CN118915102B

CN118915102B - Satellite clock error forecasting method and system

Info

Publication number: CN118915102B
Application number: CN202411159384.1A
Authority: CN
Inventors: 王德盛; 冯敏; 胡燕; 张涛; 余志程; 赵盼盼; 杜卿; 郑亚陆; 李思梦; 师会芳; 常志岐; 牛慧兰
Original assignee: Henan Vocational College Of Geology And Mineral Resources
Current assignee: Henan Vocational College Of Geology And Mineral Resources
Priority date: 2024-08-22
Filing date: 2024-08-22
Publication date: 2025-04-29
Anticipated expiration: 2044-08-22
Also published as: CN118915102A

Abstract

The invention discloses a satellite clock error forecasting method and system, wherein the method comprises the steps of performing primary error processing on clock error data of the same satellite to obtain phase data, converting the phase data into frequency data, removing abnormal epochs from the frequency data through a median method to obtain frequency incomplete data, performing complement processing on the frequency incomplete data through a Lagrange interpolation method based on the abnormal epochs to obtain a primary differential clock error data sequence, inputting the primary differential clock error data sequence into an IBOA-CNN-GRU forecasting model to perform inverse normalization and inverse differential processing to obtain satellite clock error forecasting data, wherein the IBOA-CNN-GRU forecasting model is a model obtained by performing super-parameter optimization on a CNN-GRU combined model through an improved BOA algorithm. The invention utilizes the improved Bayesian algorithm to carry out the super-parameter optimization of the combined model, and provides reliable algorithm support for improving the satellite clock error forecasting performance.

Description

Satellite clock error forecasting method and system

Technical Field

The invention belongs to the technical field of clock error data prediction, and particularly relates to a satellite clock error prediction method and system.

Background

The accuracy of the satellite-borne atomic clock directly determines the navigation, positioning and timing quality of the satellite navigation system. Because the physical characteristics of the satellite-borne atomic clock are complex and are greatly influenced by external factors, the clock error data cannot be precisely forecast by using a single forecast model under most conditions, and the establishment of a high-precision clock error forecast model is difficult. Therefore, the forecasting model with stronger adaptability, higher stability and better forecasting effect is explored by combining the advantages and characteristics of each model, and is an important research direction of future clock error forecasting.

Through researches, the clock error data has continuity, periodicity, randomness and nonlinearity on a time sequence, wherein the randomness and nonlinearity characteristics of the data have great influence on precision, and for this purpose, many scholars introduce a neural network suitable for nonlinear processing into clock error forecast, such as an EMD-SVM, a wavelet neural network (wavelet neural network WNN), a thinking evolution algorithm optimized BP neural network (MEA-BP), a radial basis function (radial basis function RBF) neural network, a cyclic neural network optimized long and short time memory model (RNN-LSTM) and the like, and the models obtain good forecast precision. Although these models can capture the nonlinear relationship of the clock error data, it is difficult to effectively extract the long-term dependency relationship of the clock error data, and the computation complexity is high, so how to improve the satellite clock error forecasting performance is a problem to be solved urgently.

The foregoing is provided merely for the purpose of facilitating understanding of the technical solutions of the present invention and is not intended to represent an admission that the foregoing is prior art.

Disclosure of Invention

The invention mainly aims to provide a satellite clock error forecasting method and system, and aims to solve the technical problem of how to improve the satellite clock error forecasting performance.

In order to achieve the above object, the present invention provides a method for forecasting satellite clock skew, which includes:

performing primary difference processing on clock difference data of the same satellite to obtain phase data, and converting the phase data into frequency data;

Removing abnormal epochs from the frequency data by a median method to obtain frequency incomplete data, and carrying out complement processing on the frequency incomplete data by a Lagrange interpolation method based on the abnormal epochs to obtain a primary differential clock difference data sequence;

And inputting the primary differential clock error data sequence into an IBOA-CNN-GRU forecasting model for inverse normalization and inverse differential processing to obtain satellite clock error forecasting data, wherein the IBOA-CNN-GRU forecasting model is obtained by performing super-parameter optimization on a CNN-GRU combined model through an improved BOA algorithm.

Optionally, before the step of inputting the primary differential clock difference data sequence into an IBOA-CNN-GRU forecasting model to perform inverse normalization and inverse differential processing to obtain satellite clock difference forecasting data, the method includes:

Determining a plurality of groups of dominant populations corresponding to the super parameters in the CNN-GRU combined model based on the upper parameter limit value and the lower parameter limit value, and determining local super parameters through a fitness function according to the plurality of groups of dominant populations;

training the CNN-GRU combined model based on the primary differential clock difference data training set and the local super parameters to obtain an IBOA-CNN-GRU combined model;

determining root mean square error corresponding to an output result of the IBOA-CNN-GRU combined model through a loss function;

And if the root mean square error is smaller than or equal to a preset error threshold value and the local super-parameter is a global optimal parameter, taking the IBOA-CNN-GRU combined model as an IBOA-CNN-GRU forecasting model.

Optionally, the step of determining a plurality of dominant populations corresponding to superparameter in the CNN-GRU combined model based on the upper parameter limit and the lower parameter limit includes:

Determining a plurality of groups of populations corresponding to super parameters in the CNN-GRU combined model based on the parameter upper limit value and the parameter lower limit value, wherein the super parameters comprise the number of hidden units, an initial learning rate and L2 regularization parameters;

And selecting a plurality of groups of dominant populations from the plurality of groups of populations through the fitness function.

Optionally, the step of selecting a plurality of dominant populations from the plurality of populations by using the fitness function includes:

Respectively determining the fitness value corresponding to each group of population through a fitness function;

the fitness function is as follows:

Wherein x _m is population, f (x _m) is fitness value corresponding to x _m, m is predicted population number, L _i is the predicted value of the ith data in the x _m population, and L3562 is the true value of the ith data in the x _m population;

Selecting a plurality of groups of populations to be optimized corresponding to the fitness value smaller than a preset fitness threshold from the plurality of groups of populations;

respectively adjusting a plurality of groups of populations to be optimized through a local search algorithm to obtain a plurality of groups of adjusted populations, and determining fitness values corresponding to the groups of adjusted populations;

And determining a plurality of groups of dominant populations according to the fitness value corresponding to each group of adjusted populations and the fitness value corresponding to each group of populations to be optimized.

Optionally, after the step of determining the root mean square error corresponding to the output result of the IBOA-CNN-GRU combined model through the loss function, the method includes:

And if the root mean square error is smaller than or equal to a preset error threshold value and the local super parameter is not the global optimal parameter, carrying out iterative updating on a plurality of groups of populations through a minimum hill climbing method and a mode ant colony algorithm according to an iterative rule, and returning to the step of selecting a plurality of groups of dominant populations from the plurality of groups of populations through a fitness function.

Optionally, the step of iteratively updating the multiple groups of populations according to the iteration rule by using a minimum hill climbing method and a mode ant colony algorithm includes:

constructing a Bayesian network structure based on a plurality of groups of dominant populations according to a scoring formula;

selecting a plurality of groups of candidate populations from a plurality of groups of dominant populations through a minimum mountain climbing method and a mode ant colony algorithm according to the Bayesian network structure;

and carrying out iterative updating on the multiple groups of populations according to the iteration rules and the multiple groups of candidate populations.

Optionally, the step of constructing the bayesian network structure based on the multiple dominant populations according to the scoring formula includes:

determining a plurality of father nodes according to the plurality of groups of dominant populations;

Determining a parent node set according to a plurality of parent nodes through a scoring formula;

The scoring formula is:

Wherein x _i is a dominant population, x _j is a father node, θ is a parameter in scoring, N is the number of individuals in the topological network structure, pi _i is a father node set, P is a likelihood function of x _i under pi _i, and D is a super parameter in x _i;

And constructing a Bayesian network structure based on the father node set.

Optionally, the step of selecting a plurality of candidate populations from a plurality of dominant populations according to the bayesian network structure through a minimum hill climbing method and a mode ant colony algorithm includes:

adjusting the Bayesian network structure, and determining a plurality of excellent solutions according to a plurality of groups of dominant populations by a minimum climbing method based on the adjusted Bayesian network structure;

constructing a pheromone matrix based on a plurality of excellent solutions, and adjusting the pheromone matrix through a mode ant colony algorithm;

and determining a plurality of groups of candidate populations according to the adjusted pheromone matrix.

In addition, in order to achieve the above object, the present invention also provides a satellite clock error forecasting system, which includes:

The data processing module is used for carrying out primary difference processing on clock difference data of the same satellite to obtain phase data and converting the phase data into frequency data;

The data processing module is further used for removing abnormal epochs from the frequency data through a median method to obtain frequency incomplete data, and carrying out completion processing on the frequency incomplete data through a Lagrange interpolation method based on the abnormal epochs to obtain a primary differential clock difference data sequence;

The model operation module is used for inputting the primary differential clock difference data sequence into an IBOA-CNN-GRU forecasting model to perform inverse normalization and inverse differential processing to obtain satellite clock difference forecasting data, wherein the IBOA-CNN-GRU forecasting model is obtained by performing super-parameter optimization on a CNN-GRU combined model through an improved BOA algorithm.

In addition, in order to achieve the above object, the invention also proposes a satellite clock difference forecasting device comprising a memory, a processor and a satellite clock difference forecasting program stored on the memory and executable on the processor, the satellite clock difference forecasting program being configured to implement the steps of the satellite clock difference forecasting method as described above.

In addition, in order to achieve the above object, the present invention also proposes a storage medium having stored thereon a satellite clock difference forecasting program which, when executed by a processor, implements the steps of the satellite clock difference forecasting method as described above.

The method comprises the steps of firstly carrying out primary difference processing on clock difference data of the same satellite to obtain phase data, converting the phase data into frequency data, then removing abnormal epochs from the frequency data through a median method to obtain frequency incomplete data, carrying out complement processing on the frequency incomplete data through a Lagrange interpolation method based on the abnormal epochs to obtain a primary difference clock difference data sequence, and then inputting the primary difference clock difference data sequence into an IBOA-CNN-GRU forecasting model for carrying out inverse normalization and inverse difference processing to obtain satellite clock difference forecasting data, wherein the IBOA-CNN-GRU forecasting model is obtained after carrying out super-parameter optimization on a CNN-GRU combined model through an improved BOA algorithm. The invention combines a convolution neural network with strong sequence feature extraction capability with GRU with a long-term memory structure, utilizes the convolution and pooling operation of CNN to automatically extract the space vector of clock error data, excavates the time sequence feature in the data, solves the problem of prediction error accumulation of a model by accurately judging a prediction result, then utilizes the GRU to extract the time feature of the clock error data, improves the performance of the model by little calculation amount, plays the data excavation capability of the model, combines the advantages of the two models, establishes a CNN-GRU combined model, utilizes a Bayesian algorithm to perform the super-parameter optimization of the combined model, effectively jumps out the problem that local extremum is difficult to select aiming at the super-parameter, and ensures the global convergence of the algorithm.

Drawings

FIG. 1 is a schematic diagram of a satellite clock error forecasting device based on a hardware operating environment according to an embodiment of the present invention;

FIG. 2 is a flowchart of a satellite clock error prediction method according to a first embodiment of the present invention;

FIG. 3 is a block diagram of an IBOA-CNN-GRU combined model according to a first embodiment of the satellite clock error forecasting method of the present invention;

FIG. 4 is a flowchart of a model process according to a first embodiment of the satellite clock error prediction method of the present invention;

Fig. 5 is a block diagram of a satellite clock error prediction system according to a first embodiment of the present invention.

The achievement of the objects, functional features and advantages of the present invention will be further described with reference to the accompanying drawings, in conjunction with the embodiments.

Detailed Description

It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.

Referring to fig. 1, fig. 1 is a schematic structural diagram of a satellite clock error prediction device in a hardware operating environment according to an embodiment of the present invention.

As shown in fig. 1, the satellite-based clock error forecasting device may include a processor 1001, such as a central processing unit (Central Processing Unit, CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Wherein the communication bus 1002 is used to enable connected communication between these components. The user interface 1003 may include a Display, an input unit such as a Keyboard (Keyboard), and the optional user interface 1003 may further include a standard wired interface, a wireless interface. The network interface 1004 may optionally include a standard wired interface, a Wireless interface (e.g., a Wireless-Fidelity (Wi-Fi) interface). The Memory 1005 may be a high-speed random access Memory (Random Access Memory, RAM) or a stable nonvolatile Memory (NVM), such as a disk Memory. The memory 1005 may also optionally be a storage system separate from the processor 1001 described above.

It will be appreciated by those skilled in the art that the structure shown in fig. 1 is not limiting of the satellite-based clock skew forecasting apparatus, and may include more or fewer components than shown, or certain components may be combined, or a different arrangement of components.

As shown in fig. 1, an operating system, a network communication module, a user interface module, and a satellite clock error forecasting program may be included in the memory 1005 as one type of storage medium.

In the satellite-based clock difference forecasting device shown in fig. 1, the network interface 1004 is mainly used for carrying out data communication with the network server, the user interface 1003 is mainly used for carrying out data interaction with a user, and the processor 1001 and the memory 1005 in the satellite-based clock difference forecasting device can be arranged in the satellite-based clock difference forecasting device, and the satellite-based clock difference forecasting device calls a satellite clock difference forecasting program stored in the memory 1005 through the processor 1001 and executes the satellite clock difference forecasting method provided by the embodiment of the invention.

The embodiment of the invention provides a satellite clock error forecasting method, and referring to fig. 2, fig. 2 is a flow chart of a first embodiment of the satellite clock error forecasting method.

In this embodiment, the satellite clock error forecasting method includes the following steps:

And S10, performing primary difference processing on clock difference data of the same satellite to obtain phase data, and converting the phase data into frequency data.

It is to be understood that the execution subject of the present embodiment may be a satellite clock error prediction system with functions of data processing, network communication, program running, etc., or may be other computer devices with similar functions, etc., and the present embodiment is not limited thereto.

In order to increase the nonlinear effect of the original clock difference data and reduce the effect of trend items in the data, the clock difference data of the same satellite needs to be subjected to primary difference processing, the primary difference data after the processing is phase data, and the phase data needs to be converted into frequency data.

And step S20, eliminating abnormal epochs from the frequency data by a median method to obtain frequency incomplete data, and carrying out completion processing on the frequency incomplete data by a Lagrange interpolation method based on the abnormal epochs to obtain a primary differential clock difference data sequence.

It is to be understood that the frequency data of primary difference is subjected to rough difference detection by using a median method, and the principle that the number of abnormal values is not more than 5% of modeling data is followed, so that the influence of the abnormal data on model forecast is avoided, and the rejection of effective information is also avoided. And (3) for the removed epoch, complementing by adopting a Lagrange interpolation method to obtain a data sequence subjected to primary difference clock difference pretreatment (namely a primary difference clock difference data sequence).

And step S30, inputting the primary differential clock difference data sequence into an IBOA-CNN-GRU forecasting model for inverse normalization and inverse differential processing to obtain satellite clock difference forecasting data, wherein the IBOA-CNN-GRU forecasting model is obtained by performing super-parameter optimization on a CNN-GRU combined model through an improved BOA algorithm.

In this embodiment, the CNN model is mainly composed of a convolution layer, a pooling layer, a full connection layer, and the like, and this structure simplifies the complexity of the network model. The convolution layer is composed of one or more convolution kernels, which are a learnable parameter matrix, and local features with different scales and directions are extracted and the depth of the features is increased by moving the convolution kernels on the data and performing dot product with the data. Since the clock difference data is one-dimensional sequence data, namely, one-dimensional convolutional neural network is selected, the calculation formula is as follows:

In the formula, Is the kth convolution map of layer i, f is the activation function,Is the weight of the layer i,The offset for the k-th convolution kernel corresponds to layer l.

The pooling layer processes the feature matrix of the convolution layer through pooling operation, reduces the dimension of clock error data features, reduces the parameters of the network, and prevents overfitting. The robustness of the features is enhanced while the data is simplified, and the local optimal features of the clock difference data on the space structure are obtained.

And finally, connecting the full-connection layer with all neurons of the upper layer, extracting all characteristics of the neurons of the upper layer, and integrating the characteristic vectors output by the pooling layer to realize a final regression task.

The CNN model is often provided with a plurality of rolling and pooling layers, and the network layers automatically generate feature vectors of data by acquiring effective information, so that difficulty in feature extraction and data reconstruction is reduced, and quality of spatial features is improved.

The CNN model can extract key information from complex sample characteristics and amplify the characteristics, so that an implicit relation between more abstract and deep characteristics is found, the operation load of a training process is reduced, the convergence speed of the model is accelerated, but the model lacks the sensing capability of front-back fluctuation of sample data when the model processes time series data.

In a specific implementation, referring to fig. 3, fig. 3 is a block diagram of a CNN-GRU combined model of a first embodiment of the satellite clock error forecasting method according to the present invention, the present invention combines CNN and GRU, which not only compensates for the problem that the CNN network cannot adapt to time dependence, but also utilizes the GRU network to extract the nonlinear characteristic of clock error data, so as to implement high-precision forecasting of clock error data. Firstly, extracting spatial characteristics of an input satellite clock error by using a CNN convolution layer, then, transmitting the extracted spatial characteristics to a GRU module to process nonlinear characteristics of the clock error, and finally, outputting forecast data of the clock error through a full connection layer.

In order to increase the operation efficiency, improvement of the BOA algorithm is required. In order to ensure the global convergence of the algorithm, the algorithm is locally adjusted by combining a minimum hill climbing method in the running process of each algorithm by establishing an adaptability function so as to lead the algorithm to trend to an optimal solution. And the generated solutions are subjected to fitness inheritance, so that the evaluation time of fitness is effectively saved, and the local extremum can be effectively jumped out by searching the local substructures of the individuals. In order to avoid useless searching in the searching process, a mode ant colony algorithm is introduced to perform structure learning, nodes are selected according to probability transition rules, and the evolution direction is guided through an optimization mode.

In specific implementation, referring to fig. 4, fig. 4 is a flow chart of model processing in a first embodiment of the satellite clock error forecasting method of the present invention, (1) data of a single satellite is adopted for each processing, and primary differential processing is performed on original clock error data to obtain primary differential data of clock error. (2) And detecting and removing the first differential clock error abnormal value by adopting a median coarse error detection method, and filling the abnormal value by using a piecewise linear interpolation method to obtain relatively complete clock error data. (3) Initializing a network model structure, normalizing a data set, dividing the data set into a training set and a testing set, selecting one subset as the testing set each time, and training the model by using the other subsets as the training set. Setting super parameters such as input dimension, convolution kernel number, convolution kernel size, pooling size, neuron number, hidden layer number, learning rate, regularization coefficient and the like of the CNN-GRU combination model. (4) Inputting the clock difference data of the steps (1) - (3) into the CNN-GRU combined model for network training, adopting a minimum hill climbing method to determine a search path in the process of training the CNN-GRU combined model, continuously searching and selecting the next group of dominant populations with the highest potential by using a mode ant colony algorithm, calculating a loss function value in each iteration to evaluate the performance of the current model, and generating candidate populations which are used for updating the model and carrying out the next new iteration. (5) And (4) repeating the step (4), stopping iteration if the dominant population of the new iteration meets the requirement or reaches the maximum iteration number, outputting the super-optimal super-parameter combination, and otherwise, returning to the step (4). (6) And 5, taking the optimal super parameters output in the step as an implicit layer, a learning rate and an L2 regularization coefficient of the IBOA-CNN-GRU combined model, then using the model to perform primary difference data forecasting of the clock difference, and performing inverse normalization and contrast division operation on the forecasting data to finally obtain a clock difference forecasting value.

In this embodiment, a global optimal parameter of a CNN-GRU combined model needs to be determined through an improved BOA algorithm, so that multiple groups of dominant populations corresponding to superparameter in the CNN-GRU combined model need to be determined based on an upper parameter limit value and a lower parameter limit value, local superparameter is determined according to multiple groups of dominant populations through a fitness function, the CNN-GRU combined model is trained based on a primary differential clock difference data training set and the local superparameter to obtain an IBOA-CNN-GRU combined model, a root mean square error corresponding to an output result of the IBOA-CNN-GRU combined model is determined through a loss function, and if the root mean square error is smaller than or equal to a preset error threshold value and the local superparameter is the global optimal parameter, the IBOA-CNN-GRU combined model is used as an IBOA-CNN-GRU forecast model.

Further, the processing mode of determining a plurality of groups of dominant populations corresponding to the super parameters in the CNN-GRU combined model based on the parameter upper limit value and the parameter lower limit value is to determine a plurality of groups of populations corresponding to the super parameters in the CNN-GRU combined model based on the parameter upper limit value and the parameter lower limit value, wherein the super parameters comprise hidden unit numbers, initial learning rate and L2 regularization parameters, and the plurality of groups of dominant populations are selected from the plurality of groups of populations through an fitness function.

It should be further noted that, at the beginning, initialization processing needs to be performed on the CNN-GRU combined model, at this time, multiple groups of initial populations x _m corresponding to super parameters in the CNN-GRU combined model can be randomly generated based on the upper boundary value and the lower boundary value of the parameters, and P (0) is a matrix corresponding to the multiple groups of initial populations.

It should also be understood that P (0) refers to the starting point at which the algorithm operates, and then sets the size M and maximum number of iterations N of its population. Meanwhile, the CNN-GRU combination model has three parameters to be optimized, namely a hidden unit number (NumOfUnits), an initial learning rate (INITIALLEARNRATE) and an L2regularization parameter (L2 Regularization), wherein the parameters set a lower bound and an upper bound, the lower bound is assumed to be lb= [10,0.0005,1e-6], the upper bound is set to be ub= [50,0.001,1e-3], and P (0) is a matrix of M x 3, and each row represents an initial population, and one initial population comprises three parameter values given above.

Assuming here that the initial population size is 20, then P (0) is a matrix of 20 x 3,

The values within this matrix are all within this range. The diversity of the population is ensured through random generation, the whole search space is covered, and a basis is provided for subsequent iterative optimization.

The formula can be expressed as:

P(0)=[x₁;x₂…x_m]

x_m＝[x_m1,x_m2…x_m3]

Further, the processing mode of selecting a plurality of groups of dominant populations from the plurality of groups of populations through the fitness function is that fitness values corresponding to the groups of populations are respectively determined through the fitness function, a plurality of groups of populations to be optimized, the fitness values of which are smaller than the preset fitness threshold, are selected from the groups of populations, the groups of populations to be optimized are respectively adjusted through a local search algorithm, a plurality of groups of adjusted populations are obtained, fitness values corresponding to the groups of adjusted populations are determined, and a plurality of groups of dominant populations are determined according to the fitness values corresponding to the groups of adjusted populations and the fitness values corresponding to the groups of populations to be optimized.

The fitness function is:

Wherein x _m is population, f (x _m) is fitness value corresponding to x _m, m is predicted population number, L _i is the true value of the ith data in the x _m population, which is the predicted value of the ith data in the x _m population.

It should be noted that P (0) generated at the beginning is random, but each subsequent generation P (t) is generated based on the steps of selection, local search, and the like of the previous generation. For example, when t is 1, the search is selected and locally searched according to P (0).

It should be appreciated that the initial population x _m is then evaluated using a fitness function, where the fitness function f (x _m) selects RMSE, the multiple groups of populations to be optimized are found by comparing fitness values, and then the multiple groups of populations to be optimized are adjusted using a local search algorithm, where the fitness of the multiple groups of populations to be optimized is mainly better adjusted, resulting in the best dominant population.

In the local search, the current population x _i to be optimized is subjected to small-amplitude parameter adjustment so as to find a better population.

Assuming that we have selected and adjusted a population to be optimized x _i, the new population to be optimized x' _i can be expressed as:

x'_i＝x_i+δ

Where δ is a floating value for adjusting x _i, the formula is:

δ_j＝random×(ub_j-lb_j)

where ub _j and lb _j are the upper and lower bounds of the j-th parameter and random is a random number between 0, 1.

It should be noted that, the new population to be optimized x '_i and the previous population to be optimized x _i are brought into the fitness function to calculate the fitness value, and the fitness value is compared to determine whether the update is needed, if the fitness of the dominant population x' _i is smaller than that of the previous dominant population x _i, that is, f (x '_i)<f(x_i), the value of x' _i is replaced by the value of x _i, so that the local individual is optimally adjusted.

Further, after the step of determining the root mean square error corresponding to the output result of the IBOA-CNN-GRU combined model through the loss function, if the root mean square error is smaller than or equal to a preset error threshold and the local super parameter is not the global optimal parameter, carrying out iterative updating on multiple groups of populations through the minimum hill climbing method and the mode ant colony algorithm according to the iterative rule, and returning to the step of selecting multiple groups of dominant populations from the multiple groups of populations through the fitness function.

The processing mode of carrying out iterative updating on a plurality of groups of population according to the iteration rule through the minimum climbing method and the mode ant colony algorithm is that a Bayesian network structure is constructed based on a plurality of groups of dominant population according to a scoring formula, a plurality of groups of candidate population are selected from the plurality of groups of dominant population according to the Bayesian network structure through the minimum climbing method and the mode ant colony algorithm, and the plurality of groups of population are subjected to iterative updating according to the iteration rule.

The method comprises the steps of establishing a Bayesian network structure according to a scoring formula based on a plurality of groups of dominant populations, determining a plurality of father nodes according to the plurality of groups of dominant populations, determining a father node set according to the plurality of father nodes according to the scoring formula, and establishing the Bayesian network structure based on the father node set.

The scoring formula is:

Wherein x _i is a dominant population, x _j is a parent node, θ is a parameter in scoring, N is the number of individuals in the topological network structure, pi _i is a parent node set, P is a likelihood function of x _i under pi _i, and D is a super parameter in x _i.

In this embodiment, a plurality of groups of dominant populations x _i are reported in the dominant population matrix, and in the dominant population matrix S (t), topological ranks of n variables are randomly generated, and these ranks determine the order of the population individuals, and these ranks provide the order of the population individuals for the subsequent scoring algorithm.

Parent nodes are added step by step according to the topological ordering of a given individual to maximize the scoring function of the network.

Wherein, the detailed steps of constructing the Bayesian network structure are as follows:

1. inputting a dominant population matrix S (t), and initializing the maximum parent node quantity lambda.

2. For x _i in the dominant population matrix S (t), initialize its parent node setWhile setting the structure matrix of the bayesian network to the 0 matrix.

3. Each dominant population x _i is processed one by one in the order of topological ordering.

4. Adding a father node:

for the dominant population currently being processed, an optimal Score (x _i,Π_i) is initialized, which is formulated as:

Score(x_i,Π_i)＝logP(x_i∣Π_i)

Where P is a likelihood function at n _i for x _i and log P (x _i∣Π_i) is a log likelihood value at n _i for x _i.

Then try to take { x ₁,x₂,…,x_i-1 } as the parent node of x _i one by one, for each parent node x _j (j < i), calculate the score after adding x _j, if x _j can increase the score, add, the new scoring formula is:

Where θ is a specific parameter in the score (i.e., a parameter in the score), and N is the number of individuals in the topological network structure.

Attempts to add parent nodes continue until the number of parent nodes reaches a maximum or the score cannot be further increased.

And finally, adding the finally determined pi _i into a Bayesian network.

The method comprises the steps of selecting a plurality of groups of candidate populations from a plurality of groups of dominant populations through a minimum climbing method and a mode ant colony algorithm according to a Bayesian network structure, adjusting the Bayesian network structure, determining a plurality of excellent solutions according to the plurality of groups of dominant populations through the minimum climbing method based on the adjusted Bayesian network structure, constructing a pheromone matrix based on the plurality of excellent solutions, adjusting the pheromone matrix through the mode ant colony algorithm, and determining a plurality of groups of candidate populations according to the adjusted pheromone matrix.

In this embodiment, a minimum mountain climbing method and a mode ant colony algorithm are adopted to search to obtain a structure diagram with highest evaluation, a candidate population matrix O (t) is generated according to a constructed bayesian network, and multiple groups of candidate populations exist in the candidate population matrix, wherein the detailed steps of generating the candidate population matrix O (t) according to the constructed bayesian network are as follows:

Climbing method, ensuring the quality of solution.

Starting from the Bayesian network structure, searching for a more excellent solution (namely, low individual fitness) in the neighborhood of the current construction individual by adding, deleting or reversing edges, then calculating the function grading value of the new solution, updating the current solution into the solution if the solution with the more excellent grading value exists in the neighborhood, and otherwise, stopping searching. The above process is repeated until there is no more optimal solution.

And the mode ant colony algorithm ensures the diversity of the search space.

Firstly, setting ant number, iteration times, pheromone importance factors, heuristic information importance factors, pheromone volatilization rate, pheromone increasing strength and other parameters.

Then the ant constructs a solution according to the pheromone matrix and heuristic information, evaluates the quality of the solution according to a scoring function, and updates the pheromone matrix according to the solution found by the ant, including global and local pheromone updates.

Pheromone matrix assuming n solutions, the pheromone matrix is a matrix of n x n, and is used for recording the relation and intensity between the solutions.

The function of the pheromone matrix is that ants can select paths according to the concentration of the pheromone when constructing solutions. The higher the concentration, the greater the probability of selecting the path.

The pheromone updating mechanism is that assuming that y ants exist, each ant finds a solution of x _i, the fitness is f (x _i), and the formula of the pheromone updating is as follows:

Volatilizing:

τ_ij＝(1-ρ)τ_ij

Wherein τ _ij represents the concentration of pheromone from the ith solution to the jth solution, represents the connection strength between the two solutions, ρ is the volatilization coefficient of the pheromone, and the value is between 0 and 1.

And (3) increasing:

In the formula, The amount of information enhancing elements on paths i to j for ant k is represented, and Q is a constant.

Thus, the pheromone matrix is dynamically adjusted through the volatilization and addition process, so that ants are guided to find a better solution, and the overall performance of the algorithm is improved.

In summary, a candidate population matrix O (t) is generated using a hill climbing algorithm and a pattern ant colony algorithm.

It should also be noted that, the populations in the t-th generation population matrix P (t) are replaced by multiple groups of candidate populations in the newly generated candidate population matrix O (t), so as to generate the next generation population matrix P (t+1), thereby improving the overall quality of the populations.

It should be appreciated that when the maximum number of iterations N or fitness value set before is reached is no longer improved, the algorithm is stopped, the current optimal solution, i.e. the optimal hyper-parameters of the CNN-GRU combined model, is output, and the model is predicted using this hyper-parameters.

As the BOA algorithm may be subjected to useless searches in the process of locally optimizing, the locally optimal extremum is trapped. The optimized BOA algorithm enhances the global optimizing capability of the BOA algorithm and improves the prediction precision of the model. The improved IBOA algorithm not only can improve the training and prediction time of the model, but also has higher stability, and shows the feasibility of the IBOA-CNN-GRU model in the aspect of short-term clock error prediction.

It should also be noted that the invention constructs a CNN-GRU combined model and optimizes the selection of super parameters on the basis of the CNN-GRU combined model. The CNN-GRU model can effectively solve the problem that the CNN model is difficult to adapt to long-sequence time dependence, and the stability and the prediction accuracy of the CNN-GRU model are improved and are higher than those of BP and LSTM models. The BOA model added with the optimized Bayesian algorithm avoids the problem that the superparameter is difficult to select, reduces the operation time, and has better prediction precision compared with CNN and CNN-GRU. The IBOA-CNN-GRU model solves the problem of error accumulation caused by the increase of a single model along with the time, and simultaneously improves useless search and local optimal solution in super-parameter selection. Compared with other neural network models, the method has larger promotion and reflects the feasibility in the process of clock error forecasting.

In the embodiment, first, primary difference processing is performed on clock difference data of the same satellite to obtain phase data, the phase data are converted into frequency data, then abnormal epochs are removed from the frequency data through a median method to obtain frequency incomplete data, the frequency incomplete data are subjected to complement processing through a Lagrange interpolation method based on the abnormal epochs to obtain a primary difference clock difference data sequence, then the primary difference clock difference data sequence is input into an IBOA-CNN-GRU forecasting model to be subjected to inverse normalization and inverse difference processing to obtain satellite clock difference forecasting data, and the IBOA-CNN-GRU forecasting model is obtained after super-parameter optimization is performed on a CNN-GRU combined model through an improved BOA algorithm. According to the embodiment, a convolutional neural network with strong sequence feature extraction capability is combined with GRU with a long-term memory structure, the space vector of clock difference data is automatically extracted by utilizing the convolution and pooling operation of CNN, the time sequence features in the data are mined, the problem of prediction error accumulation of a model is solved by accurately judging a prediction result, then the time features of the clock difference data are extracted by utilizing GRU, the model performance is improved by using small calculation amount, the data mining capability of the model is brought into play, the advantages of the two models are combined, a CNN-GRU combined model is established, the super-parameter optimization of the combined model is carried out by utilizing a Bayesian algorithm, the problem that local extremum is difficult to select aiming at the super-parameter is effectively jumped out, and the global convergence of the algorithm is ensured.

Referring to fig. 5, fig. 5 is a block diagram of a satellite clock error prediction system according to a first embodiment of the present invention.

As shown in fig. 5, the satellite clock error forecasting system provided by the embodiment of the invention includes:

the data processing module 5001 is configured to perform primary difference processing on clock difference data of the same satellite to obtain phase data, and convert the phase data into frequency data;

the data processing module 5001 is further configured to remove an abnormal epoch from the frequency data by using a median method to obtain frequency incomplete data, and perform completion processing on the frequency incomplete data by using a lagrangian interpolation method based on the abnormal epoch to obtain a primary differential clock difference data sequence;

The model operation module 5002 is configured to input the primary differential clock difference data sequence into an IBOA-CNN-GRU prediction model for inverse normalization and inverse differential processing, to obtain satellite clock difference prediction data, where the IBOA-CNN-GRU prediction model is a model obtained by performing super-parameter optimization on a CNN-GRU combination model through an improved BOA algorithm.

Other embodiments or specific implementation manners of the satellite clock error forecasting system of the present invention may refer to the above method embodiments, and are not described herein.

It should be noted that, in this document, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or system. Without further limitation, an element defined by the phrase "comprising one does not exclude the presence of other like elements in a process, method, article, or system that comprises the element.

The foregoing embodiment numbers of the present invention are merely for the purpose of description, and do not represent the advantages or disadvantages of the embodiments.

From the above description of the embodiments, it will be clear to those skilled in the art that the above-described embodiment method may be implemented by means of software plus a necessary general hardware platform, but of course may also be implemented by means of hardware, but in many cases the former is a preferred embodiment. Based on such understanding, the technical solution of the present invention may be embodied essentially or in a part contributing to the prior art in the form of a software product stored in a storage medium (e.g. read-only memory/random-access memory, magnetic disk, optical disk), comprising instructions for causing a terminal device (which may be a mobile phone, a computer, a server, or a network device, etc.) to perform the method according to the embodiments of the present invention.

The foregoing description is only of the preferred embodiments of the present invention, and is not intended to limit the scope of the invention, but rather is intended to cover any equivalents of the structures or equivalent processes disclosed herein or in the alternative, which may be employed directly or indirectly in other related arts.

Claims

1. The satellite clock error forecasting method is characterized by comprising the following steps of:

Inputting the primary differential clock difference data sequence into an IBOA-CNN-GRU forecasting model for inverse normalization and inverse differential processing to obtain satellite clock difference forecasting data, wherein the IBOA-CNN-GRU forecasting model is obtained by performing super-parameter optimization on a CNN-GRU combined model through an improved BOA algorithm;

the step of inputting the primary differential clock difference data sequence into an IBOA-CNN-GRU forecasting model for inverse normalization and inverse differential processing to obtain satellite clock difference forecasting data comprises the following steps:

2. The method of claim 1, wherein the step of determining a plurality of dominant populations for the superparameter in the CNN-GRU combined model based on the upper and lower parameter thresholds comprises:

3. The method of claim 2, wherein the step of selecting a plurality of dominant populations from the plurality of populations by a fitness function comprises:

the fitness function is as follows:

4. A method as claimed in claim 2, wherein after the step of determining the root mean square error corresponding to the output result of the IBOA-CNN-GRU combination model by a loss function, the method comprises:

5. The method of claim 4, wherein the step of iteratively updating the plurality of groups of populations by a minimum hill climbing method and a pattern ant colony algorithm according to an iteration rule comprises:

6. The method of claim 5, wherein the step of constructing a bayesian network structure from scoring formulas based on a plurality of dominant populations comprises:

The scoring formula is:

x_j∈{x₁,x₂,…,x_i-1}

And constructing a Bayesian network structure based on the father node set.

7. The method of claim 6, wherein the selecting a plurality of candidate populations from a plurality of dominant populations by a minimum hill climbing method and a pattern ant colony algorithm according to the bayesian network structure comprises:

8. A satellite clock error forecasting system, the satellite clock error forecasting system comprising:

The model operation module is used for inputting the primary differential clock difference data sequence into an IBOA-CNN-GRU forecasting model for inverse normalization and inverse differential processing to obtain satellite clock difference forecasting data, wherein the IBOA-CNN-GRU forecasting model is obtained by performing super-parameter optimization on a CNN-GRU combined model through an improved BOA algorithm;

the model operation module is further used for determining a plurality of groups of dominant populations corresponding to the superparameter in the CNN-GRU combination model based on the parameter upper limit value and the parameter lower limit value, determining local superparameter according to the plurality of groups of dominant populations through a fitness function, training the CNN-GRU combination model based on a primary differential clock difference data training set and the local superparameter to obtain an IBOA-CNN-GRU combination model, determining root mean square error corresponding to an output result of the IBOA-CNN-GRU combination model through a loss function, and taking the IBOA-CNN-GRU combination model as an IBOA-CNN-GRU forecast model if the root mean square error is smaller than or equal to a preset error threshold value and the local superparameter is a global optimal parameter.