Method for predicting orbital information of a space object and electronic device

Abstract

Embodiments of the present application provide a method for predicting orbit information of a space object and an electronic device. The method comprises: obtaining orbit error time series data between actual orbit information of the space object and predicted orbit information based on a dynamic model, wherein the orbit error time series data comprises orbit errors at a plurality of continuous historical time points, and the orbit errors comprise velocity errors and position errors. The orbit error time series data is input into an orbit error prediction model, and orbit errors at a plurality of future time points are output, the orbit error prediction model is a deep learning model based on an attention mechanism, and orbit information of the space object at the plurality of future time points predicted by the dynamic model is corrected according to the orbit errors at the plurality of future time points, to obtain corresponding orbit information at the plurality of future time points. The present application improves the accuracy of predicting orbit information of a space object.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a method and electronic device for predicting the orbital information of a space target. Background Technology

[0002] With the rapid development of artificial intelligence technology, applying it to the prediction of orbital information of space targets is of great significance.

[0003] In related technologies, the prediction of orbital information for space targets is typically achieved using two dynamic model-based methods: numerical methods and analytical methods. Specifically, numerical methods take the initial orbital state of the space target as input and predict the orbital information at any future time by gradually integrating the derivatives of the state changes. Analytical methods derive analytical expressions for the space target's trajectory based on the main perturbations affecting orbital motion and directly solve them for orbit prediction. Among these, the Simplified General Perturbations, version 4 (SGP4) is a commonly used analytical orbital prediction model. It predicts the future orbital information of space targets by inputting two-line elements (TLEs) into SGP4.

[0004] However, the above methods are inefficient at predicting the orbital information of space targets, or the predicted orbital information contains large errors. Summary of the Invention

[0005] This application provides a method and electronic device for predicting the orbital information of space targets, which can improve the efficiency and accuracy of orbital information prediction for space targets.

[0006] In a first aspect, embodiments of this application provide a method for predicting the orbital information of a space target, including:

[0007] The orbital error time series data between the actual orbital information of a space target and the predicted orbital information based on a dynamic model is obtained. The orbital error time series data includes orbital errors at multiple consecutive historical moments, and the orbital errors include velocity errors and position errors.

[0008] The orbital error time series data is input into the orbital error prediction model, and the predicted orbital errors at multiple future times are output. The orbital error prediction model is a deep learning model based on an attention mechanism.

[0009] Based on the orbital errors at the multiple future time points, the orbital information of the space target predicted by the dynamic model at the multiple future time points is corrected to obtain the corresponding orbital information at the multiple future time points.

[0010] In one possible implementation, the attention-based deep learning model includes: a first encoder based on a self-attention mechanism and a first decoder based on a temporal attention mechanism, wherein the first encoder based on the self-attention mechanism includes a first self-attention layer and a first target encoder, and the first decoder based on the temporal attention mechanism includes: a first temporal attention layer and a first target decoder.

[0011] The step of inputting the orbital error time series data into the orbital error prediction model and outputting the predicted orbital errors at multiple future times includes:

[0012] The orbital error time series data is input into the first self-attention layer, and the self-attention weight is calculated on the orbital error time series data through the first self-attention layer to determine the first orbital error time series data with added self-attention weight;

[0013] The first target encoder encodes the first orbital error time series data with added self-attention weights to obtain the first hidden state output vector at each historical moment.

[0014] The first time attention layer is used to calculate the time attention weight based on the first hidden state output vector at each historical moment to determine the first weighted hidden state output vector at multiple future moments.

[0015] The second hidden state output vector at multiple future moments is determined by the first target decoder and the first weighted hidden state output vector based on the multiple future moments.

[0016] A linear transformation is performed on the output vector of the second hidden state at the multiple future time points to output the predicted orbital errors at the multiple future time points.

[0017] In one possible implementation, the step of calculating self-attention weights on the orbital error time series data through the first self-attention layer to determine the first orbital error time series data with added self-attention weights includes:

[0018] The first self-attention layer is used to calculate the self-attention weight of the orbital error time series data to determine the first orbital error weight time series data of the orbital error time series data.

[0019] The first self-attention layer merges the first orbital error weight time series data with the orbital error time series data to determine the first orbital error time series data with added self-attention weights.

[0020] In one possible implementation, the step of calculating temporal attention weights through the first temporal attention layer and based on the first hidden state output vector at each historical moment to determine the first weighted hidden state output vector at multiple future moments includes:

[0021] Perform the following operations to determine the first weighted hidden state output vector at the t′-th future time, until t′ equals T′, where T′ is the last future time. The following operations include:

[0022] Based on the first weight score calculation algorithm preset in the first time attention layer, the weight scores of the first hidden state output vectors at each historical time corresponding to the t′-1 future time are determined according to the second hidden state output vector at the t′-1 future time. Among them, when the t′-1 future time is the initial future time, the second hidden state output vector at the t′-1 future time is the first hidden state output vector at the last historical time.

[0023] Based on the first activation function preset in the first time attention layer and the weight scores of the first hidden state output vector at each corresponding historical moment, the weights of the first hidden state output vector at each corresponding historical moment are determined.

[0024] Based on the first time attention layer, the first hidden state output vector at each historical time corresponding to the t′ future time and the weights corresponding to each first hidden state output vector are weighted and summed to obtain the first weighted hidden state output vector at the t′ future time.

[0025] In one possible implementation, determining the second hidden state output vector at multiple future time points using the first decoder and the first weighted hidden state output vector at the multiple future time points includes:

[0026] Perform the following operations to determine the second hidden state output vector at the t′-th future time step, until t′ equals T′, where T′ is the last future time step. The following operations include:

[0027] The second hidden state output vector at the t′-1 future time is obtained based on the second hidden state output vector at the t′-1 future time and the first weighted hidden state output vector at the t′-1 future time.

[0028] In one possible implementation, the attention-based deep learning model includes: a second encoder and a second decoder based on a self-attention mechanism, wherein the second encoder based on the self-attention mechanism includes a second self-attention layer and a second target encoder;

[0029] The step of inputting the orbital error time series data into the orbital error prediction model and outputting the predicted orbital errors at multiple future times includes:

[0030] The orbital error time series data is input into the second self-attention layer, and the self-attention weight is calculated on the orbital error time series data through the second self-attention layer to determine the second orbital error time series data with added self-attention weight;

[0031] The second target encoder encodes the second orbital error time series data with added self-attention weights to obtain the third hidden state output vector at the last historical moment.

[0032] The fourth hidden state output vector for multiple future moments is determined by the second decoder and based on the third hidden state output vector at the last historical moment.

[0033] A linear transformation is performed on the fourth hidden state output vector at the multiple future time points to output the predicted orbital errors at the multiple future time points.

[0034] In one possible implementation, the step of calculating self-attention weights on the orbital error time series data through the second self-attention layer to determine the second orbital error time series data with added self-attention weights includes:

[0035] The second self-attention layer is used to calculate the self-attention weight of the orbital error time series data to determine the second orbital error weight time series data of the orbital error time series data.

[0036] The second self-attention layer merges the second orbital error weight time series data with the orbital error time series data to determine the second orbital error time series data with added self-attention weights.

[0037] In one possible implementation, the attention-based deep learning model includes: a third encoder and a third decoder based on a time attention mechanism, wherein the third decoder based on the time attention mechanism includes: a second time attention layer and a third target decoder.

[0038] The step of inputting the orbital error time series data into the orbital error prediction model and outputting the predicted orbital errors at multiple future times includes:

[0039] The orbital error time series data is input into the third encoder and encoded by the third encoder to obtain the fifth hidden state output vector at each historical moment.

[0040] The second time attention layer is used to calculate the time attention weight based on the fifth hidden state output vector at each historical moment to determine the second weighted hidden state output vector at multiple future moments.

[0041] The sixth hidden state output vector at multiple future times is determined by the third decoder and the second weighted hidden state output vector based on the multiple future times.

[0042] A linear transformation is performed on the output vector of the sixth hidden state at the multiple future time points to obtain the predicted orbital error time series data at the multiple future time points.

[0043] In one possible implementation, the step of calculating time attention weights through the second time attention layer and based on the fifth hidden state output vector at each historical moment to determine the second weighted hidden state output vector at multiple future moments includes:

[0044] Perform the following operations to determine the second weighted hidden state output vector at the t′-th future time, until t′ equals T′, where T′ is the last future time. The following operations include:

[0045] Based on the second weight score calculation algorithm preset by the second time attention layer, the weight score of the fifth hidden state output vector at each historical time corresponding to the t′-1 future time is determined according to the sixth hidden state output vector at the t′-1 future time. Among them, when the t′-1 future time is the initial future time, the sixth hidden state output vector at the t′-1 future time is the fifth hidden state output vector at the last historical time.

[0046] Based on the second activation function preset in the second time attention layer and the weight score of the fifth hidden state output vector at each corresponding historical moment, the weight of the fifth hidden state output vector at each corresponding historical moment is determined.

[0047] Based on the second time attention layer, the fifth hidden state output vector at each historical time corresponding to the t′ future time and the weights corresponding to each fifth hidden state output vector are weighted and summed to obtain the second weighted hidden state output vector at the t′ future time.

[0048] In one possible implementation, determining the sixth hidden state output vector at multiple future time points using the third decoder and the second weighted hidden state output vector at the multiple future time points includes:

[0049] Perform the following operations to determine the output vector of the sixth hidden state at the t′-th future time, until t′ equals T′, where T′ is the last future time. The following operations include:

[0050] The sixth hidden state output vector at the t′-1 future time is obtained based on the sixth hidden state output vector at the t′-1 future time and the second weighted hidden state output vector at the t′-1 future time.

[0051] In one possible implementation, the training steps of the orbital error prediction model include:

[0052] Acquire training data, which includes a first orbital error time series sample and a second orbital error time series sample, wherein the second orbital error time series sample is a label, and the first orbital error time series sample is the time series sample preceding the second orbital error time series sample.

[0053] The initial orbital error prediction model is trained using the training data until a preset convergence condition is met, resulting in a trained orbital error prediction model. The initial orbital error prediction model is an attention-based initial deep learning model.

[0054] Secondly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0055] The memory stores computer-executed instructions;

[0056] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0057] Thirdly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0058] Fourthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0059] The orbital information prediction method and electronic device for space targets provided in this application acquire time-series data of orbital errors between the actual orbital information of the space target and the predicted orbital information based on a dynamic model. The orbital error time-series data includes orbital errors at multiple consecutive historical moments, and these errors include velocity and position errors. The orbital error time-series data is input into an orbital error prediction model, which outputs predicted orbital errors at multiple future moments. This orbital error prediction model is a deep learning model based on an attention mechanism. Based on the orbital errors at multiple future moments, the orbital information of the space target predicted by the dynamic model at multiple future moments is corrected to obtain the corresponding orbital information at multiple future moments. Attached Figure Description

[0060] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0061] Figure 1 A schematic diagram of a single-step prediction in an exemplary prior art provided for this application;

[0062] Figure 2 A flowchart illustrating a method for predicting the orbital information of a space target, provided in an embodiment of this application;

[0063] Figure 3 A schematic diagram of a multi-step direct prediction provided in an embodiment of this application;

[0064] Figure 4 A flowchart illustrating a method for predicting orbital errors at multiple future moments, provided in an embodiment of this application;

[0065] Figure 5 An architecture diagram of an attention-based deep learning model provided for an embodiment of this application;

[0066] Figure 6 A flowchart illustrating another method for predicting orbital errors at multiple future moments, provided in an embodiment of this application;

[0067] Figure 7 A flowchart illustrating another method for predicting orbital errors at multiple future moments, provided in an embodiment of this application;

[0068] Figure 8 A flowchart illustrating a method for training a trajectory error prediction model provided in an embodiment of this application;

[0069] Figure 9 A schematic diagram of a training process provided in an embodiment of this application;

[0070] Figure 10 A schematic diagram of a training process provided in an embodiment of this application;

[0071] Figure 11 A comparison diagram of predicted orbital error value and actual orbital error value in the x-direction is provided for an embodiment of this application;

[0072] Figure 12 A comparison diagram of predicted orbital error value and actual orbital error value in the y-direction is provided for an embodiment of this application;

[0073] Figure 13 A comparison diagram of predicted orbital error value and actual orbital error value in the z-direction is provided for an embodiment of this application;

[0074] Figure 14 A schematic diagram of the structure of a space target orbit information prediction device provided in an embodiment of this application;

[0075] Figure 15 A schematic diagram of the structure of the electronic device provided in this application.

[0076] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0077] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0078] In the description of the embodiments of this application, the terms "inner" and "outer", etc., which indicate the direction or positional relationship, are based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or component must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this application.

[0079] In the description of the embodiments of this application, unless otherwise expressly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of this application according to the specific circumstances.

[0080] Below, we will first explain the terms used in this application:

[0081] Two-line orbital elements (TLE): TLE is a standard data format used to describe the position and velocity information of satellites or space debris in Earth's orbit. This data format is widely used in the aerospace industry, especially in areas such as space target orbit calculation, prediction, astronomical observation, and space exploration. Taking satellites as an example, in TLE, the first line typically contains information such as the satellite's serial number, launch year, and orbit type. The second line contains the satellite's orbital parameters, such as the semi-major axis, eccentricity, orbital inclination, argument of perigee, right ascension of the ascending node, mean perigee, and the first and second derivatives of the mean motion. These parameters collectively define the satellite's orbital state, enabling the calculation of the satellite's orbital information at any future time using appropriate orbit prediction models. The orbital information includes both position and velocity information.

[0082] Simplified General Perturbation Model, Version 4 (SGP4), also known as the Simplified Conventional Perturbation Model, is a simplified model for calculating the orbits of space targets. It is widely used in orbit prediction and navigation applications for low Earth orbit (LEO) space targets. SGP4 considers factors such as Earth's gravity, atmospheric drag, and the gravitational pull of the Sun and Moon, and can successfully predict the orbital periods of LEO space targets with periods less than 225 minutes. The SGP4 calculation process mainly includes the following steps: First, the Time Exit (TLE) of the space target is obtained and analyzed to obtain six orbital parameters: semi-major axis, eccentricity, orbital inclination, right ascension of the ascending node, argument of perigee, and mean perigee. These orbital parameters are then input into SGP4 for calculation, ultimately yielding the orbital information of the space target at any given time.

[0083] To facilitate understanding of this application, the prior art that may be involved is described below by way of example.

[0084] In related technologies, the prediction of orbital information for space targets is typically performed using two dynamic model-based methods: numerical methods and analytical methods. Numerical methods take the initial orbital state of the space target as input and predict the orbital information at any future time by gradually integrating the derivatives of the state changes. However, using numerical methods for high-precision orbital prediction not only relies on constructing detailed models of the space target's geometric characteristics, Earth's gravity, lunar gravity, atmospheric drag, and solar radiation pressure, but also incurs high computational and time costs, resulting in low efficiency and a poor user experience in predicting space target orbital information.

[0085] Analytical methods derive analytical expressions for the trajectories of space targets based on the main perturbations affecting orbital motion, and directly solve them for orbit prediction. Taking the SGP4 analytical method as an example, the Time-Like Equation (TLE) is input into SGP4, and SGP4 outputs the predicted orbital information of the space target at any future time. The advantage of analytical methods lies in their high computational efficiency and low resource consumption. However, due to the complexity of orbital motion equations with various perturbations, the output of SGP4 is often an approximate solution, resulting in poor prediction accuracy. Currently, to improve the accuracy of predicted orbital information, deep learning-based orbital error prediction models are commonly used to correct the predicted orbital information output by traditional dynamic models. Specifically, the predicted orbital information output by the dynamic model, such as SGP4, is subtracted from the actual orbital information to obtain the orbital error. Then, the orbital error is input into the orbital error prediction model to calculate the error compensation for future time moments, obtaining the orbital error compensation value. Finally, the orbital error compensation value is added to the orbital information predicted by the dynamic model for that future time moment to obtain the error-compensated orbital information.

[0086] Specifically, the deep learning-based trajectory error prediction model can be an encoder-decoder model based on a recurrent neural network, where the recurrent neural network includes a long short-term memory network and a gated recurrent unit, or it can be an encoder-decoder model based on a convolutional neural network, etc.

[0087] Taking the encoder-decoder model based on a recurrent neural network as an example, assuming that SGP4 has predicted the orbital information at the first, second, and third time points, and based on the actual orbital information corresponding to the orbital information at the first, second, and third time points, three corresponding orbital errors are obtained. These three orbital errors are input into the encoder-decoder model, which then predicts the orbital errors at the fourth and fifth time points. The orbital error at the fourth time point is then added to the orbital information predicted by SGP4 at the fourth time point to obtain the predicted orbital information at the fourth time point. Similarly, the orbital error at the fifth time point is added to the orbital information predicted by SGP4 at the fifth time point to obtain the predicted orbital information at the fifth time point.

[0088] However, in the above methods, the encoder-decoder model typically employs a multi-step recursive prediction method when predicting orbital errors at multiple time points; that is, a simple single-step prediction method (prediction step size set to 1). Specifically, it iterates through subsequent multi-step predictions using the predicted value from the previous step as input, building upon the single-step prediction. Please refer to [link to relevant documentation]. Figure 1 , Figure 1 This application provides an exemplary prior art diagram illustrating single-step prediction. In this example, a sliding window is set with a size of 5, meaning the encoder-decoder model input contains orbital errors at five time points: orbital error P_1 at the first time point, orbital error P_2 at the second time point, ..., orbital error P_5 at the fifth time point. During the first prediction, the window contains all predicted historical orbital errors {P_1, P_2, ..., P_5}. These five time-point data are input into the encoder-decoder model, which outputs the predicted value for the next time point, i.e., the orbital error P_6 at the sixth time point. To predict the orbital error at the seventh time point, the orbital error at the first time point (P_1) is discarded, and the orbital error output by the previous model (P_6) is added to the window, with P_6 serving as the last time point data in the window. Then, the data {P_2, ..., P_6} within the window is input into the encoder-decoder model to obtain the orbital error P_7 at the seventh time point. If the forecast is for a longer time step, the window continues to slide, and the above operation is repeated until the forecast is complete.

[0089] Therefore, it is evident that multi-step recursive prediction methods suffer from error propagation, which causes the predicted orbital error to increase significantly with the increase of prediction time, resulting in a decrease in the accuracy of the predicted orbital error as the prediction time increases. Furthermore, for multi-step prediction, being able to predict the orbital error at only one future moment at a time also leads to a decrease in prediction efficiency, which in turn reduces the efficiency of obtaining the predicted space target orbital information based on the predicted orbital error.

[0090] Furthermore, since orbital errors include multiple data points—positional errors and velocity errors in different directions—the aforementioned encoder-decoder model cannot effectively select relevant data for prediction, neglecting the correlation between multiple data points, resulting in poor accuracy of the predicted orbital errors. Moreover, the performance of the aforementioned encoder-decoder model deteriorates sharply with the length of the input sequence, especially in time series analysis. This is because encoder-decoder models typically require predictions based on relatively long output and input sequences; however, the aforementioned encoder-decoder model cannot accurately capture the long-term temporal dependencies between long input and output sequences, causing the model to fail to fully consider the contextual information of the entire sequence during prediction, thus resulting in low accuracy of the predicted orbital errors.

[0091] Therefore, addressing the aforementioned technical problems in the prior art, the inventors discovered during their research that existing encoder-decoder models, i.e., orbital error prediction models, only predict the value at one future moment at a time. To predict multiple future moments, this step-by-step prediction process needs to be repeated multiple times, resulting in slow prediction efficiency. Therefore, the inventors found that predicting orbital error values ​​at multiple future moments at once can effectively improve prediction efficiency. Furthermore, because existing orbital error prediction models cannot effectively select relevant data from the multiple data points included in the orbital error, nor can they capture the long-term temporal dependencies between long input and long output sequences, the accuracy of the predicted orbital errors is low. Therefore, the inventors discovered that an orbital error prediction model based on an attention mechanism can effectively solve the above problems. Specifically, by acquiring the orbital error time-series data between the actual orbital information of the space target and the predicted orbital information based on the dynamic model, and inputting the orbital error time-series data into the orbital error prediction model, the model outputs the predicted orbital errors at multiple future moments. This orbital error prediction model is a deep learning model based on an attention mechanism. Then, based on the orbital errors at multiple future moments, the orbital information of the space target predicted by the dynamic model at multiple future moments is corrected to obtain the corresponding orbital information at multiple future moments. By simultaneously predicting the orbital information of a space target at multiple future moments, the efficiency of orbital information prediction for space targets is improved. Furthermore, by employing a deep learning model based on an attention mechanism to obtain the correlation between multiple data points involved in the orbital error and / or the long-term temporal dependence between long input sequences and long output sequences, the accuracy of orbital information prediction for space targets is improved. Based on these findings, this application proposes a method and electronic device for predicting the orbital information of space targets.

[0092] The application scenario of the method in this application can be an orbital information prediction device for a space target. This device can receive an activation command sent by a user through a client and begin periodic predictions based on the command. By executing the orbital information prediction method of this application, it obtains the predicted orbital information of the space target. After obtaining the predicted orbital information, it can analyze the possibility of collisions between satellites, either independently or through interaction with other devices, and monitor the actual launch orbit of the satellite to analyze its deviation trajectory. It is understood that the above scenarios are for illustrative purposes only and do not impose limitations on this application.

[0093] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0094] Figure 2 A flowchart illustrating a method for predicting the orbital information of a space target, as provided in an embodiment of this application, is shown below. Figure 2 As shown, the executing entity of this method can be a space target orbit information prediction device, which can be implemented by a computer program; it can also be implemented by a medium storing the relevant computer program, such as a USB flash drive and / or optical disc, or it can be implemented by a physical device integrating or installing the relevant computer program, such as a chip or electronic device, which can be a server, server cluster, smart terminal, etc. The method includes:

[0095] S201. Obtain time-series data of orbital error between the actual orbital information of the space target and the predicted orbital information based on the dynamic model.

[0096] Optionally, space targets include, but are not limited to, artificial satellites, space stations, spacecraft, and space debris.

[0097] Track error time series data refers to a data format in which track errors are arranged sequentially over time within a certain period of time.

[0098] When acquiring time-series data of orbital errors, it can be obtained by accessing a database that stores orbital errors at consecutive different times. The orbital errors in the database can be obtained in the following way:

[0099] The system acquires predicted orbit information at consecutive different time points from a dynamic model, such as SGP4, while simultaneously collecting actual orbit information at the corresponding time points. For any given time point, the difference between the predicted orbit information output by SGP4 and the actual orbit information at that time point is calculated to obtain the orbit error at that time. Thus, the database can store orbit errors at consecutive different time points.

[0100] In this embodiment, the orbital error time series data includes orbital errors at multiple consecutive historical moments. The orbital errors include position errors and velocity errors. The position errors include errors in three preset directions: x-axis, y-axis, and z-axis. The velocity errors also include errors in three preset directions: x-axis, y-axis, and z-axis.

[0101] For example, suppose the acquired orbital error time series data of the space target includes the orbital error at the first time, the second time, the third time, and the fourth time. What needs to be predicted is the orbital error at the fifth time and the sixth time in the future. Then the orbital error at the first time, the second time, the third time, and the fourth time are the orbital errors at consecutive historical time points.

[0102] S202. Input the orbital error time series data into the orbital error prediction model and output the predicted orbital error at multiple future times.

[0103] In this embodiment, after obtaining the orbital error time series data through step S201, the orbital error time series data is normalized to unify it to the same scale. A preset sliding window algorithm is used to set the window size of the normalized orbital error time series data, converting it into a dataset usable by the orbital error prediction model. This yields orbital error time series data that can be directly input into the orbital error prediction model, where the orbital error prediction model is a pre-trained model.

[0104] For example, when setting the window, the sliding window size can be set to T, i.e., the input sequence is T. The input sequence can be the orbital errors at T historical time points, and the prediction step size is T′, i.e., the output is the orbital errors at T′ future time points. The orbital errors at T historical time points, i.e., the orbital error time series data, can be represented as x=(x1,x2,···x T ), where x∈R T ×K K is the number of features in the input sequence. Let y represent the orbital error at time t, and let T′ represent the predicted orbital error at future times (y).T+1 ,y T+2 ,···y T+T′ ).

[0105] In this embodiment, the orbital error prediction model can be a deep learning model based on an attention mechanism. The deep learning model can be an encoder-decoder model composed of a long short-term memory network.

[0106] Optionally, the deep learning model based on the attention mechanism may include a first encoder based on the self-attention mechanism and a first decoder based on the temporal attention mechanism.

[0107] Optionally, the deep learning model based on the attention mechanism can also include a second encoder and a second decoder based on the self-attention mechanism.

[0108] Optionally, the deep learning model based on the attention mechanism can also include a third encoder and a third decoder based on the temporal attention mechanism.

[0109] The orbital error time series data includes multiple data points, including position errors and velocity errors in different directions. The encoder based on the self-attention mechanism can calculate the similarity between each data point in the orbital error time series data and assign different weights to each data point to extract the correlation between multiple data points.

[0110] Decoders based on time attention mechanisms can assign different importance to each moment and allocate different weights to each moment in order to capture the long-term temporal dependency between the input orbital error time series data at historical moments and the output orbital error time series data at future moments.

[0111] The orbital error time series data is input into any of the above orbital error prediction models to obtain the predicted orbital errors at multiple future times.

[0112] In this embodiment, a multi-step direct prediction method is adopted, which can obtain the orbital error at multiple future moments in a single prediction.

[0113] For a schematic diagram of multi-step direct prediction, please refer to [link / reference]. Figure 3 , Figure 3 This illustration shows a multi-step direct prediction method provided in an embodiment of this application. Assume the sliding window size is 5, meaning the input to the orbital error prediction model contains orbital errors at five time points. Specifically, the input orbital error time-series data includes the orbital error P_1 at the first time point, P_2 at the second time point, ..., P_5 at the fifth time point. The orbital error prediction model directly predicts the orbital errors P_6 at the sixth time point, P_7 at the seventh time point, and P_8 at the eighth time point in a single prediction.

[0114] S203. Based on the orbital errors at multiple future time points, correct the orbital information of the space target predicted by the dynamic model at multiple future time points to obtain the corresponding orbital information at multiple future time points.

[0115] One possible implementation is:

[0116] Obtain the orbital information of the space target at multiple future moments, such as the orbital information predicted by the SGP4 dynamic model. For any future moment, add the orbital information predicted by the SGP4 at that moment to the orbital error predicted by the orbital error prediction model at that future moment to obtain the orbital information of the space target at that future moment, and thus obtain the orbital information of the space target at multiple future moments.

[0117] In the above embodiments of this application, orbital error time-series data between the actual orbital information of a space target and the predicted orbital information based on a dynamic model is obtained. This orbital error time-series data is then input into an orbital error prediction model, which outputs predicted orbital errors at multiple future times. Based on these orbital errors, the orbital information of the space target predicted by the dynamic model at multiple future times is corrected to obtain the corresponding orbital information at multiple future times. The method of this embodiment improves the orbital information prediction efficiency of space targets because it can simultaneously predict the orbital information of a space target at multiple future times. Furthermore, because the orbital error prediction model introduces an attention mechanism, it can extract the correlation between each data point included in the orbital error time-series data and / or capture the long-term temporal dependency between the input historical orbital error time-series data and the output future orbital error time-series data, thereby improving the accuracy of space target orbital information prediction.

[0118] Furthermore, based on the above embodiments, the process of inputting orbital error time series data into the orbital error prediction model and outputting the predicted orbital errors at multiple future moments is described when the attention-based deep learning model includes a first encoder based on a self-attention mechanism and a first decoder based on a temporal attention mechanism.

[0119] Please see Figure 4 , Figure 4 A flowchart illustrating a method for predicting orbital errors at multiple future time points, provided in an embodiment of this application, includes the following steps:

[0120] S401. Input the orbital error time series data into the first self-attention layer, and calculate the self-attention weight of the orbital error time series data through the first self-attention layer to determine the first orbital error time series data with added self-attention weight.

[0121] Figure 5 An architecture diagram of an attention-based deep learning model provided for embodiments of this application is shown below. Figure 5 As shown, the following is based on Figure 5 right Figure 4 The proposed solution will be described.

[0122] In this embodiment, the first encoder based on the self-attention mechanism includes a first self-attention layer and a first target encoder, and the first decoder based on the temporal attention mechanism includes a first temporal attention layer and a first target decoder.

[0123] Optionally, the first target encoder is an LSTM-based encoder, and the first target decoder is an LSTM-based decoder.

[0124] The first self-attention layer can dynamically capture the complex dependencies in the time series data of orbital errors corresponding to multiple historical moments in the input sequence, without relying on external information. In addition, the computation process of the first self-attention layer is mainly based on matrix operations, which is fast and can be parallelized.

[0125] In this embodiment, a self-attention layer is used to emphasize the correlation between multiple data points included in the input sequence, i.e., the orbital error time series data, so that the model can better capture important data in the data, thereby improving the prediction accuracy performance of the orbital error prediction model.

[0126] One possible implementation is:

[0127] The first orbital error weight time series data is determined by calculating the self-attention weight of the orbital error time series data through the first self-attention layer.

[0128] The first orbital error weight time series data is obtained by calculating the self-attention weight based on the first self-attention layer.

[0129] Specifically, based on a preset first weight calculation algorithm, according to the orbital error time series data x=(x1,x2,···x T ), determine the first query vector q, the first key vector k, and the first value vector v.

[0130] The formulas for determining the first query vector q, the first key vector k, and the first value vector v are shown in formulas (1)-(3) respectively:

[0131] q = W q x (1)

[0132] k = W k x (2)

[0133] v = Wv x (3)

[0134] Among them, W q W represents the first learnable weight parameter; k W represents the second learnable weight parameter; v This represents the third learnable weight parameter.

[0135] Based on the first query vector and the first key vector, determine the attention weight 'a' for the orbital error time series data. self .

[0136] The correlation between the first query vector q and the first key vector k is calculated by performing a dot product, thus obtaining the degree of correlation between the data included in the orbital error time series data. A larger dot product value indicates a higher similarity between the corresponding data. To ensure the attention matrix approximately follows a standard normal distribution, the Softmax activation function is used to normalize the dot product result, resulting in better numerical stability. Furthermore, to ensure better gradient balance during backpropagation, the dot product result of q and k is divided by an adaptively set parameter. Therefore, the attention weight is calculated as shown in formula (4):

[0137]

[0138] Among them, a self Indicates attention weight; This indicates that the parameters are set adaptively.

[0139] Attention weights reflect the importance of data included in orbital error time series data and the correlation between the included data.

[0140] Based on the attention weights and first value vectors of the orbital error time series data, the first orbital error weight time series data of the orbital error time series data is determined according to the following formula (5).

[0141] x att =a self v (5)

[0142] Where, x att This represents the time series data of the first orbital error weight.

[0143] The first self-attention layer merges the first orbital error weight time series data with the orbital error time series data to determine the first orbital error time series data with added self-attention weights.

[0144] x through the first self-attention layer att Combined with the original input data x, to determine the first orbital error time series data with added self-attention weights.

[0145] S402, The first target encoder encodes the first orbital error time series data with added self-attention weights to obtain the first hidden state output vector at each historical moment.

[0146] The first target encoder uses the orbital error timing data with added self-attention weights. Reading sequentially in time order, the hidden state output vector of the LSTM in the first target encoder changes accordingly, where the hidden state output vector h at the current time t... t From input and the hidden state output vector h from the previous time step t-1 Generation. After reading is complete, the first target encoder generates an encoding vector, thereby obtaining the first hidden state output vector at each historical time point.

[0147] S403. Calculate the time attention weights through the first time attention layer and based on the first hidden state output vector at each historical moment to determine the first weighted hidden state output vector at multiple future moments.

[0148] In this embodiment, a time attention layer is used to dynamically calculate the attention weight of each time step, assigning different importance or attention to different time steps. This enables the model to effectively capture the long-term temporal dependency between the input historical orbital error time series data and the output future orbital error time series data, thereby improving the prediction accuracy performance of the orbital error prediction model.

[0149] Perform the following operations to determine the first weighted hidden state output vector at the t′-th future time, until t′ equals T′, where T′ is the last future time. One possible implementation is:

[0150] Based on the first weight score calculation algorithm preset in the first time attention layer, the weight scores of the first hidden state output vectors at each historical time corresponding to the t′ future time are determined according to the second hidden state output vector at the t′-1 future time.

[0151] Wherein, when the t′-th future time is the initial future time, the second hidden state output vector at the t′-1-th future time is the first hidden state output vector at the last historical time.

[0152] Specifically, the weight scores of the first hidden state output vector at each historical time corresponding to the t′ future time are calculated using the following formula (6).

[0153]

[0154] Where, h′ t′-1 h represents the output vector of the second hidden state at the (t′-1)th future time step in the first target decoder, which is the input at the t′th future time step; t e represents the output vector of the first hidden state at the t-th historical moment in the first target encoder; t′t V represents the weight score of the first hidden state output vector at the t-th historical time corresponding to the t′-th future time; a W represents the first learnable weight; a This represents the second learnable weight.

[0155] Based on the first activation function preset in the first time attention layer and the weight scores of the first hidden state output vector at each historical time, the weights of the first hidden state output vector at each historical time are determined.

[0156] The weights 'a' of the first hidden state output vector at each historical time step are obtained based on the first activation function Softmax preset in the first-time attention layer. t′t Therefore, at the t′ future time, for

[0157] The first hidden state output vector {h1,h2,…,h} corresponding to all historical moments in the first target encoder is used to output the first hidden state vector. t ,…,h T The assigned attention weights can be represented as a. t′ ={a t′1 ,a t′2 ,…,a t′t ,…,a t′T}

[0158] The first time attention layer performs a weighted summation of the first hidden state output vector at each historical time corresponding to the t′ future time and the weights corresponding to each first hidden state output vector to obtain the first weighted hidden state output vector at the t′ future time.

[0159] Based on the first-time attention layer, the output vector {h1,h2,…,h} of the first hidden state at each historical time corresponding to the future time t′ is obtained. t ,…,h T}, and the weights a corresponding to the output vectors of each first hidden state. t′ ={a t′1 ,a t′2 ,…,a t′t ,…,a t′T We perform a weighted summation, specifically, we obtain the first weighted hidden state output vector at the t′ future time step through the following formula (7).

[0160]

[0161] Among them, S t′ Let represent the first weighted hidden state output vector at the t′-th future time.

[0162] S404. Determine the second hidden state output vector at multiple future time points using the first target decoder and the first weighted hidden state output vector based on multiple future time points.

[0163] Specifically, the following operations are performed to determine the second hidden state output vector at the t′-th future time step, until t′ equals T′, where T′ is the last future time step. These operations include:

[0164] Based on the second hidden state output vector at the (t′-1)th future time and the first weighted hidden state output vector at the t′th future time, the second hidden state output vector at the t′th future time is obtained.

[0165] The first weighted hidden state output vector S at the t′ future time step t′ The output vector h′ of the second hidden state of the first target decoder at the (t′-1)th future time. t′-1 As input at the t′-th future time step, we obtain the second hidden state output vector h at the t′-th future time step. ′t .

[0166] S405. Perform a linear transformation on the output vector of the second hidden state at multiple future time points to output the predicted orbital error at multiple future time points.

[0167] The output vector of the second hidden state at each future time step of the first target decoder is linearly transformed through a fully connected layer (i.e., a linear layer) to generate the corresponding output sequence. The output sequence is the predicted orbital error (y) at multiple future time steps. T+1 ,y T+2 ,···y T+T′ ).

[0168] In the above embodiments of this application, orbital error time-series data is input into a first self-attention layer, and self-attention weights are calculated on the orbital error time-series data through the first self-attention layer to determine the first orbital error time-series data with added self-attention weights. The first target encoder encodes the orbital error time-series data with added self-attention weights to obtain the first hidden state output vector at each historical time. A first temporal attention layer calculates temporal attention weights based on the first hidden state output vector at each historical time to determine the first weighted hidden state output vector at multiple future time. A first target decoder determines the second hidden state output vector at multiple future time based on the first weighted hidden state output vector at multiple future time. A linear transformation is performed on the second hidden state output vector at multiple future time to output the predicted orbital error at multiple future time. The method in this embodiment calculates self-attention weights on the orbital error time series data through a first self-attention layer, which facilitates the encoder to obtain the correlation between multiple data included in the orbital error. It calculates temporal attention weights through a first temporal attention layer based on the output vector of the first hidden state at each historical moment, which facilitates the decoder to obtain the long-term temporal dependency between the long input sequence and the long output sequence. The calculation is based on the correlation between multiple data included in the orbital error and the long-term temporal dependency between the long input sequence and the long output sequence, making the predicted orbital error at future moments more accurate. As a result, the predicted orbital information of the space target obtained based on the more accurate orbital error is more accurate.

[0169] Furthermore, based on the above embodiments, the process of inputting orbital error time series data into the orbital error prediction model and outputting the predicted orbital errors at multiple future moments is described when the attention-based deep learning model includes a second encoder and a second decoder based on a self-attention mechanism.

[0170] Please see Figure 6 , Figure 6 A flowchart illustrating another method for predicting orbital errors at multiple future time points, provided as an embodiment of this application, is shown. The method includes the following steps:

[0171] S601. Input the orbital error time series data into the second self-attention layer, and calculate the self-attention weight of the orbital error time series data through the second self-attention layer to determine the second orbital error time series data with added self-attention weight.

[0172] In this embodiment, the second encoder based on the self-attention mechanism includes: a second self-attention layer and a second target encoder.

[0173] Optionally, the second target encoder is an LSTM-based encoder, and the second decoder is an LSTM-based decoder.

[0174] The second self-attention layer can dynamically capture the complex dependencies in the time series data of orbital errors corresponding to multiple historical moments in the input sequence, without relying on external information. In addition, the computation process of the second self-attention layer is mainly based on matrix operations, which is fast and can be parallelized.

[0175] In this embodiment, a self-attention layer is used to emphasize the correlation between multiple data points included in the input sequence, i.e., the orbital error time series data, so that the model can better capture important data in the data, thereby improving the prediction accuracy performance of the orbital error prediction model.

[0176] In this embodiment, the steps and functions implemented by the second self-attention layer are the same as those implemented by the first self-attention layer.

[0177] One possible implementation is:

[0178] The second self-attention layer is used to calculate the self-attention weights of the orbital error time series data to determine the second orbital error weight time series data of the orbital error time series data.

[0179] The second orbital error weight time series data is obtained by calculating the self-attention weight based on the second self-attention layer.

[0180] Specifically, based on a preset second weight calculation algorithm, according to the orbital error time series data x=(x1,x2,···x T ), determine the second query vector q, the second key vector k, and the second value vector v.

[0181] The formulas for determining the second query vector q, the second key vector k, and the second value vector v are shown in formulas (8)-(10), respectively:

[0182] q = W q x (8)

[0183] k = W k x (9)

[0184] v = W v x (10)

[0185] Among them, W q W represents the first learnable weight parameter; k W represents the second learnable weight parameter; v This represents the third learnable weight parameter.

[0186] Based on the second query vector and the second key vector, determine the attention weight 'a' for the orbital error time series data.self .

[0187] The dot product of the second query vector q and the second key vector k is calculated to obtain the correlation between q and k, i.e., the degree of correlation between the data included in the orbital error time series data. A larger dot product value indicates a higher similarity between the corresponding two data points. To make the attention matrix approximately follow a standard normal distribution, the Softmax activation function is used to normalize the dot product result, resulting in better numerical stability. Furthermore, to ensure better gradient balance during backpropagation, the dot product result of q and k is divided by an adaptively set parameter. Therefore, the attention weight is calculated as shown in formula (11):

[0188]

[0189] Among them, a self Indicates attention weight; This indicates that the parameters are set adaptively.

[0190] Attention weights reflect the importance of data included in orbital error time series data and the correlation between the included data.

[0191] Based on the attention weights and second value vectors of the orbital error time series data, the second orbital error weight time series data of the orbital error time series data is determined according to the following formula (12).

[0192] x att =a self v (12)

[0193] Where, x att This represents the time series data of the second orbital error weights.

[0194] The second self-attention layer merges the second orbital error weight time series data with the orbital error time series data to determine the second orbital error time series data with added self-attention weights.

[0195] x through the second self-attention layer att Combined with the original input data x, to determine the second orbital error time series data with added self-attention weights.

[0196] S602. The second target encoder encodes the second orbital error time series data with added self-attention weights to obtain the third hidden state output vector at the last historical moment.

[0197] The first target encoder uses the orbital error timing data with added self-attention weights. Reading sequentially in time order, the hidden state output vector of the LSTM in the first target encoder changes accordingly, where the hidden state output vector h at the current time t... t From input and the hidden state output vector h from the previous time step t-1 Generation. After reading is complete, the first target encoder generates an encoded vector, thus obtaining the first hidden state output vector at the last historical moment.

[0198] S603. Using the second decoder and based on the third hidden state output vector at the last historical moment, determine the fourth hidden state output vector at multiple future moments.

[0199] The third hidden state output vector at the last historical moment is used as the hidden state output vector of the first future moment predicted in the second decoder for iterative calculation. When the calculation reaches the t′-th future moment, the fourth hidden state output vector h′ at the t′-1-th future moment is used. t′-1 As input at the t′-th future time step, we obtain the fourth hidden state output vector h′ at the t′-th future time step. t′ This allows us to determine the fourth hidden state output vector at multiple future time points.

[0200] S604. Perform a linear transformation on the fourth hidden state output vector at multiple future time points to output the predicted orbital error at multiple future time points.

[0201] The output vector of the fourth hidden state at each future time step of the second decoder is linearly transformed through a fully connected layer (i.e., a linear layer) to generate the corresponding output sequence. The output sequence is the predicted orbital error (y) at multiple future time steps. T+1 ,y T+2 ,···y T+T′ ).

[0202] In the above embodiments of this application, orbital error time-series data is input into a second self-attention layer, and self-attention weights are calculated on the orbital error time-series data through the second self-attention layer to determine second orbital error time-series data with added self-attention weights. A second target encoder encodes the second orbital error time-series data with added self-attention weights to obtain a third hidden state output vector at the last historical moment. A second decoder, based on the third hidden state output vector at the last historical moment, determines fourth hidden state output vectors at multiple future moments. A linear transformation is performed on the fourth hidden state output vectors at multiple future moments to output the predicted orbital errors at multiple future moments. The method of this embodiment, by calculating self-attention weights on the orbital error time-series data through the second self-attention layer, facilitates the encoder in obtaining the correlation between the multiple data points included in the orbital error. Calculations based on the correlation between the multiple data points included in the orbital error make the predicted orbital errors at future moments more accurate, and thus the predicted orbital information of the space target obtained based on the more accurate orbital errors is more accurate.

[0203] Furthermore, based on the above embodiments, the process of inputting orbital error time series data into the orbital error prediction model and outputting the predicted orbital errors at multiple future moments is described when the attention-based deep learning model includes a third encoder and a third decoder based on a time attention mechanism.

[0204] Please see Figure 7 , Figure 7 A flowchart illustrating another method for predicting orbital errors at multiple future time points, provided as an embodiment of this application, is shown. This method includes the following steps:

[0205] S701. Input the orbital error time series data into the third encoder and encode the orbital error time series data through the third encoder to obtain the fifth hidden state output vector at each historical moment.

[0206] In this embodiment, the third decoder based on the time attention mechanism includes: a second time attention layer and a third target decoder.

[0207] Optionally, the third encoder is an LSTM-based encoder, and the third target decoder is an LSTM-based decoder.

[0208] The trajectory error time series data is input into the third encoder, and iterative calculation is performed through the LSTM in the third encoder. When the calculation reaches the t-th historical time, the fifth hidden state output vector h at the (t-1)-th historical time is generated. t-1 As input at the t-th historical moment, we obtain the fifth hidden state output vector h at the t-th historical moment. tThis allows us to determine the fifth hidden state output vector at multiple historical moments.

[0209] S702. The second time attention layer is used to calculate the time attention weight based on the fifth hidden state output vector at each historical moment, so as to determine the second weighted hidden state output vector at multiple future moments.

[0210] In this embodiment, a time attention layer is used to dynamically calculate the attention weight of each time step, assigning different importance or attention to different time steps. This enables the model to effectively capture the long-term temporal dependency between the input historical orbital error time series data and the output future orbital error time series data, thereby improving the prediction accuracy performance of the orbital error prediction model.

[0211] In this embodiment, the steps and functions implemented by the second time attention layer are the same as those implemented by the first time attention layer.

[0212] Perform the following operations to determine the second weighted hidden state output vector at the t′-th future time step, until t′ equals T′, where T′ is the last future time step. One possible implementation is:

[0213] Based on the second weight score calculation algorithm preset by the second time attention layer, the weight score of the fifth hidden state output vector at each historical time corresponding to the t′ future time is determined according to the sixth hidden state output vector at the t′-1 future time.

[0214] Wherein, when the t′-th future time is the initial future time, the output vector of the sixth hidden state at the t′-1-th future time is the output vector of the fifth hidden state at the last historical time.

[0215] Specifically, the weight scores of the fifth hidden state output vector at each historical time corresponding to the t′ future time are calculated using the following formula (13).

[0216]

[0217] Where, h′ t′-1 Let h represent the output vector of the sixth hidden state at the (t′-1)th future time step in the third target decoder, which is the input at the t′th future time step; t e represents the output vector of the fifth hidden state at the t-th historical moment in the third encoder; t′t V represents the weight score of the fifth hidden state output vector at the t-th historical time corresponding to the t′-th future time; a W represents the first learnable weight; a This represents the second learnable weight.

[0218] Based on the second activation function preset in the second time attention layer and the weight scores of the fifth hidden state output vector at each historical moment, the weights of the fifth hidden state output vector at each historical moment are determined.

[0219] The weights 'a' of the fifth hidden state output vector at each historical time step are obtained based on the second activation function Softmax preset by the second time attention layer. ′ Therefore, at the t′ future time, for

[0220] tt

[0221] The fifth hidden state output vector {h1,h2,…,h} corresponding to all historical moments in the third encoder. t ,…,h T The assigned attention weights can be represented as a. t′ ={a t′1 ,a t′2 ,…,a t′t ,…,a t′T}

[0222] The second time attention layer performs a weighted summation of the fifth hidden state output vector at each historical time corresponding to the t′ future time and the weights corresponding to each fifth hidden state output vector to obtain the second weighted hidden state output vector at the t′ future time.

[0223] Based on the second time attention layer, the fifth hidden state output vector {h1,h2,…,h} is generated for each historical time corresponding to the future time t′. t ,…,h T}, and the weights a corresponding to the output vectors of each fifth hidden state. t′ ={a t′1 ,a t′2 ,…,a t′t ,…,a t′T We perform a weighted summation, specifically, we obtain the second weighted hidden state output vector at the t′ future time using the following formula (14).

[0224]

[0225] Among them, S t′ Let represent the output vector of the second weighted hidden state at the t′ future time.

[0226] S703. Determine the sixth hidden state output vector at multiple future time points using the third target decoder and the second weighted hidden state output vector based on multiple future time points.

[0227] Specifically, perform the following operations to determine the output vector of the sixth hidden state at the t′-th future time step, until t′ equals K, where K is the last future time step, including the following operations:

[0228] Based on the output vector of the sixth hidden state at the (t′-1)th future time and the output vector of the second weighted hidden state at the t′th future time, the output vector of the sixth hidden state at the t′th future time is obtained.

[0229] The output vector of the second weighted hidden state at the t′-th future time is S. t′ And the output vector h′ of the sixth hidden state of the third target decoder at the (t′-1)th future time. t′-1 As input at the t′-th future time step, we obtain the sixth hidden state output vector h′ at the t′-th future time step. t′ .

[0230] S704. Perform a linear transformation on the output vector of the sixth hidden state at multiple future time points to output the predicted orbital error at multiple future time points.

[0231] The output vector of the sixth hidden state at each future time step of the third target decoder is linearly transformed through a fully connected layer (i.e., a linear layer) to generate the corresponding output sequence. The output sequence is the predicted orbital error (y) at multiple future time steps. T+1 ,y T+2 ,···y T+T′ ).

[0232] In the above embodiments of this application, orbital error time-series data is input to a third encoder and encoded to obtain a fifth hidden state output vector at each historical moment. A second time attention layer is used to calculate time attention weights based on the fifth hidden state output vectors at each historical moment to determine a second weighted hidden state output vector at multiple future moments. A third target decoder is then used to determine a sixth hidden state output vector at multiple future moments based on the second weighted hidden state output vectors at multiple future moments. A linear transformation is performed on the sixth hidden state output vectors at multiple future moments to output the predicted orbital errors at multiple future moments. The method in this embodiment, by using a second time attention layer and calculating time attention weights based on the fifth hidden state output vectors at each historical moment, facilitates the decoder in obtaining the long-term temporal dependency between the long input sequence and the long output sequence. Calculations based on this long-term temporal dependency make the predicted orbital errors at future moments more accurate, resulting in more accurate predicted orbital information for the space target.

[0233] The orbital error prediction model in this application can be trained using the following method; please refer to [link / reference]. Figure 8 , Figure 8 This application provides a flowchart illustrating a method for training an orbital error prediction model, which includes the following steps:

[0234] S801, Obtain training data.

[0235] Please see Figure 9 , Figure 9 This is a schematic diagram of a training process provided in an embodiment of the present application, which includes data acquisition, data preprocessing, model training, and model testing.

[0236] Specifically, data acquisition:

[0237] In this embodiment, the training data includes orbital error time series samples, which can be obtained in the following manner.

[0238] Collect historical TLE data from satellites, such as collecting two rows of orbital elements for a low-Earth orbit satellite over approximately 7 days, setting the sampling interval to 240 seconds, and recording a set of observations every 4 minutes.

[0239] The predicted orbit information for each observation moment within the observation period is obtained using a dynamic model such as SGP4. This predicted orbit information includes the satellite's predicted position vector (X′, Y′, Z′) and velocity vector (V′) in the x, y, and z directions. x ′,V y ′,V z ′).

[0240] The system acquires precise ephemeris data of the satellite within the observation period. This precise ephemeris data, with an accuracy down to the meter level, provides accurate satellite position information. Therefore, the actual satellite orbit value can be obtained from the precise ephemeris.

[0241] Optionally, space environment data can also be obtained, from which the geomagnetic KP index (KP), planetary equivalent amplitude (AP), international sunspot number (ISN), observed solar 10.7 cm radio current (F10.7_0BS), and the average solar 10.7 cm radio current over the past 81 days (F10.7_LAST81) can be obtained at the corresponding observation time.

[0242] The orbital error within the observation period is obtained by subtracting the predicted orbital information output by SGP4 from the actual orbital information.

[0243] The orbital errors within the observation period are integrated to form an experimental dataset.

[0244] Specifically, data preprocessing:

[0245] Since the units of different orbital errors in the experimental dataset are different, some data included in the orbital errors may dominate the learning algorithm. Therefore, the Min-Max standardization method can be used to standardize the experimental dataset, for example, normalization, which normalizes all the data in the experimental dataset to the range of [0, 1]. Specifically, this can be achieved through the following formula (15):

[0246]

[0247] in, This represents the value of the k-th data point included in the original orbital error, where K is the total number of data points included in the original orbital error. This represents the minimum value of the k-th data point; This represents the maximum value of the k-th data point; This is the value of the k-th data point after normalization.

[0248] Finally, we obtain the normalized multidimensional data. This yields the normalized experimental dataset and the time series samples of orbital errors.

[0249] The normalized experimental dataset is converted using a sliding window algorithm. For example, the sliding window size is set to 60 and the prediction step size is 45, converting it into a dataset usable by supervised machine learning models, thus obtaining the converted experimental dataset.

[0250] The experimental dataset is divided into a training set, a validation set, and a test set.

[0251] For any dataset in the training set, validation set, and test set, based on the time sequence of the orbital errors in the dataset, it is divided into a first orbital error time sequence sample and a second orbital error time sequence sample. The second orbital error time sequence sample is the label, and the first orbital error time sequence sample is the time sequence sample preceding the second orbital error time sequence sample.

[0252] S802. The initial orbital error prediction model is trained using training data until the preset convergence condition is met, and a trained orbital error prediction model is obtained. The initial orbital error prediction model is an attention-based initial deep learning model.

[0253] Specifically, model training:

[0254] The initial orbital error prediction model is trained using training data (the training set). The hyperparameters of the initial orbital error prediction model are tuned by setting the number of training iterations (e.g., 100), the learning rate (e.g., 1e-3), and the batch size (e.g., 64). Furthermore, the MSE loss function is used, and the Adam optimizer is employed to update the parameters of the initial orbital error prediction model. In the LSTM encoder and decoder, the number of neurons in the LSTM is set to 256.

[0255] The initial orbital error prediction model is trained until it converges on both the training and validation sets, resulting in a well-trained orbital error prediction model.

[0256] Specifically, model testing:

[0257] The trained orbital error prediction model was tested using a test set. To evaluate the model's predictive performance, mean absolute error (MAE) and root mean square error (RMSE) were used as performance metrics.

[0258] Among them, MAE is the expected value of the sum of absolute errors, which characterizes the average deviation between the actual value and the predicted value, and truly reflects the prediction error situation. RMSE is the square root of the expected value of the sum of squared prediction errors, which is sensitive to outliers (data points with large prediction errors) and can reflect the robustness of the model to a certain extent. The smaller the values ​​of MAE and RMSE, the better the model quality and the higher the prediction accuracy. The specific calculation is shown in the following formulas (15)-(16):

[0259]

[0260]

[0261] in, This represents the true value of test data sample j in satellite position dimension i; M represents the predicted value of test data sample j in satellite position dimension i; M represents the number of test set samples; N represents a constant of 3.

[0262] When the MAE meets the preset MAE threshold and the RMSE meets the preset RMSE threshold, the test is passed and training is complete, indicating that the trained trajectory error prediction model has high accuracy.

[0263] To facilitate understanding of the content of this embodiment, the following will be explained... Figure 10 For a brief explanation, please refer to [link / reference]. Figure 10 , Figure 10This is a schematic diagram of a training process provided in an embodiment of this application. The acquired TLE (Trajectory Error) at different times in the first time series is input into a dynamics model, such as SGP4. SGP4 outputs the predicted orbital information at the corresponding time, and simultaneously acquires the actual orbital information at the corresponding time. The difference between the predicted orbital information and the actual orbital information at the corresponding time is calculated to obtain the orbital error at that time. Correspondingly, the orbital errors at different times in the second time series are obtained in the same way. The orbital errors at different times in the first and second time series are input into an initial orbital error prediction model. The initial orbital error prediction model is trained using the orbital errors at different times in the second time series as labels. The initial orbital error prediction model can be a deep neural network model incorporating an attention mechanism. Training continues until the model converges, completing the training and obtaining a trained orbital error prediction model. The trained orbital error prediction model is then used to predict the orbital error at future times, and this orbital error is used to correct the orbital information predicted by the dynamics model at those future times.

[0264] In the above embodiments of this application, training data is acquired and used to train an initial orbital error prediction model until a preset convergence condition is met, resulting in a trained orbital error prediction model. The initial orbital error prediction model is an attention-based initial deep learning model. The method in this embodiment, by introducing an attention mechanism into the initial orbital error prediction model, enables the model to learn the correlations between multiple data points included in the input orbital error, and / or learn the long-term temporal dependency between the input orbital error time-series data and the output orbital error time-series data, thereby improving the accuracy of the trained orbital error prediction model.

[0265] Furthermore, the technical effects of applying this application are illustrated below with specific data. Please refer to Table 1, which includes encoder-decoder models without an attention mechanism (LSTM), encoder-decoder models with only a temporal attention mechanism (Attention-LSTM), and encoder-decoder models with both self-attention and temporal attention mechanisms (DualAttention-LSTM). The prediction performance of their orbital errors under different prediction durations is shown in Table 1.

[0266] Table 1

[0267]

[0268] As can be seen from Table 1, the DualAttention-LSTM model has the smallest MAE and RMSE values ​​under different prediction durations, indicating that the encoder-decoder model that introduces both self-attention and temporal attention mechanisms has more accurate predicted trajectory errors. The next best is the encoder-decoder model that only introduces temporal attention mechanisms, and the last is the encoder-decoder model that does not introduce attention mechanisms.

[0269] Please see Figure 11 , Figure 11 This application provides a comparison chart of predicted orbital error values ​​and actual orbital error values ​​in the x-direction, where the horizontal axis represents the prediction time and the vertical axis represents the orbital error in the x-direction. Figure 11 As can be seen, the encoder-decoder model that introduces both self-attention and temporal attention mechanisms predicts orbital errors that are closer to the actual orbital error values. This is followed by the encoder-decoder model that only introduces temporal attention mechanisms, and finally the encoder-decoder model that does not introduce attention mechanisms.

[0270] Accordingly, please see Figure 12 , Figure 12 This application provides a comparison chart of predicted orbital error values ​​and actual orbital error values ​​in the y-direction, where the horizontal axis represents the prediction time and the vertical axis represents the orbital error in the y-direction. Figure 12 As can be seen, the encoder-decoder model that introduces both self-attention and temporal attention mechanisms predicts orbital errors that are closer to the actual orbital error values. This is followed by the encoder-decoder model that only introduces temporal attention mechanisms, and finally the encoder-decoder model that does not introduce attention mechanisms.

[0271] Accordingly, please see Figure 13 , Figure 13 This application provides a comparison chart of predicted orbital error values ​​and actual orbital error values ​​in the z-direction, where the horizontal axis represents the prediction time and the vertical axis represents the orbital error in the z-direction. Figure 13 As can be seen, the encoder-decoder model that introduces both self-attention and temporal attention mechanisms predicts orbital errors that are closer to the actual orbital error values. This is followed by the encoder-decoder model that only introduces temporal attention mechanisms, and finally the encoder-decoder model that does not introduce attention mechanisms.

[0272] In summary, the encoder-decoder model in this application, which incorporates both self-attention and temporal attention mechanisms, predicts more accurate orbital errors.

[0273] This application also provides a device for predicting the orbital information of a space target; please refer to [link to relevant documentation]. Figure 14 , Figure 14 This is a schematic diagram of the structure of a space target orbit information prediction device provided in an embodiment of this application, as shown below. Figure 14 As shown, it includes:

[0274] The acquisition module 141 is used to acquire orbital error time series data between the actual orbital information of the space target and the predicted orbital information based on the dynamic model. The orbital error time series data includes orbital errors at multiple consecutive historical moments, and the orbital errors include velocity errors and position errors.

[0275] The processing module 142 is used to input the orbital error time series data into the orbital error prediction model and output the predicted orbital error at multiple future times. The orbital error prediction model is a deep learning model based on the attention mechanism.

[0276] The prediction module 143 is used to correct the orbital information of the space target predicted by the dynamic model at multiple future times based on the orbital error at multiple future times, so as to obtain the corresponding orbital information at multiple future times.

[0277] One possible implementation is as follows: the attention-based deep learning model includes a first encoder based on a self-attention mechanism and a first decoder based on a temporal attention mechanism. The first encoder based on the self-attention mechanism includes a first self-attention layer and a first target encoder, and the first decoder based on the temporal attention mechanism includes a first temporal attention layer and a first target decoder. Processing module 142 is specifically used for:

[0278] The orbital error time series data is input into the first self-attention layer, and the self-attention weight is calculated on the orbital error time series data through the first self-attention layer to determine the first orbital error time series data with added self-attention weight.

[0279] The first target encoder encodes the first orbital error time series data with added self-attention weights to obtain the first hidden state output vector at each historical moment.

[0280] The first time attention layer is used to calculate the time attention weights based on the first hidden state output vector at each historical moment, so as to determine the first weighted hidden state output vector at multiple future moments.

[0281] The second hidden state output vector at multiple future time points is determined by the first target decoder and the first weighted hidden state output vector based on multiple future time points.

[0282] A linear transformation is performed on the output vector of the second hidden state at multiple future time points to output the predicted orbital error at multiple future time points.

[0283] One possible implementation is: processing module 142, specifically used for:

[0284] The first orbital error weight time series data is determined by calculating the self-attention weight of the orbital error time series data through the first self-attention layer.

[0285] The first self-attention layer merges the first orbital error weight time series data with the orbital error time series data to determine the first orbital error time series data with added self-attention weights.

[0286] One possible implementation is: processing module 142, specifically used for:

[0287] Perform the following operations to determine the first weighted hidden state output vector at the t′-th future time, until t′ equals T′, where T′ is the last future time. The following operations include:

[0288] Based on the first weight score calculation algorithm preset in the first time attention layer, the weight scores of the first hidden state output vectors at each historical time corresponding to the t′-1 future time are determined according to the second hidden state output vector at the t′-1 future time. Specifically, when the t′-1 future time is the initial future time, the second hidden state output vector at the t′-1 future time is the first hidden state output vector at the last historical time.

[0289] Based on the first activation function preset in the first time attention layer and the weight scores of the first hidden state output vector at each historical time, the weights of the first hidden state output vector at each historical time are determined.

[0290] The first time attention layer performs a weighted summation of the first hidden state output vector at each historical time corresponding to the t′ future time and the weights corresponding to each first hidden state output vector to obtain the first weighted hidden state output vector at the t′ future time.

[0291] One possible implementation is: processing module 142, specifically used for:

[0292] Perform the following operations to determine the output vector of the second hidden state at the t′-th future time, until t′ equals T′, where T′ is the last future time. The following operations include:

[0293] Based on the second hidden state output vector at the (t′-1)th future time and the first weighted hidden state output vector at the t′th future time, the second hidden state output vector at the t′th future time is obtained.

[0294] One possible implementation is as follows: the attention-based deep learning model includes a second encoder and a second decoder based on a self-attention mechanism. The second encoder based on the self-attention mechanism includes a second self-attention layer and a second target encoder. The processing module 142 is specifically used for:

[0295] The orbital error time series data is input into the second self-attention layer, and the self-attention weight is calculated on the orbital error time series data through the second self-attention layer to determine the second orbital error time series data with added self-attention weight.

[0296] The second target encoder encodes the second orbital error time series data with added self-attention weights to obtain the third hidden state output vector at the last historical moment.

[0297] The fourth hidden state output vector for multiple future moments is determined by using the second decoder and based on the third hidden state output vector at the last historical moment.

[0298] A linear transformation is performed on the fourth hidden state output vector at multiple future time points to obtain the predicted orbital error time series data at multiple future time points.

[0299] One possible implementation is: processing module 142, specifically used to: calculate the self-attention weight of the orbital error time series data through the second self-attention layer, so as to determine the second orbital error weight time series data of the orbital error time series data.

[0300] The second self-attention layer merges the second orbital error weight time series data with the orbital error time series data to determine the second orbital error time series data with added self-attention weights.

[0301] One possible implementation is as follows: the attention-based deep learning model includes a third encoder and a third decoder based on a temporal attention mechanism. The third decoder based on the temporal attention mechanism includes a second temporal attention layer and a third target decoder. Processing module 142 is specifically used for:

[0302] The orbital error time series data is input into the third encoder and encoded by the third encoder to obtain the fifth hidden state output vector at each historical moment.

[0303] The second time attention layer is used to calculate the time attention weights based on the fifth hidden state output vector at each historical moment, so as to determine the second weighted hidden state output vector at multiple future moments.

[0304] The sixth hidden state output vector at multiple future time points is determined by using the third target decoder and the second weighted hidden state output vector based on multiple future time points.

[0305] A linear transformation is performed on the output vector of the sixth hidden state at multiple future time points to output the predicted orbital error at multiple future time points.

[0306] One possible implementation is: processing module 142, specifically used for:

[0307] Perform the following operations to determine the second weighted hidden state output vector at the t′-th future time step, until t′ equals T′, where T′ is the last future time step. The following operations include:

[0308] Based on the second weight score calculation algorithm preset by the second time attention layer, the weight scores of the fifth hidden state output vectors at each historical time corresponding to the t′-1 future time are determined according to the sixth hidden state output vector at the t′-1 future time. Specifically, when the t′-1 future time is the initial future time, the sixth hidden state output vector at the t′-1 future time is the fifth hidden state output vector at the last historical time.

[0309] Based on the second activation function preset in the second time attention layer and the weight scores of the fifth hidden state output vector at each historical moment, the weights of the fifth hidden state output vector at each historical moment are determined.

[0310] The second time attention layer performs a weighted summation of the fifth hidden state output vector at each historical time corresponding to the t′ future time and the weights corresponding to each fifth hidden state output vector to obtain the second weighted hidden state output vector at the t′ future time.

[0311] One possible implementation is: processing module 142, specifically used for:

[0312] Perform the following operations to determine the output vector of the sixth hidden state at the t′-th future time, until t′ equals T′, where T′ is the last future time. The following operations include:

[0313] Based on the output vector of the sixth hidden state at the (t′-1)th future time and the output vector of the second weighted hidden state at the t′th future time, the output vector of the sixth hidden state at the t′th future time is obtained.

[0314] One possible implementation is that processing module 142 is also used for:

[0315] Acquire training data, which includes first orbital error time series samples and second orbital error time series samples. The second orbital error time series samples are the labels, and the first orbital error time series samples are the time series samples preceding the second orbital error time series samples.

[0316] The initial orbital error prediction model is trained using training data until the preset convergence condition is met, resulting in a trained orbital error prediction model. The initial orbital error prediction model is an attention-based initial deep learning model.

[0317] The orbital information prediction device for space targets provided in this embodiment can execute the method provided in the above-described method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0318] Figure 15 A schematic diagram of the structure of the electronic device provided in this application. Figure 15 As shown, the electronic device provided in this embodiment includes at least one processor 1501 and a memory 1502. Optionally, the device 150 also includes a communication component 1503. The processor 1501, memory 1502, and communication component 1503 are connected via a bus 1504.

[0319] In a specific implementation, at least one processor 1501 executes computer execution instructions stored in memory 1502, causing at least one processor 1501 to perform the above-described method.

[0320] The specific implementation process of processor 1501 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0321] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0322] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0323] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0324] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0325] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0326] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0327] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0328] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0329] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0330] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0331] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0332] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0333] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for predicting the orbital information of a space target, characterized in that, include: The orbital error time series data between the actual orbital information of a space target and the predicted orbital information based on a dynamic model is obtained. The orbital error time series data includes orbital errors at multiple consecutive historical moments, and the orbital errors include velocity errors and position errors. The orbital error time series data is input into the orbital error prediction model, and the predicted orbital errors at multiple future times are output. The orbital error prediction model is a deep learning model based on an attention mechanism. Based on the orbital errors at the multiple future time points, the orbital information of the space target predicted by the dynamic model at the multiple future time points is corrected to obtain the corresponding orbital information at the multiple future time points.

2. The method according to claim 1, characterized in that, The attention-based deep learning model includes: a first encoder based on a self-attention mechanism and a first decoder based on a temporal attention mechanism. The first encoder based on the self-attention mechanism includes: a first self-attention layer and a first target encoder. The first decoder based on the temporal attention mechanism includes: a first temporal attention layer and a first target decoder. The step of inputting the orbital error time series data into the orbital error prediction model and outputting the predicted orbital errors at multiple future times includes: The orbital error time series data is input into the first self-attention layer, and the self-attention weight is calculated on the orbital error time series data through the first self-attention layer to determine the first orbital error time series data with added self-attention weight; The first target encoder encodes the first orbital error time series data with added self-attention weights to obtain the first hidden state output vector at each historical moment. The first time attention layer is used to calculate the time attention weight based on the first hidden state output vector at each historical moment to determine the first weighted hidden state output vector at multiple future moments. The second hidden state output vector at multiple future moments is determined by the first target decoder and the first weighted hidden state output vector based on the multiple future moments. A linear transformation is performed on the output vector of the second hidden state at the multiple future time points to output the predicted orbital errors at the multiple future time points.

3. The method according to claim 2, characterized in that, The step of calculating self-attention weights on the orbital error time series data through the first self-attention layer to determine the first orbital error time series data with added self-attention weights includes: The first self-attention layer is used to calculate the self-attention weight of the orbital error time series data to determine the first orbital error weight time series data of the orbital error time series data. The first self-attention layer merges the first orbital error weight time series data with the orbital error time series data to determine the first orbital error time series data with added self-attention weights.

4. The method according to claim 2, characterized in that, The step of calculating time attention weights through the first time attention layer and based on the first hidden state output vector at each historical moment to determine the first weighted hidden state output vector at multiple future moments includes: Perform the following operations to determine the t-th ′ The first weighted hidden state output vector at each future time step, up to t ′ equal to T ′ Up to now, T ′ For the last future moment, the following operations include: Based on the first weight score calculation algorithm preset in the first time attention layer, according to the t-th ′ The output vector of the second hidden state at -1 future time step determines the t-th time step. ′ The weight scores of the first hidden state output vectors at each historical time corresponding to each future time; where, when the t-th time... ′ When the t-th future time is the initial future time, the t-th future time... ′ The output vector of the second hidden state at -1 future time is the output vector of the first hidden state at the last historical time. Based on the first activation function preset in the first time attention layer and the weight score of the first hidden state output vector at each corresponding historical moment, the weight of the first hidden state output vector at each corresponding historical moment is determined. Based on the first time attention layer, the t-th... ′ The first hidden state output vectors at each historical time corresponding to each future time step are weighted and summed with their corresponding weights to obtain the output vector at the t-th time step. ′ The first weighted hidden state output vector at each future time.

5. The method according to claim 2, characterized in that, The step of determining the second hidden state output vector at multiple future time points using the first decoder and the first weighted hidden state output vector at the multiple future time points includes: Perform the following operations to determine the t-th ′ The output vector of the second hidden state at each future time step, up to t ′ equal to T ′ Up to now, T ′ For the last future moment, the following operations include: According to the tth ′ The output vector of the second hidden state at time -1 in the future and the t-th time ′ The first weighted hidden state output vector at the t-th future time step yields the t-th time step. ′ The output vector of the second hidden state at each future time.

6. The method according to claim 1, characterized in that, The attention-based deep learning model includes: a second encoder and a second decoder based on a self-attention mechanism, wherein the second encoder based on the self-attention mechanism includes: a second self-attention layer and a second target encoder; The step of inputting the orbital error time series data into the orbital error prediction model and outputting the predicted orbital errors at multiple future times includes: The orbital error time series data is input into the second self-attention layer, and the self-attention weight is calculated on the orbital error time series data through the second self-attention layer to determine the second orbital error time series data with added self-attention weight; The second target encoder encodes the second orbital error time series data with added self-attention weights to obtain the third hidden state output vector at the last historical moment. The fourth hidden state output vector for multiple future moments is determined by the second decoder and based on the third hidden state output vector at the last historical moment. A linear transformation is performed on the fourth hidden state output vector at the multiple future time points to output the predicted orbital errors at the multiple future time points.

7. The method according to claim 6, characterized in that, The step of calculating self-attention weights on the orbital error time series data through the second self-attention layer to determine the second orbital error time series data with added self-attention weights includes: The second self-attention layer is used to calculate the self-attention weight of the orbital error time series data to determine the second orbital error weight time series data of the orbital error time series data. The second self-attention layer merges the second orbital error weight time series data with the orbital error time series data to determine the second orbital error time series data with added self-attention weights.

8. The method according to claim 1, characterized in that, The attention-based deep learning model includes a third encoder and a third decoder based on a time attention mechanism, wherein the third decoder based on the time attention mechanism includes a second time attention layer and a third target decoder. The step of inputting the orbital error time series data into the orbital error prediction model and outputting the predicted orbital errors at multiple future times includes: The orbital error time series data is input into the third encoder and encoded by the third encoder to obtain the fifth hidden state output vector at each historical moment. The second time attention layer is used to calculate the time attention weight based on the fifth hidden state output vector at each historical moment, so as to determine the second weighted hidden state output vector at multiple future moments. The sixth hidden state output vector at multiple future moments is determined by the third target decoder and the second weighted hidden state output vector based on the multiple future moments. A linear transformation is performed on the output vector of the sixth hidden state at the multiple future time points to output the predicted orbital errors at the multiple future time points.

9. The method according to claim 8, characterized in that, The step of calculating time attention weights through the second time attention layer and based on the fifth hidden state output vector at each historical moment to determine the second weighted hidden state output vector at multiple future moments includes: Perform the following operations to determine the t-th ′ The second weighted hidden state output vector at each future time step, up to t ′ equal to T ′ Up to now, T ′ For the last future moment, the following operations include: Based on the second weight score calculation algorithm preset in the second time attention layer, according to the t-th ′ The output vector of the sixth hidden state at -1 future time step determines the t-th time step. ′ The weight scores of the fifth hidden state output vector at each historical time corresponding to each future time; where, when the t-th time... ′ When the t-th future time is the initial future time, the t-th future time... ′ The output vector of the sixth hidden state at -1 future time is the output vector of the fifth hidden state at the last historical time. Based on the second activation function preset in the second time attention layer and the weight score of the fifth hidden state output vector at each corresponding historical moment, the weight of the fifth hidden state output vector at each corresponding historical moment is determined. Based on the second time attention layer, the t-th... ′ The output vectors of the fifth hidden state at each historical time corresponding to each future time are weighted and summed with their corresponding weights to obtain the output vector of the t-th time. ′ The output vector of the second weighted hidden state at each future time.

10. The method according to claim 8, characterized in that, The step of determining the sixth hidden state output vector at multiple future time points using the third decoder and the second weighted hidden state output vector at multiple future time points includes: Perform the following operations to determine the t-th ′ The output vector of the sixth hidden state at each future time step, up to t. ′ equal to T ′ Up to now, T ′ For the last future moment, the following operations include: According to the tth ′ The output vector of the sixth hidden state at -1 future time and the t-th time... ′ The output vector of the second weighted hidden state at the t-th future time step yields the t-th time step. ′ The output vector of the sixth hidden state at a future time.

11. The method according to claim 1, characterized in that, The training steps for the orbital error prediction model include: Acquire training data, which includes a first orbital error time series sample and a second orbital error time series sample, wherein the second orbital error time series sample is a label, and the first orbital error time series sample is the time series sample preceding the second orbital error time series sample. The initial orbital error prediction model is trained using the training data until a preset convergence condition is met, resulting in a trained orbital error prediction model. The initial orbital error prediction model is an attention-based initial deep learning model.

12. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-11.

Images