Satellite orbit prediction method based on attention mechanism long-short term memory network

Abstract

The invention discloses a satellite orbit prediction method based on an attention mechanism long-short term memory network, and relates to the technical field of satellite orbit prediction. The method comprises the following steps: preprocessing broadcast ephemeris data and precise ephemeris data, subtracting the two processed data to obtain a correction number sequence, and normalizing the correction number sequence to obtain standardized data; inputting the standardized data into a satellite orbit prediction model to obtain a multi-step correction number prediction result; and performing broadcast ephemeris correction according to a correction number prediction result to obtain final predicted orbit information. Wherein the satellite orbit prediction model adopts an Attention-LSTM (Long Short Term Memory) model, and captures an implicit relationship among input characteristics and a non-linear uncertain time trend of a time sequence for a plurality of fully connected layers of a periodic term. According to the method, the real-time orbit data is calculated by using the precise ephemeris data and the broadcast ephemeris, and high-precision and real-time prediction of the satellite orbit is realized according to the satellite orbit prediction model.

Description

Technical Field

[0001] This invention relates to the field of satellite orbit prediction technology, and in particular to a satellite orbit prediction method based on attention mechanism long short-term memory networks. Background Technology

[0002] Satellite orbits refer to the paths along which artificial satellites orbit the Earth, and are fundamental to space applications such as satellite navigation, remote sensing, and communication. Predicting satellite orbits is crucial for satellite attitude control, orbit maintenance, collision warning, and improving navigation and positioning accuracy. Satellite orbits are influenced by a variety of complex factors, including the non-uniformity of the Earth's gravitational field, atmospheric drag, solar radiation pressure, the geomagnetic field, and the gravitational pull of other celestial bodies, causing orbital parameters to change dynamically over time. Therefore, orbit prediction has become a critical and challenging technical task in space missions.

[0003] The core objective of satellite orbit prediction is to accurately calculate a satellite's position and velocity over a future period based on existing orbital data and dynamic models. Traditional orbit prediction methods primarily rely on physical models and numerical integration techniques to simulate orbital evolution by precisely calculating various disturbances experienced by the satellite. While these methods have a solid theoretical foundation, they are highly dependent on the accuracy of environmental models and initial conditions, and their computational complexity is high, making it difficult to meet the demands of real-time and large-scale orbit prediction.

[0004] With the development of global satellite navigation systems and space monitoring networks, acquiring high-precision satellite orbit observation data has become possible. In particular, the fusion of multi-source observation data (such as laser ranging, radio frequency measurements, and onboard sensor data) provides rich information for orbit prediction. Orbit determination and prediction methods based on observation data have gradually become a research hotspot. Typical methods include extended Kalman filter (EKF), unscented Kalman filter (UKF), and particle filter (PF) algorithms. These methods can dynamically correct orbital parameters and improve prediction accuracy. However, the evolution of satellite orbits exhibits highly nonlinear and complex time-varying characteristics and is significantly affected by space environment disturbances. Traditional filtering and physical modeling methods have limited performance in the face of sudden disturbances and model uncertainties. Summary of the Invention

[0005] The purpose of this invention is to provide a satellite orbit prediction method based on attention mechanism long short-term memory networks, aiming to solve or improve at least one of the above-mentioned technical problems.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A satellite orbit prediction method based on attention mechanism long short-term memory networks includes:

[0008] Preprocessing of broadcast ephemeris data and precise ephemeris data yields processed broadcast ephemeris orbital positions and precise ephemeris orbital positions;

[0009] The difference between the two processed data is used to obtain the corrected number sequence;

[0010] The corrected number sequence is normalized to obtain standardized data;

[0011] A satellite orbit prediction model is constructed, and the standardized data is input into the satellite orbit prediction model for processing to obtain multi-step corrected prediction results; wherein, the satellite orbit prediction model adopts the Attention-LSTM model, and captures the implicit relationship between input features and the nonlinear uncertain time trend of the time series for multiple fully connected layers of the periodic term;

[0012] Based on the predicted corrections, broadcast ephemeris corrections are performed to obtain the final predicted orbit information.

[0013] Optionally, the preprocessing of the broadcast ephemeris data and precise ephemeris data to obtain the processed broadcast ephemeris orbital positions and precise ephemeris orbital positions specifically includes:

[0014] Acquire precise ephemeris data and perform compliance processing on the precise ephemeris data;

[0015] The compliant data of each satellite is processed into a coordinate file separately. The broadcast ephemeris data for the corresponding number of days is downloaded. The satellite's geocentric and geofixed coordinate system position is calculated based on the broadcast ephemeris data, and antenna phase center correction is performed. The data is then converted into SP3 format containing satellite position information and timestamps.

[0016] Optionally, the compliance processing includes outlier detection, missing value imputation, and stationarity testing, used to remove outliers in the orbital observation data and ensure data continuity.

[0017] Optionally, the sequence of correction numbers is specifically represented as:

[0018] δ(t)=r brd (t)-r prec (t)

[0019] Where δ(t) represents the three-dimensional position residual vector of the correction number sequence, r brd (t) represents the orbital position of the broadcast ephemeris, r prec (t) represents the precise ephemeris orbital position.

[0020] Optionally, the normalization process is specifically expressed as:

[0021]

[0022] Among them, z i It is the normalized value, x i σ represents the original data points, μ is the mean of the data, and σ is the standard deviation of the data.

[0023] Optionally, the Attention-LSTM model consists of an input layer, an LSTM layer, an Attention layer, and an output layer; wherein the LSTM layer is a two-layer LSTM.

[0024] Optionally, the dual-layer LSTM uses Dropout of 0.2 to reduce overfitting, and the hidden units are set to 256 layers. The input and output are the correction numbers for the first day and the correction numbers for the second day, respectively. The output layer periodically constructs the corresponding fully connected layers according to the data, and outputs the results of each fully connected layer in sequence. Finally, they are combined to form a multi-step correction number prediction result.

[0025] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0026] This invention discloses a satellite orbit prediction method based on an attention-based long short-term memory network. The method includes preprocessing broadcast ephemeris data and precise ephemeris data; subtracting the processed data to obtain a correction sequence; normalizing this sequence to obtain standardized data; inputting the standardized data into a satellite orbit prediction model to obtain multi-step correction prediction results; and correcting the broadcast ephemeris data based on the correction prediction results to obtain the final predicted orbit information. The satellite orbit prediction model employs an Attention-LSTM model and uses multiple fully connected layers in the periodic term to capture the implicit relationships between input features and the nonlinear, uncertain time trend of the time series. This invention utilizes precise ephemeris data and real-time orbit data calculated from broadcast ephemeris data, and achieves high-precision, real-time prediction of satellite orbits based on the satellite orbit prediction model. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the 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.

[0028] Figure 1 This is a flowchart of the satellite orbit prediction method in this embodiment;

[0029] Figure 2This is a diagram of the Attention-LSTM model structure in this embodiment;

[0030] Figure 3 This is a diagram of the LSTM neuron structure in this embodiment. Detailed Implementation

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0032] The purpose of this invention is to provide a satellite orbit prediction method based on attention mechanism long short-term memory networks, aiming to solve or improve at least one of the above-mentioned technical problems.

[0033] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] like Figure 1-Figure 3 As shown, this invention provides a satellite orbit prediction method based on attention mechanism long short-term memory networks, including:

[0035] The broadcast ephemeris data and the precise ephemeris data are preprocessed to obtain the processed broadcast ephemeris orbital positions and precise ephemeris orbital positions.

[0036] The difference between the two processed data is used to obtain the corrected number sequence.

[0037] The corrected number sequence is normalized to obtain standardized data.

[0038] A satellite orbit prediction model is constructed, and the standardized data is input into the satellite orbit prediction model for processing to obtain multi-step corrected prediction results. The satellite orbit prediction model adopts the Attention-LSTM model and captures the implicit relationships between input features and the nonlinear uncertain time trend of the time series for multiple fully connected layers of the periodic term.

[0039] Based on the predicted corrections, broadcast ephemeris corrections are performed to obtain the final predicted orbit information.

[0040] As one specific implementation method, the above steps will be explained.

[0041] 1. Data sources for deep learning models

[0042] For satellite position information, various analytical institutions worldwide possess precise post-hoc position information, known as precise ephemeris, which offers centimeter-level positioning accuracy and high temporal continuity. This invention uses precise ephemeris as historical satellite orbit observation data, undergoing outlier detection, missing value imputation, and stationarity checks to ensure data quality and the predictability of the time series. Broadcast ephemeris data is released in real-time by the satellite navigation system. For each satellite, its broadcast ephemeris orbit is calculated and differiated from the corresponding actual satellite orbit observation data to obtain a correction sequence. This correction sequence reflects the deviation between the broadcast ephemeris orbit and the actual orbital state, serving as a crucial input for subsequent model training and prediction.

[0043] 2. Data preprocessing and model building

[0044] Step 1: Data Preprocessing

[0045] After obtaining the detailed post-hoc ephemeris, the data for each satellite is processed into a separate coordinate file. The broadcast ephemeris for the corresponding number of days is downloaded. The satellite's geocentric and Earth-fixed coordinate system position is calculated from the broadcast ephemeris, and antenna phase center correction is performed. The data is then converted into SP3 format, containing satellite position information and a timestamp. The processed broadcast ephemeris orbital position r is then determined. brd (t) and precise ephemeris position r prec By subtracting (t) from the original value, we obtain the correction sequence δ(t) = r. brd (t)-r prec δ(t) is the three-dimensional position residual vector. Since the value range and physical dimensions of different input data are inconsistent, the data feature 1 needs to be normalized before being input into the model.

[0046]

[0047] Among them, z i It is the normalized value, x i σ represents the original data points, μ is the mean of the data, and σ is the standard deviation of the data.

[0048] Step 2: Use the Attention-LSTM model for data prediction

[0049] This invention employs a prediction method based on the Attention-LSTM model, which consists of four main parts: an input layer, an LSTM layer, an Attention layer, and an output layer. Historical correction values ​​are input into the model, and after prediction, the resulting corrected prediction value is added back to the satellite position calculated from the corresponding broadcast ephemeris to compensate for the error. LSTM (Long Short-Term Memory) is an improvement on traditional Recurrent Neural Networks (RNNs), aiming to address the short-term memory limitations of RNNs during forward propagation and the gradient vanishing and gradient exploding problems that may occur during backward propagation. LSTM significantly alleviates these two problems by introducing a gating mechanism. An LSTM neural unit includes three key parts: an input gate, a forget gate, and an output gate. The structure of LSTM and its related calculation formulas are as follows (Formulas (2) to (7)).

[0050] i t =σ(W i ·[h t-1 ,x t ]+b i (2)

[0051]

[0052] f t =σ(W f ·[h t-1 ,x t ]+b f (4)

[0053]

[0054] h t =o t ·tanh(C t (6)

[0055] o t =σ(W o ·[h t-1 ,x t ]+b o (7)

[0056] Equations (2) and (3) indicate that the input gate determines the current input x. t The impact on the current state. The input gate consists of two parts: a gating operation i through the sigmoid function. t The other is candidate state C. t , C t Indicates new candidate memory content, i tIt is the control signal of the input gate; Formula (4) indicates that the forget gate determines how much of the previous memory to be forgotten, and the output f of the forget gate is... t W is calculated using a sigmoid activation function. f This is the weight matrix of the forget gate, b f σ is the bias term, and σ represents the sigmoid activation function. Equation (5) indicates that the cell state update is achieved by combining the forget gate and the input gate. t C represents the current cell state. t-1 Given the cell state at the previous time step, equations (6) and (7) represent the output gates, which are based on the current cell state C. t and the current input x t This determines the current output. The output gate calculates the output control signal o through the sigmoid activation function. t It determines which cell state information will be passed to the output, h t This represents the hidden state at the current moment, containing current memory information, and serves as the input or final output for the next moment. t The output of the output gate controls the current hidden state h. t That is, the final output.

[0057] The main role of the Attention layer is to further enhance the model's ability to focus on important time steps based on the LSTM network. While the LSTM model can capture long-term dependency information, it may still lose information when processing long sequences of data. The Attention mechanism assigns different weights to different time steps in the input sequence, allowing the model to "focus" on different information at each time step, thus better learning the features of important moments. In the output of the LSTM model, the hidden state at each time step contains a portion of the information from the input sequence, and the Attention layer determines which time steps are more important to the final output by calculating the attention weights for each hidden state. Specifically, the Attention mechanism is implemented by the following three formulas:

[0058] score t =V T tanh(Wh t +Uh t-1 +b) (8)

[0059]

[0060] Formula (8) is used to calculate the attention score, where W, U, V are the learned weight matrices, and ht This is the output of the LSTM at time t, b is the bias term, and the attention weights are calculated using formula (9), where α t Let represent the attention weight at time t, which is normalized by softmax. Finally, the final attention output is obtained by weighted summation using formula (10).

[0061] To adapt to the dataset structure and reduce overfitting, this invention also uses a two-layer LSTM with a Dropout of 0.2 to reduce overfitting. The hidden units are 256 layers, and the input and output are the correction numbers for the first day and the correction numbers for the second day, respectively. The output layer constructs corresponding fully connected layers according to the periodicity of the data, and outputs the results of each fully connected layer in sequence. Finally, they are combined to form a multi-step prediction result.

[0062] Model prediction result evaluation:

[0063] The prediction results are inversely normalized to obtain the original data dimensions. The data is then corrected to the broadcast ephemeris orbit to calculate the predicted satellite position. The prediction results are typically evaluated using the average error (bias), mean squared error (MSE), and root mean squared error (RMSE) between the predicted and actual values.

[0064]

[0065] Where N represents the total number of samples, indicating the number of comparisons between the predicted and actual locations. i represents the sample index, ranging from i = 1, 2, ..., N represents the i-th sample. This represents the predicted position of the i-th sample, an estimate given by the model. This represents the true position of the i-th sample, derived from actual data.

[0066] Therefore, the present invention has the following beneficial effects:

[0067] 1. Improve orbit prediction accuracy: By differentially correcting the broadcast ephemeris and the actual orbit data, and combining the efficient prediction of the correction number by the deep learning model, the orbit prediction error is reduced, thereby improving the accuracy of satellite orbit prediction.

[0068] 2. Make full use of historical orbit data and broadcast ephemeris information, and adopt a method that combines broadcast ephemeris preprocessing with corrected number sequence prediction. Make full use of existing broadcast ephemeris information, and at the same time, dynamically compensate for its errors to achieve efficient fusion of orbit predictions.

[0069] 3. Enhance the model's ability to capture time series nonlinearity and uncertainty. The deep learning model, which combines the Attention mechanism with LSTM, can automatically identify key temporal dependencies and implicit relationships in the input features, effectively capture the complex nonlinear dynamic changes of orbital data, and improve the robustness and stability of predictions.

[0070] 4. Data preprocessing ensures the quality of model input. Through multiple steps of data preprocessing, such as outlier detection, missing value imputation, and stationarity testing, the integrity and temporal characteristics of the input data are guaranteed, providing a high-quality data foundation for model training and prediction.

[0071] 5. The model has a simple and easily extensible structure. It adopts the Attention-LSTM model structure, which avoids the complexity of multi-model fusion, facilitates model deployment and maintenance, and has good extensibility, which can easily integrate more track-related features to further improve performance.

[0072] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0073] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A satellite orbit prediction method based on a long short-term memory network with an attention mechanism, characterized in that, include: Preprocessing of broadcast ephemeris data and precise ephemeris data yields processed broadcast ephemeris orbital positions and precise ephemeris orbital positions; The difference between the two processed data is used to obtain the corrected number sequence; The corrected number sequence is normalized to obtain standardized data; A satellite orbit prediction model is constructed, and the standardized data is input into the satellite orbit prediction model for processing to obtain multi-step corrected prediction results; wherein, the satellite orbit prediction model adopts the Attention-LSTM model, and captures the implicit relationship between input features and the nonlinear uncertain time trend of the time series for multiple fully connected layers of the periodic term; Based on the predicted corrections, broadcast ephemeris corrections are performed to obtain the final predicted orbit information.

2. The satellite orbit prediction method based on attention mechanism long short-term memory network according to claim 1, characterized in that, The preprocessing of broadcast ephemeris data and precise ephemeris data to obtain processed broadcast ephemeris orbital positions and precise ephemeris orbital positions specifically includes: Acquire precise ephemeris data and perform compliance processing on the precise ephemeris data; The compliant data of each satellite is processed into a coordinate file separately. The broadcast ephemeris data for the corresponding number of days is downloaded. The satellite's geocentric and geofixed coordinate system position is calculated based on the broadcast ephemeris data, and antenna phase center correction is performed. The data is then converted into SP3 format containing satellite position information and timestamps.

3. The satellite orbit prediction method based on attention mechanism long short-term memory network according to claim 2, characterized in that, The compliance processing includes outlier detection, missing value imputation, and stationarity testing, which are used to remove outliers in orbital observation data and ensure data continuity.

4. The satellite orbit prediction method based on attention mechanism long short-term memory network according to claim 1, characterized in that, The sequence of correction numbers is specifically represented as follows: δ(t)=r brd (t)-r prec (t) Where δ(t) represents the three-dimensional position residual vector of the correction number sequence, r brd (t) represents the orbital position of the broadcast ephemeris, r prec (t) represents the precise ephemeris orbital position.

5. The satellite orbit prediction method based on attention mechanism long short-term memory network according to claim 1, characterized in that, The normalization process is specifically expressed as follows: Among them, z i It is the normalized value, x i σ represents the original data points, μ is the mean of the data, and σ is the standard deviation of the data.

6. The satellite orbit prediction method based on attention mechanism long short-term memory network according to claim 1, characterized in that, The Attention-LSTM model consists of an input layer, an LSTM layer, an Attention layer, and an output layer; wherein, the LSTM layer adopts a two-layer LSTM.

7. The satellite orbit prediction method based on attention mechanism long short-term memory network according to claim 6, characterized in that, The dual-layer LSTM uses Dropout of 0.2 to reduce overfitting, with 256 hidden units. The input and output are the correction numbers for the first day and the second day, respectively. The output layer periodically constructs corresponding fully connected layers based on the data, and outputs the results of each fully connected layer in sequence. Finally, they are combined to form a multi-step correction number prediction result.

Images