Satellite orbit single-point state forecasting method based on deep learning

Abstract

The invention discloses a satellite orbit single-point state forecasting method based on deep learning. An orbit precision ephemeris is used as training data, a six-dimensional position speed state at a current moment is used as input, a state at a next moment is used as a target value, a prediction network in which LSTM and a one-dimensional CNN are connected in series is constructed, and a CBAM attention module is introduced to realize adaptive weighting of channel and time sequence features. And designing a weighted loss function, carrying out weighted fusion on a mean square error of the model to a real ephemeris, an HPOP equal-step prediction result and a residual mean square error output by the model, and improving prediction rationality through physical constraints. And training and evaluating through the training set, the verification set and the test set, and outputting a predicted ephemeris. According to the method, a historical sequence or TLE is not needed, high-precision and rapid forecasting can be achieved only through a single-point state, and unmodeled perturbative force is implicitly compensated.

Description

Technical Field

[0001] This invention relates to the field of satellite orbit determination and prediction, and specifically to a method for single-point state prediction of satellite orbit based on deep learning. Background Technology

[0002] With the development of aerospace technology, human exploration and utilization of space have become increasingly in-depth, leading to an explosive increase in the number of space targets (satellites, rocket debris, failed payloads, and space debris, etc.). The space environment is becoming increasingly complex, making potential security threats and other issues increasingly prominent. For space debris, there are two main countermeasures: active and passive. These include deorbiting and disposal, active removal, monitoring and early warning, and collision avoidance and protection. However, all of these countermeasures rely heavily on high-precision orbit prediction. Orbit prediction can provide the precise position, velocity, and future orbital trends of space targets.

[0003] Existing orbit prediction technologies are mainly divided into two categories:

[0004] 1) The method based on a precise dynamic model: This method predicts orbits by establishing dynamic equations that incorporate perturbations such as Earth's non-spherical gravity, lunar and solar gravity, atmospheric drag, and solar radiation pressure, and then performing numerical integration. Its accuracy heavily depends on the completeness of the model and the accuracy of environmental parameters. For example, the uncertainty of the atmospheric density model is a major source of prediction errors for low-Earth orbit satellites. Furthermore, the computational burden of numerical integration for complex models is large, making it difficult to meet the needs of autonomous and rapid onboard mission planning.

[0005] 2) Data-driven machine learning and deep learning methods: In recent years, artificial intelligence technology has been introduced to improve forecast accuracy and efficiency. Most researchers focus on error correction of parameters closely related to orbit forecasting, such as using machine learning methods to predict the forecast errors of traditional dynamic models and then correcting those errors. This method still essentially relies on the underlying dynamic model. Some scholars use long-term historical observation data as input, rather than directly extrapolating from a single point. Another approach is to correct atmospheric density, space environment indices, etc., to minimize the forecast errors of traditional dynamic models. However, the space environment is affected by multiple factors, and the factors influencing orbit forecasting are also undergoing complex dynamic changes. Therefore, the final forecast accuracy may accumulate significantly in some cases. Summary of the Invention

[0006] Based on the above analysis, this invention presents a deep learning-based method for predicting the state of a single point in a satellite's orbit, comprising the following steps:

[0007] S1. Obtain precise orbital ephemeris data and use the precise orbital ephemeris data as training data. The precise orbital ephemeris data includes the state of a single point in the satellite orbit at a certain moment.

[0008] S2. Divide the training data into current time data and next time data, using the current time data as the input group and the next time data as the target value group;

[0009] S3. Construct a network model, which includes LSTM (Long Short-Term Memory) and CNN (Convolutional Neural Network). The LSTM is connected to the CNN. The LSTM is used to capture long-term dependencies in the training data, and the CNN is used to extract local features from the training data.

[0010] S4. Optimize the network model, which also includes a CBAM (Convolutional Block Attention Module). The CBAM is connected to the CNN and provides an intelligent weighting strategy for the network model.

[0011] S5. Design the loss function by weighting and summing the orbit prediction results of HPOP (High-Precision Orbit Prediction Model) with the error of the model prediction and the mean square error loss function to form the loss function.

[0012] S6. Train the network model by inputting the input group into the network model, training the network model and outputting predicted ephemeris data, and evaluating the predicted ephemeris data based on the data of the target value group.

[0013] Preferably, the state of the satellite orbit point at a certain moment includes the position of the satellite orbit point in the x-axis direction, the position in the y-axis direction, the position in the z-axis direction, the velocity in the x-axis direction, the velocity in the y-axis direction, and the velocity in the z-axis direction at that moment.

[0014] Preferably, the LSTM includes time intervals. Memory cells State Hidden Vector ,enter Forgotten Gate Update door Output gate and sigmoid function The forget gate The calculation formula is:

[0015]

[0016] in, , These represent the weights between the input and the forget gate, and between the hidden state from the previous time step and the forget gate, respectively. This is the bias vector.

[0017] The update gate The calculation formula is:

[0018]

[0019]

[0020]

[0021] in, , , , These are all weights of the corresponding parts at the current moment. , This is the corresponding bias vector.

[0022] The output gate The calculation formula is:

[0023]

[0024]

[0025] in, , For the corresponding weights, This is the bias vector.

[0026] Preferably, the CNN employs a one-dimensional convolutional neural network.

[0027] Preferably, the CNN includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer.

[0028] Preferably, the CBAM includes CA (Channel Attention) and SA (Spatial Attention). The CA is used to assign higher weights to perturbation features related to orbital state in the network model, and the SA is used to guide the network model to focus on temporal features that have decisive information about changes in orbital state.

[0029] Preferably, the loss function includes the measured mean squared error of the model and the predicted mean squared error of the model, wherein the formula for calculating the measured mean squared error of the model is:

[0030]

[0031] in, This is the actual value of the orbital state at the next ephemeris time. It is the predicted orbital state at the next ephemeris time;

[0032] The model prediction mean square error is first predicted using HPOP to forecast the single-point state of the input satellite orbit with the same step size as the training data. The residual is then calculated by comparing the predicted orbit state value at the next ephemeris time with the HPOP prediction result, as shown in the following formula:

[0033] in, It is the orbital status at the next ephemeris predicted by HPOP;

[0034] The final loss function formula is:

[0035]

[0036] Wherein, the loss function of the present invention The weighted sum of the two types of errors mentioned above. and For the corresponding weights.

[0037] Preferably, the training network model in S6 includes dividing the training data into a training set, a validation set, and a test set according to a certain ratio. The training set is used to train the network model, the validation set is used to verify the training results of the model in each round during the training process, and the test set is used to evaluate the performance of the finally trained network model.

[0038] Preferably, the trained network model in step S6 further includes a normalization process, the specific formula of which is as follows:

[0039]

[0040] .

[0041] Preferably, the evaluation metrics for the trained network model in S6 are MSE (mean squared error), MAE (mean absolute error), MAPE (mean absolute percentage error), and R2 (coefficient of determination).

[0042] The beneficial effects of this invention are as follows:

[0043] (1) This invention directly learns high-dimensional, nonlinear orbital evolution mapping through deep neural networks, which can implicitly compensate for the perturbation forces that are not modeled or are not modeled accurately in traditional dynamic models, thereby obtaining higher prediction accuracy than traditional methods.

[0044] (2) When using this invention, only the position and velocity state at a single moment is required as input, without the need for historical observation sequences or TLE, making data acquisition convenient.

[0045] (3) By introducing HPOP orbit prediction as a constraint and weighted loss, this invention makes the model's prediction results not only conform to the statistical laws of data, but also conform to the basic physical principles, avoiding the physically unreasonable predictions that may be generated by the pure data-driven model, and significantly improving the reliability and safety of on-orbit application. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart from the present invention;

[0048] Figure 2 This is a schematic diagram of the long short-term memory network in this invention;

[0049] Figure 3 This is a schematic diagram of the one-dimensional convolutional layer operation process in this invention;

[0050] Figure 4 This is a schematic diagram of the maximum pooling operation process in this invention;

[0051] Figure 5 This is a schematic diagram of the average pooling operation process in this invention;

[0052] Figure 6 This is a diagram of the CBAM network structure in this invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] like Figure 1 As shown, this invention proposes a satellite orbit single-point state prediction method based on deep learning. It uses a deep learning network model to predict the orbit by utilizing the single-point state of the satellite orbit. Acquiring the dataset is relatively easy and will not be elaborated upon in detail here. The focus is on sample feature selection, network construction, and loss error design, as detailed below:

[0055] S1. Obtain precise orbital ephemeris

[0056] Precise ephemeris data of satellite orbits is used as training data, and the position and velocity of a single point in the satellite orbit are used to achieve the final mission. The ephemeris data includes the position and velocity of a single point in the satellite orbit at a certain moment, including the position, position, and velocity in the x-axis, y-axis, and z-axis directions.

[0057] S2. Adjust the training data and group the training data.

[0058] The training data is divided into an input group and a target group. The current time, position along the x-axis, position along the y-axis, position along the z-axis, velocity along the x-axis, velocity along the y-axis, and velocity along the z-axis are used as the input group. The next time, position along the x-axis, position along the y-axis, position along the z-axis, velocity along the x-axis, velocity along the y-axis, and velocity along the z-axis are used as the target group for model training.

[0059] S3. Constructing a network model

[0060] The network model uses LSTM (Long Short-Term Memory) as one of its basic units because orbital data itself is time-series data, and orbital prediction cannot be separated from the temporal characteristics of the data. Using this unit can capture long-term dependencies in the data and characterize the influence of various factors on the orbital state. However, the model of this invention differs from pure time-series prediction tasks. This invention does not require inputting a series of data over a period of time for prediction, but rather relies solely on the state of a single point to predict the future state of the satellite.

[0061] In addition, a one-dimensional CNN (Convolutional Neural Network) is used as another basic unit. One-dimensional convolution has the ability to extract local features. Since the perturbation factors affecting orbit prediction are complex and extremely difficult to simulate, it is necessary to automatically and efficiently extract multi-level, abstract local features from the data and transform them into higher-level, more compact feature representations, so that they can better capture the relationship between fluctuations in influencing factors and orbital state. Specifically:

[0062] 1) Long Short-Term Memory Network

[0063] LSTM is a special type of recurrent neural network (RNN) specifically designed to process and predict long-term dependencies in sequential data.

[0064] The overall operation of an RNN can be represented by the following formula:

[0065]

[0066]

[0067] in, Let the hidden state vector of the RNN be _____. and These are the input and output of the RNN, respectively. For the time step, there is also a shared weight matrix in RNN. In RNNs, the hidden state vector obtained at each time step contains information from the previous time step, so the output only needs the hidden state vector at the last time step as the final state. However, during training, RNNs use backpropagation to update the weights. When dealing with long sequences, the impact of errors on the weights needs to be backpropagated step by step along the time step. In this process, the gradient is continuously multiplied by the same weight matrix. If the eigenvalues ​​of the weight matrix are less than 1, the gradient will shrink exponentially until it becomes negligible, i.e., gradient vanishing, causing the model to fail to learn early information; if the eigenvalues ​​are greater than 1, the gradient will grow exponentially until it overflows, i.e., gradient explosion, which is the fundamental flaw of RNNs.

[0068] LSTM networks incorporate forget gates to handle the retention and removal of long-term information, thus resolving the problem of long-term data dependencies. For example... Figure 2 As shown, LSTM includes time... Memory cells State Hidden Vector ,enter Forgotten Gate Update door Output gate and sigmoid function .

[0069] In LSTM, the forget gate mimics the memory patterns of the human brain. When the human brain memorizes information from a long time ago, it selectively forgets some information. Similarly, the forget gate in LSTM decides to forget information that is less relevant when processing data. The calculation formula can be expressed as:

[0070]

[0071] in, , These represent the weights between the input and the forget gate, and between the hidden state from the previous time step and the forget gate, respectively. This is the bias vector.

[0072] like Figure 2 As shown, The update gate processes the current input information and extracts highly relevant information to update the current state. The calculation formula is:

[0073]

[0074]

[0075]

[0076] in, The cells are in an activated state. , , , These are all weights of the corresponding parts at the current moment. , This is the corresponding bias vector.

[0077] In addition, LSTM also includes output gates. The information integrated by the input gate and the forget gate, after passing through the output gate, can save all the relevant information of the current time step and past time steps, i.e., the hidden state vector. Therefore, like RNNs, LSTMs only need to output the final hidden state vector as the final vector. (Output gate) The calculation formula can be expressed as:

[0078]

[0079]

[0080] in, , For the corresponding weights, This is the bias vector.

[0081] 2) One-dimensional convolutional neural network

[0082] Classical convolutional neural networks typically use two-dimensional convolution to process image data, effectively capturing the spatial features of images. However, one-dimensional convolution is more suitable for processing one-dimensional signals or sequential data. One-dimensional convolution operations can effectively extract local features from temporal data, which is crucial for identifying short-term and long-term dependencies in a sequence. Furthermore, compared to the matrix convolution operations of two-dimensional convolution, one-dimensional convolution has lower computational complexity and fewer model parameters. Structurally, a one-dimensional convolutional network can be divided into five parts: input layer, convolutional layer, pooling layer, fully connected layer, and output layer.

[0083] ① Input Layer: The input layer is the first layer of a deep learning model, and its main function is to take the training data as input to the model. The dimension of the input layer corresponds to the number of features in the input data.

[0084] ② Convolutional Layer: The convolutional layer is the core layer of a CNN, extracting local and global features from the input data during training. It then performs convolution operations by setting the number of filters, kernel size, stride, and whether padding is used to extract feature vectors. The operation process of the convolutional layer is as follows: Figure 3 As shown, its convolution kernel is 3, stride is 1, and no padding is used.

[0085] ③ Pooling Layer: The pooling method of the pooling layer is similar to that of the convolution layer, performing pooling operations on the input matrix data through window movement. The main purpose of pooling is to downsample the input data, reducing the spatial dimensionality of the data after convolution, removing redundant information, and preventing the network from fitting unimportant information, thus preventing overfitting. This reduces the number of network parameters and the network's memory usage. Pooling is crucial for building deep networks. Currently, the most commonly used pooling methods are max pooling and average pooling, and the calculation process is as follows: Figure 4 and Figure 5 As shown below, the max pooling and average pooling methods in one-dimensional convolutional networks are expressed by formulas:

[0086]

[0087]

[0088] in, For pooling step size, The length of the pooling kernel, For sequence number, For the first The data matrix of the layer, The first pooled window represents the pooled window. Next move.

[0089] ④ Fully connected layer: The fully connected layer is located at the end of the CNN. By setting one or more fully connected layers, multiple linear transformations are performed on the flattened one-dimensional feature vector, so that the parameters are adjusted by weight and the features are combined, thereby establishing the intrinsic relationship between the output and the features, and finally outputting the prediction result.

[0090] ⑤ Output layer: The output layer is the last layer of the CNN. Its main function is to output the final prediction result. The design of the output layer should take into account various factors such as the characteristics of the problem, the attributes of the data, and the requirements of the task.

[0091] S4. Optimize the network model

[0092] During network optimization, CBAM (Convolutional Block Attention Module) was added to the network. CBAM is a convolutional block attention module designed to enhance the feature representation capability of convolutional neural networks by explicitly modeling attention in both channel and spatial dimensions. Its structure is as follows: Figure 6 As shown.

[0093] CBAM provides an intelligent weighting strategy for the network model. CBAM consists of two sub-modules: CA (Channel Attention) and SA (Spatial Attention). CA enables the network model to assign higher weights to perturbation features closely related to orbital state, while SA guides the network model to focus on temporal features that provide decisive information about changes in orbital state. This results in an orbital prediction model with local perception, intelligent weighting, and awareness of temporal fluctuations. Details are as follows:

[0094] CA (Capture Attention) is used to model the dependencies between channels, focusing attention on the key content of a given input. It performs global max pooling or average pooling on the feature maps of different channels to obtain max-pooled channel attention vectors and average channel attention vectors, both of which have a specific size. Then, these two vectors are input into an MLP that shares weights across only one hidden layer, where the shape of the hidden layer's weight vector is... This process yields two processed channel attention vectors. Finally, these two vectors are summed element-wise and processed with the Sigmoid activation function. The elements are then multiplied by the original feature map to obtain a new feature map.

[0095] The formula for calculating CA is as follows:

[0096]

[0097] in, This represents a 1D channel attention map. The sigmoid activation function is used, and MLP represents a multilayer perceptron. and This represents the weight values ​​of the two-layer MLP. and These represent the features after average pooling and max pooling along the channel axis, respectively. This indicates the percentage reduction in the middle channel.

[0098] SA is used to model dependencies between spatial locations. SA focuses on the location of key information in the input, performs max pooling and average pooling operations on the input feature map, compresses it at the channel level to obtain two feature maps, concatenates them according to the channel dimension to form a feature map with 2 channels, and then performs a convolution operation of a single convolutional layer to make the processed feature map consistent with the input in spatial dimension.

[0099] The formula for calculating SA is as follows:

[0100]

[0101]

[0102] in, Representing a spatial attention map, It is the Sigmoid activation function. Represents the convolution kernel. and These represent the features after average pooling and max pooling on the spatial axis, respectively.

[0103] S5. Design the loss function

[0104] For general time series forecasting or regression forecasting tasks, the loss functions used are related to mean squared error and mean squared error. However, if only mean squared error is used as the loss function, it will lead to divergent or unreliable results for long-term orbit forecasting tasks, which will limit its application in scenarios with high reliability requirements.

[0105] Therefore, this invention introduces the HPOP orbit prediction method into the loss function, adding a layer of physical constraint to the loss function. That is, the HPOP orbit prediction result and the error of the model prediction are weighted and summed with the mean square error loss function to form the final loss function. When the network weight parameters are continuously adjusted during training, a boundary can be added to the range of model prediction, enabling the network model to converge faster, while preventing the model from deviating from objective kinematic laws in order to achieve the minimum final loss.

[0106] The mean squared error (MSE) measures the sum of the squares of the differences between predicted and actual values. The formula is as follows:

[0107]

[0108] in, This is the actual value. It is a predicted value. This refers to the sample data volume. This invention predicts the position and velocity of a single point, with a total of 6 values. Therefore, the formula for calculating the mean squared error is:

[0109]

[0110] in, This is the actual value of the orbital state at the next ephemeris time. It is the predicted orbital state at the next ephemeris time.

[0111] In addition, the HPOP orbit prediction method is introduced. Specifically, HPOP is first used to predict the single-point state of the input satellite orbit at the same step size as the training data. The residual is calculated by comparing the predicted orbit state at the next ephemeris time with the HPOP prediction result. The specific formula is as follows:

[0112] in, This is the orbital status for the next ephemeris predicted by HPOP.

[0113] Finally, the loss function designed in this invention is:

[0114]

[0115] Wherein, the loss function of the present invention The weighted sum of the two types of errors mentioned above. and As corresponding weights, in this invention, The value is set to 0.7 because the purpose of this invention is to predict the orbital operating state more accurately and to pay more attention to the error between the model prediction results and the actual state. HPOP is added to make the model prediction more consistent with the laws of orbital dynamics, so the weight of the latter error is relatively small.

[0116] S6, Training the network model

[0117] 1) Network Description

[0118] For the training dataset, the orbital states at two consecutive time points need to be processed as input and output, with the state at the previous time point as the input and the state at the next time point as the output. Secondly, the training dataset is divided into a training set, a validation set, and a test set in a 7:1:2 ratio. The training set is used to train the network model, the validation set is used to validate the model's training results at each round, and the test set is used to evaluate the performance of the final trained model file. Furthermore, the aforementioned loss function is used during training.

[0119] 2) Normalization Processing Instructions

[0120] This invention preprocesses data according to its characteristics by performing vector modulo operations. This method further enhances the correlation between values ​​during data processing, namely:

[0121]

[0122]

[0123] 3) Explanation of evaluation indicators

[0124] The evaluation metrics used in this invention are "MSE (mean squared error)," "MAE (mean absolute error)," "MAPE (mean absolute percentage error)," and "R² (ratio of determination)." Testing can be performed using the test set partitioned during training, or evaluation can be conducted using other orbital data not used for training.

[0125] The obtained orbital precise ephemeris format is as follows:

[0126]

[0127] The ephemeris time step size needs to be consistent, and the unit of measurement for the position and velocity of a single point on the satellite also needs to be consistent, such as m, m / s or km, km / s. In the data shown above, when the ephemeris time is in UTC, the step size is 4 minutes, and the unit of the position and velocity of a single point is km and km / s.

[0128] The specific steps to transform the acquired raw data into training data are as follows: 1) First, separate two adjacent time points from the raw data; 2) Then, organize the data from the two adjacent time points into input and output, with the specific format as follows:

[0129]

[0130] The constructed neural network's input and output are both position and velocity information at the corresponding time. When the network inputs, the position and velocity are normalized, while the prediction results need to be denormalized to obtain the final output results. The network is trained according to this instruction, and a model file is formed after the network is trained.

[0131] When performing model inference, after loading the model file, the orbital state of a single point (position and velocity of the single point) is input, and the orbital state at the next moment can be predicted. The desired prediction period is set, and the final result can be predicted iteratively.

[0132] The prior description of the invention is provided to enable any person skilled in the art to make or use the invention. Various modifications to the invention will be apparent to those skilled in the art, and the general principles defined herein can be applied to other variations without departing from the spirit or scope of the invention. Thus, the invention is not intended to be limited to the examples and designs described herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0133] The embodiments of the present invention have been described in detail above. The description of the embodiments above is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A satellite orbit single-point state prediction method based on deep learning, characterized in that, Includes the following steps: S1. Obtain precise orbital ephemeris data and use the precise orbital ephemeris data as training data. The precise orbital ephemeris data includes the state of a single point in the satellite orbit at a certain moment. S2. Divide the training data into current time data and next time data, using the current time data as the input group and the next time data as the target value group; S3. Construct a network model, which includes LSTM (Long Short-Term Memory) and CNN (Convolutional Neural Network). The LSTM is connected to the CNN. The LSTM is used to capture long-term dependencies in the training data, and the CNN is used to extract local features from the training data. S4. Optimize the network model, which also includes a CBAM (Convolutional Block Attention Module). The CBAM is connected to the CNN and provides an intelligent weighting strategy for the network model. S5. Design the loss function by weighting and summing the orbit prediction results of HPOP (High-Precision Orbit Prediction Model) with the error of the model prediction and the mean square error loss function to form the loss function. S6. Train the network model by inputting the input group into the network model, training the network model and outputting predicted ephemeris data, and evaluating the predicted ephemeris data based on the data of the target value group.

2. The satellite orbit single-point state prediction method according to claim 1, characterized in that, The state of a single point in the satellite orbit at a certain moment includes the position of the single point in the satellite orbit along the x-axis, the position along the y-axis, the position along the z-axis, the velocity along the x-axis, the velocity along the y-axis, and the velocity along the z-axis at that moment.

3. The satellite orbit single-point state prediction method according to claim 1, characterized in that, The LSTM includes time. Memory cells State Hidden Vector ,enter Forgotten Gate Update door Output gate and sigmoid function The forget gate The calculation formula is: in, , These represent the weights between the input and the forget gate, and between the hidden state from the previous time step and the forget gate, respectively. This is the bias vector. The update gate The calculation formula is: in, , , , These are all weights of the corresponding parts at the current moment. , This is the corresponding bias vector. The output gate The calculation formula is: in, , For the corresponding weights, This is the bias vector.

4. The satellite orbit single-point state prediction method according to claim 1, characterized in that, The CNN uses a one-dimensional convolutional neural network.

5. The satellite orbit single-point state prediction method according to claim 1, characterized in that, The CNN includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer.

6. The satellite orbit single-point state prediction method according to claim 1, characterized in that, The CBAM includes CA (Channel Attention) and SA (Spatial Attention). CA is used to give higher weights to perturbation features related to orbital state in the network model, and SA is used to guide the network model to focus on temporal features that have decisive information about changes in orbital state.

7. The satellite orbit single-point state prediction method according to claim 1, characterized in that, The loss function includes the measured mean squared error of the model and the predicted mean squared error of the model, wherein the formula for calculating the measured mean squared error of the model is: in, This is the actual value of the orbital state at the next ephemeris time. It is the predicted orbital state at the next ephemeris time; The model prediction mean square error is first predicted using HPOP to forecast the single-point state of the input satellite orbit with the same step size as the training data. The residual is then calculated by comparing the predicted orbit state value at the next ephemeris time with the HPOP prediction result, as shown in the following formula: in, It is the orbital status at the next ephemeris predicted by HPOP; The final loss function formula is: Wherein, the loss function of the present invention The weighted sum of the two types of errors mentioned above. and For the corresponding weights.

8. The satellite orbit single-point state prediction method according to claim 1, characterized in that, S6 The training network model includes dividing the training data into a training set, a validation set, and a test set according to a certain ratio. The training set is used to train the network model, the validation set is used to verify the training results of the model in each round during the training process, and the test set is used to evaluate the performance of the finally trained network model.

9. The satellite orbit single-point state prediction method according to claim 1, characterized in that, The training network model described in S6 also includes a normalization process, the specific formula of which is as follows: 。 10. The satellite orbit single-point state prediction method according to claim 1, characterized in that, The evaluation metrics for the trained network model described in S6 are MSE (mean squared error), MAE (mean absolute error), MAPE (mean absolute percentage error), and R2 (coefficient of determination).

Images