CN117454963A - SGP4 model precision improvement method and system based on GA-BP neural network - Google Patents

SGP4 model precision improvement method and system based on GA-BP neural network Download PDF

Info

Publication number
CN117454963A
CN117454963A CN202311407886.7A CN202311407886A CN117454963A CN 117454963 A CN117454963 A CN 117454963A CN 202311407886 A CN202311407886 A CN 202311407886A CN 117454963 A CN117454963 A CN 117454963A
Authority
CN
China
Prior art keywords
neural network
activity index
atmospheric resistance
resistance modulation
modulation coefficient
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202311407886.7A
Other languages
Chinese (zh)
Inventor
黄卫权
田露
王毅博
丛榕
刘妍
黄昊钰
王然
王振旭
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Harbin Engineering University
Original Assignee
Harbin Engineering University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Harbin Engineering University filed Critical Harbin Engineering University
Priority to CN202311407886.7A priority Critical patent/CN117454963A/en
Publication of CN117454963A publication Critical patent/CN117454963A/en
Pending legal-status Critical Current

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/084Backpropagation, e.g. using gradient descent
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00Computer-aided design [CAD]
    • G06F30/20Design optimisation, verification or simulation
    • G06F30/27Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00Computer-aided design [CAD]
    • G06F30/20Design optimisation, verification or simulation
    • G06F30/28Design optimisation, verification or simulation using fluid dynamics, e.g. using Navier-Stokes equations or computational fluid dynamics [CFD]
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/048Activation functions
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/086Learning methods using evolutionary algorithms, e.g. genetic algorithms or genetic programming
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F2113/00Details relating to the application field
    • G06F2113/08Fluids
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/14Force analysis or force optimisation, e.g. static or dynamic forces
    • YGENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02TCLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T90/00Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Geometry (AREA)
  • Computer Hardware Design (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Physiology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Algebra (AREA)
  • Fluid Mechanics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses an SGP4 model precision improvement method and system based on a GA-BP neural network, and relates to the technical field of machine learning. The technical key points of the invention include: acquiring historical flight data and a corresponding space activity index of a satellite, and acquiring an optimal atmospheric resistance modulation coefficient; the historical flight data, the spatial activity index and the optimal atmospheric resistance modulation coefficient form a training data set; constructing a BP neural network, optimizing the BP neural network by utilizing a genetic algorithm, and inputting a training data set into the BP neural network optimized by the genetic algorithm for training; and inputting the real-time satellite flight data and the spatial activity index into a trained prediction model for prediction, and obtaining an atmospheric resistance modulation coefficient prediction result. And replacing the atmospheric resistance modulation coefficient in the TLE data with the atmospheric resistance modulation coefficient prediction result so as to use the new TLE data for satellite orbit prediction. The invention effectively improves the forecasting precision and the robustness of the SGP4 orbit forecasting model.

Description

SGP4 model precision improvement method and system based on GA-BP neural network
Technical Field
The invention relates to the technical field of machine learning, in particular to an SGP4 model precision improvement method and system based on a GA-BP neural network.
Background
The SGP4 orbit prediction model is a satellite orbit prediction model issued by North America air defense united commander (NORAD), simplifies complex perturbation items involved in satellite orbit prediction, considers the influence of four perturbation forces of global non-spherical perturbation, atmospheric resistance, solar pressure and solar and lunar attraction on orbit prediction, and is mainly used for near-earth target orbit prediction with a period less than 225 min. The SGP4 orbit prediction model is a space analysis model, and the principle is to reconstruct TLE data and then predict the orbit by the perturbation of short-period long-period items and long-period items, and the SGP4 orbit prediction model is one of the most widely used satellite orbit prediction models at present due to the characteristics of convenient data acquisition and high calculation speed. However, the SGP4 orbit prediction model has a large error, especially for satellites with orbit heights below 500km, the average prediction error is about 5km, and when space activity is active, the maximum error can be above 20 km. It is therefore practical to make improvements to the SGP4 model using appropriate methods.
At present, a plurality of students at home and abroad carry out improvement on the SGP4 model in the precision direction. The literature 'microsatellite orbit prediction research based on GPS and SGP4 model' fits the data measured by the GPS into double-row metadata, and the instantaneous root number is used as an initial value, so that the fitting error of the double-row metadata is reduced. Literature on research on satellite Doppler frequency shift compensation method based on SGP4 model, correction compensation is carried out on ground station by utilizing Doppler frequency shift, and forecasting result is superior to that of the traditional SGP4 model. However, both methods require re-compensation calculation each time, and have certain limitations. The document NORAD TLE Conversion from Osculating Orbital Element and the document Examination ofSGP along-track errors for initially circular orbits both mention that selecting a suitable atmospheric density coefficient B can effectively improve the prediction accuracy of the SGP4 model, but neither provides an efficient selection method. The document 'resistance coefficient self-adaptive modulation method based on precise ephemeris' uses a clustering algorithm to complete self-adaptive adjustment of the atmospheric density coefficient.
However, the disadvantage of the improvement of the SGP4 orbit prediction model is that the orbit prediction accuracy of the SGP4 orbit prediction model for the low orbit satellite is not high, so that it is urgent how to improve the SGP4 orbit prediction model in both the accuracy and the speed to meet the requirement of the orbit prediction of the low orbit networking satellite.
Disclosure of Invention
Therefore, the invention provides an SGP4 model precision improvement method and system based on a GA-BP neural network, so as to improve the orbit forecasting precision of a low orbit satellite.
According to an aspect of the present invention, there is provided an SGP4 model accuracy improvement method based on a GA-BP neural network, the method including the steps of:
step 1, acquiring historical flight data of a satellite and a corresponding space activity index, and acquiring an optimal atmospheric resistance modulation coefficient; the historical flight data, the spatial activity index and the optimal atmospheric resistance modulation coefficient form a training data set;
step 2, constructing a BP neural network, optimizing the BP neural network by utilizing a genetic algorithm, inputting the training data set into the BP neural network optimized by the genetic algorithm for training, and obtaining a prediction model of an atmospheric resistance modulation coefficient;
and step 3, inputting real-time satellite flight data and the spatial activity index into a trained prediction model to predict, and obtaining an atmospheric resistance modulation coefficient prediction result.
Further, the method further includes step 4 of replacing the atmospheric resistance modulation coefficient in the TLE data with the atmospheric resistance modulation coefficient prediction result to perform satellite orbit prediction using the TLE data including the atmospheric resistance modulation coefficient prediction result.
Further, the historical flight data in the step 1 comprises six orbits of the satellite, wherein the six orbits comprise an orbit semi-long axis, an orbit eccentricity, an orbit inclination angle, an ascending intersection point right ascent, a near-place amplitude angle and a flat-near point angle; the spatial activity index includes a solar activity index F10.7 and a geomagnetic activity index AP.
Further, parameters of the BP neural network constructed in step 2 include: input-output dimension, number of neural network layers, number of hidden layer neurons, initial weight and threshold, activation function, learning rate, momentum factor and global error function.
Further, the input dimension parameters of the BP neural network constructed in the step 2 comprise six tracks in TLE data of the previous day, a solar activity index and a geomagnetic activity index before two days of forecast, a solar activity index and a geomagnetic activity index before one day of forecast, and an optimal atmospheric resistance modulation coefficient before one day of forecast; the output dimension parameter is an atmospheric resistance modulation coefficient to be predicted; the number of the neural network layers is three; the number of hidden layer neurons is calculated according to the input and output dimensions.
Further, the initial weight and the threshold value of the BP neural network are optimized by utilizing a genetic algorithm, and the optimization process comprises the following steps:
initializing corresponding populations according to a coding scheme; calculating individual fitness according to the fitness function; the fitness function is as follows:
wherein k represents an adjustment coefficient; n represents the total number of training sets; t is t i True data representing the ith sample; x is x i A predicted value representing an i-th sample;
selecting, crossing and mutating; generating a new generation population; judging whether the maximum iteration times are reached, if the maximum iteration times are reached, giving output to the BP neural network, and if the maximum iteration times are not satisfied, returning to the step of calculating the individual fitness.
Further, the selecting, crossing and mutating process comprises the following steps:
1) Selecting excellent individuals from the population according to the following probability formula:
wherein P is i Representing the probability that the ith individual is selected;F i representing individual fitness values; n represents the population size;
2) Crossover operations are performed to create new individuals according to the following formula:
wherein a is mk ,a nk Respectively represent the mth individual a m And the nth individual a n Crossing at random position k; b represents [0,1]]Random numbers in between;
3) The mutation operation is performed according to the following formula:
wherein a is max Representing the maximum value of the code; a, a min Representing the minimum value of the code; b, r represents [0,1]]Random numbers in between; numdd represents the number of iterations; max GA represents the maximum iteration number.
According to another aspect of the present invention, there is provided an SGP4 model accuracy improvement system based on a GA-BP neural network, the system including:
the data acquisition module is configured to acquire historical flight data of the satellite and a corresponding space activity index, and acquire an optimal atmospheric resistance modulation coefficient; the historical flight data, the spatial activity index and the optimal atmospheric resistance modulation coefficient form a training data set; the historical flight data comprises six orbits of the satellite, wherein the six orbits comprise an orbit semi-long axis, an orbit eccentricity, an orbit inclination angle, an ascending intersection point right ascent, a near-place amplitude angle and a near-plane point angle; the spatial activity index comprises a solar activity index F10.7 and a geomagnetic activity index AP;
the model training module is configured to construct a BP neural network, optimize the BP neural network by utilizing a genetic algorithm, input the training data set into the BP neural network optimized by the genetic algorithm for training, and acquire a prediction model of an atmospheric resistance modulation coefficient; parameters of the BP neural network constructed include: input/output dimension, number of neural network layers, number of hidden layer neurons, initial weight and threshold, activation function, learning rate, momentum factor and global error function;
the prediction module is configured to input real-time satellite flight data and a spatial activity index into a trained prediction model for prediction, and obtain an atmospheric resistance modulation coefficient prediction result.
Further, the system also includes a satellite orbit prediction module configured to replace the barometric pressure modulation factor in TLE data with the barometric pressure modulation factor prediction result to perform satellite orbit prediction using TLE data containing the barometric pressure modulation factor prediction result.
Further, the model training module optimizes the initial weight and the threshold value of the BP neural network by using a genetic algorithm, and the optimization process comprises the following steps:
initializing corresponding populations according to a coding scheme; calculating individual fitness according to the fitness function; the fitness function is as follows:
wherein k represents an adjustment coefficient; n represents the total number of training sets; t is t i True data representing the ith sample; x is x i A predicted value representing an i-th sample;
selecting, crossing and mutating; generating a new generation population; wherein the selecting, crossing and mutating processes comprise:
1) Selecting excellent individuals from the population according to the following probability formula:
wherein P is i Representing the probability that the ith individual is selected;F i representing individual fitness values; n represents the population size;
2) Crossover operations are performed to create new individuals according to the following formula:
wherein a is mk ,a nk Respectively represent the mth individual a m And the nth individual a n Crossing at random position k; b represents [0,1]]Random numbers in between;
3) The mutation operation is performed according to the following formula:
in the method, in the process of the invention,a max representing the maximum value of the code; a, a min Representing the minimum value of the code; b, r represents [0,1]]Random numbers in between; numdd represents the number of iterations; max GA represents the maximum iteration number;
judging whether the maximum iteration times are reached, if the maximum iteration times are reached, giving output to the BP neural network, and if the maximum iteration times are not satisfied, returning to the step of calculating the individual fitness.
The beneficial technical effects of the invention are as follows:
the invention provides an SGP4 orbit prediction model improvement method and system based on a GA-BP neural network, wherein the GA-BP neural network is used as a model training algorithm to establish a prediction model of an atmospheric resistance modulation coefficient B, and the predicted coefficient is used for replacing the atmospheric resistance modulation coefficient in original double-row metadata to improve the prediction precision of the SGP4 orbit prediction model. Finally, the feasibility of the scheme is verified through simulation results of the satellite 37820. As shown by simulation results, the invention can improve the average track prediction accuracy of the SGP4 track prediction model by more than 40%. Compared with the prior art, the method and the device effectively improve the forecasting precision of the SGP4 orbit forecasting model, and the forecasting model can be used for a long time after training. The invention also has the advantages of clear and easily understood logic, high flexibility, strong robustness and the like.
Drawings
The invention may be better understood by reference to the following description taken in conjunction with the accompanying drawings, which are included to provide a further illustration of the preferred embodiments of the invention and to explain the principles and advantages of the invention, together with the detailed description below.
Fig. 1 is a flowchart of an SGP4 model accuracy improvement method based on a GA-BP neural network according to an embodiment of the present invention;
FIG. 2 is a BP neural network structure and neuron propagation diagram;
FIG. 3 is a partial flow chart of a BP neural network;
FIG. 4 is a flow chart of a GA-BP neural network according to an embodiment of the present invention;
FIG. 5 is a flowchart of a genetic algorithm according to an embodiment of the present invention;
FIG. 6 is a diagram showing a comparison between a GA-BP neural network and a conventional BP neural network according to an embodiment of the present invention;
FIG. 7 is a graph showing the prediction of the atmospheric resistance modulation factor of the GA-BP neural network according to an embodiment of the present invention;
FIG. 8 is a graph of the orbit prediction result using the prediction value of the GA-BP neural network according to the embodiment of the present invention.
Detailed Description
In order that those skilled in the art will better understand the present invention, exemplary embodiments or examples of the present invention will be described below with reference to the accompanying drawings. It is apparent that the described embodiments or examples are only implementations or examples of a part of the invention, not all. All other embodiments or examples, which may be made by one of ordinary skill in the art without undue burden, are intended to be within the scope of the present invention based on the embodiments or examples herein.
The invention provides an SGP4 model precision improvement method and system based on a GA-BP neural network, which are used for improving the forecasting precision of an SGP4 orbit forecasting model. The overall realization idea of the GA-BP neural network is as follows: firstly, determining a network architecture of a traditional BP neural network, wherein the network architecture comprises input and output dimensions, the number of layers of the neural network, an activation function, a learning rate, a momentum factor, the maximum iteration number and the like; according to the structure of the traditional BP neural network, using a GA algorithm to perform optimizing operation on an initial weight and a threshold value in the BP neural network; and finally, further training the BP neural network by using the optimal weight and the threshold value obtained by the GA algorithm, and finishing training after the training finishing condition is met.
Errors of the SGP4 orbit prediction model are mainly divided into input double-row metadata errors and perturbation modeling errors in the model. The perturbation modeling error is an inherent error when the SGP4 orbit prediction model is built, and the perturbation model used in the SGP4 orbit prediction model is accurate, so that the error of the part cannot be eliminated; the error of the double-row metadata mainly comes from two parts, one part is the error of the measurement data for fitting the double-row metadata, and the other part is the error of the fitting algorithm.
The main error of the SGP4 orbit prediction model comes from the speed direction, and the prediction accuracy of the vertical orbit surface and the orbit radius direction is higher, and the error is smaller, so the error of the speed direction is the main factor causing the error of the SGP4 orbit prediction model. Meanwhile, by combining with the orbit perturbation analysis of the low-orbit satellite, the perturbation which has the greatest influence on the orbit forecast of the low-orbit satellite is the global aspheric perturbation and the atmospheric resistance, and the SGP4 orbit forecast model is more accurate by using the global aspheric perturbation model, and cannot be effectively optimized and improved under the condition of not changing the model. The use of double row metadata for track prediction does not take into account that the atmospheric density may cause a change in atmospheric resistance during the prediction process, so it can be deduced that the atmospheric resistance modulation factor B in the double row metadata may be the cause of the velocity direction error.
Therefore, the invention takes the atmospheric resistance modulation coefficient B in the double-row metadata as an access point, and improves the low-orbit networking satellite orbit prediction expansion precision based on the SGP4 model.
The embodiment of the invention provides an SGP4 model precision improvement method based on a GA-BP neural network, which comprises the following steps:
step 1, acquiring historical flight data of a satellite and a corresponding space activity index, and acquiring an optimal atmospheric resistance modulation coefficient; the historical flight data, the spatial activity index and the optimal atmospheric resistance modulation coefficient form a training data set;
step 2, constructing a BP neural network, optimizing the BP neural network by utilizing a genetic algorithm, inputting the training data set into the BP neural network optimized by the genetic algorithm for training, and obtaining a prediction model of an atmospheric resistance modulation coefficient;
step 3, inputting real-time satellite flight data and a spatial activity index into a trained prediction model to predict, and obtaining an atmospheric resistance modulation coefficient prediction result;
and 4, replacing the atmospheric resistance modulation coefficient in the TLE data with the atmospheric resistance modulation coefficient prediction result so as to use the TLE data containing the atmospheric resistance modulation coefficient prediction result to carry out satellite orbit prediction.
Fig. 1 shows a flowchart of an SGP4 model accuracy improvement method based on a GA-BP neural network according to an embodiment of the present invention.
The method starts in step 1. Firstly, step 1 is executed to obtain historical flight data of satellites and corresponding space activity indexes. Wherein the historical flight data of the satellite includes six orbits of the satellite; the spatial activity index comprises a solar activity index F10.7 and a geomagnetic index AP; and obtaining an optimal atmospheric resistance modulation coefficient B0 in TLE data through simulation. The historical flight data, the spatial activity index and the optimal atmospheric resistance modulation coefficient form a training data set.
According to the embodiment of the invention, the six track numbers comprise a track semi-long axis, a track eccentricity, a track inclination angle, an ascending intersection point right ascent, a near-place amplitude angle and a flat near-point angle. The atmospheric resistance modulation coefficient B in the double-row metadata (TLE data) of the day before forecasting is changed within the range of 0.1-5B, the step length is 0.01B, the atmospheric resistance modulation coefficient simulating the minimum satellite orbit change is denoted as B0, and the minimum orbit change refers to the orbit with the minimum semi-long axis attenuation in one period. Then, the relationship between B0 and the solar activity index f10.7 and the geomagnetic activity index Ap is compared.
And then, executing the step 2, constructing a BP neural network, optimizing the BP neural network by utilizing a genetic algorithm, inputting the training data set into the BP neural network optimized by the genetic algorithm for training, and obtaining a prediction model of the atmospheric resistance modulation coefficient.
According to the embodiment of the invention, the error back propagation neural network is composed of a plurality of layers of neural networks, namely BP neural networks, and each layer of neural network internally comprises a plurality of neurons as shown in figure 2. Wherein x is 1 ~x n Representing the input signal quantity omega 0 ~ω n The weights representing the input signal values, b, are model bias parameters. The input signals are input into the calculated output quantity of the neuron after weight and paranoid, and the output quantity is subjected to nonlinear mapping through Activation Function (transfer/activation function) to obtain fitted output quantityOutput signal, y 0 Is the output of the current neuron, which propagates down as the input to the next layer of neural network.
The BP neural network is a model trained in a supervised learning mode, continuously learns based on an input training set, outputs a target fitting value, is an input-output mapping model without establishing complex constraint conditions, and when the input training set is provided to the neural network as the training set, data is transmitted from an input layer to an output layer through a hidden layer by neuron fitting and weighting. If the actual output deviates from the expected output, the BP neural network can perform error back propagation from the output layer to the input layer based on the idea of gradient descent, and in the process, the neurons can adjust weights and neuron thresholds according to the feedback error values so as to gradually reduce the errors. Along with the input of the input set, the error of the output value and the expected value of the network is continuously reduced, the accuracy is continuously increased, and finally the expected optimal model can be reached.
As shown in fig. 3, taking historical flight data of satellites for one year as a data set of the neural network, and determining the input and output dimensions of the BP neural network; after determining the input and output dimensions of the BP neural network, selecting parameters of the BP neural network, wherein the parameters comprise the number of layers of the neural network, the number of neurons of an hidden layer, the initial weight and threshold value of the network, the learning rate, an activation function and a momentum factor; and (3) completing the construction of the BP neural network, and evaluating the BP neural network by using an evaluation method. The evaluation method refers to a global error function, the global error E is the sum of squares of differences between the expected output value and the actual output value, the smaller the E value is, the smaller the difference between the predicted value and the expected value is, the better the model training effect is, and the worse the model training effect is proved to be. The specific selection method of the parameters is as follows:
1) Input-output dimension of BP neural network: optimal atmospheric resistance modulation coefficient B obtained by simulation according to correlation 0 * With a characteristic of 1-2 days in advance, 11 input parameters are determined, namely six tracks in TLE data of the previous day and a solar activity index F two days before the forecast day 10.7 And geomagnetic activity index Ap, solar activity before one dayIndex F 10.7 And geomagnetic activity index Ap, and optimal atmospheric resistance modulation coefficient B before one day 0 * All the parameters can be obtained before the prediction is started, and the dimension of the neuron of the output layer is 1, namely the predicted atmospheric resistance modulation coefficient;
2) Layer number of neural network: practice proves that any rational function can be fitted through a three-layer BP neural network. In order to reduce fitting time and avoid the problem of complex structure caused by a multi-layer network when the network is actually constructed, the number of neurons in an implicit layer is generally more prone to be added instead of the number of layers of the neural network;
3) Hidden layer neuron number: the evaluation of the neural network model is generally carried out from two aspects of output precision and fitting time, the proper number of neurons in an hidden layer can effectively reduce the output fitting time and output errors, the small number of neurons can cause insufficient fitting precision, the large number of neurons can cause complex model structure and reduced generalization capability. The general formula for the number of neurons is as follows:
wherein: n represents the input layer dimension; m represents the output layer dimension; a represents a paranoid integer, and is generally 1 to 10.
4) Initial weights and thresholds: the initial weight and the threshold are initial values of model training, and proper initial values are selected to shorten the fitting time of the model training and prevent the model from falling into local optimum, wherein the weight exists between neural network layers, the threshold exists between neurons, the two weights are set randomly, so that the equidirectional adjustment of each step of weight can be prevented, the initial set value is set to be a random number between [ -1,1] or adjusted according to the input characteristic number, and the initial set value can be a random number of [ -2.4/n,2.4/n ]. Wherein n is the number of input features, the initial weight and the threshold value of the network are selected by adopting a GA algorithm, and then the initial weight and the threshold value are given to the neural network;
5) Activation function: the common activation functions include log sig, tan sig and purlin, and the function of the activation functions is to map an input signal into an output value through a threshold value, the range of the input signal is from minus infinity to plus infinity, the range of the output signal is different according to different activation function ranges, and finally, the nonlinear mapping of the signal is realized, wherein the specific formula is as follows:
log function:
tan sig function:
purelin function:
f(x)=x
6) Learning rate eta: during the training of the model, the weights are dynamically adjusted based on the learning rate. Proper learning compensation can improve the stability of the model, reduce the training time of the model, and generally, the eta value is generally between [0,1] during initialization in order to ensure the stability of the model;
7) Momentum factor mc: mc is a main factor for determining the convergence rate of the neural network, and generally has a value not greater than the learning rate eta;
8) Global error function E: the global error E is the sum of squares of differences between the expected output value and the actual output value, the smaller the E value is, the smaller the difference between the predicted value and the expected value is, the better the model training effect is, and the worse the model training effect is proved, three different BP neural network models are formed aiming at the three activation functions, the model training test is carried out by utilizing the data set to finish tuning the activation functions, and the activation function with the best effect is selected to establish the BP neural network model.
However, the conventional BP neural network adopts a gradient decreasing mode for adjusting the weight and the threshold value during training, and the previous gradient direction is not considered, so that the problem that the convergence speed is low and local optimization is easy to fall into is unavoidable.
Therefore, this phenomenon is optimized by combining a global search algorithm-Genetic Algorithm (GA) on the basis of the conventional BP neural network, and fig. 4 is a flowchart of the GA-BP neural network in this embodiment; the genetic algorithm part flow is shown in fig. 5.
According to the embodiment of the invention, as shown in fig. 5, a genetic algorithm is adopted to optimize the initial weight and the threshold value of the BP neural network, specifically:
step 21: initializing a corresponding population P (t) according to a coding scheme;
step 22: generating a suitable individual Fitness Fitness (i) by a Fitness function, the Fitness representing the advantage or disadvantage of an individual in the population, the probability of the individual inheriting the next generation in a selected operation being determined in accordance with the Fitness;
step 23: selecting, crossing and mutating;
step 24: generating a new generation population;
step 25: judging whether the maximum iteration number is reached, if the maximum iteration number is reached, ending the algorithm, assigning the output to the BP neural network, and if the maximum iteration number is not satisfied, repeating the step 22.
The fitness function in step 22 is:
wherein: k represents an adjustment coefficient; n represents the number of training set samples; t is t i True data representing the ith sample; y is i Representing the predicted value of the i-th sample.
The rules of selection, crossover and mutation of the genetic algorithm in the step 23 are specifically as follows:
1) Selection of
The main function of the selection operation is to select the optimal individual from the population, typically the higher fitness individual as the parent of the next generation population. Selection of individuals from a population by roulette selection algorithm, due to the fitness of individuals F in the present invention i Is the difference between the predicted and expected values in the BP neural network, so the probability that an individual is selected is inversely proportional to the fitness value,the larger the fitness value, the smaller the probability of being selected and vice versa. The specific operation is as follows:
wherein: f (F) i Representing individual fitness values; n represents the population size; p (P) i Representing the probability that the ith individual is selected.
2) Crossover
Crossover operations of genetic algorithms in order to generate new individuals, the choice of crossover algorithm is largely dependent on the individual's coding scheme, and there is a need to efficiently generate new individuals while retaining excellent individuals during crossover. Because the encoding mode of the invention adopts a real number encoding mode, the use formula is as follows.
Wherein: a, a mk ,a nk Represents the mth individual a m And the nth individual a n Crossing at random position k; b represents [0,1]]Is a random number of (a) in the memory.
3) Variation of
The mutation operation in the genetic algorithm is to prevent the algorithm from being premature, and the algorithm converges to a local optimal solution, so that the diversity of the population can be effectively ensured, and the aim of global optimization is fulfilled. The formula is used as follows:
wherein a is max Representing the maximum value of the code; a, a min Representing the minimum value of the code; b, r represents [0,1]]Random numbers of (a); numdd represents the number of iterations; max GA represents the maximum iteration number.
And then, executing the step 3, and inputting the real-time satellite flight data and the spatial activity index into a trained prediction model to predict, thereby obtaining an atmospheric resistance modulation coefficient prediction result.
And finally, executing step 4, and replacing the atmospheric resistance modulation coefficient in the TLE data with the atmospheric resistance modulation coefficient prediction result so as to use the TLE data containing the atmospheric resistance modulation coefficient prediction result to carry out satellite orbit prediction.
Aiming at the feasibility problem of the method in application, software simulation is performed as follows.
Constructing a BP neural network by using historical flight data and a spatial environment index of a satellite 37820 as data sets, predicting an atmospheric resistance modulation coefficient B in TLE data, and then optimizing the BP neural network by using a genetic algorithm, wherein a comparison diagram of the GA-BP neural network and a traditional BP neural network is shown in FIG. 6; FIG. 7 shows a prediction graph of the atmospheric resistance modulation factor of a GA-BP neural network; FIG. 8 shows a graph of the orbit forecast results using the GA-BP neural network predictors. As shown by the forecasting result, the track is forecasted by using the traditional BP neural network forecasting value, the forecasting days are 76 days, the average error is reduced from 4.0322km to 2.766km, the average progress is improved to 31.4%, the improvement fails in 11 days, and the improvement success rate is 85.3%. The prediction value of the GA-BP neural network is used for track prediction, the prediction days are 76 days, the average error is not corrected and fails, the average error is reduced from 4.0322km to 2.1014km, the average prediction accuracy is improved to 47.8%, the effect is better than that of 31.4% of that of the conventional BP neural network, and no improvement failure exists. The specific improvements are shown in the following table.
Another embodiment of the present invention proposes an SGP4 model accuracy improvement system based on a GA-BP neural network, the system comprising:
the data acquisition module is configured to acquire historical flight data of the satellite and a corresponding space activity index, and acquire an optimal atmospheric resistance modulation coefficient according to the historical flight data and the space activity index in a simulation mode; the historical flight data, the spatial activity index and the optimal atmospheric resistance modulation coefficient form a training data set; the historical flight data comprises six orbits of the satellite, wherein the six orbits comprise an orbit semi-long axis, an orbit eccentricity, an orbit inclination angle, an ascending intersection point right ascent, a near-place amplitude angle and a near-plane point angle; the spatial activity index comprises a solar activity index F10.7 and a geomagnetic activity index AP;
the model training module is configured to construct a BP neural network, optimize the BP neural network by utilizing a genetic algorithm, input the training data set into the BP neural network optimized by the genetic algorithm for training, and acquire a prediction model of an atmospheric resistance modulation coefficient; parameters of the BP neural network constructed include: input/output dimension, number of neural network layers, number of hidden layer neurons, initial weight and threshold, activation function, learning rate, momentum factor and global error function;
the prediction module is configured to input real-time satellite flight data and a spatial activity index into a trained prediction model for prediction, and obtain an atmospheric resistance modulation coefficient prediction result.
Further, the system also includes a satellite orbit prediction module configured to replace the barometric pressure modulation factor in TLE data with the barometric pressure modulation factor prediction result to perform satellite orbit prediction using TLE data containing the barometric pressure modulation factor prediction result.
Further, the model training module optimizes the initial weight and the threshold value of the BP neural network by using a genetic algorithm, and the optimization process comprises the following steps:
initializing corresponding populations according to a coding scheme; calculating individual fitness according to the fitness function; the fitness function is as follows:
wherein k represents an adjustment coefficient; n represents the total number of training sets; t is t i True data representing the ith sample; x is x i A predicted value representing an i-th sample;
selecting, crossing and mutating; generating a new generation population; wherein the selecting, crossing and mutating processes comprise:
1) Selecting excellent individuals from the population according to the following probability formula:
wherein P is i Representing the probability that the ith individual is selected;F i representing individual fitness values; n represents the population size;
2) Crossover operations are performed to create new individuals according to the following formula:
wherein a is mk ,a nk Respectively represent the mth individual a m And the nth individual a n Crossing at random position k; b represents [0,1]]Random numbers in between;
3) The mutation operation is performed according to the following formula:
wherein a is max Representing the maximum value of the code; a, a min Representing the minimum value of the code; b, r represents [0,1]]Random numbers in between; numdd represents the number of iterations; max GA represents the maximum iteration number;
judging whether the maximum iteration times are reached, if the maximum iteration times are reached, giving output to the BP neural network, and if the maximum iteration times are not satisfied, returning to the step of calculating the individual fitness.
The function of the SGP4 model accuracy improvement system based on the GA-BP neural network in the embodiment of the present invention may be described by the foregoing SGP4 model accuracy improvement method based on the GA-BP neural network, so that the system embodiment is not described in detail, and reference may be made to the above method embodiment, which is not described herein.
While the invention has been described with respect to a limited number of embodiments, those skilled in the art, having benefit of the above description, will appreciate that other embodiments are contemplated within the scope of the invention as described herein. The disclosure of the present invention is intended to be illustrative, but not limiting, of the scope of the invention, which is defined by the appended claims.

Claims (10)

1. The SGP4 model precision improvement method based on the GA-BP neural network is characterized by comprising the following steps of:
step 1, acquiring historical flight data of a satellite and a corresponding space activity index, and acquiring an optimal atmospheric resistance modulation coefficient; the historical flight data, the spatial activity index and the optimal atmospheric resistance modulation coefficient form a training data set;
step 2, constructing a BP neural network, optimizing the BP neural network by utilizing a genetic algorithm, inputting the training data set into the BP neural network optimized by the genetic algorithm for training, and obtaining a prediction model of an atmospheric resistance modulation coefficient;
and step 3, inputting real-time satellite flight data and the spatial activity index into a trained prediction model to predict, and obtaining an atmospheric resistance modulation coefficient prediction result.
2. The SGP4 model accuracy improvement method based on GA-BP neural network according to claim 1, further comprising step 4 of replacing the atmospheric resistance modulation coefficients in TLE data with the atmospheric resistance modulation coefficient prediction result to perform satellite orbit prediction using TLE data containing the atmospheric resistance modulation coefficient prediction result.
3. The SGP4 model accuracy improvement method based on GA-BP neural network according to claim 1 or 2, wherein in step 1, the historical flight data includes six orbits of the satellite, the six orbits including a semi-long axis of orbit, an orbit eccentricity, an orbit inclination, an ascending intersection point right ascent, a near-to-earth amplitude, and a near-to-earth angle; the spatial activity index includes a solar activity index F10.7 and a geomagnetic activity index AP.
4. The SGP4 model accuracy improvement method based on GA-BP neural network according to claim 1 or 2, wherein the parameters of the BP neural network constructed in step 2 include: input-output dimension, number of neural network layers, number of hidden layer neurons, initial weight and threshold, activation function, learning rate, momentum factor and global error function.
5. The method for improving the precision of an SGP4 model based on a GA-BP neural network according to claim 4, wherein the input dimension parameters of the BP neural network constructed in the step 2 comprise six tracks in TLE data of the previous day, a solar activity index and a geomagnetic activity index before two days of forecast, a solar activity index and a geomagnetic activity index before one day of forecast, and an optimal atmospheric resistance modulation factor before one day of forecast; the output dimension parameter is an atmospheric resistance modulation coefficient to be predicted; the number of the neural network layers is three; the number of hidden layer neurons is calculated according to the input and output dimensions.
6. The method for improving the precision of an SGP4 model based on a GA-BP neural network according to claim 5, wherein the initial weight and threshold of the BP neural network are optimized by using a genetic algorithm, and the optimization process comprises:
initializing corresponding populations according to a coding scheme; calculating individual fitness according to the fitness function; the fitness function is as follows:
wherein k represents an adjustment coefficient; n represents the total number of training sets; t is t i True data representing the ith sample; x is x i A predicted value representing an i-th sample;
selecting, crossing and mutating; generating a new generation population; judging whether the maximum iteration times are reached, if the maximum iteration times are reached, giving output to the BP neural network, and if the maximum iteration times are not satisfied, returning to the step of calculating the individual fitness.
7. The method for improving the precision of the SGP4 model based on the GA-BP neural network as set forth in claim 6, wherein the selecting, crossing, and mutating steps include:
1) Selecting excellent individuals from the population according to the following probability formula:
wherein P is i Representing the probability that the ith individual is selected;F i representing individual fitness values; n represents the population size;
2) Crossover operations are performed to create new individuals according to the following formula:
wherein a is mk ,a nk Respectively represent the mth individual a m And the nth individual a n Crossing at random position k; b represents [0,1]]Random numbers in between;
3) The mutation operation is performed according to the following formula:
wherein a is max Representing the maximum value of the code; a, a min Representing the minimum value of the code; b, r represents [0,1]]Random numbers in between; numdd represents the number of iterations; maxGA represents the maximum iteration number.
8. An SGP4 model accuracy improvement system based on a GA-BP neural network, comprising:
the data acquisition module is configured to acquire historical flight data of the satellite and a corresponding space activity index, and acquire an optimal atmospheric resistance modulation coefficient; the historical flight data, the spatial activity index and the optimal atmospheric resistance modulation coefficient form a training data set; the historical flight data comprises six orbits of the satellite, wherein the six orbits comprise an orbit semi-long axis, an orbit eccentricity, an orbit inclination angle, an ascending intersection point right ascent, a near-place amplitude angle and a near-plane point angle; the spatial activity index comprises a solar activity index F10.7 and a geomagnetic activity index AP;
the model training module is configured to construct a BP neural network, optimize the BP neural network by utilizing a genetic algorithm, input the training data set into the BP neural network optimized by the genetic algorithm for training, and acquire a prediction model of an atmospheric resistance modulation coefficient; parameters of the BP neural network constructed include: input/output dimension, number of neural network layers, number of hidden layer neurons, initial weight and threshold, activation function, learning rate, momentum factor and global error function;
the prediction module is configured to input real-time satellite flight data and a spatial activity index into a trained prediction model for prediction, and obtain an atmospheric resistance modulation coefficient prediction result.
9. The SGP4 model accuracy improvement system based on a GA-BP neural network of claim 8, further comprising a satellite orbit prediction module configured to replace the atmospheric resistance modulation coefficients in TLE data with the atmospheric resistance modulation coefficient predictions to make satellite orbit predictions using TLE data containing the atmospheric resistance modulation coefficient predictions.
10. The SGP4 model accuracy improvement system based on GA-BP neural network as set forth in claim 8, wherein the model training module optimizes the initial weight and the threshold of the BP neural network by using a genetic algorithm, and the optimizing process includes:
initializing corresponding populations according to a coding scheme; calculating individual fitness according to the fitness function; the fitness function is as follows:
wherein k represents an adjustment coefficient; n represents the total number of training sets; t is t i True data representing the ith sample; x is x i A predicted value representing an i-th sample;
selecting, crossing and mutating; generating a new generation population; wherein the selecting, crossing and mutating processes comprise:
1) Selecting excellent individuals from the population according to the following probability formula:
wherein P is i Representing the probability that the ith individual is selected;F i representing individual fitness values; n represents the population size;
2) Crossover operations are performed to create new individuals according to the following formula:
wherein a is mk ,a nk Respectively represent the mth individual a m And the nth individual a n Crossing at random position k; b represents [0,1]]Random numbers in between;
3) The mutation operation is performed according to the following formula:
wherein a is max Representing the maximum value of the code; a, a min Representing the minimum value of the code; b, r represents [0,1]]Random numbers in between; numdd represents the number of iterations; maxGA represents the maximum iteration number;
judging whether the maximum iteration times are reached, if the maximum iteration times are reached, giving output to the BP neural network, and if the maximum iteration times are not satisfied, returning to the step of calculating the individual fitness.
CN202311407886.7A 2023-10-27 2023-10-27 SGP4 model precision improvement method and system based on GA-BP neural network Pending CN117454963A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202311407886.7A CN117454963A (en) 2023-10-27 2023-10-27 SGP4 model precision improvement method and system based on GA-BP neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202311407886.7A CN117454963A (en) 2023-10-27 2023-10-27 SGP4 model precision improvement method and system based on GA-BP neural network

Publications (1)

Publication Number Publication Date
CN117454963A true CN117454963A (en) 2024-01-26

Family

ID=89588685

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202311407886.7A Pending CN117454963A (en) 2023-10-27 2023-10-27 SGP4 model precision improvement method and system based on GA-BP neural network

Country Status (1)

Country Link
CN (1) CN117454963A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117724128A (en) * 2024-02-07 2024-03-19 中南大学 Low-orbit satellite orbit prediction method, system, terminal and medium
CN117743275A (en) * 2024-02-19 2024-03-22 天津云遥宇航科技有限公司 Star-masking orbit data application system and method based on SGP4 forecasting model

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117724128A (en) * 2024-02-07 2024-03-19 中南大学 Low-orbit satellite orbit prediction method, system, terminal and medium
CN117724128B (en) * 2024-02-07 2024-04-30 中南大学 Low-orbit satellite orbit prediction method, system, terminal and medium
CN117743275A (en) * 2024-02-19 2024-03-22 天津云遥宇航科技有限公司 Star-masking orbit data application system and method based on SGP4 forecasting model
CN117743275B (en) * 2024-02-19 2024-05-28 天津云遥宇航科技有限公司 Method for occultation orbit data application system based on SGP4 forecasting model

Similar Documents

Publication Publication Date Title
CN117454963A (en) SGP4 model precision improvement method and system based on GA-BP neural network
CN111351488B (en) Intelligent trajectory reconstruction reentry guidance method for aircraft
CN110455287A (en) Adaptive Unscented kalman particle filter method
CN105180935A (en) Integrated navigation data fusion method suitable for weak signals of GNSS
CN111767517B (en) BiGRU multi-step prediction method, system and storage medium applied to flood prediction
CN111783233B (en) On-orbit backup scheme optimization design method for middle-orbit Walker navigation constellation
CN109145434A (en) A method of broadcast ephemeris orbit error is predicted using improved BP
CN111797573A (en) Ionized layer electron concentration total content time sequence prediction method based on deep learning technology
CN108764475A (en) The Gyro Random error compensating method and system of genetic wavelet neural network
CN112149230B (en) Method for predicting comfort deterioration of wind-induced train of strong wind railway
Suparta et al. Modeling of zenith path delay over Antarctica using an adaptive neuro fuzzy inference system technique
CN114199248B (en) AUV co-location method for optimizing ANFIS based on mixed element heuristic algorithm
CN108805287A (en) The Gyro Random error compensating method and system of genetic wavelet neural network
CN115859789A (en) Method for improving inversion accuracy of polar atmosphere temperature profile
Yang et al. Prediction of significant wave heights based on CS-BP model in the south China sea
CN116151102A (en) Intelligent determination method for space target ultra-short arc initial orbit
CN116090688B (en) Moving target traversal access sequence planning method based on improved pointer network
CN112183721B (en) Construction method of combined hydrological prediction model based on self-adaptive differential evolution
CN116781139A (en) Flow prediction satellite path selection method and system based on reinforcement learning
CN109862625A (en) A kind of shortwave radio monitor flexible networking method based on deep learning
CN115545159A (en) Average sea surface temperature forecasting method of deep neural network
Sarma et al. Modelling of foF 2 using neural networks at an equatorial anomaly station
CN114509825A (en) Strong convection weather prediction method and system for improving three-dimensional confrontation generation neural network based on hybrid evolution algorithm
CN115826013B (en) Beidou satellite positioning method based on light reinforcement learning under urban multipath environment
CN116523154B (en) Model training method, route planning method and related devices

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination