsystem. The clock error prediction model subsystem is used to predict clock errors using the prediction method corresponding to the selection result.

2. The adaptive low-Earth orbit satellite clock bias real-time prediction system based on multi-method fusion according to claim 1, characterized in that, The adaptive selection subsystem includes an availability judgment module and a prediction accuracy comparison module; The availability judgment module is used to select a preliminary prediction method that meets the computing resource conditions and data availability conditions from the first prediction method based on the polynomial periodic term model, the second prediction method based on the pre-trained and fine-tuned Autoformer model, and the third prediction method based on the wavelet analysis LSTM model, based on the availability information. The forecast accuracy comparison module is used to select the preliminary forecast method with the highest forecast accuracy from all preliminary forecast methods that meet the data availability conditions as the available forecast method, and to feed back the selection result to the clock difference forecast model subsystem.

3. The adaptive low-Earth orbit satellite clock bias real-time prediction system based on multi-method fusion according to claim 2, characterized in that, The availability determination module is specifically used for: Determine whether there is a prediction method among the first prediction method based on a polynomial periodic term model, the second prediction method based on a pre-trained and fine-tuned Autoformer model, and the third prediction method based on a wavelet analysis LSTM model that requires no more than a computational resource threshold. If so, determine the prediction method as the preliminary prediction method. For each preliminary forecast method, if the preliminary forecast method has historical clock difference data for input and the number of historical clock difference data exceeds a predetermined threshold, then the preliminary forecast method is determined to meet the data availability condition. Otherwise, mark the preliminary forecasting method as unavailable.

4. The adaptive low-Earth orbit satellite clock bias real-time prediction system based on multi-method fusion according to claim 3, characterized in that, The forecast accuracy comparison module is specifically used for: S1, if the preliminary forecast method that meets the data availability condition is the first forecast method, then the first forecast method will be determined as the forecast method to be screened. S2, if the preliminary forecast method that meets the data availability condition is the second forecast method, then determine whether the second forecast method has a corresponding fine-tuned Autoformer model; if the second forecast method does not have a corresponding fine-tuned Autoformer model, then generate a fine-tuned Autoformer model; check whether the fine-tuned Autoformer model needs to be updated, if so, update the fine-tuned Autoformer model to obtain the latest fine-tuned Autoformer model, and then determine the second forecast method as the forecast method to be screened; if it does not need to be updated, then directly determine the second forecast method as the forecast method to be screened. S3. If the preliminary forecast method that meets the data availability conditions is the third forecast method, then determine whether there is a corresponding LSTM model for the third forecast method; if there is no LSTM model, then generate an LSTM model; check whether the LSTM model needs to be updated, and if so, update the LSTM model to obtain the latest LSTM model, and then determine the third forecast method as the forecast method to be screened. If no update is required, the third forecasting method will be directly selected as the forecasting method to be screened. S4. For each forecast method to be screened, calculate the forecast accuracy of the forecast method to be screened based on the difference between the historical forecast clock difference data and the actual clock difference data of the forecast method to be screened. S5. For each forecasting method to be screened, set a priority based on the forecast accuracy of the forecasting method to be screened, and select the forecasting method with the highest priority as the available forecasting method.

5. The adaptive low-Earth orbit satellite clock bias real-time prediction system based on multi-method fusion according to claim 4, characterized in that, After obtaining the available forecasting methods, the forecast accuracy comparison module uses the clock difference data predicted by the available forecasting methods and the actual future clock difference data to calculate the new forecast accuracy, thereby updating the priority.

6. The adaptive low-Earth orbit satellite clock bias real-time prediction system based on multi-method fusion according to claim 4, characterized in that, S2 include: If the preliminary forecast method that meets the data availability condition is the second forecast method, then determine whether the second forecast method has a corresponding fine-tuned Autoformer model; If the second forecast method does not have a corresponding fine-tuned Autoformer model, then it is determined whether there is a pre-trained Autoformer model corresponding to the second forecast method. If not, a pre-trained Autoformer model is trained using long-term clock bias data. If a pre-trained Autoformer model exists, the existing pre-trained Autoformer model is retrained using the latest clock bias data to obtain a fine-tuned Autoformer model. After generation, the second forecast method is identified as the forecast method to be screened. If the second forecasting method has a fine-tuned Autoformer model, then check whether the fine-tuned Autoformer model needs to be updated. If so, update it and then determine the second forecasting method as a forecasting method to be screened after the update. The condition for whether an update is needed is whether the generation time of the fine-tuned Autoformer model is within the predetermined validity period.

7. The adaptive low-Earth orbit satellite clock bias real-time prediction system based on multi-method fusion according to claim 3, characterized in that, The clock error prediction model subsystem is specifically used for: Receive the selection result; If the selected result is that the first forecasting method is an available forecasting method, then the clock error forecasting is performed using the forecasting method based on the polynomial periodic term forecasting model. If the second prediction method is selected as the available prediction method, then the clock error prediction is performed using the prediction method based on the pre-trained and fine-tuned Autoformer model. If the third forecasting method is selected as the available forecasting method, then the clock error forecasting method based on the LSTM model of wavelet analysis is used.

8. The adaptive low-Earth orbit satellite clock bias real-time prediction system based on multi-method fusion according to claim 7, characterized in that, The process of clock error prediction using the prediction method based on a polynomial periodic term prediction model includes: Polynomial fitting was performed using historical clock error data from low-Earth orbit satellites to obtain the fitting residuals; The periodic term information of the clock error is obtained by using the least squares estimation of the periodic term coefficients using the fitted residuals. A polynomial periodic term prediction model for clock error is established using the periodic term information of the clock error. The predicted clock error data is obtained by using the polynomial periodic term prediction model.

9. The adaptive low-Earth orbit satellite clock bias real-time prediction system based on multi-method fusion according to claim 7, characterized in that, The process of clock error prediction using the prediction method based on wavelet analysis and LSTM model includes: Historical clock bias data of low-Earth orbit satellites are acquired and preprocessed to obtain clock bias training data; The clock difference training data are used to form a clock difference sequence, and the clock difference sequence is decomposed by wavelet transform to obtain the wavelet components of each layer. For each wavelet component, an independent LSTM model is constructed, and the prediction results of the wavelet component are generated using the LSTM model. The predicted clock error data is obtained by reconstructing the prediction results of all wavelet components using wavelet refactoring.

10. The adaptive low-Earth orbit satellite clock bias real-time prediction system based on multi-method fusion according to claim 7, characterized in that, The process of clock error prediction using the prediction method based on a pre-trained fine-tuned Autoformer model includes: The predicted clock error data is obtained by using the prediction method based on the pre-trained and fine-tuned Autoformer model and the corresponding fine-tuned Autoformer model for clock error prediction, and is represented as follows: ； in, This represents existing historical clock difference data. This indicates the parameters of the latest fine-tuned Autoformer model.

Images