A method for joint prediction of multiple optical ground station link states and candidate station determination

Abstract

This invention discloses a method for joint prediction of link status and candidate station determination across multiple optical ground stations. The method acquires historical atmospheric observation data from multiple optical ground stations, constructs a historical time-series input sequence, and extracts shared time-series features from multiple stations using a shared time-series encoder. Based on this, a cloud cover prediction branch and a conditionally coupled atmospheric turbulence prediction branch are constructed, and physical consistency constraint weights are generated using auxiliary features to jointly train the joint prediction model. Then, based on the cloud cover prediction results and atmospheric turbulence prediction results, future link availability evaluation results are generated, and a candidate station set and station priority ranking results are output. This invention can improve the accuracy, stability, and physical rationality of link status prediction in multi-station scenarios, enhance the generalization ability under different climate zone conditions, reduce the errors caused by relying solely on single-station or single-cloud cover predictions, and improve the support capability and engineering application value of candidate station recommendation results for collaborative access of multiple optical ground stations.

Description

Technical Field

[0001] This invention belongs to the field of satellite optical communication and intelligent prediction technology, specifically relating to a method for joint prediction of link status of multiple optical ground stations and determination of candidate stations. Background Technology

[0002] Free-space optical communication (FSO) is a crucial technology for achieving high-speed data transmission between satellites and the ground. With the large-scale deployment of low-Earth orbit satellite constellations and the continuous growth in demand for satellite-to-ground data transmission, the problem of link interruptions and insufficient system availability caused by local weather conditions affecting a single optical ground station has become increasingly prominent. Therefore, employing multiple optical ground stations for coordinated reception and improving overall link availability through site diversity has become the mainstream technical approach in multi-optical ground station collaborative free-space optical communication systems. In this scenario, cloud cover determines link connectivity, and atmospheric turbulence determines link transmission quality. Accurately predicting the future link status of each station is fundamental to achieving multi-station collaborative access control.

[0003] Currently, link state prediction mainly includes two technical approaches: one is the numerical prediction method based on atmospheric physical mechanisms, which solves for future weather conditions by establishing physical models of cloud evolution and boundary layer changes, and has the characteristic of clear physical meaning; the other is the machine learning prediction method based on historical observation data, which achieves prediction by learning the mapping relationship between meteorological characteristics and link state, and has the characteristics of flexible modeling and high computational efficiency.

[0004] However, existing technologies still have significant shortcomings in collaborative applications of multiple optical ground stations. Current methods employ a single-station independent modeling approach, failing to fully utilize the commonalities and correlations in the temporal evolution and spatial distribution of atmospheric data from multiple stations. This results in insufficient parameter reuse capabilities, limited prediction accuracy for small sample stations, and the need for model retraining when adding new stations. Most existing methods only predict cloud cover, making it difficult to simultaneously output multi-dimensional atmospheric turbulence parameters characterizing link transmission quality, and failing to jointly characterize link connectivity and quality. Numerical prediction methods based on atmospheric physics mechanisms suffer from high computational complexity, poor real-time performance, and difficulty adapting to the local microclimate of ground stations under different terrain and altitude conditions; accuracy decreases significantly with increasing prediction time steps. While purely data-driven machine learning methods offer high flexibility, their generalization ability is limited by the coverage of training samples, leading to large prediction biases under extreme weather conditions and prone to abnormal outputs that do not conform to the laws of atmospheric physics evolution. Furthermore, existing prediction results only provide raw numerical and probabilistic information, failing to effectively connect with the requirements of multi-station collaborative access control and unable to directly generate candidate station determination results.

[0005] Therefore, existing technologies are still insufficient to simultaneously meet the requirements of high precision, multiple parameters, long timeliness, consistent mechanism, and decision-friendly operation in multi-station collaborative scenarios. There is an urgent need for a technology for joint prediction of link status and candidate station determination for multiple optical ground stations, so as to provide a reliable and forward-looking decision-making basis for multi-station collaborative access control of space-ground free space optical communication systems. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes a method for joint prediction of multi-optical ground station link states and candidate station determination, which improves the accuracy, stability, and physical rationality of multi-station atmospheric state prediction, and generates a set of candidate stations and a station priority ranking result.

[0007] The above objectives are achieved through the following technical solutions:

[0008] Step 1: Acquire historical atmospheric observation data from multiple optical ground stations, perform time alignment, time feature construction, and sliding window extraction on the historical atmospheric observation data, and construct the historical time series input sequence for each station.

[0009] Step 2: Based on the historical time-series input sequences of each site constructed in Step 1, each site is uniformly encoded by a shared time-series encoder to obtain the time-series feature representations corresponding to each site.

[0010] Step 3: Based on the temporal feature representations of each station obtained in Step 2, construct a cloud cover prediction branch and a conditionally coupled atmospheric turbulence prediction branch for multiple preset time steps in the future; use the cloud cover condition information output by the cloud cover prediction branch to conditionally correct the atmospheric turbulence prediction branch to form a coupled dual-branch output structure of the joint prediction model.

[0011] Step 4: Based on the cloud cover condition information output in Step 3 and the physical consistency constraint weights generated from the auxiliary features in the historical atmospheric observation data obtained in Step 1, the cloud cover prediction results and atmospheric turbulence prediction results are aligned at the same future time step. The consistency relationship between the trend of cloud cover prediction results and the trend of atmospheric turbulence prediction results is weighted to obtain a conditional physical consistency constraint term. The conditional physical consistency constraint term is then incorporated into the joint loss function to train the joint prediction model.

[0012] Step 5: Based on the joint prediction model trained in Step 4, perform joint predictions for multiple optical ground stations at multiple future time steps to generate cloud cover predictions and atmospheric turbulence predictions for each station at multiple future time steps.

[0013] Step 6: Based on the cloud cover prediction results and atmospheric turbulence prediction results generated in Step 5, perform a joint evaluation of the link status at multiple future time steps to generate future link availability evaluation results for each site.

[0014] Step 7: Based on the future link availability evaluation results obtained in Step 6, generate a candidate site set and site priority ranking results.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] 1. By uniformly modeling historical atmospheric observation data from 45 optical ground stations and extracting shared temporal features from multiple stations using a shared time-series encoder, the problem of insufficient parameter reuse capability and weak cross-site transfer capability under single-station independent modeling is overcome, thereby improving the modeling efficiency, sample utilization efficiency and cross-site generalization capability in multi-station scenarios.

[0017] 2. By constructing a cloud cover prediction branch and a conditionally coupled atmospheric turbulence prediction branch in the same joint prediction model, and using the cloud cover condition information output by the cloud cover prediction branch to perform conditional correction on the atmospheric turbulence prediction branch, the coupled joint characterization of link connectivity factors and link transmission quality factors is realized. This overcomes the problem in the existing technology of simply setting up two prediction tasks in parallel, which makes it difficult to reflect the comprehensive state of the link, thereby improving the completeness and engineering applicability of future link state prediction results.

[0018] 3. By introducing conditional physical consistency constraints modulated by auxiliary features, and using these constraints to jointly train the shared temporal encoder and dual prediction branches, the problem of inconsistent output of physical evolution laws in pure data-driven prediction is overcome, thereby improving the physical rationality, stability and credibility of the prediction results.

[0019] 4. By jointly outputting and uniformly evaluating the cloud cover prediction results and atmospheric turbulence prediction results at multiple prediction time steps of 1 hour, 3 hours, 6 hours and 12 hours in the future, a future link availability score is formed. This overcomes the problem that existing technologies only output raw prediction values ​​and cannot directly support link status analysis, thereby improving the adaptability of multi-time step prediction results to space-ground free space optical communication scenarios.

[0020] 5. By generating a candidate station set and station priority ranking results based on the future link availability evaluation results, the output results are extended from prediction results to decision-level results. This overcomes the problem that existing technologies only output raw numerical and probability information and cannot directly support multi-station collaborative access, thereby improving the practicality of the output format for multi-optical ground station collaborative access scenarios.

[0021] 6. By Figure 4 and Figure 5The corresponding comparative verification results show that the prediction error of the method of the present invention is lower than that of the single-station independent prediction method and the single cloud cover prediction method in multiple prediction time steps in the future. It also maintains good prediction performance in three climate zones: inland Europe, the Mediterranean and the Tibetan Plateau. This indicates that the method of the present invention has good prediction accuracy, cross-regional generalization ability and engineering application value. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0023] Figure 2 This is a schematic diagram of the overall structure of the joint prediction model in this invention.

[0024] Figure 3 This is a schematic diagram of the process for generating future link availability evaluation and candidate station determination results in this invention.

[0025] Figure 4 This is a comparison chart of the prediction errors of the method of the present invention and other prediction methods over multiple prediction time steps in the future.

[0026] Figure 5 This is a comparison chart of the prediction results for the three climate zones. Figure 5 (a) shows a comparison of prediction errors in different climate zones, and (b) shows the prediction trends at multiple time steps in different climate zones. Detailed Implementation

[0027] The present invention will now be described in detail with reference to a specific embodiment. This embodiment is used to illustrate the present invention.

[0028] This embodiment focuses on a ground station network consisting of 45 optical ground stations, constructing a joint prediction and candidate station recommendation method using multiple optical ground stations. These 45 optical ground stations are distributed across three climate zones: inland Europe, the Mediterranean, and the Tibetan Plateau, with 15 stations in each climate zone. The method includes the following steps.

[0029] Step 1: Acquire historical atmospheric observation data from multiple optical ground stations, perform time alignment, time feature construction, and sliding window truncation on the historical atmospheric observation data to construct the historical time series input sequence for each station.

[0030] like Figure 1 As shown, this embodiment first performs the process of acquiring historical atmospheric observation data and constructing historical time-series input sequences.

[0031] First, atmospheric observation data from multiple optical ground stations over a historical time period were acquired. This atmospheric observation data includes cloud cover data arranged chronologically, auxiliary features characterizing the near-surface atmospheric state, and time information corresponding to the observation data at each time step. To enhance the model's ability to represent periodic weather changes, the intraday times and dates within the year corresponding to each time step were periodically encoded based on the time information, forming time feature data.

[0032] In this embodiment, the historical atmospheric observation data used is ERA5 reanalysis data, with a time range from 2010 to 2023, a temporal resolution of 1 hour, a spatial resolution of 0.25°, and a total time step of 122712.

[0033] In this embodiment, the time feature data includes:

[0034] Periodic encoding of intraday timeframes;

[0035] Periodic coding of dates within the year.

[0036] The auxiliary features used to characterize the near-surface atmospheric state include auxiliary meteorological features.

[0037] In this embodiment, auxiliary meteorological features include surface air pressure, air temperature, surface temperature difference with a 2-meter radius, wind speed, and boundary layer height.

[0038] Subsequently, continuous observation data is extracted within a fixed time window to construct a historical time-series input sequence. In this embodiment, the data from the past 24 hours is used as an input sample window to form the historical time-series input sequence for each station.

[0039] In this embodiment, the training set corresponds to historical atmospheric observation data from 2010 to 2018, with a sample time step of 78,840; the validation set corresponds to historical atmospheric observation data from 2019 to 2020, with a sample time step of 17,520; and the test set corresponds to historical atmospheric observation data from 2021 to 2023, with a sample time step of 26,280.

[0040] In this embodiment, for the first The historical time-series input sequence expression for each optical ground station is as follows: in, Indicates the first Historical time-series input sequences of an optical ground station. Indicates the optical ground station number. Indicates the first The input feature vector of an optical ground station at time t Indicates the first An optical ground station at time The input feature vector, Indicates the current moment. Indicates the length of the historical time window. This represents the dimension of the input features at a single time step.

[0041] Step 2: As Figure 2 As shown, after constructing the historical time series input sequence for each site in step 1, the historical time series input sequence is input into the shared time series encoder to extract the time series feature representation corresponding to each site.

[0042] Based on the historical time-series input sequences of each site constructed in step 1, the historical time-series input sequences are input into the shared time-series encoder. The shared time-series encoder uses a unified parameter configuration for multiple sites to share time-series modeling capabilities among different sites and to perform unified modeling of the historical time-series input sequences of multiple sites.

[0043] In this embodiment, the shared temporal encoder employs a bidirectional long short-term memory network. This network utilizes both forward and reverse contextual information of the time series to encode the input sequence and outputs temporal feature representations corresponding to each station.

[0044] For the Each optical ground station, its historical time-series input sequence is denoted as... The temporal feature representation of the output of the shared temporal encoder is denoted as... The encoding relationship is as follows: in, Indicates the first Temporal characteristics of each optical ground station. This represents the shared timing encoder mapping function.

[0045] When the shared sequential encoder uses a bidirectional long short-term memory network, the above equation can be further written as: in, This represents the bidirectional long short-term memory network encoding function. This represents the dimension of the temporal feature representation. In this embodiment, the hidden state dimension in each direction of the bidirectional Long Short-Term Memory network is 128, and the temporal feature representation dimension after bidirectional concatenation is... .

[0046] By employing a shared temporal encoder, we avoid building a separate prediction model for each site, thereby reducing model complexity and improving generalization ability in multi-site scenarios.

[0047] Step 3: Construct the coupled dual-branch output structure of the joint prediction model.

[0048] like Figure 2 As shown, the joint prediction model connects the cloud cover prediction branch and the atmospheric turbulence prediction branch at the output of the shared time encoder, and uses the cloud cover condition information output by the cloud cover prediction branch to perform condition correction on the atmospheric turbulence prediction branch, thereby forming a coupled dual-branch output structure.

[0049] Based on the temporal feature representations of each site obtained in step 2, a cloud cover prediction branch and a conditionally coupled atmospheric turbulence prediction branch are constructed for multiple preset time steps in the future, forming a coupled dual-branch output structure for the joint prediction model. By introducing the cloud cover prediction branch after sharing the temporal feature representations, and utilizing the cloud cover condition information output by the cloud cover prediction branch to participate in the conditional correction of the atmospheric turbulence prediction branch, coupled joint modeling of link connectivity factors and link transmission quality factors is achieved. Specifically, the cloud cover prediction branch is used to form the cloud cover prediction mapping relationship for multiple future time steps for each site and output the cloud cover condition information, while the atmospheric turbulence prediction branch is used to form the atmospheric turbulence prediction mapping relationship for the corresponding future time steps with the participation of the cloud cover condition information.

[0050] In this embodiment, the multiple preset time steps in the future include: The next hour; The next 3 hours; The next 6 hours; The next 12 hours.

[0051] In this embodiment, the number of multiple preset time steps in the future .

[0052] The cloud cover prediction branch adopts a multilayer perceptron structure. Its output is activated and mapped to a predetermined cloud cover value range, generating cloud cover condition information for the atmospheric turbulence prediction branch to use. The atmospheric turbulence prediction branch also adopts a multilayer perceptron structure. Its input time step setting is consistent with that of the cloud cover prediction branch. Based on the shared temporal feature representation, the cloud cover condition information is further introduced, thereby ensuring that the two prediction branches correspond one-to-one at the same future time step and have a clear conditional coupling relationship.

[0053] In this embodiment, the cloud cover condition information expression is: in, Indicates the first Cloud cover information corresponding to each optical ground station. This represents the conditional feature extraction mapping function in the cloud cover prediction branch.

[0054] Correspondingly, the output expression of the cloud cover prediction branch is: in, Indicates the first Cloud cover predictions for multiple time steps from an optical ground station. This represents the output mapping parameters of the cloud cover prediction branch. This represents an activation function that maps the output to a predetermined range of cloud cover. , This indicates the number of future prediction time steps.

[0055] Correspondingly, the output expression for the atmospheric turbulence prediction branch is: in, Indicates the first Atmospheric turbulence predictions for multiple time steps from an optical ground station This represents the multilayer perceptron mapping function in the atmospheric turbulence prediction branch. This represents the conditional input after concatenating time-series feature representations with cloud cover information. In this embodiment, .

[0056] The key technical point of this step is that instead of simply setting up the cloud cover prediction branch and the atmospheric turbulence prediction branch side by side, the cloud cover condition information output by the cloud cover prediction branch is used to modify the atmospheric turbulence prediction branch, so that the two prediction branches form a clear coupling relationship based on the shared time series feature representation obtained in step 2, thereby providing a unified two-parameter prediction basis for subsequent conditional physical consistency constraint training and link state joint evaluation.

[0057] Step 4: Introduce conditional physical consistency constraints and perform joint training.

[0058] like Figure 2 As shown, after constructing the cloud cover prediction branch and the conditionally coupled atmospheric turbulence prediction branch in step 3, a conditional physical consistency constraint is further introduced and a joint loss function is constructed to train the joint prediction model.

[0059] A conditional physical consistency constraint is introduced, based on the changing trends of cloud cover prediction results output by the cloud cover prediction branch and atmospheric turbulence prediction results output by the atmospheric turbulence prediction branch in step 3, to improve the physical inconsistency problem in pure data-driven prediction. Specifically, the cloud cover prediction results and atmospheric turbulence prediction results are aligned at the same future time step; simultaneously, based on the auxiliary features in the historical atmospheric observation data obtained in step 1, the auxiliary features within the current historical time window are aggregated to obtain an auxiliary feature aggregation vector, and then constraint weights are generated based on the auxiliary feature aggregation vector. A two-parameter changing trend constraint term is constructed under the modulation of the constraint weights, and the constraint term is used to jointly train the joint prediction model composed of the shared temporal encoder, the cloud cover prediction branch, and the atmospheric turbulence prediction branch.

[0060] The key technical point of this step is that instead of training cloud cover prediction and atmospheric turbulence prediction as two independent output tasks, a physical correlation between their changing trends is established at the same future time step. The physical correlation is then conditionally modulated using auxiliary features, and the conditional physical correlation is introduced into the joint training process of the shared temporal encoder and the dual prediction branches, thereby improving the reliability and consistency of the joint prediction results.

[0061] In this embodiment, the total loss function for joint training is written as: in, This represents the total loss function of the joint prediction model. This represents the cloud cover prediction loss term. This represents the turbulence prediction loss term. This represents a conditional physical consistency constraint term. and The term represents the trade-off coefficient; the cloud cover prediction loss term and the turbulence prediction loss term are used to optimize the cloud cover prediction results and the turbulence prediction results, respectively, and the conditional physical consistency constraint term is used to constrain the consistency relationship between the two trends.

[0062] The expression for the cloud cover prediction loss term is as follows: in, Indicates the number of optical ground stations. Indicates the future time step number. Indicates the first The first optical ground station Cloud cover forecast for the next future time step. This represents the corresponding actual cloud cover value. This represents the mean squared error function. In this embodiment, .

[0063] The expression for the turbulence prediction loss term is: in, Indicates the first The first optical ground station Atmospheric turbulence predictions for a future time step. This represents the actual value of the corresponding atmospheric turbulence.

[0064] Furthermore, based on the auxiliary features in the historical atmospheric observation data obtained in step 1, the auxiliary features within the current historical time window are aggregated to obtain an auxiliary feature aggregation vector, the expression of which is: in, Indicates the first Auxiliary feature sequences of an optical ground station within the current historical time window. This represents the auxiliary feature aggregation function. Indicates the first The auxiliary feature aggregation vector corresponding to each optical ground station.

[0065] Furthermore, constraint weights are constructed based on the auxiliary feature aggregation vector, and their expression is as follows: in, Indicates the first The physical consistency constraint weights corresponding to each optical ground station. This represents a mapping function that maps auxiliary feature aggregation vectors to constraint weights.

[0066] Furthermore, the expression for the conditional physical consistency constraint term is: in, This represents the time difference operator. This represents the correlation calculation function. Represents the linear rectified function. Indicates the first Differential sequence of cloud cover predictions for multiple time steps from an optical ground station Indicates the first Differential sequences of atmospheric turbulence predictions for multiple time steps from an optical ground station.

[0067] In this embodiment, the tradeoff coefficient in the loss function is taken as . , .

[0068] In this embodiment, the joint prediction model is trained using the AdamW optimizer with a learning rate of [missing information]. The weight decay coefficient is The batch size is 128, and a cosine annealing learning rate scheduling strategy and an early stopping strategy are adopted, with an early stopping patience value of 10 training rounds.

[0069] This joint training method simultaneously improves the practicality, stability, and physical rationality of cloud cover prediction and atmospheric turbulence prediction.

[0070] Step 5: Generate two-parameter prediction results using the joint prediction model.

[0071] like Figure 2 As shown, the joint prediction model trained in step 4 maintains the shared temporal encoder and coupled bi-branch prediction structure, and outputs cloud cover prediction results and atmospheric turbulence prediction results in the application stage.

[0072] Based on the joint prediction model trained in step 4, joint predictions are made for multiple optical ground stations over several future time steps. This generates cloud cover prediction results and atmospheric turbulence prediction results for each station over several future time steps, output by the cloud cover prediction branch in step 3 and the atmospheric turbulence prediction branch corrected by cloud cover condition information, respectively. The cloud cover prediction results and atmospheric turbulence prediction results are arranged in the same future time step order, providing input for the joint evaluation of link status in step 6.

[0073] In this embodiment, the joint prediction model simultaneously outputs cloud cover predictions and atmospheric turbulence predictions for each optical ground station for the next 1, 3, 6, and 12 hours. The atmospheric turbulence predictions are obtained using... Output in format.

[0074] The key technical point of this step is to use the joint prediction model trained with conditional physical consistency constraints to simultaneously output the cloud cover prediction results and atmospheric turbulence prediction results formed by the coupled dual-branch structure in step 3, thereby improving the consistency and availability of the dual-parameter prediction results in multi-station scenarios.

[0075] Step 6: Perform a joint evaluation of the two-parameter prediction results and generate future link availability evaluation results.

[0076] like Figure 3 As shown, after obtaining the cloud cover prediction results and atmospheric turbulence prediction results output in step 5, the two-parameter prediction results are further jointly evaluated to generate future link availability evaluation results.

[0077] Based on the cloud cover prediction and atmospheric turbulence prediction results generated in step 5, the link status at multiple future time steps is jointly evaluated to generate future link availability evaluation results for each site. Specifically, the cloud cover prediction and atmospheric turbulence prediction results at the same future time step are mapped to link status evaluation results, and the link status evaluation results at multiple future time steps are integrated to generate future link availability evaluation results for each site; the future link availability evaluation results are site availability scores.

[0078] In this embodiment, for the first The future link availability score for an optical ground station is expressed as follows: in, Indicates the first Future link availability score for each optical ground station This represents an evaluation function that converts cloud cover forecasts and atmospheric turbulence forecasts at the same future time step into a link availability score.

[0079] Furthermore, based on the future obtained in step 5... The cloud cover prediction results at each time step are used to construct the prediction link quality for that time step: in, Indicates the first The first optical ground station Predicted link quality for one future time step. This indicates the cloud cover threshold.

[0080] In this embodiment, cloud cover threshold .

[0081] Then, the predicted link quality is fused with the turbulence penalty generated corresponding to the atmospheric turbulence prediction results at the same time step in step 5, and a future link availability score is generated based on the fusion results at multiple future time steps: in, Indicates the first The weights corresponding to each future time step This represents the turbulence penalty coefficient. This represents a mapping function that converts atmospheric turbulence prediction results into a penalty amount. Indicates the first The first optical ground station Atmospheric turbulence predictions for the next future time step.

[0082] In this embodiment, for the four prediction time steps of 1 hour, 3 hours, 6 hours, and 12 hours, the weights are... Set according to the predicted duration, increasing monotonically and decreasing, and satisfying... The weights for the four prediction time steps are set to 0.4, 0.3, 0.2, and 0.1, respectively. Turbulence penalty coefficient. In this embodiment, a fixed value is adopted after optimization using the validation set. .

[0083] The key technical point of this step is that it does not stop at the level of the original prediction results, but further transforms the joint prediction results into link availability evaluation results.

[0084] Step 7: Generate a set of candidate sites and the site priority ranking results.

[0085] like Figure 3 As shown, after obtaining the future link availability evaluation results generated in step 6, multiple optical ground stations are sorted to generate a candidate station set and a site priority ranking result.

[0086] To address the issue of existing technologies only outputting raw predicted values, a result generation method tailored to multi-site collaborative access requirements is introduced. Based on the future link availability evaluation results obtained in step 6, a candidate station set and a station priority ranking result corresponding to the multi-optical ground station collaborative access requirements are generated. Specifically, multiple optical ground stations are ranked according to the future link availability evaluation results of each station, and the top-ranked stations are selected to form a candidate station set. A complete station priority ranking result is then output, where the candidate station set is ranked according to a preset order. It consists of a famous optical ground station.

[0087] The key technical point of this step is that it does not stop at the level of the original prediction results and a single score value, but further forms a candidate station set and site priority ranking results that are directly oriented towards multi-site collaborative scenarios, so that the joint prediction results have an application expression form for access decision-making.

[0088] In this embodiment, the candidate station set expression is: in, Represents the set of candidate stations. This indicates the threshold for the number of candidate stations.

[0089] In this embodiment, the threshold for the number of candidate stations .

[0090] In the comparative verification, the method of the present invention was compared with single-site independent prediction methods and single cloud cover prediction methods, and the prediction results of the method of the present invention were verified in three climate zones: inland Europe, the Mediterranean, and the Tibetan Plateau. Figure 4 As shown, at the four prediction time steps of 1 hour, 3 hours, 6 hours, and 12 hours, the prediction errors of different methods all tend to increase with the increase of prediction duration. However, the error of the method of this invention at each prediction time step is always lower than that of the single-station independent prediction method and the single cloud cover prediction method. This indicates that by uniformly modeling 45 optical ground stations, extracting shared temporal features of multiple stations using a shared temporal encoder, and adopting a coupled dual-branch structure in which the cloud cover prediction branch participates in the conditional correction of the atmospheric turbulence prediction branch, as well as joint training with conditional physical consistency constraints modulated by auxiliary features, it is possible to maintain high accuracy in both short-term and long-term prediction, thereby improving the stability and reliability of multi-time-step prediction tasks.

[0091] like Figure 5 As shown, the method of this invention maintains good prediction performance in three climate zones: inland Europe, the Mediterranean, and the Tibetan Plateau. There are certain differences in error among the different climate zones. The Mediterranean region has the lowest prediction error, indicating that in areas with relatively stable weather evolution, the method of this invention can fully leverage the advantages of multi-station shared modeling, cloud cover information participating in turbulence prediction, and joint training with conditional physical consistency constraints. The prediction error in inland Europe is relatively high, indicating that cloud cover changes in this region are more complex, but the method of this invention can still maintain stable prediction results, demonstrating its good adaptability to complex weather changes. The prediction error in the Tibetan Plateau region is between the two, indicating that the method of this invention is also applicable to high-altitude areas with strong convective characteristics. The above comparative results show that the method of this invention is not only superior to independent prediction methods that rely solely on single-station historical data, but also superior to prediction methods that only predict cloud cover as a single factor, and has good generalization ability and engineering application value under different climate zone conditions.

[0092] Through the above steps, this embodiment realizes joint prediction and candidate station recommendation for multi-optical ground station scenarios, forming a candidate station set and site priority ranking results that are forward-looking and physically reasonable.

Claims

1. A method for joint prediction of link status and candidate station determination of multiple optical ground stations, characterized in that, The method includes the following steps: Step 1: Acquire historical atmospheric observation data from multiple optical ground stations, perform time alignment, time feature construction, and sliding window truncation on the historical atmospheric observation data, and construct the historical time series input sequence for each station; Step 2: Based on the historical time-series input sequences of each site constructed in Step 1, each site is uniformly encoded by a shared time-series encoder to obtain the time-series feature representations corresponding to each site; Step 3: Based on the temporal feature representations of each station obtained in Step 2, construct a cloud cover prediction branch and a conditionally coupled atmospheric turbulence prediction branch for multiple preset time steps in the future. The cloud cover condition information output by the cloud cover prediction branch is used to modify the atmospheric turbulence prediction branch to form a coupled dual-branch output structure of the joint prediction model. Step 4: Based on the cloud cover condition information output in Step 3 and the physical consistency constraint weights generated from the auxiliary features in the historical atmospheric observation data obtained in Step 1, the cloud cover prediction results and atmospheric turbulence prediction results are aligned at the same future time step. The consistency relationship between the trend of cloud cover prediction results and the trend of atmospheric turbulence prediction results is weighted to obtain a conditional physical consistency constraint term. The conditional physical consistency constraint term is then incorporated into the total loss function of the joint prediction model to train the joint prediction model. Step 5: Based on the joint prediction model trained in Step 4, perform joint predictions for multiple optical ground stations at multiple future time steps to generate cloud cover prediction results and atmospheric turbulence prediction results for each station at multiple future time steps. Step 6: Based on the cloud cover prediction results and atmospheric turbulence prediction results generated in Step 5, jointly evaluate the link status at multiple future time steps and generate future link availability evaluation results for each site. Step 7: Based on the future link availability evaluation results obtained in Step 6, generate a candidate site set and site priority ranking results.

2. The method for joint prediction of link status and candidate station determination of multiple optical ground stations according to claim 1, characterized in that, In step 1, the historical atmospheric observation data includes cloud cover data arranged in chronological order, auxiliary features characterizing the near-surface atmospheric state, and time information corresponding to the observation data at each time step. Based on this time information, time feature data is constructed. Specifically, the intraday time and intrayear date corresponding to each time step are periodically encoded to form time feature data. For each optical ground station, continuous observation data is extracted according to a preset time window length to construct a historical time-series input sequence. For the first... The historical time-series input sequence expression for each optical ground station is as follows: in, Indicates the first Historical time-series input sequences of an optical ground station. Indicates the optical ground station number. Indicates the first The input feature vector of an optical ground station at time t Indicates the first An optical ground station at time The input feature vector, Indicates the current moment. Indicates the length of the historical time window. This represents the dimension of the input features at a single time step.

3. The method for joint prediction of link status and candidate station determination of multiple optical ground stations according to claim 1 or 2, characterized in that, In step 2, the historical time series input sequences of each site constructed in step 1 are input into the shared time series encoder; the shared time series encoder uses a bidirectional long short-term memory network to uniformly model the historical time series input sequences of multiple sites, thereby outputting the time series feature representations corresponding to each site; For the Each optical ground station, its historical time-series input sequence is denoted as... The temporal feature representation output by the shared temporal encoder using a bidirectional long short-term memory network is denoted as... Then we have: in, Indicates the first Temporal characteristics of each optical ground station This represents the shared timing encoder mapping function. This represents the bidirectional long short-term memory network encoding function. The temporal features are represented by the dimension.

4. The method for joint prediction of link status and candidate station determination of multiple optical ground stations according to claim 3, characterized in that, In step 3, based on the temporal feature representations of each station obtained in step 2, a cloud cover prediction branch and a conditionally coupled atmospheric turbulence prediction branch are constructed respectively. The cloud cover prediction branch is used to form a cloud cover prediction mapping relationship for each station in multiple future time steps and output cloud cover condition information. The atmospheric turbulence prediction branch is used to form an atmospheric turbulence prediction mapping relationship for future time steps corresponding to the cloud cover prediction mapping relationship with the cloud cover condition information, thereby forming a coupled dual-branch output structure of the joint prediction model. For the For each optical ground station, the cloud cover condition information expression output by the cloud cover prediction branch is: in, Indicates the first Cloud cover information corresponding to each optical ground station. This represents the conditional feature extraction mapping function in the cloud cover prediction branch; The output expression for the cloud cover prediction branch is: in, Indicates the first Cloud cover predictions for multiple time steps from an optical ground station. This represents the output mapping parameters of the cloud cover prediction branch. This represents the activation function that maps the output to a predetermined cloud cover range; The cloud cover prediction results satisfy: in, Indicates the number of prediction time steps; The output expression for the atmospheric turbulence prediction branch is: in, Indicates the first Atmospheric turbulence predictions for multiple time steps from an optical ground station This represents the multilayer perceptron mapping function in the atmospheric turbulence prediction branch. This represents the conditional input after concatenating the temporal feature representation with cloud cover information.

5. The method for joint prediction of link status and candidate station determination of multiple optical ground stations according to claim 4, characterized in that, The specific method for step 4 is as follows: Step 4-1. For the first For each optical ground station, based on the auxiliary features in the historical atmospheric observation data obtained in step 1, the auxiliary features within the current historical time window are aggregated to obtain an auxiliary feature aggregation vector, the expression of which is: in, Indicates the first Auxiliary feature sequences of an optical ground station within the current historical time window. This represents the auxiliary feature aggregation function. Indicates the first Auxiliary feature aggregation vectors corresponding to each optical ground station; Step 4-2. Construct constraint weights based on the auxiliary feature aggregation vector obtained in Step 4-1, the expression of which is: in, Indicates the first The physical consistency constraint weights corresponding to each optical ground station. This represents a mapping function that maps auxiliary feature aggregation vectors to constraint weights; Step 4-3. Align the cloud cover prediction results with the atmospheric turbulence prediction results at the same future time step, and apply a weighted constraint to the consistency relationship between the changing trends of the cloud cover prediction results and the changing trends of the atmospheric turbulence prediction results to obtain a conditional physical consistency constraint term. , represented as: in, Indicates the number of optical ground stations. This represents the time difference operator. This represents the correlation calculation function. Represents the linear rectified function. Indicates the first Differential sequence of cloud cover predictions for multiple time steps from an optical ground station Indicates the first The atmospheric turbulence prediction differential sequence of an optical ground station for multiple future time steps; the constraint term is used to limit the deviation of cloud cover change trend from turbulence change trend from a preset physical mechanism under auxiliary feature modulation; Step 4-4. Incorporate the conditional physical consistency constraint term into the total loss function of the joint prediction model. The expression for the total loss function of the joint prediction model is as follows: in, This represents the total loss function of the joint prediction model. This represents the cloud cover prediction loss term. This represents the turbulence prediction loss term. This represents a conditional physical consistency constraint term. and Indicates the trade-off coefficient. , The expression for the cloud cover prediction loss term is: The expression for the turbulence prediction loss term is: in, Indicates the future time step number. Indicates the first The first optical ground station Cloud cover forecast for a future time step. This represents the corresponding actual cloud cover value. Represents the mean square error function; Indicates the first The first optical ground station Atmospheric turbulence predictions for one future time step. This represents the actual value of the corresponding atmospheric turbulence.

6. The method for joint prediction of link status and candidate station determination of multiple optical ground stations according to claim 5, characterized in that, In step 5, the joint prediction model trained in step 4 is used to jointly predict multiple future time steps of multiple optical ground stations, generating cloud cover prediction results and atmospheric turbulence prediction results for each station at multiple future time steps, respectively output by the cloud cover prediction branch and atmospheric turbulence prediction branch in step 3; the cloud cover prediction results and atmospheric turbulence prediction results are arranged in the same future time step order.

7. The method for joint prediction of link status and candidate station determination of multiple optical ground stations according to claim 6, characterized in that, In step 6, based on the cloud cover prediction results and atmospheric turbulence prediction results generated in step 5, the cloud cover prediction results and atmospheric turbulence prediction results at the same future time step are jointly evaluated, and the evaluation results at multiple future time steps are integrated to generate the future link availability score for each site. For the The first optical ground station, based on cloud cover prediction results, was constructed in the [missing information]. Predicted link quality for each future time step: in, Indicates the first The first optical ground station Predicted link quality for one future time step. Indicates the cloud cover threshold; Future Link Availability Score The expression is: in, Indicates the first Future link availability score for each optical ground station Indicates the first The weights corresponding to each future time step This represents the turbulence penalty coefficient. This represents a mapping function that converts atmospheric turbulence prediction results into a penalty amount. Indicates the first The first optical ground station Atmospheric turbulence predictions for the next future time step.

8. The method for joint prediction of link status and candidate station determination of multiple optical ground stations according to claim 7, characterized in that, The candidate station set mentioned in step 7 The expression is: in, Represents the set of candidate stations. This indicates the threshold for the number of candidate stations.

Images