Inter-station satellite bidirectional time forecasting method

Abstract

The invention provides an inter-station satellite bidirectional time forecasting method, and belongs to the technical field of time calibration and forecasting. According to the method, a synchronous orbit GEO satellite is used as a forwarding node of a space segment, two stations respectively carry out BPSK modulation on local time mark signals to form frequency signals, and the frequency signals are transmitted to an opposite end through power amplification. In the data processing process, an inter-station time comparison result is calculated according to an existing model, then modeling and learning training are carried out on measured data by using an LSTM model, and finally a predicted inter-station time comparison sequence is output to be used for correcting inter-station time deviation of the next period and forecast of various clock differences.

Description

Technical Field

[0001] The invention belongs to the technical field of time calibration and prediction, and in particular relates to an inter-station satellite two-way time prediction method. Background Art

[0002] Currently, a variety of methods are available for achieving high-precision time and frequency synchronization, comparison, or calibration between two locations. These include fiber-optic time and frequency transfer, satellite common view technology, 1588-PTP timing, and satellite bidirectional time and frequency transfer. Fiber-optic time and frequency transfer technology originated in the 1970s and has been developed due to the widespread use of fiber-optic communications. Its advantages lie in its high precision and stability, with time synchronization accuracy reaching 0.1ns and frequency stability reaching 1E-15 / second. Optical fiber transmission has low power loss and can extend transmission distances through relays, making it a primary method for time synchronization. Satellite common-view technology, originally known as GPS one-way time transfer and GPS common-view time transfer, emerged with the development of satellite navigation systems. It is relatively easy to implement and maintain, enabling long-distance time synchronization between two locations as long as there is visible satellite. After smoothing, time synchronization accuracy reaches approximately 5ns. However, it is inherently fragile, highly susceptible to interference, and exhibits poor short-term stability. 1588-PTP is a network-based, high-precision time synchronization protocol. The first version of the protocol standard, released in 2002, was recognized by the IEEE, and the third version began revision in 2020. It utilizes a master-slave approach to time synchronization, achieving nanosecond-level synchronization accuracy. It can be used in existing local and wide area networks (LANs), supports a variety of network topologies, and has strong security protections against malicious attacks and tampering. PTP time synchronization requires specialized, high-performance hardware, requiring significant expertise to deploy and maintain, and has not yet gained widespread adoption. Satellite Two-Way Time and Frequency Transfer (TWSTFT) technology originated in the 1960s. The application of spread spectrum technology significantly improved measurement accuracy, leading to a long-term two-way satellite comparison link established by the United States and Germany in 1983 achieving an accuracy of 1 nanosecond. Statistics show that over two-thirds of the atomic clocks and real-time physical signals used in UTC / TAI calculations are compared using TWSTFT. This technology has become the primary method for generating and maintaining international UTC / TAI and the primary means for achieving long-distance, precise time and frequency synchronization in satellite navigation systems and large-scale time and frequency systems, playing a central role. However, in practical use, TWSTFT is limited by device latency, propagation path latency, and satellite motion latency, which, to a certain extent, affects inter-station time comparison results and traceability. Further improving the accuracy or prediction precision of satellite two-way time and frequency comparison results can reduce measurement uncertainty, play a significant role in future UTC / TAI joint timekeeping, national integrated PNT time and frequency unification, and enhance the performance of satellite navigation systems, potentially yielding significant benefits.

[0003] The time delays generated during the two-way satellite time and frequency transmission process can be mainly divided into three categories: the first is equipment delay, which mainly includes the delay of ground transmitting and receiving equipment, time counter delay, modem delay and satellite transponder delay; the second is propagation path delay, which mainly includes atmospheric propagation delay such as troposphere and ionosphere; the third is motion delay, which mainly refers to the measurement delay caused by the Sagnac effect caused by satellite motion and earth rotation and the asymmetry of satellite geometric path. The above delays cause the clock difference between the two stations measured by TWSTFT technology to always have deviations or non-absolute periodic fluctuations. In particular, the error introduced by the motion of the satellite relative to the ground will increase the measurement uncertainty by hundreds of picoseconds. With the increasing demand for time synchronization accuracy, TWSTFT technology and its transmission accuracy improvement have always been an important topic of research and concern for scholars in the field of time and frequency at home and abroad. A variety of TWSTFT data processing methods or result correction models have been proposed around it.

[0004] In the current research on TWSTFT technology, scholars mostly focus on using mathematical models to analyze and model the propagation path delay, Sagnac effect and device delay, and continuously iterate and optimize these models to eliminate as much as possible the noise and errors introduced by them in the application of TWSTFT technology, thereby improving the accuracy of two-way time-frequency comparison. However, this approach requires the use of a large number of prior data models for analysis. In order to extract periodic and non-periodic terms, it is necessary to undergo long-term and continuous observations, and to do a large amount of complex simulation modeling and calculations. Even so, the final comparison accuracy is difficult to significantly improve. In addition, the model established through individual cases cannot be applied to the fluctuations in long-term TWSTFT measurement in different regions and different signal bands.

[0005] In fact, if we can use the more mature neural networks or intelligent algorithms to model and predict TWSTFT measurement results, we can eliminate the influence of periodic or non-periodic terms caused by various time delays and Sagnac effects, and obtain accurate forecast results. We can use it to directly correct the measurement value of the next period, eliminate multiple error terms in the TWSTFT comparison process, and use it to calculate the trend item parameters of the slave station, thereby effectively improving the traceability and time synchronization accuracy of remote stations. Summary of the Invention

[0006] TWSTFT technology has been able to achieve a high time synchronization accuracy (≤0.5ns), which can basically meet the current usage needs. However, considering the use of future time and frequency synchronization technology, reducing the error introduced by the measurement uncertainty between remote stations, and further improving the remote time synchronization capability and joint timekeeping capability, it is necessary to conduct in-depth research on the existing TWSTFT measurement results. To this end, the present invention proposes a two-way satellite time prediction scheme between stations.

[0007] A first aspect of the present invention provides a method for satellite bidirectional time prediction between stations, the method comprising:

[0008] Step S1, using TWSTFT technology to measure the time difference between the master station and the slave station separated by two locations;

[0009] Step S2: Continuously adjust the time scale signal of the slave station based on the time scale signal of the master station to complete the frequency offset calibration of the master station and the slave station;

[0010] Step S3: Process the TWSTFT measurement data to obtain a smooth and orderly sequence of inter-station time differences;

[0011] Step S4: standardize and serialize the processed data and build a neural network based on the LSTM model;

[0012] Step S5: Output the LSTM prediction result, and compare the calculated LSTM prediction result with the actual measurement result.

[0013] In step S1, the time difference between the master station and the slave station separated by two locations is measured using the TWSTFT technology. The master station and the slave station simultaneously observe the same GEO geostationary orbit satellite, use the C or Ku band to modulate, transmit and measure the time signal, and complete the original solution of the time difference between the stations through data exchange.

[0014] In step S2: Based on the inter-station time comparison result measured by the current TWSTFT, the time signal of the master station is used as a reference to continuously adjust the time signal of the slave station until it is adjusted to a frequency deviation that can be maintained by the two stations, thereby eliminating the inherent frequency deviation of the time signals of the two stations and avoiding frequency inconsistency between the two stations.

[0015] In step S3: When the time deviation between the two stations remains stable, the original TWSTFT measurement data is extracted, and the missing points, outliers and singular points are analyzed and processed to obtain a smooth and orderly time difference series between stations. The deviation of this time series only includes the inherent deviation of the two stations and multiple delay errors in the TWSTFT measurement process. The delay error includes equipment delay error, propagation path delay error and satellite motion delay error.

[0016] In step S4: TWSTFT measurement data is learned by setting data dimensions and specifying training options; the LSTM model is used to effectively learn long-time dependencies in the sequence to perform data prediction tasks based on time correlation; the TWSTFT measurement comparison results of the first two days after data processing are used to build a model and calculate the inter-station time comparison results for the next day.

[0017] In step S5: the calculated LSTM prediction result is compared with the actual measurement result, and the time difference, frequency accuracy and root mean square error between the two are calculated; if the predicted frequency accuracy is ≤1e-15 and the root mean square error RMSE is ≤0.3, it is determined to be a valid prediction result. Based on the prediction result, the frequency deviation adjustment amount of the slave station in the next period is calculated, and the slave station time signal is adjusted in advance to complete the high-precision time synchronization of the two stations. At the same time, other external clock differences observed locally are predicted.

[0018] In the method:

[0019] Using geostationary orbit GEO satellites as forwarding nodes in the space segment, the two stations each perform BPSK modulation on the local time-scale signal to form a frequency signal, which is then amplified and transmitted to the other end.

[0020] During the data processing process, the inter-station time comparison results are first solved according to the existing model, and then the measured data are modeled and trained using the LSTM model, and finally the predicted inter-station time comparison sequence is output; the comparison results are used to correct the inter-station time deviation and various clock error forecasts for the next period.

[0021] A second aspect of the present invention provides an inter-station satellite two-way time prediction system, the system comprising a processing unit configured to execute:

[0022] Use TWSTFT technology to measure the time difference between the master station and the slave station separated by two locations;

[0023] The time scale signal of the slave station is continuously adjusted based on the time scale signal of the master station to complete the frequency offset calibration of the master station and the slave station;

[0024] The TWSTFT measurement data are processed to obtain a smooth and orderly series of time differences between stations;

[0025] Standardize and serialize the processed data and build a neural network based on the LSTM model;

[0026] Output the LSTM prediction result and compare the calculated LSTM prediction result with the actual measurement result.

[0027] The third aspect of the present invention discloses an electronic device. The electronic device includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it implements the inter-station satellite two-way time prediction method of the first aspect of the present disclosure.

[0028] A fourth aspect of the present invention discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the inter-station satellite two-way time prediction method of the first aspect of the present disclosure.

[0029] In view of the fact that TWSTFT technology still has the problem of fluctuation of periodic or non-periodic terms in the two-way comparison results due to equipment delay, propagation path delay and satellite motion delay in actual use, the present invention proposes to use the Long Short-Term Memory (LSTM) model in the neural network to construct an inter-station LSTM time series model. After model training, the measurement results of the next cycle of TWSTFT are predicted. The results can be used for: first, improving the accuracy of inter-station time synchronization. The slave station can calculate and adjust the output time scale signal of the next cycle in advance to maintain a high-precision time synchronization level and enhance the timeliness of synchronization; second, correcting the inter-station measurement error. By using the LSTM model prediction results to correct the error terms in the TWSTFT comparison process, the inter-station time difference can be corrected in a large-scale time synchronization system (such as a satellite navigation system), the satellite-to-ground measurement error introduced by multi-station tracking satellites can be deducted, and the satellite-to-ground clock difference reduction result can be further corrected, effectively improving the satellite clock difference prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0031] Figure 1 2 is a diagram showing the overall architecture of inter-station satellite two-way time prediction according to an embodiment of the present invention.

[0032] Figure 2 The figure is a schematic diagram of the implementation principle of the method for predicting the time comparison results between stations based on the LSTM model according to an embodiment of the present invention.

[0033] Figure 3 Schematic diagram of the time difference between the master station and the slave station measured by TWSTFT according to an embodiment of the present invention.

[0034] Figure 4 LSTM model structure diagram according to an embodiment of the present invention.

[0035] Figure 5 Schematic diagram of actual forecast value and frequency deviation according to an embodiment of the present invention.

[0036] Figure 6 Schematic diagram of prediction residual and root mean square error according to an embodiment of the present invention. DETAILED DESCRIPTION

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0038] Definitions of Abbreviations and Key Terms:

[0039] ARMA: Autoregressive MovingAverrage, autoregressive moving average model.

[0040] ARIMA: Autoregressive Integrated MovingAverrage, autoregressive integrated moving average model.

[0041] BIPM: Bureau International des Poids et Mesures, International Bureau of Weights and Measures.

[0042] BPSK: Binary Phase Shift Keying, binary phase shift keying.

[0043] ECEF: Earth-Centered Earth-Fixed, Earth-centered Earth-fixed coordinate system.

[0044] GEO: Geostationary Earth Orbit Satellite.

[0045] GPS: Global Positioning System.

[0046] IGSO: Inclined GeoSynchronous Orbit, inclined geosynchronous orbit.

[0047] IEEE: Institute of Electrical and Electronics Engineers.

[0048] LSTM: Long Short-Term Memory, long short-term memory network.

[0049] NIST: National Institute of Standards and Technology, National Institute of Standards and Technology of the United States.

[0050] NSTC: National Time Service Center, National Time Service Center of the Chinese Academy of Sciences.

[0051] PNT: Positing, Navigation and Timing, positioning, navigation and timing system.

[0052] PPS: Pulse Per Second, pulse per second.

[0053] PTB: Physikalisch-Technische Bundesanstalt, Germany's national metrology institute.

[0054] PTP: Precision Time Protocol.

[0055] RMSE: Root Mean Square Error, root mean square error.

[0056] RNN: Recurrent Neural Network, recurrent neural network.

[0057] SARIMA: SeasonalARIMA, seasonal autoregressive integrated moving average model.

[0058] TAI: Temps Atomique International, International Atomic Time.

[0059] TWSTFT: Two Way Satellite Time and Frequency Transfer, satellite two-way time and frequency transfer.

[0060] UTC: Universal Time Coordinated.

[0061] Many models exist for processing time series information, such as the Autoregressive Moving Average (ARMA), the Autoregressive Integrated Moving Average (ARIMA), and the Seasonal ARIMA (SARIMA). These models are highly effective for processing short- and medium-term data, leveraging their strengths to achieve accurate forecasts. However, they lack accuracy for long-term data. Since most TWSTFT links continuously measure and calculate 1PPS pulse-per-second time-frequency signals, the sheer volume of time difference data can lead to these three models misidentifying long-term data and preventing the development of an effective model. Taking into account the characteristics of TWSTFT's continuous time difference measurements and combining them with currently popular neural network technology, a Long Short-Term Memory (LSTM) model was chosen. This model can retain long-term dependencies and handle sequence data of varying lengths, making it suitable for modeling and forecasting inter-station differences measured by TWSTFT.

[0062] The LSTM model, a variant of the recurrent neural network (RNN), is designed for applications in time series processing, natural language processing, speech recognition, and other fields. It addresses the long-term dependencies inherent in RNNs by introducing a gating strategy. The LSTM model primarily consists of three gate structures: a forget gate, an input gate, and an output gate; and two states: a cell state and a hidden state. Through this structure and state processing, the LSTM model possesses a certain degree of memory capacity, acting as a memory manager that selectively retains useful information and forgets useless information, capturing dependencies and ultimately outputting ideal results. In engineering practice, the LSTM model can be extended and applied to time comparison measurement systems in the time-frequency domain. It can model and train TWSTFT measurement results from two remote stations, ultimately outputting high-precision predictions for inter-station time comparisons. Based on this research approach, a method for forecasting satellite two-way time comparisons based on the LSTM model is proposed.

[0063] The TWSTFT inter-station time synchronization system used in a laboratory is used as the research object. The system uses a synchronous orbit GEO satellite as a forwarding node in the space segment. The two stations perform BPSK modulation on the local time scale signal to form a frequency signal, which is then transmitted to the other end after power amplification. In the data processing process, the inter-station time comparison result is first solved according to the existing model, and then the measured data is modeled and trained using the LSTM model, and finally the predicted inter-station time comparison sequence is output. The result is used to correct the inter-station time deviation and various clock error predictions for the next period. The overall architecture is as follows: Figure 1 shown.

[0064] The specific implementation process can be divided into the following steps (such as Figure 2 shown):

[0065] The first step is to use TWSTFT technology to measure the time difference between the master station and the slave station that separate the two locations. This requires that the two locations can simultaneously observe the same GEO geostationary orbit satellite, use the C or Ku band to modulate, transmit and measure the time signal, and complete the original solution of the time difference between the stations through data exchange.

[0066] The second step is to calibrate the master and slave station frequency offsets. To eliminate the inherent frequency deviation of the two station time-scale signals and avoid the trend effect caused by frequency inconsistency or excessive frequency deviation between the two stations, the slave station's time-scale signal is continuously adjusted based on the inter-station time comparison results measured by TWSTFT, using the master station's time-scale signal as a reference, until a frequency deviation is achieved that can be maintained between the two stations.

[0067] The third step is TWSTFT measurement data processing. While the time offset between the two stations remains stable, the original TWSTFT measurement data is extracted. After analyzing and processing missing points, outliers, and singular points, a smooth and orderly series of inter-station time differences is obtained. The deviations in this time series only include the inherent offset between the two stations and multiple delay errors in the TWSTFT measurement process, including equipment delay error, propagation path delay error, and satellite motion delay error.

[0068] The fourth step is modeling using the LSTM method. After data processing, the data is standardized and serialized, and a neural network based on the LSTM model is built. The TWSTFT measurement data is learned by setting the data dimensions and specifying training options. The LSTM model has the advantage of effectively learning long-span dependencies in a sequence, making it widely applicable to data prediction tasks with strong temporal correlations. However, its disadvantages are high computational complexity and limited computational efficiency. Therefore, calculations are performed on a daily basis as the minimum period. The model is built using the TWSTFT measurement comparison results from the first two days of data processing to calculate the inter-station time comparison results for the next day.

[0069] Step 5: Output the LSTM prediction results. Compare the calculated LSTM prediction results with the actual measurement results to calculate the time difference, frequency accuracy, and root mean square error (RMSE). If the predicted frequency accuracy is ≤1e-15 and the root mean square error (RMSE) is ≤0.3, the prediction result is considered valid. Based on this prediction result, the frequency offset adjustment for the slave station in the next period can be calculated. The slave station's time-stamp signal can be adjusted in advance to achieve high-precision time synchronization between the two stations. This also allows for the prediction of other external clock errors observed locally.

[0070] Example

[0071] 1. Frequency deviation calibration

[0072] A laboratory has two widely separated timekeeping sites, referred to as the master and slave. To synchronize the system's internal time, the master and slave use TWSTFT technology to measure the time difference between the two locations. The slave's clock is adjusted based on the measurement results to maintain a high time difference with the master. Due to the significant initial frequency deviation between the two sites, the slave's time-frequency signal is adjusted based on historical measurement data to reduce the fluctuation caused by the inherent frequency deviation. This keeps the frequency deviation between the slave and master within 2e-14. At this level, the time difference between the two sites due to the inherent frequency deviation does not exceed 1.73ns per day, thus minimizing the impact of the inherent frequency deviation between the sites.

[0073] 2. Data Processing

[0074] (1) Data collection and cleaning

[0075] The laboratory uses TWSTFT technology to obtain the original observation data of the two stations, and obtains the time difference between the two stations through formula calculation, time delay deduction and data format conversion. The continuous effective time difference data of the two stations for 9 days is extracted, and the time difference value for one week (7 days) can be predicted with days as the smoothing unit time. Since there are many missing values, jump values ​​or wild values ​​in the data, the time difference data needs to be smoothed to reduce the impact of these outliers on the LSTM model training. The 9-day time difference data of the master and slave stations after processing is as follows Figure 3 shown.

[0076] The above figure directly illustrates the characteristics and quality of the time difference data between the master and slave stations. The frequency deviation between the two stations is small, maintaining at 1.51E-15. The daily fluctuation in the time difference due to the inherent frequency deviation between the master and slave stations can be assumed to be 0.1ns. However, due to equipment delay, propagation path delay, and Sagnac effect delay, the maximum time difference between the two stations over a period of 9 days is -0.05ns, the minimum is -5.08ns, and the average is -2.49ns. This indicates the presence of significant periodic and aperiodic noise between the two locations, which can affect time synchronization and the unified calculation of the external time difference.

[0077] (2) Data standardization

[0078] The LSTM model is sensitive to the scale of the input data, so the data needs to be standardized.

[0079] Standardization is the process of transforming data into a distribution with a mean of 0 and a standard deviation of 1.

[0080]

[0081] Where X represents the original data set, μ represents the mean, and σ represents the standard deviation.

[0082] (3) Data partitioning

[0083] The data was divided into a training set and a test set. The training set was used to train the model, while the test set was used to adjust model parameters and evaluate model performance. Three consecutive days of inter-station TWSTFT measurements were selected. The first two days' data were used to predict the next day's data, with the data split into a 2 / 3 training set and a 1 / 3 test set. Nine days of TWSTFT measurements were sufficient to predict a week's worth of results.

[0084] (4) Data serialization processing

[0085] LSTM models need to process sequential data. If the data is not in time series format, it must first be converted to a sequence format and then alternated by one time step. In LSTM, the time step is the basic unit of sequence data processing; each time step represents an element in the sequence. The LSTM model constructs a holistic understanding of the sequence by gradually processing these time steps. Alternating time steps can transform sequence data into a static state, by using a=1 differencing to remove growth trends in the data. It can also transform time series problems into supervised learning problems, using the observations from the previous time step as input to predict the observations from the current time step.

[0086] 3. Build LSTM model

[0087] Modeling the three gates and two states of the LSTM model. The forget gate, input gate, and output gate primarily control the flow and storage of information, enabling LSTM to effectively handle long-term dependencies in sequence data. The forget gate determines how much of the cell state from the previous time step is retained in the cell state of the current time step, the input gate determines how much of the input information of the current time step is incorporated into the cell state, and the output gate determines how much of the cell state of the current time step is output to the hidden state. The cell state and hidden state of the LSTM model are responsible for long-term and short-term memory, respectively. The cell state not only remembers the information of a specific time step but also maintains a relatively stable memory of the entire time series, representing a long-term "memory"; the hidden state remembers the output of the current and previous time steps, representing a short-term "memory."

[0088] The LSTM model structure is as follows Figure 4 As shown, at time t, there are three inputs: cell state Ct-1, hidden state Ht-1, and input Xt; there are two outputs: cell state Ct, hidden state, and output Ht. The work content is as follows:

[0089] (i) The information of the previous time step and all previous prediction data Ct-1 is transmitted on the top line. The prediction result Ht-1 of the previous time step at time t and the TWSTFT measurement time difference Xt will appropriately modify the LSTM model prediction result Ct and then be transmitted to the next time step;

[0090] (ii) The previous time step and all previous prediction data Ct-1 participate in the calculation of the prediction result Ht of the current time step at time t;

[0091] (iii) The information of the prediction result Ht-1 at the previous time step modifies the cell state through the "gate" structure and participates in the output calculation.

[0092] (1) Building a neural network

[0093] The LSTM model can process multi-dimensional data sequences, such as two-dimensional or three-dimensional data. For the inter-station time difference sequence measured by TWSTFT, it is only necessary to set the input data dimension and output data dimension of the LSTM model to one dimension, and then set the number of hidden units. The hidden units are usually set according to the complexity of the data, usually to a multiple of 32. Due to the large amount of data and high complexity of the inter-station time difference, it is 288 here.

[0094] The neural network built using the LSTM model consists of four layers: the input layer, the LSTM layer, the fully connected layer, and the regression layer. The input layer converts input data into a format that can be processed internally by the neural network; the LSTM layer processes sequential data and memorizes long-term dependencies; the fully connected layer connects all nodes in the previous layer to every node in the next layer, enabling the model to learn and extract input features more deeply; and the regression layer generates regression predictions.

[0095] (2) Specify training options

[0096] The LSTM model's solver was set to the Adaptive Moment Estimation (Adam) method. This solver calculates adaptive learning rates for each parameter, resulting in faster convergence and more effective learning compared to other neural network optimizer algorithms. Due to the large amount of data, the neural network training rate is slow. To maintain effective, convenient, and accurate training, the number of training epochs was set to 30. The gradient threshold was set to 1, and the initial learning rate was specified to be 0.005. After 15 epochs of training, the learning rate was reduced by a factor of 0.2.

[0097] 4. Training Results

[0098] Denormalize the predicted value after training with the LSTM model, output the forecast result and calculate the time difference, frequency deviation, forecast residual and root mean square error between it and the actual measured value, such as Figure 5-6 shown.

[0099] right Figure 5-6The forecast data are statistically analyzed and evaluated, including the actual frequency deviation, forecast frequency deviation and forecast accuracy, as well as the maximum, minimum, mean and root mean square error of the forecast residuals. The specific results are shown in Table 1.

[0100] Table 1: LSTM model prediction statistics and evaluation results

[0101]

[0102] Therefore, the LSTM model can be used to predict the time difference between the TWSTFT measurements of the two stations. In the 7-day forecast data, the actual forecast accuracy is ≤0.9e-15 (the indicator requirement is ≤2e-15). Through calculation, it can be obtained that the time difference between stations caused by forecast uncertainty does not exceed 0.078ns per day, the maximum forecast residual is 0.97ns (the indicator requirement is ≤1.5ns), the minimum is -1.48ns, the average is -0.02ns, and the root mean square error is ≤0.2 (the indicator requirement is ≤0.3). Through specific embodiments, it is found that the use of the LSTM model can effectively predict the time difference between TWSTFT measurement stations with high accuracy, with a forecast accuracy better than 2e-15 and a forecast root mean square error better than 0.3. The frequency deviation of the master and slave stations can be corrected or adjusted according to the results obtained by the LSTM intelligent model forecast, and it can also be used to correct the forecast clock difference with the external clock, providing an accurate and reliable intelligent forecasting method for high-precision station time unification.

[0103] In summary, in practice, TWSTFT technology still suffers from periodic and aperiodic fluctuations in bidirectional comparison results due to device latency, propagation path delay, and satellite motion delay. This study proposes using the Long Short-Term Memory (LSTM) model within a neural network to construct an inter-station LSTM time series model. After model training, this model can provide a high-precision forecast of the next TWSTFT measurement result. For time synchronization systems using TWSTFT technology, this LSTM-based prediction method can effectively predict inter-station time differences. The results can be used to: first, improve inter-station time synchronization accuracy, allowing slave stations to pre-calculate and adjust their output time-scale signals for the next cycle, maintaining high-precision time synchronization; and second, correct inter-station measurement errors. The prediction results can be used to correct inter-station time differences in large-scale time synchronization systems (such as satellite navigation systems), offsetting satellite-to-ground measurement errors introduced by multiple stations tracking satellites, and effectively improving the accuracy of satellite clock error prediction. This approach has practical engineering applications.

[0104] Please note that the technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification. The above embodiments only express several implementation methods of the present application. The description is relatively specific and detailed, but it cannot be understood as a limitation on the scope of the invention patent. It should be pointed out that for ordinary technicians in this field, without departing from the concept of this application, several variations and improvements can be made, which all fall within the scope of protection of this application. Therefore, the scope of protection of the patent in this application shall be based on the attached claims.

Claims

1. A method for satellite bidirectional time prediction between stations, characterized in that: The method comprises: Step S1, using TWSTFT technology to measure the time difference between the master station and the slave station separated by two locations; Step S2: Continuously adjust the time scale signal of the slave station based on the time scale signal of the master station to complete the frequency offset calibration of the master station and the slave station; Step S3: Process the TWSTFT measurement data to obtain a smooth and orderly sequence of inter-station time differences; Step S4: standardize and serialize the processed data and build a neural network based on the LSTM model; Step S5: Output the LSTM prediction result, and compare the calculated LSTM prediction result with the actual measurement result.

2. The method for predicting satellite time between stations according to claim 1, wherein: In step S1, the time difference between the master station and the slave station separated by two locations is measured using the TWSTFT technology. The master station and the slave station simultaneously observe the same GEO geostationary orbit satellite, use the C or Ku band to modulate, transmit and measure the time signal, and complete the original solution of the time difference between the stations through data exchange.

3. The method for predicting satellite time between stations according to claim 2, wherein: In step S2: Based on the inter-station time comparison result measured by the current TWSTFT, the time signal of the master station is used as a reference to continuously adjust the time signal of the slave station until it is adjusted to a frequency deviation that can be maintained by the two stations, thereby eliminating the inherent frequency deviation of the time signals of the two stations and avoiding frequency inconsistency between the two stations.

4. The method for predicting satellite time between stations according to claim 3, wherein: In step S3: When the time deviation between the two stations remains stable, the original TWSTFT measurement data is extracted, and the missing points, outliers and singular points are analyzed and processed to obtain a smooth and orderly time difference series between stations. The deviation of this time series only includes the inherent deviation of the two stations and multiple delay errors in the TWSTFT measurement process. The delay error includes equipment delay error, propagation path delay error and satellite motion delay error.

5. The method for predicting satellite time between stations according to claim 4, wherein: In step S4: learning the TWSTFT measurement data by setting the data dimension and specifying the training options; The LSTM model is used to effectively learn the dependencies of long time spans in the sequence to perform data prediction tasks based on time correlation. The TWSTFT measurement comparison results of the first two days after data processing are used to build a model and calculate the inter-station time comparison results for the next day.

6. The method for satellite bidirectional time prediction between stations according to claim 5, characterized in that: In step S5: the calculated LSTM prediction result is compared with the actual measurement result, and the time difference, frequency accuracy and root mean square error between the two are calculated; if the predicted frequency accuracy is ≤1e-15 and the root mean square error RMSE is ≤0.3, it is determined to be a valid prediction result. Based on the prediction result, the frequency deviation adjustment amount of the slave station in the next period is calculated, and the slave station time signal is adjusted in advance to complete the high-precision time synchronization of the two stations. At the same time, other external clock differences observed locally are predicted.

7. The method for predicting satellite time between stations according to any one of claims 1 to 6, characterized in that: In the method: Using geostationary orbit GEO satellites as forwarding nodes in the space segment, the two stations each perform BPSK modulation on the local time-scale signal to form a frequency signal, which is then amplified and transmitted to the other end. During the data processing process, the inter-station time comparison results are first solved according to the existing model, and then the measured data are modeled and trained using the LSTM model, and finally the predicted inter-station time comparison sequence is output; the comparison results are used to correct the inter-station time deviation and various clock error forecasts for the next period.

8. An inter-station satellite two-way time prediction system, characterized in that: The system includes a processing unit configured to perform: Use TWSTFT technology to measure the time difference between the master station and the slave station separated by two locations; The time scale signal of the slave station is continuously adjusted based on the time scale signal of the master station to complete the frequency offset calibration of the master station and the slave station; The TWSTFT measurement data are processed to obtain a smooth and orderly series of time differences between stations; Standardize and serialize the processed data and build a neural network based on the LSTM model; Output the LSTM prediction result and compare the calculated LSTM prediction result with the actual measurement result.

9. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, it implements the inter-station satellite two-way time prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for two-way time prediction between satellite stations according to any one of claims 1 to 7 is implemented.

Images