Selective collection and flow prediction method for low earth orbit satellite flow data

Abstract

The invention provides a selective collection and flow prediction method for flow data of a low earth orbit satellite. Relates to one or more controllers and a plurality of satellite nodes. In a giant low-orbit satellite network, ground users communicate through a satellite relay. The method comprises the following steps: a central controller satellite generates and issues a collection scheme to a data plane, the data plane uploads traffic data based on the scheme, and the controller satellites perform complementation and prediction in sequence based on the incomplete traffic. According to the method, the data collection amount is actively reduced, so that the processing and calculation pressure of a controller satellite is relieved. Meanwhile, the invention further provides a traffic data complementation scheme and a traffic prediction scheme. Compared with a traditional method, the strategy greatly improves the communication efficiency and saves satellite resources while guaranteeing the prediction accuracy.

Description

Technical Field

[0001] This invention relates to the fields of low-orbit satellite networks, software-defined networks, and deep neural networks. Specifically, it provides a method for selectively collecting and predicting low-orbit satellite traffic data. This method addresses the communication management needs of large-scale low-orbit satellite constellations. By combining the centralized control capabilities of the SDN framework with the spatiotemporal modeling advantages of deep neural networks, it enables efficient traffic data collection, completion, and prediction. Background Art

[0002] Low-orbit satellite networks have demonstrated the potential to provide global communications services, offering significant advantages in low latency and reduced launch costs. Due to their lower orbital altitude, low-orbit satellites are able to provide shorter communication latencies and are relatively inexpensive to launch compared to traditional geostationary satellites. However, achieving global coverage requires the deployment of large-scale satellite constellations and precise coordination with densely deployed ground stations, which undoubtedly introduces significant complexity to the management and optimization of ultra-large-scale networks. In particular, when these networks operate on a global scale, the communication and coordination requirements between satellites and ground stations will become even more complex. Without an effective management strategy, low-orbit satellite networks are susceptible to problems such as link outages and uneven communication demand, which can not only lead to congestion on inter-satellite links (ISLs), but also increase the risk of end-to-end latency and packet loss, affecting communication quality and overall system performance. These challenges highlight the urgent need to improve communication efficiency and optimize management strategies in such large-scale network environments.

[0003] Current satellite management strategies are approached from multiple perspectives. One key approach is to apply the software-defined networking (SDN) framework to low-orbit satellite constellations. This framework enables controllers to collect real-time satellite network status information and, based on this data, issue dynamic routing policies to precisely schedule traffic. Furthermore, some research has leveraged traffic prediction techniques to proactively avoid paths prone to congestion or link disruptions, ensuring smooth data transmission. Introducing a traffic prediction module within the SDN framework is an effective approach for managing low-orbit satellite communications, as it can proactively identify potential network bottlenecks and enable dynamic adjustments. However, while these solutions have theoretical validity, their direct application to low-orbit satellite networks faces numerous challenges. Specifically, as satellite constellations continue to expand, inter-satellite link bandwidth and computing resources become increasingly limited, posing a significant challenge to SDN controllers. In large-scale satellite network environments, efficiently processing and managing massive amounts of satellite data and optimizing resource allocation remain key challenges. Summary of the Invention

[0004] To overcome the deficiencies of the above-mentioned existing technologies, the present invention proposes a method for selectively collecting and predicting low-earth orbit satellite traffic data, which involves one or more controllers and multiple satellite nodes. The central controller satellite is responsible for generating and distributing a collection plan to the data plane, and the data plane selectively uploads traffic data according to this plan. The controller satellite sequentially completes and predicts based on the uploaded incomplete traffic data to supplement the missing or unavailable traffic information. This process can not only significantly reduce the amount of data collection, but also effectively relieve the computing and processing pressure on the controller satellite, and avoid performance bottlenecks caused by excessive real-time data processing. The technical solution of the present invention is as follows:

[0005] (1) Traffic collection module:

[0006] Step 1, initialize the collection plan: If there are N satellites in the network, the controller initializes the collection plan for the next T time intervals This matrix consists of 0s and 1s and will guide each satellite to upload traffic data at the specified time

[0007] Step 2, distribute the plan to the satellites: The controller distributes the collection plan to each satellite in the data plane, and this plan will guide the traffic data collection of the satellites in the next T time intervals

[0008] Step 3, collect traffic data: Denote the satellite traffic data selectively collected in Step 1 as a matrix

[0009] (2) Traffic matrix completion module:

[0010] Step 4, feature extraction: Input part of the observed data X and the adjacency matrix A containing the relationships between satellites into the feature extraction module, use a neural network to extract key features and generate a low-dimensional embedding representation to capture the internal structure information of the data, denoted as the feature extractor f1. Denote the hidden features containing time and space features as h, then the relationship is as follows:

[0011] h = f1(X, A)

[0012] Step 5, traffic matrix reconstruction: Use the hidden features h learned from the low-dimensional representation obtained in Step 4 to learn the completion layer f2, and use a neural network to complete the missing traffic data. If the completed traffic data is denoted as the completion matrix Then the relationship is as follows:

[0013] Ι = f2(h, A)

[0014] Step 6, evaluate the completion result: Compare the reconstructed matrix Ι with the real traffic matrix Y, and use various evaluation indicators to evaluate the completion effect. For example, the goal of this step is to minimize the difference between the completed data and the original data, that is:

[0015] min||Y - Ι||

[0016] (3) Flow prediction module:

[0017] Step 7, Temporal and spatial feature extraction: Based on the temporal and spatial feature extraction methods, from the completed traffic data Ι and the network structure information obtained in Step 5, use a neural network to extract the key features in the data, capture the spatio-temporal correlation and change trend of the traffic, denoted as the feature extractor f3.

[0018] Step 8, Prediction model construction and training: Construct a traffic prediction model based on the features mentioned in Step 7, and use the completed traffic matrix Ι and the adjacency matrix A obtained in Step 5 to predict the traffic data for the next T f steps, denoted as the learnable function. The relationship is as follows:

[0019] ψ = g(f3(Ι), A)

[0020] Step 9, Prediction result evaluation: Compare the predicted traffic with the real traffic X P and the goal of this step is to minimize the difference between the completed data and the original data, that is:

[0021] min||X P - ||

[0022] (4) Collection plan update module:

[0023] Step 10, Evaluate the collection plan: Based on the completion evaluation result in Step 6 and the prediction evaluation result in Step 9, comprehensively measure the effect of the current collection plan. The specific steps are as follows:

[0024] (10 - 1) Evaluate the accuracy of matrix completion and traffic prediction. For example, use the mean square error as an index, and its formula is as follows:

[0025]

[0026] where Y is the original traffic matrix data, X P is the true value of the predicted traffic, N is the number of completed entries, and M is the number of predicted entries.

[0027] (10 - 2) Establish a comprehensive evaluation index. In order to comprehensively measure the overall performance of the collection plan, combine the matrix completion accuracy and the prediction accuracy to construct an error function L:

[0028] L = αE C + (1 - α)E P

[0029] Among them, α ∈ (0, 1) is a weight coefficient for balancing the accuracy of complementation and the accuracy of prediction.

[0030] Step 11: Update the collection scheme. Based on the error L obtained in Step 10, make a targeted adjustment to the collection scheme according to the error analysis result to preferentially collect data in areas with higher reconstruction errors, and obtain a new generation scheme for the next time period. The joint optimization method based on gradient descent is adopted, and the specific steps are as follows:

[0031] (11 - 1) Perform backpropagation and gradient descent update on the parameters of the traffic collection module based on the joint loss function L obtained by the traffic matrix complementation module and the traffic prediction module until the loss converges.

[0032] (11 - 2) Threshold the obtained continuous collection metrics to generate the final discretized collection scheme M for subsequent data collection tasks.

[0033] Step 12: Online update. Repeat Steps 2 to 11 until the specified number of learning rounds is completed.

[0034] The controller contains multiple neural network models, which are respectively used to generate and adjust the collection scheme, complement missing data, and predict future traffic. The model parameters are obtained through training to achieve reasonable design of the collection scheme, accurate estimation of missing data, and accurate prediction of future traffic. The high-performance neural network models mentioned above include but are not limited to GCN, GAT, and GRU models.

[0035] Compared with the prior art, the technical effects of the present invention are as follows:

[0036] 1) The collection scheme generated by the controller through the generator accurately guides the efficient collection of traffic data from the specified ISLs. By selectively collecting data only from the labeled ISLs, this scheme not only significantly reduces the data collection burden on the SDN controller but also reduces unnecessary consumption of computing resources, further optimizing the overall performance of the system. In addition, this method also shows obvious advantages in improving the real-time performance and response speed of the system, providing an innovative solution idea for efficient traffic management in large-scale dynamic networks.

[0037] 2) Further enhance the performance of the collection scheme generator through matrix completion algorithms. This algorithm fully exploits the inherent temporal and spatial correlations in the ISL traffic data and combines the powerful learning ability of high-performance neural networks (taking the GCN-based autoencoder as an example) to accurately infer and reconstruct missing values from the existing data. This method not only significantly improves the integrity and reliability of the data but also provides more comprehensive and high-quality data support for downstream tasks. With this scheme, the decision-making ability and overall performance of the system are effectively enhanced, and its adaptability in complex dynamic network environments is further improved, providing a solid technical guarantee for efficient traffic management.

[0038] 3) By leveraging the spatial correlations of traffic data in satellite networks and empowered by high-performance neural networks (taking the GAT- and GRU-based residual network as an example), in-depth modeling of temporal dependencies can accurately capture the patterns and potential trends of traffic evolution over time. By comprehensively considering spatial and temporal characteristics, this method significantly improves the accuracy and robustness of traffic prediction, adapts to complex dynamic communication environments, and ensures that the prediction results can support more optimal network resource allocation and management strategies, providing strong guarantees for the efficient operation of satellite networks. Brief Description of the Drawings

[0039] Figure 1 is a schematic diagram of low-earth orbit satellite communication based on the software-defined network framework.

[0040] Figure 2 is the overall flowchart of the traffic prediction method based on selective collection of low-earth orbit satellite traffic data.

[0041] Figure 3 is a schematic diagram of the performance of the traffic collection module proposed in the present invention at different collection rates

[0042] Figure 4 is a schematic diagram of the performance of the traffic matrix completion module proposed in the present invention at different collection rates.

[0043] Figure 5 is a schematic diagram of the performance of the traffic prediction module proposed in the present invention at different collection rates. Detailed Embodiments

[0044] The following further explains the technical solutions of the present invention in conjunction with the drawings and embodiments, but the protection scope of the present invention should not be limited thereby.

[0045] As Figure 1As shown in the figure, in the low-earth orbit (LEO) satellite communication architecture based on software-defined networking (SDN), the SDN framework enables the controller to manage the entire LEO satellite network in a centralized manner through the decoupled design of the control plane and the data plane. Under this architecture, each satellite continuously collects network status data such as inter-satellite link (ISL) traffic and uploads it to the controller in real time, so that the controller can comprehensively analyze the network status and make decisions. However, in a large-scale LEO satellite communication network, the massive data transmission and processing between satellites impose a heavy burden on the controller, which may lead to problems such as insufficient communication bandwidth and scarce computing resources.

[0046] To ensure the normal operation of the software-defined network function, this embodiment adopts a method for selective collection and traffic prediction of LEO satellite traffic data. By deploying a central controller satellite in the SDN control plane, the central controller satellite is responsible for generating and distributing the collection scheme to the data plane, and the data plane selectively uploads traffic data according to this collection scheme. The controller satellite sequentially completes and predicts based on the uploaded incomplete traffic data to supplement the missing or unavailable traffic information. This embodiment can reduce the data acquisition and computing burden of the SDN controller; combine matrix completion algorithms to mine the spatio-temporal correlation of traffic data, improving the integrity and reliability of the data; model the time and space dependencies through a deep learning model to achieve high-precision traffic prediction. Finally, it provides innovative solutions and technical guarantees for efficient traffic management and resource optimization in complex LEO satellite network environments.

[0047] As Figure 2 shown, the method for selective collection and traffic prediction of LEO satellite traffic data can be divided into four major modules, including a traffic collection module, a traffic matrix completion module, a traffic prediction module, and a collection scheme update module. Each module operates in coordination to form a closed-loop optimization process. The specific implementation steps are as follows:

[0048] (1) Traffic collection module: Based on the collection scheme designed by the controller, guide the satellite to upload traffic data as needed

[0049] Step 1, initialize the collection scheme: The controller generates an initial collection scheme according to the number of satellites N in the network and the prediction time window T This matrix consists of 0s and 1s and will guide each satellite to upload traffic data at a specified moment.

[0050] Step 2, distribute the collection scheme to the satellites: The controller distributes the collection scheme M to each satellite in the data plane. This scheme will guide the traffic data collection of the satellite in the future T time interval. After receiving the mask, the satellite node parses the local collection task. For example, if M i,t = 1, then satellite i needs to upload ISL traffic data at time t.

[0051] Step 3, Collect traffic data: The satellite uploads satellite traffic data as needed according to the instructions of collection scheme M, and the data in non-acquisition time slices is marked as NaN (missing value). The controller integrates all the traffic data uploaded by the satellites to construct a partially observed traffic matrix.

[0052] (II) Traffic matrix completion module:

[0053] Step 4, Spatiotemporal feature extraction: Input the partially observed data X and the adjacency matrix A containing the relationships between satellites into the feature extraction module. Use the graph convolutional network (GCN) to extract key features and generate a low-dimensional embedding representation to capture the inherent structural information of the data, denoted as the feature extractor f1. And denote the hidden features containing time and space features as h, then the relationship is as follows:

[0054] h = f1(X, A)

[0055] Step 5, Traffic matrix reconstruction: Use the hidden features h learned from the low-dimensional representation obtained in Step 4 to learn the completion layer f2, and use an autoencoder based on the graph attention mechanism to complete the missing traffic data. If the completed traffic data is denoted as the completion matrix Then the relationship is as follows:

[0056] Ι = f2(h, A)

[0057] Step 6, Completion result evaluation: Compare the reconstructed matrix Ι with the real traffic matrix Y, and use multiple evaluation indicators to evaluate the completion effect. For example, the goal of this step is to minimize the difference between the completed data and the original data, that is:

[0058] min||Y - Ι||

[0059] (III) Traffic prediction module:

[0060] Step 7, Temporal and spatial feature extraction: Based on the temporal and spatial feature extraction methods, from the completed traffic data Ι obtained in Step 5 and the structural information of the network, use the graph attention mechanism and the gated recurrent unit to extract the key features in the data, capture the spatiotemporal correlation and change trend of the traffic, denoted as the feature extractor f2.

[0061] Step 8, Prediction model construction and training: Construct a traffic prediction model based on the features mentioned in Step 7, and use the completed traffic matrix Ι and the adjacency matrix A obtained in Step 5 to predict the traffic data for the next T f steps, denoted as the learnable function. The relationship is as follows:

[0062] = (f2(Ι),)

[0063] Step 9, Prediction result evaluation: Compare the predicted traffic with the actual traffic X P In this step, the goal is to minimize the difference between the completed data and the original data, that is:

[0064] min||X P -||

[0065] (4) Collection scheme update module:

[0066] Step 10, Evaluate the collection scheme: Based on the completion evaluation result in Step 6 and the prediction evaluation result in Step 9, comprehensively measure the effectiveness of the current collection scheme. The specific steps are as follows:

[0067] (10-1) Evaluate the accuracy of matrix completion and traffic prediction. For example, use the mean square error as an indicator, and its formula is as follows:

[0068]

[0069] where Y is the original traffic matrix data, X P is the true value of the predicted traffic, N is the number of completed entries, and M is the number of predicted entries.

[0070] (10-2) Establish a comprehensive evaluation index. To comprehensively measure the overall performance of the collection scheme, combine the matrix completion accuracy and prediction accuracy to construct an error function L:

[0071] L = αE C +(1-α)E P

[0072] where α∈(0,1) is the weight coefficient that balances the completion accuracy and prediction accuracy.

[0073] Step 11, Update the collection scheme: Based on the error L obtained in Step 10, make a targeted adjustment to the collection scheme according to the error analysis result to preferentially collect data in the area with a higher reconstruction error, and obtain a new generation scheme for the next time period. Adopt a joint optimization method based on gradient descent. The specific steps are as follows:

[0074] (11-1) Perform backpropagation and gradient descent update on the parameters of the traffic collection module based on the joint loss function L obtained by the traffic matrix completion module and the traffic prediction module until the loss converges.

[0075] (11-2) Threshold the continuous collection indicators obtained by training to generate the final discretized collection scheme M for subsequent data collection tasks.

[0076] Step 12, Online update: Repeat Steps 2 to 11 until the specified number of learning rounds has been completed.

[0077] Figure 3 It shows the performance of the data collection scheme proposed in this patent, which shows the cosine similarity between the data collected by this partial collection scheme at different collection rates and the original matrix, indicating that the downsampled matrices obtained by this scheme at different collection rates can be as consistent with the original matrix as possible.

[0078] Figure 4 It shows the performance of matrix completion by adopting the structure of the autoencoder based on the graph attention network in the controller, proving the effectiveness of the matrix completion module in this scheme at different collection rates.

[0079] Figure 5 It shows the performance of traffic prediction by adopting the structure combining the graph attention network and the gated recurrent unit in the controller, proving the effectiveness of the traffic prediction module in this scheme at different collection rates.

Claims

1. A method for selectively collecting and predicting traffic data of low-earth orbit satellites, characterized in that It includes the following steps: Step 1: Traffic collection module (1) The controller generates a mask matrix M ∈ B based on the number N of satellite networks and the prediction time window T N×T , where the matrix elements are 0 or 1, and are used to indicate whether each satellite node uploads traffic data at a specified moment; Among them, the initial distribution of the mask is defined by the historical traffic heat map, giving priority to covering high-value spatio-temporal regions; (2) The controller distributes the collection scheme matrix M to each satellite node through the inter-satellite directional broadcast protocol. The satellite node uploads the inter-satellite link (ISL) traffic data at the specified time slice according to the position of "1" in the matrix to reduce the transmission overhead; (3) The controller integrates and uploads the data to construct a partially observed traffic matrix X ∈ R N×T , and the missing values are marked as NaN; Step 2: Traffic matrix completion module (4) Input the partially observed traffic matrix X and the adjacency matrix A of the satellite network into the hierarchical graph neural network f1. Use the graph attention network to extract key features and generate a low-dimensional embedding representation to extract spatio-temporal features, and generate a low-dimensional hidden feature h = f1(X, A); (5) Based on the hidden feature h, using the completion neural network f2, reconstruct the complete traffic matrix Ι ∈ R N×T , then Ι = f2(h, A); (6) Evaluate the completion effect by minimizing the difference between the completion matrix Ι and the real traffic matrix Y; Step 3: Traffic prediction module (7) Based on the completed traffic matrix Ι, use the neural network f3 to extract spatio-temporal features, combined with the adjacency matrix A, Construct the prediction model g; (8) Using the prediction model, output the traffic prediction matrix Ψ = g(f3(I), A) for the future T f time slices; (9) By minimizing the difference between the prediction matrix Ψ and the true traffic X P to evaluate the prediction accuracy; Step 4: Collection scheme update module (10) Based on the completion error E C and the prediction error E P , construct the joint loss function L = αE C +(1 - α)E P , where α ∈ (0, 1) is the weight coefficient for balancing the completion accuracy and the prediction accuracy; (11) Optimize the joint loss function through gradient descent, dynamically adjust the collection scheme matrix M, and preferentially collect data in areas with high reconstruction error or prediction error; (12) Repeat steps 1 to 4 to complete multiple rounds of iterative updates until the joint loss L converges or reaches the preset number of iterations.

2. The method according to claim 1, wherein Step 1 (1) Initialize the collection scheme through the historical traffic heat map, giving priority to covering high-value spatio-temporal regions.

3. The method according to claim 1, characterized in that, The encoder based on the multi-layer graph attention network (GAT) in Step 2 (4).

4. The method according to claim 1, characterized in that, In Step 3 (7), the spatio-temporal feature extraction adopts the combination of the gated recurrent unit (GRU) and the graph attention mechanism, which is used to model the temporal dependence and dynamic spatial correlation of traffic data.

5. The method according to claim 1, characterized in that, The specific method for dynamically adjusting the collection scheme matrix M in Step 4 (11) includes: (a) Threshold the continuously optimized collection metrics to generate a discretized 0-1 mask; (b) Adaptively increase the data collection frequency in high-error regions according to the error distribution.

6. The method according to claim 1, wherein The evaluation metrics in Step 2 (6) and Step 3 (9) include mean square error (MSE), mean absolute error (MAE), or cosine similarity.

7. The method according to claim 1, wherein The completion neural network in Step 2 (5) is an autoencoder based on the multi-layer graph attention network, which is used to jointly optimize the feature representation of the completion and prediction tasks.

8. The method according to any one of claims 1 to 9, characterized in that The adjacency matrix A between satellite nodes is dynamically generated based on the orbital position, link status, routing scheme, or historical communication load of the satellites.

Images