Low earth orbit satellite orbit state prediction method based on artificial intelligence

Abstract

The invention belongs to the technical field of satellite orbit prediction, and provides a low-orbit satellite orbit state prediction method based on artificial intelligence, which comprises the following steps of: dividing step lengths of a predicted orbit according to a time window, calculating a resultant acceleration of each step length, performing stress input on an initial position and speed, and obtaining a nominal orbit state sequence; forming an augmented state vector with aerodynamic synthesis parameters; inputting a time sequence neural network model, and performing rolling updating on the augmented state according to the output; obtaining an original standard deviation through mapping by using uncertainty measurement, calculating a standardized residual error, solving a right side weighted quantile according to a normalized time weight, comparing the right side weighted quantile with a preset threshold value, and carrying out adaptive deviation adjustment; and calculating risk measurement by using the acquired data, comparing the risk measurement with a set threshold value and a priority rule, automatically executing a judgment action on the risk measurement, recording a trigger reason and the judgment action, and updating parameters in the time sequence neural network model.

Description

Technical Field

[0001] This invention relates to the field of satellite orbit prediction technology, specifically a method for predicting the orbital state of low-Earth orbit satellites based on artificial intelligence. Background Technology

[0002] Orbit prediction is a technique and process that estimates and outputs the position, velocity, and uncertainties of a satellite within a future time window based on current and historical physical models and data.

[0003] Chinese patent application number CN202510162446.2 discloses a method, apparatus, and device for predicting the orbital state of a low-Earth orbit (LEO) satellite. The method includes: defining a pseudo-drag coefficient based on an atmospheric drag perturbation formula; performing a decomposition operation on space environment parameters based on pseudo-drag coefficient correlation analysis using a VMD algorithm to obtain space environment characteristic parameters, including solar radiation flux index and geomagnetic index; training an initial SVM model using the space environment characteristic parameters and a preset equivalent drag coefficient to predict and train the pseudo-drag coefficient, resulting in an optimized SVM model; predicting the pseudo-drag coefficient at future times using the optimized SVM model to obtain pseudo-drag prediction values; and calculating the LEO satellite orbital state based on the pseudo-drag prediction values ​​to obtain the state prediction result.

[0004] In the field of satellite orbit prediction technology, although there are technical solutions that use VMD algorithms to correlate space environment parameters with pseudo drag coefficients to construct pseudo drag coefficient prediction models and ensure the accuracy of orbit state prediction results, existing technologies struggle to handle sudden changes in atmospheric density and space weather in real time during short-duration GNSS observation interruptions. This leads to a rapid amplification of position and time errors within short time windows. Therefore, there is a need for an AI-based method that does not rely on continuous GNSS and can provide reliable orbit state predictions and address the uncertainties of low-Earth orbit satellites within short-duration GNSS observations. Summary of the Invention

[0005] This invention provides an artificial intelligence-based method for predicting the orbital state of low-Earth orbit satellites, aiming to solve the problem of amplified position and time errors within a short time window due to GNSS observation interruptions.

[0006] The technical solution adopted by this invention to solve the above-mentioned technical problems is as follows: A method for predicting the orbital state of low-Earth orbit satellites based on artificial intelligence is provided, comprising:

[0007] Calculating the nominal orbital state sequence and constructing the augmented state vector involves dividing the predicted orbit into time windows, calculating the composite acceleration for each step, inputting the force into the initial position and velocity, obtaining the nominal orbital state sequence, and then combining it with the aerodynamic composite parameters to form the augmented state vector.

[0008] The augmented state is updated by rolling over the input time-series neural network model. This involves inputting the nominal orbital state sequence, environmental quantities, and the difference between observations and nominal values ​​into the time-series neural network model, and updating the augmented state based on the output.

[0009] Mapping and calibrating confidence intervals, calculating bias to trigger adaptation involves obtaining the original standard deviation through mapping using uncertainty measures, calculating standardized residuals, obtaining the right-hand weighted quantile value according to normalized time weights, comparing it with a preset threshold, and performing adaptive bias adjustment.

[0010] The risk assessment and task determination process, along with the iterative update of the state prediction process, involves using acquired data to calculate risk metrics, comparing them with set thresholds and priority rules, automatically performing judgment actions on the risk metrics, recording the triggering reasons and judgment actions, and updating the parameters in the temporal neural network model.

[0011] As a preferred implementation, the specific steps for calculating the nominal orbit state sequence and constructing the augmented state vector are as follows: When GNSS is interrupted, firstly, determine the starting and ending points of the current orbit prediction. Divide the time step according to the short-term time window length between the starting and ending points. Using the position and velocity of the starting point as the initial state, establish an equally spaced time grid for the orbit from the starting point to the ending point. Prepare environmental quantities for each moment in the time grid. In each time step, define the composite acceleration in terms of gravity, air resistance, and other weak disturbances, which is used as the sole force input for propulsion. Update the position and velocity for the entire time window to obtain a time-sequential sequence of nominal position and nominal velocity. Introduce air... The aerodynamic composite parameters are placed in the same state vector as position and velocity to form an augmented state vector. The altitude is obtained from the current position and velocity, and the current position is unified to the Earth-fixed coordinate system. The intersection of the normals at this position is determined using the Earth's reference ellipsoid as a reference. The altitude is taken as the normal distance from the current position to the reference ellipsoid surface. The atmospheric density at the current time step is obtained by querying environmental indicators based on the altitude. The atmospheric linear velocity is calculated using the Earth's rotation angular velocity and the current position. The relative airflow velocity is obtained by subtracting the atmospheric linear velocity from the current velocity. Using the aerodynamic composite parameters, relative airflow velocity, and atmospheric density, the drag acceleration at the current time step is calculated. This drag acceleration is then added to the gravity and weak disturbances to obtain the composite acceleration, as shown in the formula:

[0012] a sum =a g +a d +a weak ,

[0013] Where a sum a represents the resultant acceleration. g a represents the acceleration due to gravity. da represents the acceleration due to air resistance. weak Indicates weak perturbation acceleration;

[0014] The initial values ​​of the aerodynamic synthesis parameters are set based on ground calibration values. At each time step, the air drag component is calculated using the atmospheric density obtained at altitude and the relative velocity corrected for Earth's rotation. The formula is as follows:

[0015]

[0016] Where a d Let ρ represent the air drag component, ε represent the atmospheric density at the current altitude, ε represent the aerodynamic composition parameters, v represent the current velocity, r represent the current position, and ω represent the current position. e It represents the angular velocity of Earth's rotation.

[0017] As a preferred embodiment, the specific steps for rolling updates of the augmented state in the input temporal neural network model are as follows: The nominal position and nominal velocity obtained based on the synthetic acceleration propulsion, along with environmental quantities at the same time and the differences between the nominal and observed values ​​at the corresponding time, are concatenated in chronological order to form an input sequence, which is then input into the temporal neural network model. The model outputs a value using [Δr], based on the residual correction, the slow drift of the aerodynamic synthesis parameters, and the uncertainty measure. k ,Δv k ,Δβ k ,s k ] represents, where k represents the current time, Δr k , represents the position residual correction value, Δv k This represents the velocity residual correction value, Δβ. k The slow drift of the aerodynamic composition parameters, s k It represents the uncertainty measure and updates the augmented state based on the output vector; it performs normalization and standardization on the input sequence, sets the sliding window length and stride, divides the input sequence into continuous time segments, and uses them as processing units of the temporal neural network model. It inputs the time segments into the temporal encoding unit of the temporal neural network model in time sequence, generates a context vector representing the dynamic information of the current time segment according to the time, and synchronously generates residual correction, slow drift of aerodynamic synthesis parameters and uncertainty measure based on the context vector.

[0018] Starting from the augmented state of the previous time step, using z k-1 =[r k-1 ,v k-1 ,β k-1[] represents the augmented state at time k-1. The residual correction of the temporal neural network at time k is added to the residual correction of the prior state. The slow drift of the aerodynamic synthesis parameters of the prior state is addressed by implementing small-step updates and physical upper and lower bound pruning, based on the uncertainty metric s of the temporal neural network model. k Adjust the process noise intensity of the current step to form the process noise covariance matrix, as shown in the formula:

[0019] noise k =noise base ·g(s k ),

[0020] noise k Let g(s) represent the process noise covariance matrix. k ) represents a monotonically amplifying function, noise base This represents the baseline process noise covariance matrix, without any adaptive adjustments.

[0021] As a preferred embodiment, the specific steps of mapping and calibrating the confidence interval and calculating the deviation to trigger adaptation are as follows: at the calculation time, read the uncertainty measure s output by the temporal neural network. k , will s k According to u k =as k +b restores the data to the training numerical domain, where a and b are the linear scaling factor and the linear bias term, respectively, and u k Represents an integer metric, which is transformed through a mapping function. k The original standard deviation σ k A target confidence level g is pre-set to calculate the projection of the augmented state's position and velocity at time step p, and the corresponding basis vectors obtained are denoted as the current point prediction p. k , and the mapped standard deviation σ k Construct the original interval [p k ±z g u k ], where z g Let represent the critical multiple of the standard normal distribution, and g represent the target confidence level. A scaling factor is obtained by comparing the weighted quantile of the standardized residuals with the Gaussian critical value. This factor is used to calibrate the interval. For each target confidence level, empirical coverage and coverage deviation are calculated, along with the comprehensive deviation index and the covariance entropy growth rate. The standardized residuals for each sample are calculated by dividing the absolute value of the difference between the true value and the predicted point by the corresponding standard deviation. The standardized residuals of each sample are sorted in ascending order, and their corresponding weights are accumulated. The residual value corresponding to the first time the accumulated weight reaches 1-g is set as the right-hand weighted quantile. This value is then compared with the Gaussian critical multiple z. gIn contrast, the ratio is used as a scaling factor to calibrate the current time interval proportionally.

[0022] As a preferred implementation, the specific steps of the risk assessment and task determination, and the iterative update of the state prediction process are as follows: The acquired state prediction and uncertainty information, real-time observation data, and environmental service data from external sources are used to pre-define risk sets for each indicator. A risk metric is calculated based on the data. The risk metric is compared with the set threshold and priority rules. If the risk metric exceeds the threshold, a corresponding judgment action is executed according to the trigger level. Simultaneously, the triggering reason and the proposed action are written into the operation log and reported. Newly acquired observations and telemetry data are used to update the parameters in the time-series neural network model in small steps, and the state prediction process is re-run to form a closed-loop iteration until the risk is reduced to an acceptable range. For each predefined risk indicator, the required point prediction and covariance of the risk indicator are first obtained. The corresponding violation event of the risk indicator is defined as a subset in the state space using an inequality. Assuming the prediction error is approximately a multivariate normal distribution, a fast numerical approximation is used to calculate the probability of the violation event. The probability is compared with the threshold and weighted and aggregated to form the final risk score.

[0023] The beneficial effects of this invention are as follows:

[0024] 1. This invention constructs a nominal orbit sequence and augmented state, introduces aerodynamic synthesis parameters, and uses a time-series neural network model to correct the residuals online. During GNSS interruptions, it rapidly compensates for transient errors caused by sudden changes in atmospheric density, thereby suppressing the rapid amplification of position and time errors within a short time window.

[0025] 2. This invention uses calibrated point estimates and covariance for probabilistic risk measurement, and automatically generates action suggestions by combining threshold and priority rules, which can trigger protective measures earlier and reduce task risks.

[0026] Legend

[0027] Figure 1 This is a flowchart of a method for predicting the orbital state of low-Earth orbit satellites based on artificial intelligence.

[0028] Figure 2 This is a comparison chart showing the effectiveness of an artificial intelligence-based method for predicting the orbital state of low-Earth orbit satellites. Detailed Implementation

[0029] To make the technical means, creative features, achieved objectives, and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention.

[0030] Example 1, such as Figure 1 This is an artificial intelligence-based method for predicting the orbital state of low-Earth orbit satellites, comprising the following steps:

[0031] Calculate the nominal orbital state sequence and construct the augmented state vector;

[0032] Input a temporal neural network model and continuously update the augmented state;

[0033] Map and calibrate confidence intervals, calculate bias to trigger adaptation;

[0034] Assess risks and determine tasks, and iteratively update the status prediction process.

[0035] The specific implementation steps are as follows: A method for predicting the orbital state of low-Earth orbit satellites based on artificial intelligence, wherein the specific steps for calculating the nominal orbital state sequence and constructing the augmented state vector are as follows:

[0036] When the Global Navigation Satellite System (GNSS) is interrupted, the starting and ending points of the current orbit prediction are first determined. The time step is divided according to the short time window length between the starting and ending points. The position and velocity after time alignment of the most recent valid observation are used as the initial state. An equally spaced time grid is established for the orbit from the starting point to the ending point. Environmental quantities are prepared for each moment in the time grid, including Earth's gravity parameters, atmospheric density, and atmospheric rotational linear velocity. In each time step, the composite acceleration is defined by gravity, air resistance, and other weak perturbations and used as the sole force input for propulsion. The position and velocity of the entire time window are updated to obtain a time-series sequence of nominal position and nominal velocity. Aerodynamic composite parameters are introduced to describe the strong and weak fluctuations of air resistance caused by attitude, light angle, and changes in the external environment. The aerodynamic composite parameters, position, and velocity are placed in the same state vector to form an augmented state vector.

[0037] Specifically, the composite acceleration is calculated by first determining the altitude from the current position and velocity. This involves unifying the current position to a Earth-fixed coordinate system, using the Earth's reference ellipsoid as a reference to determine the intersection of the normals at that position, and taking the altitude as the normal distance from the current position to the reference ellipsoid. The atmospheric density at the current time step is obtained by querying environmental indicators based on the altitude. The atmospheric linear velocity is then calculated using the Earth's rotational angular velocity and the current position. The relative airflow velocity is obtained by subtracting the atmospheric linear velocity from the current velocity. Using aerodynamic composite parameters, relative airflow velocity, and atmospheric density, the drag acceleration at the current time step is calculated. This drag acceleration is then added to the gravity and weak disturbances to obtain the composite acceleration, as shown in the formula:

[0038] a sum =a g +a d +a weak ,

[0039] Where a sum a represents the resultant acceleration. g a represents the acceleration due to gravity. d a represents the acceleration due to air resistance. weak Indicates weak perturbation acceleration;

[0040] Specifically, the aerodynamic synthesis parameters are introduced, with initial values ​​set based on ground calibration. At each time step, the air drag component is calculated using the aerodynamic synthesis parameters based on the atmospheric density obtained at altitude and the relative velocity corrected for Earth's rotation. The formula is as follows:

[0041]

[0042] Where a d Let ρ represent the air drag component, ε represent the atmospheric density at the current altitude, ε represent the aerodynamic composition parameters, v represent the current velocity, r represent the current position, and ω represent the current position. e This represents the Earth's angular velocity of rotation;

[0043] The specific steps for inputting a temporal neural network model and continuously updating the augmented state are as follows:

[0044] The nominal position and nominal velocity obtained based on synthetic acceleration propulsion, along with environmental quantities at the same time and the differences between the nominal and observed values ​​at the corresponding time, are concatenated in chronological order into an input sequence. This sequence is then input into a time-series neural network model. The model outputs residual corrections, slow drift of aerodynamic synthesis parameters, and uncertainty measures, expressed as [Δr]. k ,Δv k ,Δβ k ,s k ] represents, where k represents the current time, Δr k , represents the position residual correction value, Δv kThis represents the velocity residual correction value, Δβ. k The slow drift of the aerodynamic composition parameters, s k It represents a measure of uncertainty and updates the augmented state;

[0045] Specifically, the residual correction, slow drift of aerodynamic synthesis parameters, and uncertainty measure output by the temporal neural network model are all related to the fact that a temporal neural network is a neural network model that encodes inputs organized in chronological order, utilizes cross-time dependencies and state memory, and outputs predictions or estimates that evolve over time. First, the input sequence is normalized and standardized, such as through interpolation and denoising. The sliding window length and stride are set, and the input sequence is divided into continuous time segments, which serve as the smallest unit for processing by the temporal neural network model. These time segments are then input sequentially into the temporal encoding unit of the temporal neural network model. A context vector representing the dynamic information of the current time segment is generated according to the time interval. Based on the context vector, the residual correction, slow drift of aerodynamic synthesis parameters, and uncertainty measure are generated synchronously. The residual correction represents a small additive correction to the nominal position and nominal velocity; the slow drift of aerodynamic synthesis parameters represents the smooth evolution of the aerodynamic synthesis parameters in adjacent time intervals; and the uncertainty measure represents the scale of uncertainty in the prediction at the current time.

[0046] Specifically, the augmented state update first uses synthetic acceleration to obtain the augmented state at time k of the prior state. Then, based on the augmented state of the previous time step, z... k-1 =[r k-1 ,v k-1 ,β k-1 [] represents the augmented state at time k-1. The residual correction of the temporal neural network at time k is added to the residual correction of the prior state. The slow drift of the aerodynamic synthesis parameters of the prior state is addressed by implementing small-step updates and physical upper and lower bound pruning, based on the uncertainty metric s of the temporal neural network model. k Adjust the process noise intensity of the current step to form the process noise covariance matrix, as shown in the formula:

[0047] noise k =noise base ·g(s k ),

[0048] noise k Let g(s) represent the process noise covariance matrix. k ) represents a monotonically amplifying function, noise base This represents the baseline process noise covariance matrix, without any adaptive adjustments.

[0049] The specific steps for mapping and calibrating confidence intervals and calculating bias to trigger adaptation are as follows:

[0050] At the current computation time, read the uncertainty measure s from the output of the temporal neural network. k , will s k According to u k =as k +b restores the data to the training numerical domain, where a and b are the linear scaling factor and the linear bias term, respectively, and u k Represents an integer metric, which is transformed by a mapping function such as softplus. k The original standard deviation σ k A target confidence level g is preset, and the corresponding basis vectors obtained by projecting the position and velocity of the augmented state are denoted as the current point prediction p. k , and the mapped standard deviation σ k Construct the original interval [p k ±z g u k ], where z g The critical multiple of the standard normal distribution is represented by g, which represents the target confidence level. The scaling factor is obtained by the ratio of the weighted quantile of the standardized residual to the Gaussian critical value. The interval is calibrated. At each target confidence level, the empirical coverage and coverage deviation are calculated. At the same time, the comprehensive deviation index and the covariance entropy growth rate are calculated. When the preset threshold is exceeded, adaptive measures are automatically triggered. Late samples are given a decay weight according to the lag time. The upper and lower bounds and slight smoothing are set for key parameters. Each calibration and adaptation is recorded in the operation log to ensure numerical stability, controllable coverage, and process traceability.

[0051] Specifically, the scaling factor is obtained by comparing the weighted quantile of the standardized residuals with the Gaussian critical value. First, the standardized residuals of each sample are calculated. The absolute value of the difference between the true value and the predicted point is divided by the corresponding standard deviation. The standardized residuals of each sample are sorted in ascending order, and their corresponding weights are accumulated. The residual value corresponding to the first time the accumulated weight reaches 1-g is set as the right-hand weighted quantile. This weighted quantile is then compared with the Gaussian critical factor z. g In contrast, the ratio is used as a scaling factor to calibrate the current time interval proportionally.

[0052] The specific steps for assessing risks and determining tasks, and iteratively updating the state prediction process are as follows:

[0053] The acquired state prediction and uncertainty information, real-time observation data, and environmental service data from external sources are used to pre-set risk sets for each indicator. Risk measures are calculated based on the data and compared with the set thresholds and priority rules. If the risk measures exceed the threshold, corresponding judgment actions are executed according to the trigger level. At the same time, the triggering reasons and the proposed actions are written into the operation log and reported. The parameters in the time-series neural network model are updated in small steps using newly acquired observations and telemetry, and the state prediction process is rerun to form a closed loop iteration until the risk is reduced to an acceptable range.

[0054] Specifically, the risk metric is calculated based on a pre-defined risk set of indicators. For each predefined risk indicator, such as communication interruption or missed task time window, the required point prediction and covariance of the risk indicator are first obtained. The corresponding violation event of the risk indicator is defined as a subset in the state space using an inequality. Under the assumption that the prediction error is approximately a multivariate normal distribution, the probability of the violation event is calculated using a fast numerical approximation. The probability is compared with a threshold and weighted and aggregated to form the final risk score.

[0055] like Figure 2 This is a comparison chart of the effects of an artificial intelligence-based method for predicting the orbital state of low-Earth orbit satellites. The table on the left of the chart lists the original indicators used for comparison, including position error, 95% confidence interval coverage deviation, and calculation delay per step. The bar chart on the right shows the scores after standardizing these original indicators for horizontal comparison. The average of each method in five aspects is calculated as the comprehensive score.

[0056] Example 2, based on Example 1 above, describes the application of an AI-based method for predicting the orbital state of low-Earth orbit satellites in short-term navigation scenarios where GNSS observations are interrupted. The specific scheme is as follows:

[0057] Step 1: Using the most recent valid GNSS positioning result as the initial state, including position and velocity, obtain the atmospheric density for this step by looking up a table or interpolating based on the current position and the atmospheric linear velocity under the Earth-fixed system. Combine the three vectors of gravity, air resistance, and weak disturbance to obtain the acceleration. Obtain the nominal orbital state sequence by the step size of the time window. Combine the nominal potential velocity at each moment with the corresponding environmental quantity, the combined acceleration value and the initial aerodynamic parameters to form the augmented state vector. During the GNSS interruption, the nominal state sequence serves as the main baseline for short-term prediction until the new observation arrives.

[0058] Step 2: At each calculation time, the historical segments with equal intervals of time window step are assembled into an input sequence, including nominal velocity, position, and the deviation between nominal and observed values. The residual correction, synthetic parameter drift, and uncertainty scale are obtained through previous inference. The augmented state of the posterior at the previous time is advanced by synthetic acceleration to obtain the augmented state of the prior. Then, the residual output by the network is added to the prior to obtain the corrected prior. The posterior is solidified and the time index is moved forward. The above steps are repeated until the required prediction period is covered or GNSS is recovered.

[0059] Step 3: At each time step, after destandardization and linear shaping of the uncertainty scale of the network output, apply monotonic positive value mapping to obtain the original standard deviation. Through weighted median estimation and shifting the interval center, calculate the standardized residual for each sample, and obtain the right-hand weighted quantile value according to the normalized time weight. Continuously calculate the coverage deviation, comprehensive deviation index, and covariance entropy growth rate. When any index continuously exceeds the preset threshold, trigger adaptive measures.

[0060] Step four involves calculating a series of risk metrics based on calibrated point estimates and covariance, and determining the task response accordingly. Each risk probability is weighted according to task priority or compared to a preset threshold using the maximum value. If the threshold is exceeded, a response action is automatically selected based on the trigger level, and the action and trigger reason are recorded. After the action is executed or manually confirmed, the system re-runs the steps to form a closed loop, using new observations and telemetry data as input to adjust the model parameters.

[0061] The embodiments of the present invention described above are subject to modification and change of method by those skilled in the art without departing from the embodiments and broader aspects of the present invention. The appended claims are intended to include all such modifications and changes of method that do not depart from the present invention.

Claims

1. A method for predicting a low earth orbit satellite orbit state based on artificial intelligence, the method comprising: The method comprises the following steps: calculating a nominal orbit state sequence, constructing an augmented state vector, dividing the predicted orbit into time window steps, calculating the synthetic acceleration of each step, inputting the initial position and velocity as force input, obtaining the nominal orbit state sequence, and constructing the augmented state vector with the air dynamic synthesis parameter; inputting a time sequence neural network model and rolling updating the augmented state, inputting the nominal orbit state sequence, environmental quantity, observation and nominal difference into the time sequence neural network model, and rolling updating the augmented state according to the output; mapping and calibrating the confidence interval, calculating the deviation trigger self-adaption, obtaining the original standard deviation by mapping the uncertainty metric, calculating the standardized residual, calculating the right weighted quantile value according to the normalized time weight, comparing with the preset threshold, and performing self-adaptive deviation adjustment; evaluating the risk and task judgment, and iteratively updating the state prediction process, comparing the obtained data with the set threshold and priority rule, automatically performing the judgment action on the risk metric, recording the trigger reason and judgment action, and updating the parameters in the time sequence neural network model.

2. The low earth orbit satellite orbit state prediction method based on artificial intelligence according to claim 1, characterized in that: The specific steps of calculating the nominal orbit state sequence and constructing the augmented state vector are as follows: When the GNSS is interrupted, firstly, the starting point and the ending point of the current orbit prediction are determined, the time step is set according to the length of the short-term time window between the starting point and the ending point, the position and velocity of the starting point are taken as the initial state, the time grid with equal intervals from the starting point to the ending point is established, the environmental quantity at each time point on the time grid is prepared, in each time step, the synthetic acceleration is defined in the form of gravity, air resistance and other weak disturbances, which is used as the only force input for propulsion, the position and velocity of the entire time window are updated, the nominal position and velocity sequence arranged in time sequence are obtained, the air dynamic synthesis parameter is introduced, the air dynamic synthesis parameter and the position and velocity are placed in the same state vector to form the augmented state vector.

3. The low earth orbit satellite orbit state prediction method based on artificial intelligence according to claim 2, characterized in that: The specific steps of calculating the nominal orbit state sequence and constructing the augmented state vector further comprise: the height at the current position and velocity is calculated, the current position is unified to the geostationary coordinate system, the normal intersection of the position is calculated based on the earth reference ellipsoid, the height is taken as the normal distance from the current position to the reference ellipsoid surface, the atmospheric density of the current time step is obtained by querying the environmental index according to the height, the atmospheric linear velocity is calculated based on the earth rotation angular velocity and the current position, the relative airflow velocity is obtained by subtracting the atmospheric linear velocity from the current velocity, the resistance acceleration of the current time step is calculated by using the air dynamic synthesis parameter, the relative airflow velocity and the atmospheric density, the synthetic acceleration is obtained by adding the gravity and the weak disturbance, and the formula is: a sum = a g + a d + a weak , where a sum represents the synthetic acceleration, a g represents the gravitational acceleration, a d represents the air resistance acceleration, a weak represents the weak disturbance acceleration.

4. The low earth orbit satellite orbit state prediction method based on artificial intelligence according to claim 2, characterized in that: The specific steps of calculating the nominal orbit state sequence and constructing the augmented state vector further comprise: the initial value of the air dynamic synthesis parameter is set according to the ground calibration value, at the calculation time of each time step, the air resistance component is calculated by calling the air dynamic synthesis parameter according to the atmospheric density obtained by the height and the relative velocity after the earth rotation correction, and the formula is: where a d represents the air resistance component, p represents the atmospheric density at the current altitude, e represents the aerodynamic synthesis parameter, v represents the current speed, r represents the current position, and w e represents the earth rotation angular velocity.

5. The low earth orbit satellite orbit state prediction method based on artificial intelligence according to claim 1, characterized in that: The specific steps of inputting the time sequence neural network model and rolling updating the augmented state are as follows: The nominal position and nominal velocity obtained based on the synthetic acceleration propulsion are spliced into an input sequence in time sequence with the environmental quantity at the same time, and the difference between the nominal value and the observation value at the corresponding time, and input to a time sequence neural network model. The output is expressed by [Δr k ,Δv k ,Δβ k ,s k ], wherein k represents the current time, Δr k represents the position residual correction value, Δv k represents the velocity residual correction value, Δβ k represents the slow drift of the air force synthesis parameter, s k represents the uncertainty measure, and the augmented state is updated according to the output vector.

6. The low earth orbit satellite orbit state prediction method based on artificial intelligence according to claim 5, characterized in that: The specific steps of the input time sequence neural network model, the rolling update and the extended state further include: The input sequence is normalized and standardized, the sliding window length and the step are set, the input sequence is divided into continuous time segments as the time sequence neural network model processing unit, the time segments are input into the time sequence encoding unit of the time sequence neural network model in time sequence, the context vector representing the dynamic information of the current time segment is generated according to the time, and the residual correction amount, the slow drift amount of the aerodynamic synthesis parameter and the uncertainty measure are synchronously generated based on the context vector; Based on the augmented state of the last time as the starting point, the augmented state of k-1 time is represented as z k-1 = [r k-1 , v k-1 , β k-1 ] , the residual correction of the time sequence neural network at k time is added to the residual correction of the prior state, the slow drift of the aerodynamic synthetic parameter of the prior state is updated by a small step and the physical upper and lower bounds are clipped, and the process noise intensity of the current step is adjusted according to the uncertainty measurement s k of the time sequence neural network model to form the process noise covariance matrix, and the formula is: noise k = noise base • g(s k ), where noise k represents the process noise covariance matrix, g(s k ) represents a monotonic amplification function, noise base represents the baseline process noise covariance matrix, which is not adapted at all.

7. The low earth orbit satellite orbit state prediction method based on artificial intelligence according to claim 1, characterized in that: The specific steps of the mapping and calibrating confidence interval and calculating deviation trigger adaptation are: The uncertainty measure s is read from the output of the temporal neural network at the computation time. k , will s k According to u k =as k +b restores the data to the training numerical domain, where a and b are the linear scaling factor and the linear bias term, respectively, and u k Represents an integer metric, which is transformed through a mapping function. k The original standard deviation σ k A target confidence level g is pre-set to calculate the projection of the augmented state's position and velocity at time step p, and the corresponding basis vectors obtained are denoted as the current point prediction p. k , and the mapped standard deviation σ k Construct the original interval [p k ±z g u k ], where z g The critical multiple of the standard normal distribution is represented by , g represents the target confidence level, and the scaling factor is obtained by the ratio of the weighted quantile of the standardized residuals to the Gaussian critical value. The interval is calibrated, and at each target confidence level, the empirical coverage and coverage deviation are calculated. At the same time, the comprehensive deviation index and the covariance entropy growth rate are calculated.

8. The low earth orbit satellite orbit state prediction method based on artificial intelligence according to claim 7, characterized in that: The specific steps of the mapping and calibrating confidence interval and calculating deviation trigger adaptation further include: The standardized residual of each sample is calculated by dividing the absolute value of the difference between the true value and the point prediction by the corresponding standard deviation. The standardized residuals of the samples are sorted in ascending order, and the cumulative weight corresponding to each sample is calculated. When the cumulative weight first reaches 1-g, the residual value corresponding to the cumulative weight is set as the right weighted quantile value. The Gaussian critical multiple z g In contrast, the ratio is taken as the scaling coefficient, which is used to proportionally calibrate the interval constructed at the current time. 9.The low earth orbit satellite orbit state prediction method based on artificial intelligence according to claim 1, wherein: The specific steps of the risk evaluation and task decision and iterative update of state prediction process are: The obtained state prediction and uncertainty information, real-time observation data, and environment service data from the outside are used to set the risk set of each index, the risk measure is calculated according to the data, the risk measure is compared with the set threshold value and priority rule, if the risk measure exceeds the threshold value, the corresponding decision action is executed according to the trigger level, the trigger reason and the action to be taken are written into the operation log and reported, the newly obtained observation and telemetry are used to update the parameters in the time sequence neural network model in small steps and the state prediction process is run again to form a closed loop iteration until the risk is reduced to an acceptable range.

10. The low earth orbit satellite orbit state prediction method based on artificial intelligence according to claim 9, characterized in that: The specific steps of the risk evaluation and task decision and iterative update of state prediction process further include: For each pre-defined risk index, the required point prediction and covariance of the risk index are obtained, the corresponding violation event of the risk index is defined as a subset in the state space by using inequality, the probability of the violation event is calculated by using fast numerical approximation on the premise of assuming that the prediction error is approximately multivariate normal distribution, the probability is compared with the threshold value and weighted aggregation to form the final risk score.

Images