Satellite clock error prediction method and system

Abstract

The invention relates to the technical field of satellite navigation and time synchronization, in particular to a satellite clock error prediction method and system. According to the method, a satellite clock error prediction model is constructed, multi-scale features of input data are extracted, sequence decomposition is performed on the input data, decomposed seasonal items are subjected to self-correlation processing through an encoder and a decoder, and the data are subjected to sequence decomposition in the encoder and the decoder; and the season item output by the decoder and the trend item after each time of sequence decomposition are subjected to dimension superposition and then converted into clock error prediction data. The method comprises multiple times of sequence decomposition, adopts a progressive decomposition architecture to peel noise interference step by step, purifies trend components in seasonal terms through time convolution operation, and optimizes a trend accumulation process through trend term superposition by utilizing a residual feedback mechanism, so that the purity of trend extraction is remarkably improved, and the method is suitable for large-scale popularization and application. And meanwhile, the phase integrity of the seasonal components is kept, so that the model can accurately model the multi-scale characteristics of the clock error data.

Description

Technical Field

[0001] The present invention relates to the technical field of satellite navigation and time synchronization, and in particular to a satellite clock error prediction method and system. Background Art

[0002] Satellite clock error prediction is one of the key technologies in satellite navigation systems, directly affecting the positioning accuracy. Traditional satellite clock error prediction methods are mainly based on offline learning, using historical data for model training, and cannot adapt to the changes in real-time data streams. Existing methods have significant deficiencies in prediction accuracy and real-time performance, and it is difficult to meet the requirements of heavy-haul train group positioning control.

[0003] Traditional statistical models, such as ARIMA and Kalman filtering, are widely used in time series prediction. The ARIMA model can capture linear trends and seasonal term changes, while the Kalman filter is suitable for state estimation of dynamic systems and can update model parameters in real time. However, these methods assume that the data has a linear relationship and are difficult to handle complex non-linear time series. In addition, traditional statistical models have high requirements for the stationarity of data distribution, have poor prediction effects on non-stationary data, and cannot adapt to the changes in real-time data streams.

[0004] Machine learning models, such as support vector machines (SVM) and random forests, process non-linear data through kernel functions and ensemble learning methods, and are suitable for medium and small-scale data sets. However, these methods have high computational complexity when dealing with large-scale data and are difficult to meet the real-time prediction requirements. In addition, machine learning models usually can only capture local time dependencies, are difficult to handle long-distance dependence problems, and the model performance depends on manually designed features, and the feature engineering process is complex and time-consuming.

[0005] Deep learning models, such as LSTM and Transformer, capture long-distance dependence relationships through gating mechanisms and self-attention mechanisms, and are suitable for large-scale data. The LSTM network captures the long-term dependence relationships of time series through gating mechanisms, while the Transformer model captures global dependence relationships through self-attention mechanisms. However, the computational complexity of these models is high and it is difficult to meet the real-time prediction requirements. In addition, existing deep learning models usually adopt single-scale feature extraction, are difficult to capture short-term fluctuations and long-term trends simultaneously, and the model parameters are usually fixed and cannot be dynamically adjusted to adapt to the changes in data distribution.

[0006] In summary, existing methods have significant deficiencies in prediction accuracy, computational complexity, and real-time performance, and it is difficult to meet high-precision satellite positioning requirements such as heavy-haul train group positioning control. Therefore, there is an urgent need for a satellite clock error prediction method that can adapt to dynamic data changes. Summary of the Invention

[0007] To overcome the deficiencies of high computational complexity, insufficient prediction accuracy, and difficulty in adapting to real-time data streams in the existing satellite clock error prediction algorithms in the above-mentioned prior art, the present invention proposes a satellite clock error prediction method. By combining data preprocessing and prediction model optimization, it can efficiently and accurately predict satellite clock errors, thereby improving the accuracy and real-time performance of satellite positioning.

[0008] A satellite clock error prediction method proposed by the present invention first constructs a satellite clock error prediction model. Its input is the satellite clock error window data after differential processing and amplification processing, and its output is the clock error prediction data for the time period after the input data. The satellite clock error prediction model extracts multi-scale features from the input data and performs sequence decomposition. The decomposed seasonal term undergoes autocorrelation processing through an encoder and a decoder. Both the encoder and the decoder internally decompose the data into sequences. The seasonal term output by the decoder is dimensionally stacked with the trend term after each sequence decomposition and then converted into clock error prediction data. Then, real-time satellite clock error data is collected and windowed. Differential processing and amplification processing are performed on each window data, and the processed window data is input into the satellite clock error prediction model to obtain clock error prediction data.

[0009] Preferably, the satellite clock error prediction model includes a data embedding module, a multi-scale feature extraction module, a sequence decomposition module, an encoder, a decoder, a dimension stacking unit, and an output module connected in sequence. The multi-scale feature extraction module uses convolution kernels of different sizes to perform convolution processing on the input data and then splices them to generate multi-scale features. The multi-scale features are decomposed into seasonal terms and trend terms by the sequence decomposition module. The seasonal term is encoded by the encoder and then decomposed into seasonal terms and trend terms again. The seasonal term output by the encoder is processed by the decoder and then decomposed into seasonal terms and trend terms again. The dimension stacking unit dimensionally stacks the seasonal term output by the decoder with the trend term after each sequence decomposition and then inputs it into the output module for processing into clock error prediction data.

[0010] Preferably, the data embedding module adopts a time encoding embedding method.

[0011] Preferably, the decoder includes a temporal convolutional network, a first sequence decomposition layer, an autocorrelation network, a second sequence decomposition layer, a feed-forward network layer, and a third sequence decomposition layer connected in sequence. Each sequence decomposition layer decomposes the input into sequences. The decomposed seasonal term propagates backward, and the seasonal term output by the third sequence decomposition layer serves as the seasonal term output by the decoder. The trend terms of each sequence decomposition are dimensionally stacked to form the trend term output by the decoder.

[0012] Preferably, the encoder includes an autocorrelation layer, a feed-forward network layer, and a sequence decomposition network connected in sequence.

[0013] Preferably, the output module adopts a fully connected layer.

[0014] Preferably, the training process of the satellite clock error prediction model includes the following steps: Collect historical clock error data and perform windowing processing to obtain 2m consecutive window data X1, X2, …, Xj, …, Xm, Y1, Y2, …, Yj, …, Ym; construct data samples {Xj, Yj|1≤j≤m}; X and Y have the same number of steps; Perform difference processing and amplification processing on Xj in the data samples {Xj, Yj|1≤j≤m}, convert Xj into preprocessed data Zj, and obtain training samples {Zj, Yj|1≤j≤m}; Extract the data Z1, Z2, …, Zj, …, Zm of multiple training samples and input them into the satellite clock error model to obtain predicted data Y'1, Y'2, …, Y'j, …, Y'm; compare the data Y1, Y2, …, Yj, …, Ym with the predicted data Y'1, Y'2, …, Y'j, …, Y'm, calculate the model loss, and use the gradient descent optimization method to update the model parameters; repeat this step until the model reaches the convergence condition.

[0015] Preferably, the model convergence condition is: the number of model iterations reaches a set value, or the absolute value of the difference between the model losses in at least the last three rounds is less than a set threshold.

[0016] A satellite clock error prediction system proposed by the present invention includes a data preprocessing module, a memory, and a processor. The memory stores a computer program and a satellite clock error prediction model. The data preprocessing module is used to obtain real-time collected satellite clock error data and perform difference processing and amplification processing. The processor is connected to the memory and the data preprocessing module, and the processor is used to execute the computer program to implement the satellite clock error prediction method.

[0017] A storage medium proposed by the present invention stores a computer program, and the computer program is used to implement the satellite clock error prediction method when executed.

[0018] The advantages of the present invention are as follows: (1) The present invention includes multiple sequence decompositions, adopts a progressive decomposition architecture to gradually strip noise interference, purifies the trend component in the seasonal term through time convolution operation, and utilizes the residual feedback mechanism through trend term superposition to optimize the trend accumulation process, thereby significantly improving the purity of trend extraction while maintaining the phase integrity of the seasonal component, enabling the model to accurately model the multi-scale characteristics of clock error data. Compared with traditional single decomposition methods, the present invention has stronger anti-noise ability and physical consistency, can adapt to the clock error prediction requirements of satellites with different orbital characteristics, and performs excellently in terms of computational efficiency and long-term prediction stability.

[0019] (2) The present invention performs amplification processing on the known satellite clock error data after differential processing, amplifying the data from an extremely small value to a form where the differences can be clearly manifested, improving the prediction accuracy, avoiding the influence of the dimensional differences between different features on model training, and optimizing the numerical stability. Through experiments, it is proved that the magnitude of the original clock error difference value is about 1e-7 to 1e-5 nanoseconds. After being amplified by 100,000 times, the utilization rate of significant figures is increased from 17% to 100%, and the effective numerical value can enter the optimal processing range (0.1 - 10.0) of a typical deep learning model through 1e5 times amplification, while maintaining a strict correspondence with the physical dimension. At the same time, the data after differential and amplification processing ensures high efficiency and low latency. Combined with data windowing processing, it can effectively cope with the concept drift problem in the data stream and ensure the long-term stability of the model.

[0020] (3) Through time encoding, the present invention can directly map the real physical environment in which the atomic clock works. For example, the year and month encoding reflects the frequency aging characteristic of the atomic clock at 1e-13 per year; the day and hour encoding matches the periodic temperature change caused by the earth's rotation; the minute and second encoding captures the high-frequency noise of orbital vibration. Secondly, multi-scale feature decoupling can be performed. For example, the year and month layers model the long-term drift, i.e., the trend term; the day and hour layers extract the orbital period feature, i.e., the seasonal term; the minute and second layers process the short-term jitter, i.e., the residual term. Compared with the situation where the original models such as Autoformer use position encoding embedding, the use of time encoding embedding in the present invention has better physical interpretability in time series data processing, can improve the robustness during on-orbit deployment, and support on-board autonomous time reference maintenance.

[0021] (4) During the sequence decomposition before the output module, perform relevant operations such as time series convolution and sequence decomposition of the seasonal term, purify the trend term, and perform accumulation. Finally, superimpose the dimension of the purified trend term and the decomposed seasonal term, and finally use it as the input to enter the output module for sequence prediction.

[0022] (5) The multi-scale feature extraction module extracts convolution features through convolution kernels of different sizes to comprehensively capture local details and global trends; the encoder and decoder cooperate to implement a lightweight Autoformer module, using the autocorrelation mechanism to capture long-distance dependence relationships, avoiding the limitations of single-scale feature extraction and traditional models in the prior art that are difficult to process non-linear and non-stationary data, and being able to make full use of multi-scale feature and long-distance dependence information, effectively improving the prediction accuracy and real-time performance while maintaining low computational complexity.

[0023] (6)The satellite clock error prediction method and system proposed by the present invention significantly improve the significance of time series features through data preprocessing, comprehensively capture the clock error change pattern by combining multi-scale modeling with local details and global trends; enhance the model's parsing ability for periodic laws and improve the physical interpretability of the model by separating and purifying seasonal terms and trend terms. Moreover, in the present invention, the lightweight network design and dynamic window division meet the real-time prediction requirements. The present invention has important application value in the fields of satellite navigation, time synchronization, etc., and especially performs better than traditional methods in complex dynamic environments. Description of the Drawings

[0024] Figure 1 It is a module diagram of the satellite clock error prediction model proposed by the present invention; Figure 2 is Figure 1 the encoder architecture diagram in Figure 3 is Figure 1 the decoder architecture diagram in Figure 4 is the Autoformer network architecture; Figure 5 It is a flowchart of the satellite clock error prediction method proposed by the present invention; Figure 6 It is a line graph of the original data of a certain satellite in the example dataset; Figure 7 is Figure 6 the quasi-linear trend characteristic presented by the original data in Figure 8 It is a comparison graph of the loss function results of three models in the example; Figure 9 It is a comparison line graph of the predicted values and true values of the model in the example. Detailed Implementation Modes

[0025] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0026] As Figure 1 shown, this embodiment proposes a satellite clock error prediction model, including a data embedding module, a multi-scale feature extraction module, a sequence decomposition module, an encoder, a decoder, a dimension stacking unit, and an output module connected in sequence.

[0027] The input of this model is the satellite clock error data after data preprocessing, and the output is the clock error prediction data. The data preprocessing includes differential processing and amplification processing to generate preprocessed data with significant features.

[0028] Specifically, when implementing, the collected window data can be denoted as X = {x1, x2,..., x i ,..., x N}, and the future clock error data corresponding to X, that is, the prediction data, is denoted as Y = {y1, y2,..., y i ,..., y N}, where x i represents the i-th historical clock error data, and y i represents the future clock error data corresponding to x i , 1 ≤ i ≤ N, and N represents the window length; that is, x1, x2, x N respectively represent the 1st, 2nd, and Nth historical clock error data in X, and y1, y2, y N respectively represent the future clock error data corresponding to x1, x2, x N . Assuming the time interval between data x i and x i+1 is t, then the time interval between y i and x i is T, and T ≥ Nt.

[0029] The preprocessing process of the window data X is as follows: First, perform differential processing on the window data X to obtain the corresponding differential data ΔX = {Δx1, Δx2,..., Δx i ,..., Δx N}, where Δx1, Δx2, Δx i , Δx N are the differential calculation results corresponding to x1, x2, x i , x N respectively, and Δx i = x i - x i-1 ; x i-1 (i = 1) represents the historical clock error data collected at a time interval of t before x1; Then, perform amplification processing on the differential data ΔX to obtain the amplified data Z = {z1, z2,..., z i ,..., z N}, where z i = Δx i × σ, and σ is the amplification multiple. Considering that the satellite clock error is in nanoseconds, in specific implementation, in order to highlight the data trend, it may need to be amplified by more than ten thousand times, that is, σ ≥ 10 4, which can specifically take values in the range of 50,000 to 100,000, and the specific value in the subsequent embodiments is one hundred thousand.

[0030] The satellite clock error prediction model is trained on the dataset {Z, Y}.

[0031] The data embedding module embeds the preprocessed data Z into the set feature space using time encoding embedding to generate embedded features, facilitating the convolutional processing of the multi-scale feature extraction module.

[0032] The multi-scale feature extraction module includes convolutional networks with multiple different-sized convolutional kernels. Each convolutional network performs convolutional processing on the embedded features output by the data embedding module to generate convolutional data of different sizes, and the convolutional data is spliced into multi-sized feature data through a splicing network. For example, in the subsequent embodiments, the multi-scale feature extraction module uses convolutional networks with 1×1, 3×3, and 5×5 convolutional kernels to perform convolutional processing on the features output by the data embedding module. The data embedding module embeds the data Z into the specified data space, and after convolutional processing, dimension splicing is performed to form multi-scale features.

[0033] The sequence decomposition module decomposes the multi-sized feature data into seasonal terms and trend terms and outputs them; the seasonal terms are encoded by the encoder and then decomposed into seasonal terms and trend terms again, and the seasonal terms output by the encoder are processed by the decoder and then decomposed into seasonal terms and trend terms again; The dimension stacking unit stacks the trend terms output by the sequence decomposition module, the trend terms output by the encoder, the trend terms output by the decoder, and the seasonal terms output by the decoder. The stacked features are converted into clock error prediction data by the output module. The output module can specifically use an activation function or a fully connected layer.

[0034] In this embodiment, the encoder and decoder use self-correlation mechanisms and feed-forward neural networks to capture the long-range dependencies of time series.

[0035] Refer to Figure 2 , the encoder includes a self-correlation layer, a feed-forward network layer, and a sequence decomposition network connected in sequence; the self-correlation layer is used to extract the attention features of the input data, that is, the seasonal terms output by the sequence decomposition module, and the number of attention heads is set to 8; the attention features are further separated into seasonal terms and trend terms by the sequence decomposition network after being processed by the feed-forward network.

[0036] Refer to Figure 3, the decoder includes a sequentially connected temporal convolutional network, a first sequence decomposition layer, an autocorrelation network, a second sequence decomposition layer, a feed-forward network layer, and a third sequence decomposition layer. The input data of the decoder, which is the seasonal term output by the encoder, is processed by the temporal convolutional network and then further decomposed into a seasonal term and a trend term by the first sequence decomposition layer; the seasonal term extracts attention features through the autocorrelation layer, and the attention features are further separated into a seasonal term and a trend term by the second sequence decomposition layer; the seasonal term output by the second sequence decomposition layer is processed by the feed-forward network layer and then further decomposed into a seasonal term and a trend term by the third sequence decomposition layer.

[0037] In this embodiment, the encoder and the decoder cooperate to form a lightweight Autoformer network architecture. Compared with Figure 4 the Autoformer network architecture, multiple sequence decompositions are performed in the decoder to achieve the purification of the seasonal term and the trend term, and then the trend terms extracted by the encoder and the decoder are accumulated through the dimension stacking unit, realizing the purification and then stacking of the seasonal term and the trend term, so that the data features are more prominent and the time features are more obvious. Combined with the time encoding method adopted by the data embedding module, better physical interpretability is achieved, thereby improving the accuracy of the time series prediction task.

[0038] Obviously, in the satellite clock error prediction model, the data preprocessing module is fixed, and the parts of the model to be trained include the data embedding module, the multi-scale feature extraction module, the sequence decomposition module, the encoder, the decoder, the dimension stacking unit, and the output module.

[0039] Referring to Figure 5 , this embodiment also proposes a satellite clock error prediction method, including the following steps: St1, crop the real-time collected satellite clock error data stream into multiple time windows X1, X2,..., Xj,..., Xm of equal length, and perform difference processing and amplification processing on each time window to obtain the preprocessed data Z1, Z2,..., Zj,..., Zm; 1 ≤ j ≤ m; m is the number of windows, and N is the window length; Xj = {x (j-1)N+1 , x (j-1)N+2 ,..., x (j-1)N+i ,..., x jN}; Zj = {z (j-1)N+1 , z (j-1)N+2 ,..., z (j-1)N+i ,..., z jN}; x (j-1)N+1 , x (j-1)N+2 , x (j-1)N+i and x jN respectively represent the 1st, 2nd, jth, and Nth satellite clock error data in Xj, z(j-1)N+1 , z (j-1)N+2 , z (j-1)N+i and z jN respectively represent the pre - processed data after differencing and amplification corresponding to x (j-1)N+1 , x (j-1)N+2 , x (j-1)N+i and x jN ; 1 ≤ i ≤ N.

[0040] In this embodiment, before and after the pre - processing of a certain piece of data, as shown in Figure 6 and Figure 7 , it can be seen that through the differencing and amplification processing, the data trend becomes more obvious.

[0041] St2. Input the pre - processed data Z1, Z2, …, Zj, …, Zm into the trained satellite clock error prediction model, and the model outputs the prediction data {y1, y2,..., y k ,..., y mN}; y1, y2, y k and y mN respectively represent the satellite clock error data predicted for the 1st, 2nd, kth, and mNth.

[0042] The following combines specific embodiments to train and verify the above - mentioned satellite clock error prediction model.

[0043] In this embodiment, sampling is carried out from the data center of a certain university, and a total of about 60,000 pieces of satellite clock error data are obtained. The data lengths include three cases: 15 minutes, half an hour, and 1 hour.

[0044] In this embodiment, a data sample is constructed for each piece of data, and the specific method is as follows: First, determine the window length N for window processing of the data. Divide the data window into 2m consecutive window data X1, X2, …, Xj, …, Xm, Y1, Y2, …, Yj, …, Ym; construct the data sample {Xj, Yj|1 ≤ j ≤ m}; Xj = {x (j-1)N+1 , x (j-1)N+2 ,..., x (j-1)N+i ,..., x jN}; Yj = {y (j-1)N+1 , y (j-1)N+2 ,..., y (j-1)N+i ,..., y jN}; 1 ≤ j ≤ m.

[0045] Then, perform differencing and amplification processing on the data {Xj, 1 ≤ j ≤ m} in each data sample to obtain the processing result Zj corresponding to the data Xj, and the amplification factor is selected as 100,000; Zj = {z (j-1)N+1 , z (j-1)N+2 ,..., z (j-1)N+i ,..., z jN}; Construct the data sample {Zj, Yj|1 ≤ j ≤ m}.

[0046] In this embodiment, the data sample {Zj, Yj|1 ≤ j ≤ m} is divided into a training set and a test set, and the above satellite clock error prediction model and the given comparison models Autoformer and Transformer are trained on the training set.

[0047] The satellite clock error prediction model in this embodiment is as Figures 1-3 shown.

[0048] The model training process is as follows: S1. Initialize the model; S2. Extract learning samples from the training set, input the data Z1, Z2,..., Zj,..., Zm of the learning samples into the satellite clock error prediction model, and obtain the model prediction results Y'1, Y'2,..., Y'j,..., Y'm; Y'j = {y' (j-1)N+1 , y' (j-1)N+2 ,..., y' (j-1)N+i ,..., y' jN}; y' (j-1)N+1 , y' (j-1)N+2 , y' (j-1)N+i , y' jN respectively represent the (j - 1)N + 1, (j - 1)N + 2, (j - 1)N + i, jNth prediction data.

[0049] S3. Calculate the model loss on the learning samples, and use the gradient descent optimization method to update the model parameters by minimizing the model loss; the model loss can specifically use the mean square error loss MSE; ; {Yj = {y (j-1)N+1 , y (j-1)N+2 ,..., y (j-1)N+i ,..., y jN |1 ≤ j ≤ m}; Y'j = {y' (j-1)N+1 , y' (j-1)N+2 ,..., y' (j-1)N+i ,..., y' jN |1 ≤ j ≤ m}; S4. Repeat the above steps S2 - S3 until the model converges.

[0050] In this embodiment, it is set that the model converges after 15 iterations; in specific implementation, the model convergence condition can also be set as the convergence of the loss function, and an early stopping mechanism can be introduced, etc.

[0051] In this embodiment, the trained model is tested on the test set, and the MSE index and MAE index are evaluated respectively. The test results are shown in Table 1 and Figure 8 as follows.

[0052] ;

[0053] Table 1 Comparison table of loss function results of three models ; As can be seen from Table 1, the satellite clock error prediction model proposed in the present invention (abbreviated as the model of the present invention) improves by 4.29% in the MAE index and 1.69% in the MSE compared with Autoformer; compared with Transformer, it improves by 1.32% in the MAE index and 0.26% in the MSE.

[0054] From the test results, the improvement in MSE is lower than that in MAE, which may be because outliers have been removed and missing values have been imputed in the data source, improving the smoothness of the data.

[0055] However, the test results have also proved that the satellite clock error prediction model proposed in the present invention is more accurate in satellite clock error prediction than the existing models.

[0056] In this embodiment, a sample is selected from the test set, and the prediction results of the model of the present invention are further compared with the true values. The results are shown in Table 2 and Figure 9 as follows.

[0057] Table 2 List of error values for predicting 16 time steps ; The average error at 16 time steps: 0.000424975 ns, and the error reaches the order of 10e - 5, which proves the accuracy of the prediction of the present invention.

[0058] An ablation experiment is also carried out in this embodiment. The decoder in the satellite clock error prediction model shown in Figure 1 is replaced with the decoder shown in Figure 4 . The dimension stacking unit still stacks the trend terms output by the sequence decomposition module, the trend terms output by the encoder, the trend terms output by the decoder, and the seasonal terms output by the decoder to obtain the ablation model.

[0059] The results of the ablation experiment are shown in Table 3.

[0060] Table 3 Results of Ablation Experiments ; Through ablation experiments, it can be proved that the operation of repeatedly decomposing and purifying the trend term in the present invention significantly improves the accuracy of the prediction results.

[0061] In this embodiment, the data samples {Xj, Yj|1≤j≤m} are also divided into a training set and a test set. The model of the present invention and the ablation model are trained on the training set {Xj, Yj|1≤j≤m}, and then the model accuracy is tested on the test set {Xj, Yj|1≤j≤m} to compare the influence of data preprocessing on the prediction accuracy. The results are shown in Table 4.

[0062] Table 4 Results of Dataset Experiments ; It can be seen that the model trained with the dataset {Zj, Yj|1≤j≤m} has smaller MAE and MSE, that is, the model performance is better, further proving the advantages of the differential and amplification processing of the data in the present invention.

[0063] Of course, for those skilled in the art, the present invention is not limited to the details of the above exemplary embodiments, but also includes the same or similar structures that can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be construed as limiting the claimed rights.

[0064] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0065] The technologies, shapes, and structures not detailedly described in the present invention are all well-known technologies.

Claims

1. A satellite clock error prediction method, characterized in that, First, a satellite clock error prediction model is constructed. Its input is the satellite clock error window data after differential processing and amplification processing, and its output is the clock error prediction data for the time period after the input data; the satellite clock error prediction model extracts multi-scale features from the input data and performs sequence decomposition. The decomposed seasonal term is subjected to autocorrelation processing through an encoder and a decoder. Both the encoder and the decoder internally decompose the data into sequences. The seasonal term output by the decoder is dimensionally stacked with the trend term after each sequence decomposition and then transformed into clock error prediction data; Then, real-time satellite clock error data is collected and windowed. Differential processing and amplification processing are performed on each window data. The processed window data is input into the satellite clock error prediction model to obtain clock error prediction data.

2. The satellite clock error prediction method according to claim 1, characterized in that, The satellite clock error prediction model includes a data embedding module, a multi-scale feature extraction module, a sequence decomposition module, an encoder, a decoder, a dimension stacking unit, and an output module connected in sequence; the multi-scale feature extraction module uses convolution kernels of different sizes to perform convolution processing on the input data and then splices them to generate multi-scale features; The multi-scale features are decomposed by the sequence decomposition module into seasonal terms and trend terms; the seasonal term is encoded by the encoder and then decomposed again into seasonal terms and trend terms. The seasonal term output by the encoder is processed by the decoder and then decomposed again into seasonal terms and trend terms; the dimension stacking unit dimensionally stacks the seasonal term output by the decoder with the trend term after each sequence decomposition and then inputs it into the output module to be processed into clock error prediction data.

3. The satellite clock error prediction method according to claim 2, characterized in that, The data embedding module uses time coding embedding.

4. The satellite clock error prediction method according to claim 2, wherein, The decoder includes a temporal convolutional network, a first sequence decomposition layer, an autocorrelation network, a second sequence decomposition layer, a feedforward network layer, and a third sequence decomposition layer connected in sequence; Each sequence decomposition layer decomposes the input into sequences. The decomposed seasonal term propagates backward. The seasonal term output by the third sequence decomposition layer is used as the seasonal term output by the decoder; The trend terms of each sequence decomposition are dimensionally stacked to form the trend term output by the decoder.

5. The satellite clock error prediction method according to claim 2, wherein The encoder includes an autocorrelation layer, a feedforward network layer, and a sequence decomposition network connected in sequence.

6. The satellite clock error prediction method according to claim 2, characterized in that The output module uses a fully connected layer.

7. The satellite clock error prediction method according to any one of claims 1-6, characterized in that, The training process of the satellite clock error prediction model includes the following steps: Collect historical clock error data and perform windowing to obtain 2m consecutive window data X1, X2, …, Xj, …, Xm, Y1, Y2, …, Yj, …, Ym; construct data samples {Xj, Yj|1 ≤ j ≤ m}; the number of steps of X and Y is equal; Perform differential processing and amplification processing on Xj in the data samples {Xj, Yj|1 ≤ j ≤ m}, and transform Xj into preprocessed data Zj to obtain training samples {Zj, Yj|1 ≤ j ≤ m}; Extract the data Z1, Z2, …, Zj, …, Zm of multiple training samples and input them into the satellite clock model to obtain prediction data Y'1, Y'2, …, Y'j, …, Y'm; compare the data Y1, Y2, …, Yj, …, Ym and the prediction data Y'1, Y'2, …, Y'j, …, Y'm, calculate the model loss, and use the gradient descent optimization method to update the model parameters; repeat this step until the model reaches the convergence condition.

8. The satellite clock error prediction method according to claim 7, wherein, The model convergence condition is: the number of model iterations reaches the set value, or the absolute value of the model loss range of at least the last three rounds is less than the set threshold.

9. A satellite clock error prediction system, characterized in that, It includes a data preprocessing module, a memory and a processor, wherein the memory stores a computer program and a satellite clock error prediction model, the data preprocessing module is used to obtain satellite clock error data collected in real time and perform differential processing and amplification processing, the processor is connected to the memory and the data preprocessing module, and the processor is used to execute the computer program to implement the satellite clock error prediction method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, A computer program is stored, and when the computer program is executed, it is used to implement the satellite clock error prediction method as described in any one of claims 1-7.

Images