Intelligent prediction method and system for lunar satellite orbit

Abstract

The application discloses a kind of lunar satellite orbit intelligent prediction method and system, the method includes: constructing basic motion model;Construct prediction error dataset;Deep learning model is constructed;Deep depth learning model is trained;Error estimation is carried out using basic motion model and deep learning model;According to the error of estimation, final orbit prediction result is generated.The system includes: model construction unit, dataset construction unit, model training unit, data acquisition unit, initial prediction unit, error prediction unit and final output unit.Through using the application, high-precision orbit prediction data can be provided for lunar satellite.The application can be widely applied in satellite orbit prediction field.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of satellite orbit prediction, in particular to a method and system for intelligent prediction of lunar satellite orbit. BACKGROUND

[0002] Currently, the satellite orbit determination of the lunar satellite is mainly achieved by constructing a satellite dynamics model based on ground observation results. However, the orbit prediction technology of the traditional dynamics model for the position of the lunar satellite at a future time still has great limitations in real-time and accuracy. In order to solve the problem of large error of the dynamics model, related researches construct more refined dynamics models to achieve high-precision prediction of the lunar satellite orbit. However, the space environment around the moon is complex, and there are nonlinear perturbation factors, and there are factors that cannot be modeled by the dynamics model. These factors limit the orbit prediction accuracy of the dynamics model, so high-frequency tracking measurement of the ground measurement station is needed to update the perturbation parameters in the dynamics model to ensure the orbit determination accuracy of the lunar satellite. However, the ground measurement time is affected by the lunar satellite orbit and the moon blocking, and it is impossible to track the lunar satellite for a long time, which leads to a decrease in the orbit determination accuracy of the lunar satellite and cannot meet the requirement of providing high-precision ephemeris broadcast for lunar surface equipment navigation and positioning. SUMMARY

[0003] Therefore, in order to solve the problem of low prediction accuracy in the existing lunar satellite orbit prediction method, in the first aspect, the present application provides an intelligent prediction method for the lunar satellite orbit, which comprises the following steps:

[0004] A two-body system model is used as the basic motion model of the lunar satellite orbit;

[0005] The position prediction is performed based on the basic motion model of the lunar satellite orbit, and the prediction error data set of the satellite two-body dynamics model is constructed by combining the accurate orbit data of the ground tracking station;

[0006] A deep learning model based on self-supervised random mask is constructed, including a one-dimensional convolution feature extractor, a random mask structure, an encoder part based on prospare self-attention, and a decoder part composed of a prospare self-attention layer and a linear projection layer;

[0007] After the data set and the model are constructed, the prediction error data set is used to train the deep learning model, and the mean square error (MSE) between the input and the output is used to calculate the model loss, so as to obtain the trained deep learning module;

[0008] According to the latest orbit position information obtained before the communication with the ground station is interrupted as initial prediction parameters, the satellite position calculation formula in the designed basic motion model of the lunar satellite orbit is used to start the prediction of the satellite orbit, and an initial orbit prediction result is generated;

[0009] After obtaining the initial orbit prediction result, the initial orbit prediction result is subtracted from the real orbit data provided by the ground tracking station to obtain an orbit prediction error result;

[0010] The orbit prediction error result is taken as the input of the deep learning model, and after the model is calculated, the model error estimation result, that is, the corresponding error value of the two-body model prediction data, is output;

[0011] The initial orbit prediction result generated by the lunar satellite two-body dynamics model is combined with the model error estimation result to obtain the final accurate satellite orbit prediction result.

[0012] In a second aspect, the present application further provides a lunar satellite orbit intelligent prediction system, which comprises:

[0013] A model construction unit adopts a two-body system model as a basic motion model of the lunar satellite orbit, and constructs a deep learning model based on self-supervised random mask, which comprises a one-dimensional convolution feature extractor, a random mask structure, an encoder part based on prospareself-attention, and a decoder part composed of a prospare self-attention layer and a linear projection layer;

[0014] A data set construction unit constructs a prediction error data set of the satellite two-body dynamics model by combining the accurate orbit data of the ground tracking station with the position prediction of the basic motion model of the lunar satellite orbit;

[0015] A model training unit, after the data set and the model are constructed, trains the deep learning model by using the prediction error data set, and calculates the model loss by using the mean square error (MSE) between the input and the output, to obtain the trained deep learning module;

[0016] A data acquisition unit, according to the latest orbit position information obtained before the communication with the ground station is interrupted as initial prediction parameters;

[0017] An initial prediction unit, which uses the satellite position calculation formula in the designed basic motion model of the lunar satellite orbit to start the prediction of the satellite orbit, and generates an initial orbit prediction result;

[0018] The error prediction unit obtains an initial orbit prediction result, and obtains an orbit prediction error result by subtracting true orbit data provided by a ground tracking station from the initial orbit prediction result; the orbit prediction error result is input into a deep learning model, and a model error estimation result, i.e., a corresponding error value of the two-body model prediction data, is output after calculation of the model.

[0019] The final output unit is used for combining the initial orbit prediction result generated by the two-body dynamics model of the moon-orbiting satellite and the model error estimation result to obtain an accurate satellite final orbit prediction result.

[0020] The application further provides a moon-orbiting satellite orbit intelligent prediction device, which comprises:

[0021] At least one processor;

[0022] At least one memory for storing at least one program;

[0023] When the at least one program is executed by the at least one processor, the at least one processor implements the above-mentioned moon-orbiting satellite orbit intelligent prediction method.

[0024] Based on the above scheme, the application provides a moon-orbiting satellite orbit intelligent prediction method and system, which calculates the error between the dynamics model and the observation result by using the historical orbit data observed on the ground, learns the change rule of the error by using the neural network model of the application, and uses the error correction result calculated by the network model to correct the calculation result of the satellite dynamics model when the moon-orbiting satellite cannot receive the ground measurement result, so that the orbit prediction result of the moon-orbiting satellite dynamics model is accurately corrected without increasing additional sensors, occupying additional space inside the satellite and using external equipment, high-precision orbit prediction data is provided for the moon-orbiting satellite, the autonomous navigation and positioning accuracy of the moon-orbiting satellite is improved, and accurate ephemeris broadcasting is provided for the lunar surface equipment. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 is a step flowchart of a moon-orbiting satellite orbit intelligent prediction method of the application;

[0026] Figure 2 is a data flow direction schematic diagram of a deep learning model of a specific embodiment of the application;

[0027] Figure 3 is a structure block diagram of a moon-orbiting satellite orbit intelligent prediction system of the application. DETAILED DESCRIPTION

[0028] The application provides a high-precision autonomous orbit prediction scheme for a lunar satellite, which mainly includes two parts: one is a lunar satellite dynamics model prediction error correction method, which uses a time series neural network based on self-supervised random mask to correct the error of the satellite dynamics orbit calculation result, thereby improving the autonomous orbit prediction accuracy of the satellite; the second is a complete data acquisition, neural network model training, model updating, orbit prediction, and satellite autonomous orbit prediction scheme, which, in combination with the proposed lunar satellite autonomous orbit prediction method, can achieve accurate orbit prediction results for the lunar satellite itself under the condition of no additional sensors and no additional auxiliary measurement and control auxiliary information, thereby greatly improving the autonomous navigation and orbit prediction capability of the lunar satellite.

[0029] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0030] It should be noted that, for the convenience of description, only the parts related to the application are shown in the drawings. The embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0031] It should be understood that the "system", "device", "unit" and / or "module" used in the present application is a method for distinguishing different components, elements, parts, portions or assemblies at different levels. However, if other words can achieve the same purpose, the word can be replaced by other expressions.

[0032] As shown in the present application and claims, unless the context clearly indicates otherwise, "one", "a", "an" and / or "the" do not refer to a single number, but also include plural. Generally, the terms "include" and "contain" only indicate the inclusion of the steps and elements explicitly identified, and these steps and elements do not constitute an exclusive list, and the method or device can also include other steps or elements. The element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, product or device including the element.

[0033] In the description of the embodiments of the present application, "multiple" refers to two or more than two. The following terms "first", "second" are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features.

[0034] In addition, flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or subsequent operations are not necessarily performed in sequence. On the contrary, each step can be processed in reverse order or simultaneously. Meanwhile, other operations can be added to these processes, or one or more steps of operations can be removed from these processes.

[0035] Reference Figure 1 The flowchart of an optional example of the intelligent prediction method of the lunar satellite orbit proposed in the present application can be applied to a computer device. The orbit prediction method proposed in the present embodiment can include but is not limited to the following steps:

[0036] Step S1, constructing a lunar satellite two-body dynamics model;

[0037] Step S2, generating a prediction error data set based on the lunar satellite two-body dynamics model and combining the data of the ground tracking station;

[0038] Step S3, constructing a deep learning model based on self-supervised random mask;

[0039] Step S4, training the deep learning model based on the prediction error data set to obtain a trained deep learning model;

[0040] Step S5, obtaining current orbit position information;

[0041] Step S6, generating an initial orbit prediction result according to the current orbit position information and the lunar satellite two-body dynamics model;

[0042] Step S7, subtracting the initial orbit prediction result from the data of the ground tracking station to obtain an orbit prediction error result;

[0043] Step S8, inputting the orbit prediction error result into the trained deep learning model to obtain a model error estimation result;

[0044] Step S9, adding the model error estimation result to the initial orbit prediction result to obtain a final orbit prediction result.

[0045] In some possible embodiments, step S1 specifically includes:

[0046] First, a dynamic model of the lunar orbital satellite is established. This invention uses a two-body system model as the basic motion model for the lunar orbital satellite. The acceleration experienced by the satellite in the lunar center inertial coordinate system is as follows:

[0047]

[0048] Where G is the gravitational constant; r is the distance from the lunar center to the satellite, and r = ||r||. Let the satellite's position vector in the lunar inertial coordinate system be r = (x, y, z), then the satellite's acceleration along each axis can be written in the following component form:

[0049]

[0050] in, The accelerations of the satellite along its three axes in the inertial coordinate system are given respectively. Using the acceleration formula, the coordinates of the lunar orbiting satellite at any future moment can be calculated. The formula for the lunar orbiting satellite's position is as follows:

[0051]

[0052] Where, x t y t z t Let x be the position of the lunar orbiting satellite in the lunar inertial coordinate system at time t in the future. t-1 y t-1 z t-1 These represent the satellite's position at the previous moment. These represent the average velocities of the lunar orbiter along its three axes at the previous moment. In practical applications, the position and velocity of the lunar orbiter can be obtained through measurement and calculation by a ground station. After obtaining the initial values ​​through ground station calculations, the satellite's position and velocity at a future moment in the two-body model can be calculated using the above formulas.

[0053] In some feasible embodiments, step S2 specifically includes:

[0054] A post-hoc precise orbit dataset for lunar orbiting satellites is defined by ground measurement stations. The post-hoc corrected dataset is the precise position information obtained by comprehensively calculating and evaluating the measurement results from multiple ground stations, taking into account factors such as ionospheric errors, tropospheric errors, and clock deviations observed by surrounding meteorological stations. Because the influence of multiple variables on the observation results needs to be comprehensively considered, the precise position dataset can only be obtained through comprehensive calculation after ground station observations; it is a post-hoc calculation result. The post-hoc precise orbit dataset for ground measurement stations is defined as follows:

[0055]

[0056] Among them, (X) G YG Z G ) represent the precise positions of the three axes calculated by the ground station after post-processing. i represents the total length of the observation arc, that is, the total number of coordinate points obtained by the ground station after comprehensive calculation within an observation arc.

[0057] Based on the above formulas and the post-hoc calculation results from the ground station, we can begin to construct the satellite two-body dynamics model prediction error dataset, which is calculated as follows:

[0058]

[0059] in:

[0060] This allows us to obtain the error datasets ΔX, ΔY, and ΔZ between the lunar orbiter's dynamic model and its precise orbit, which represent the error sets of the lunar orbiter along three axes.

[0061] Combining steps S1 and S2, the training process specifically involves the following: After the lunar orbiting satellite enters the tracking and control range of the ground control station, the ground control station begins tracking and controlling the lunar orbiting satellite. After calculation, a set of precise satellite position coordinates is obtained. Simultaneously, using the two-body model of the lunar orbiting satellite, the precise positions calculated by each ground control station are used as the initial prediction parameters, and the satellite orbital position is extrapolated using the lunar orbiting satellite position formula. The ground control station provides the calculated precise position coordinates, while the model training server is responsible for the extrapolation calculation of position coordinates under the two-body model of the lunar orbiting satellite and for organizing the training dataset.

[0062] In some feasible embodiments, step S3 specifically includes:

[0063] The deep learning model designed in this invention consists of the following parts: a one-dimensional convolutional feature extractor, a random mask structure, and an encoder part that is based on prospare self-attention. The decoder part consists of a prospare self-attention layer and a linear projection layer. The decoder and encoder together form the autoencoder structure for satellite orbit error prediction. The input to the model during training is the time series of satellite orbit errors. By training the model, the error sequence of the satellite orbit data is recovered at the point level, so that the model can accurately predict the satellite orbit prediction error.

[0064] like Figure 2 The diagram shown illustrates the structure and data flow of the self-supervised random mask deep learning model proposed in this patent.

[0065] The data processing procedure for this deep learning model structure is as follows:

[0066] The present application uses a window slicing operation to split the original time series data into continuous non-overlapping subsequences, specifically, for a given time series Slicing refers to extracting information of a segment in time series data, defined as x i:σ = {x i ,x i+1 ,…,x i+σ}, where i:σ represents the size of the slicing window, and the length of the time series is T as known from the foregoing, so the data length becomes The slicing method used by the present application is a non-sliding window method, but a non-overlapping method is used to extract the original time series, that is, a same one-dimensional convolution kernel K 1D is used to perform convolution operation on the original data, and the size of the data for each convolution is the length of the sliced data above, that is, (σ-i). The convolution result of the input X is represented as where d represents the dimension of the data after one-dimensional convolution, and S is used to represent the length of the data after one-dimensional convolution S v , S m represent the visible sequence and the masked sequence respectively. Then, position encoding is performed on the data, and the position encoding assigns a unique code to each variable in the sequence using absolute position encoding, and the formula is as follows:

[0067]

[0068] where pos represents the position of the input sequence, and d model represents the dimension of the model. In order to enable the model to learn the change characteristics of the entire time series, the present application uses a random mask strategy to mask the input signal of the encoder, and it needs to be emphasized that for each subsequence unit, the proportion of the masked unit to the total input data quantity of the entire unit is the same, which is 73% in the present application. In order to improve the number of model training samples and ensure that the model can completely learn the change rule of the entire time series data, dynamic random masking is used while masking, that is, before masking each subsequence, 27% of the data in the subsequence is randomly selected as unmasked data, and the masked sequence is represented as , the visible sequence is represented as , and

[0069] In the encoder, after slicing, position encoding and random masking of the input time series, the visible sequence Z vThe input encoder maps the visible sequence data to a latent representation. The encoder part adopts a standard Informer network architecture, which uses Prosparse self-attention mechanism instead of traditional attention mechanism, greatly reduces the amount of calculation while ensuring the model's feature extraction ability unchanged. At the same time, the multi-head attention mechanism is used to capture the temporal information of different subspaces. Each head can capture different aspects or patterns of the input data, improving the model's understanding and representation ability of the orbit error data. After the Informer network calculation, the output intermediate features are obtained

[0070] In the decoder, a linear layer is first added to reduce the dimension of the input to d' to improve training efficiency, and then the output of the encoder and the masked sequence are combined to form a new input sequence. The masked sequence Z m When combined with the latent representation output by the encoder, the masked sequence is filled into the intermediate features output by the encoder according to the original position. A dropout layer (dropout rate 0.1) is added at the end of the Informer module to prevent model overfitting, and a linear projection layer is added at the end of the decoder to directly output the point-level reconstruction result of the satellite orbit error sequence. The whole process can be represented by the following formula:

[0071]

[0072] Where is the intermediate feature output by the encoder, with dimension d' and length l,χ l,d' represents the input vector, and the final output result is represented by .

[0073] It should be noted that the data processing process of the model is not much different between the training process and the application process.

[0074] In some possible embodiments, step S4 specifically includes:

[0075] After completing the model dataset and model construction, the model training and deployment are started. The mean square error (MSE) between the input and output is used to calculate the model loss L in the training process, and the calculation formula is as follows:

[0076]

[0077] After the model is trained, it enters the deployment stage. The ground tracking station waits for the circumlunar satellite to run into the observable range of the ground station again. After the circumlunar satellite enters the observable range of the ground station, the ground tracking station measures the orbit of the circumlunar satellite and performs post-solution on the real orbit position of the circumlunar satellite according to the measurement result. According to the solution result, the initial state parameters {x t-1 , yt-1 t-1}, the satellite orbit position prediction result is calculated by using the initial orbit parameter and a two-body model calculation formula, the two-body prediction result is subtracted from the ground station observation result to obtain a lunar-orbiting satellite orbit error time sequence {DeltaX, DeltaY, DeltaZ}, the error sequence is used as model training data set to complete the training of the model, and finally the model weight parameters are saved, waiting for the ground station and the lunar-orbiting satellite to complete the model weight parameter uploading again.

[0078] In combination with step S3 and step S4, the training process is specifically as follows: after the construction of the training data set is completed, the data set is input into the self-supervised random mask deep learning model to start the model parameter training, and the whole training process is performed on the model training server. After the model converges, the final model weight parameter file is generated. The weight parameter file is returned to the ground tracking station and waits for the lunar-orbiting satellite and the ground tracking station to communicate again. When the ground tracking station observes the lunar-orbiting satellite again, the position of the lunar-orbiting satellite is solved and saved, and the data transmission and communication function of the ground station and the lunar-orbiting satellite is used to upload the trained final model parameter file to the lunar-orbiting satellite host. After the satellite host receives the model parameter file, the deep learning model weight parameters in the lunar-orbiting satellite on-board computer are updated, so that the network model under the latest weight parameters is obtained.

[0079] In some possible embodiments, steps S5-S7 specifically include:

[0080] After the lunar-orbiting satellite host completes the model parameter updating, the latest orbit position information obtained before the communication with the ground station is interrupted is used as the initial prediction parameter, and a satellite position calculation formula is used to start the prediction of the satellite orbit. After the position result calculated by using the two-body model is obtained, the orbit prediction result of the two-body model is subtracted from the real orbit data provided by the ground station to obtain the orbit prediction error value as the model input. After the model is solved, the model error estimation result of the future time, that is, the corresponding error value of the two-body model prediction data, is output.

[0081] The on-board computer of the lunar-orbiting satellite adds the two-body model prediction result and the error result to obtain the final accurate orbit prediction result of the satellite. The lunar-orbiting satellite uses the calculated final accurate orbit prediction result to determine the satellite orbit position at a future time in combination with the clock system of the satellite. Meanwhile, the ground station continues to search for the lunar-orbiting satellite and waits for the lunar-orbiting satellite to enter the observation field of view. After the lunar-orbiting satellite enters the observation field of view again, the above steps are repeated, the satellite orbit data is reacquired, the network model is trained again, the network model parameters are updated, and the final model parameter file is uploaded again, so that the accurate orbit self-prediction of the lunar-orbiting satellite is realized.

[0082] Based on the above specific embodiments, the beneficial effects of the present application are specifically as follows:

[0083] ​Firstly, the application proposes a neural network based on self-supervised random mask. Compared with the traditional orbit prediction algorithm based on dynamic model, the method adopts random mask strategy to construct the training data set, which can mine the deeper internal relationship and characteristics in the data, so that the network can more accurately learn the internal law of the lunar satellite orbit error data, and the final prediction result of the network is more in line with the actual change of the lunar satellite orbit prediction error. At the same time, the application proposes a deployment scheme of lunar satellite orbit prediction algorithm, which can realize the accurate orbit self-prediction of the lunar satellite without adding additional detection equipment, and the orbit prediction accuracy is higher than that of the traditional scheme, and the cost is lower (because the satellite does not need to add additional sensors and measurement equipment, and does not change the working process of the existing ground measurement station, only one model training server is added on the ground), and the existing ground measurement resources are fully utilized, and the real-time performance of the satellite orbit self-prediction is higher (because the network model training process is completed on the ground, and only the model parameter file after training is deployed on the lunar satellite host).

[0084] As Figure 3 The application relates to an intelligent lunar satellite orbit prediction system, which comprises:

[0085] A model construction unit is used for constructing a lunar satellite two-body dynamic model and a deep learning model based on self-supervised random mask.

[0086] A data set construction unit is used for generating a prediction error data set based on the lunar satellite two-body dynamic model and the data of a ground measurement station.

[0087] A model training unit is used for training the deep learning model based on the prediction error data set, so as to obtain a trained deep learning model.

[0088] A data acquisition unit is used for acquiring current orbit position information.

[0089] An initial prediction unit is used for generating an initial orbit prediction result based on the current orbit position information and the lunar satellite two-body dynamic model.

[0090] An error prediction unit is used for subtracting the initial orbit prediction result from the data of the ground measurement station, so as to obtain an orbit prediction error result; and the orbit prediction error result is input into the trained deep learning model, so as to obtain a model error estimation result.

[0091] A final output unit is used for adding the error value to the initial orbit prediction result, so as to obtain a final orbit prediction result.

[0092] The contents in the method embodiments are applicable to the system embodiments, the system embodiments specifically implement the same functions as the method embodiments, and achieve the same beneficial effects as the method embodiments.

[0093] An intelligent lunar satellite orbit prediction device:

[0094] At least one processor;

[0095] At least one memory for storing at least one program;

[0096] When the at least one program is executed by the at least one processor, the at least one processor implements the intelligent lunar satellite orbit prediction method.

[0097] The contents in the method embodiments are applicable to the device embodiments, the device embodiments specifically implement the same functions as the method embodiments, and achieve the same beneficial effects as the method embodiments.

[0098] A storage medium, wherein the storage medium stores processor-executable instructions, and the processor-executable instructions, when executed by a processor, are used to implement the intelligent lunar satellite orbit prediction method.

[0099] The contents in the method embodiments are applicable to the storage medium embodiments, the storage medium embodiments specifically implement the same functions as the method embodiments, and achieve the same beneficial effects as the method embodiments.

[0100] The above is a specific description of the preferred embodiments of the present application, but the present application is not limited to the embodiments, and those skilled in the art can make various equivalent modifications or replacements without departing from the spirit of the present application, and these equivalent modifications or replacements are all included in the scope defined by the claims of the present application.

Claims

1. A method for intelligent prediction of lunar satellite orbits, characterized in that, Includes the following steps: Construct a two-body dynamic model for a lunar orbiting satellite; Based on the two-body dynamics model of the lunar orbiting satellite, a prediction error dataset is generated by combining data from ground tracking and control stations. Construct a deep learning model based on self-supervised random masks; The deep learning model is trained based on the prediction error dataset to obtain the trained deep learning model. Obtain current orbital position information; Based on the current orbital position information and the two-body dynamics model of the lunar orbiting satellite, an initial orbital prediction result is generated; The orbit prediction error is obtained by subtracting the initial orbit prediction result from the data from the ground control station. The trajectory prediction error result is input into the trained deep learning model to obtain the model error estimation result; The model error estimation result is added to the initial trajectory prediction result to obtain the final trajectory prediction result; The deep learning model includes a convolutional feature extractor, a random mask structure, an encoder and decoder based on prospare self-attention; The step of inputting the orbit prediction error result into the trained deep learning model to obtain the model error estimation result specifically includes: Based on the convolutional feature extractor, feature data is obtained by extracting features from the sequence in the orbit prediction error result; The feature data is positionally encoded to obtain the encoded data; Based on the random masking structure, the encoded data is randomly masked to obtain a visible sequence and a masked sequence; The encoder output is obtained by mapping the visible sequence to a latent representation based on the encoder. The encoder output and the masking sequence are used to construct a new input sequence; The new input sequence is decoded using the decoder to obtain the model error estimation result.

2. The intelligent prediction method for lunar satellite orbits according to claim 1, characterized in that, In the two-body dynamics model of the lunar orbiting satellite, the position formula of the lunar orbiting satellite is as follows: in, Let t be the position of the lunar orbiting satellite in the lunar center inertial coordinate system at time t in the future. These represent the satellite's position at the previous moment. These represent the average velocities of the three axes of the lunar satellite at the previous moment. These correspond to the accelerations of the satellite along the three axes in the inertial coordinate system.

3. The intelligent prediction method for lunar satellite orbits according to claim 1, characterized in that, Before the step of training the deep learning model based on the prediction error dataset to obtain the trained deep learning model, the following steps are also included: The original time series in the prediction error dataset is extracted using a non-overlapping method to obtain continuous non-overlapping subsequences.

4. The intelligent prediction method for lunar satellite orbits according to claim 1, characterized in that, During the training process of the deep learning model, the mean squared error is used to calculate the model loss.

5. The intelligent prediction method for lunar satellite orbits according to claim 1, characterized in that, Also includes: The trained deep learning model is then transmitted back to the ground control station.

6. A smart prediction system for lunar satellite orbits, characterized in that, A method for executing a lunar satellite orbit intelligent prediction method as described in claim 1 includes: The model building unit is used to build a two-body dynamic model of a lunar orbiting satellite and a deep learning model based on a self-supervised random mask. The dataset construction unit generates a prediction error dataset based on the lunar orbiting satellite two-body dynamics model and combined with data from ground tracking and control stations. The model training unit trains the deep learning model based on the prediction error dataset to obtain the trained deep learning model. The data acquisition unit is used to acquire the current orbital position information; The initial prediction unit is used to generate an initial orbit prediction result based on the current orbital position information and the two-body dynamics model of the lunar orbiting satellite; An error prediction unit is used to subtract the initial orbit prediction result from the data from the ground tracking and control station to obtain an orbit prediction error result; and to input the orbit prediction error result into the trained deep learning model to obtain a model error estimation result. The final output unit is used to add the model error estimation result to the initial trajectory prediction result to obtain the final trajectory prediction result.

7. A smart device for predicting the orbit of a lunar satellite, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the intelligent prediction method for lunar satellite orbits as described in any one of claims 1-5.

Images