Satellite internet traffic prediction method, system, device and medium based on federated learning framework
By using a federated learning framework that trains models locally on satellites and uploads the model parameters, the high latency and data leakage problems in satellite internet traffic management are solved, enabling efficient and secure traffic prediction and routing optimization, adapting to network changes, and protecting user privacy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIDIAN UNIV
- Filing Date
- 2023-08-28
- Publication Date
- 2026-06-02
AI Technical Summary
Satellite internet traffic management and optimization face challenges such as high latency, data leakage risks, high computation and storage requirements, and poor adaptability of prediction algorithms. Traditional centralized methods are difficult to meet the needs of its complex network topology and heterogeneous users.
Using a federated learning framework, the satellite trains a traffic prediction model locally and uploads the model parameters to a ground station server for aggregation. Through multiple iterations, a global traffic prediction model is established, and the LSTM algorithm and differential privacy technology are used to protect data privacy.
It achieves efficient and secure traffic prediction, reduces data transmission costs and latency, has good accuracy and robustness, protects user privacy, adapts to network changes, and provides timely routing decision support.
Smart Images
Figure CN117135090B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of artificial intelligence and spatial information technology, and specifically relates to a method, system, device and medium for predicting satellite internet traffic based on a federated learning framework. Background Technology
[0002] With the development of satellite communication technology, satellite internet has become an important component of the integrated space-air-ground network. Satellite internet is widely used in military, aerospace, communications, and meteorological fields, providing crucial support for remote sensing data transmission, emergency communications, and global coverage. However, the high latency, dynamic link conditions, and complex network topology of satellite internet mean that traffic management and optimization remain challenges.
[0003] Traffic forecasting is a key issue in satellite internet traffic management and optimization. Accurately forecasting satellite internet traffic can help network operators and service providers better manage network resources, plan capacity, and provide more stable and high-quality services. Due to the wide geographical distribution, complex network topology, and heterogeneity among different users, traditional centralized traffic forecasting methods face several challenges. First, traditional methods typically require centralized storage of satellite internet traffic data on a central server for processing and analysis, which may involve a large amount of user privacy information, easily leading to data leaks and security risks. Second, transmitting all satellite internet traffic data to a central server requires significant network bandwidth and time, and the central server needs powerful computing capabilities and storage resources to process and analyze massive amounts of data. Third, centralized methods usually use fixed forecasting algorithms and models, making it difficult to adapt to changes in different regions, network topologies, and user needs, which limits the accuracy and adaptability of traffic forecasting.
[0004] Y. Bie et al. studied the problem of satellite network traffic prediction, and their combined prediction model, which utilizes the self-similarity of satellite network traffic, improved the accuracy of prediction (A Combined Forecasting Model for Satellite Network Self-Similar Traffic, in IEEE Access, vol.7, pp.152004-152013, 2019), did not consider the risk of data leakage. L. Yang et al. proposed a novel satellite network traffic prediction method based on a combination of graph convolutional neural network and gated recursive unit (GCN-GRU), which improved the prediction effect of satellite internet traffic (A Noval Satellite Network Traffic Prediction Method Based on GCN-GRU, 2020 International Conference on Wireless Communications and Signal). Processing (WCSP), Nanjing, China, 2020, pp. 718-723”), however, it relies on a central server to process large amounts of data, which not only requires high computing and storage capabilities, but also consumes a lot of time to transmit data, thus increasing the transmission latency of satellite internet data. Summary of the Invention
[0005] To overcome the shortcomings of the prior art, the present invention aims to propose a satellite internet traffic prediction method, system, device, and medium based on a federated learning framework. The federated learning framework enables each satellite in the satellite internet to train a model using local traffic data, and the trained model parameters are uploaded to a ground station server for aggregation. Through multiple iterations of model updates and parameter aggregation, a global traffic prediction model can be obtained, thereby improving the accuracy and security of satellite internet traffic prediction.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0007] Satellite internet traffic prediction methods based on a federated learning framework include:
[0008] Step 1: Construct an orbital satellite internet model, which includes ground stations, satellites, ground control centers, and users;
[0009] Step 2: Based on the orbital satellite internet model constructed in Step 1, determine the satellite internet traffic transmission process;
[0010] Step 3: Establish a federated learning framework using satellites as the client and ground stations as the server.
[0011] Step 4: Train the traffic prediction model using the satellite internet traffic transmission process from Step 2 and the federated learning framework established in Step 3.
[0012] Step 5: Implement real-time satellite internet traffic prediction based on the traffic prediction model trained in Step 4.
[0013] In step 1, the satellite constellation orbital parameters are represented by (N,M,α,h,S), where N represents the number of orbits, M represents the number of satellites in each orbit, α represents the orbital inclination, h represents the satellite orbital altitude, and S represents the number of ground stations.
[0014] If the orbital period of a satellite is T, and each orbit contains M satellites, with a total number of satellites C = N × M, then the period of change in the satellite internet topology is:
[0015]
[0016] The latitude and longitude of a satellite are represented by j. s w s The latitude and longitude of the ground station are represented by j. g w g express.
[0017] The specific process of step 2 is as follows:
[0018] Step 2.1: Calculate the visibility relationship between M satellites and S ground stations, and generate a satellite-to-ground visibility matrix. The row information is the satellite number, and the column information is the ground station number. 1 represents visible, and 0 represents invisible. The formula for calculating the visibility relationship is as follows:
[0019] F1 = cos(w) s cos(w) g cos(j) s -j g )+sin(w s sin(w) g )
[0020]
[0021] Among them, w s ,w g These are the latitudes of the satellite and the ground station, respectively. s ,j gThese represent the longitudes of the satellite and the ground station, respectively; R is the Earth's radius; h represents the satellite's orbital altitude; and θ represents the satellite's elevation angle.
[0022] When F1≥F2, the satellite and ground station are visible to each other, and the flag is set to 1.
[0023] When F1 < F2, the satellite and ground station are not visible to each other, and the flag is 0;
[0024] Step 2.2: The ground station control center collects satellite trajectory information, queue and neighbor information to generate network topology, and uses the shortest path algorithm to calculate the static routing table, setting the satellite and ground station visibility relationship and route to be updated once in each time slot;
[0025] Step 2.3: The satellite periodically sends HELLO packets to neighboring satellites to detect link status, propagation delay, and queue length information in neighboring satellites, generates a delay matrix, and transmits the delay matrix to the ground station control center through a static routing table. After generating the delay matrix, the ground station control center uses the interpolation method (Floyd algorithm) to generate a Quality of Service (QoS) routing table for the transmission of traffic data packets generated by the initial user.
[0026] In step 3, satellites act as clients in the federated learning process, with a total of C clients, and ground stations act as servers, with a total of S servers. When the satellites and ground stations are visible to each other, the satellite clients upload their trained model parameters to the ground station servers. The satellite clients minimize a global objective function, enabling the ground station servers and satellite clients to collaboratively train the traffic prediction model without sharing training data. The global objective function is as follows:
[0027]
[0028] Where, m c =|D c |, Represents the entire dataset.
[0029] The local objective function is as follows:
[0030]
[0031] Here, loss(w;x) represents the training loss of model parameters w and data points x.
[0032] The specific process of training the traffic prediction model in step 4 is as follows:
[0033] Step 4.1: The satellite client trains the traffic prediction model parameters based on the local objective function:
[0034] When the satellite connects to the ground station, the ground station server sends the traffic prediction model parameters to the satellite client. The satellite client uses the model parameters to initialize the local model and uses the Long Short-Term Memory (LSTM) deep learning algorithm to train the local dataset locally. After the model converges, the updated local parameters are uploaded to the ground station server.
[0035] Step 4.2: The ground station server aggregates all traffic prediction model parameters from all participating satellites:
[0036] The ground station server uses a federated averaging algorithm to aggregate the model parameters to obtain a global model parameter g, as shown in the following formula:
[0037]
[0038] Where d represents the number of satellite samples involved, w i This represents the model parameters for the i-th satellite.
[0039] Step 4.3: The ground station server transmits the new global model parameters to the participating satellite clients:
[0040] The ground station server sends the new global model parameters aggregated at the current moment back to each participating satellite that uploaded the model parameters. It uses the validation dataset to verify the performance of the current model. If the current model performance is lower than the model error of the best model in the previous round, the distributed global model parameters are used as the initial model parameters for the next round of training. The model parameters are then updated to adapt to the features and changes of the validation dataset. Otherwise, the model parameters from the previous round are used for the next step of judgment.
[0041] Step 4.4: Train the prediction model iteratively until the required number of iterations is reached or the model converges:
[0042] Through periodic aggregation and distribution of model parameters, the satellite continuously iterates through local dataset training and model updates until the model converges or reaches the preset number of training rounds.
[0043] Step 5 uses the traffic data of all participating satellites at the current moment as input to the model to predict the traffic data that the satellites will need to forward at the next moment. Based on the traffic forwarding situation of the satellites at the next moment, the satellite service quality routing table is updated in a timely manner to achieve load balancing of the orbital satellite network and improve the overall service quality of satellite internet. The specific steps are as follows:
[0044] Step 5.1: Traffic monitoring and analysis. Monitor the traffic load of each satellite in the satellite internet in real time, including data transmission volume and transmission rate. Based on the traffic monitoring data, analyze the load of each satellite and identify the satellites with high load and low load.
[0045] Step 5.2: Real-time data migration is achieved by migrating a portion of the data from satellites with high loads to satellites with lower loads in real time to achieve load balancing. Data migration is achieved by adjusting the data transmission path and delay control.
[0046] Step 5.3: Real-time monitoring and adjustment. Monitor satellite traffic load in real time, continuously update routing strategies based on new traffic forecasts and network topology information, adjust transmission paths in real time to adapt to dynamic changes in the network environment, and ensure optimal routing and stability.
[0047] Step 5.4: Expand the satellite network according to needs. After the overall function is realized, expand the satellite Internet according to the data transmission scale. The expanded satellites participate in federated learning in a timely manner and exchange model parameters with the ground station server.
[0048] The specific training traffic prediction model in step 4 is evaluated and improved as follows;
[0049] In step 4.1, data sampling and label sharing are introduced for the local dataset. For public, non-sensitive label information, it can be shared between satellites to reduce the problem of label data imbalance.
[0050] The satellite client uses differential privacy technology to add noise to local data before uploading the noise-added data to the ground station server.
[0051] The aggregation effect of the ground station is judged based on whether the global model parameters issued by the ground station server in step 4.3 are consistent with the training effect of most local satellites. When there are many problems with the aggregation model, an aggregation algorithm with dynamic weight allocation is adopted to adjust the weights according to the contribution of each satellite in order to obtain a better global model.
[0052] A satellite internet traffic prediction system based on a federated learning framework includes:
[0053] Satellite Internet Model Module: Constructs an orbital satellite internet model, which includes ground stations, satellites, ground station control centers, and users;
[0054] Traffic transmission module: Based on the orbital satellite internet model constructed by the satellite internet model construction module, determine the satellite internet traffic transmission process;
[0055] Federated Learning Framework Module: Establishes a federated learning framework by using satellites as clients and ground stations as servers;
[0056] Model Training Module: The federated learning framework module is used to train the traffic prediction model.
[0057] Real-time traffic prediction module: Real-time satellite internet traffic prediction is achieved based on the traffic prediction model obtained from the model training module.
[0058] A satellite internet traffic prediction device based on a federated learning framework includes:
[0059] Memory: A computer program for storing the satellite internet traffic prediction method based on the federated learning framework;
[0060] Processor: Used to implement a satellite internet traffic prediction method based on a federated learning framework when executing the computer program.
[0061] A satellite internet traffic prediction medium based on a federated learning framework is provided, wherein the computer-readable storage medium stores a computer program that, when executed by a processor, enables a satellite internet traffic prediction method based on a federated learning framework.
[0062] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0063] 1. This invention uses a federated learning framework to train a satellite internet traffic prediction model. The satellite trains the model locally without sharing the original data. Furthermore, because the data is distributed across various satellites, the federated learning framework has a certain degree of resistance to attacks. Even if a satellite is attacked or malfunctions, the entire system can continue to work without causing a global failure.
[0064] 2. In this invention, each satellite only uploads processed model parameters instead of raw data. Compared to transmitting large amounts of raw data, this method reduces the amount of data and computational load, improves data efficiency, avoids the transmission and centralized storage of large amounts of data in traditional centralized methods, reduces data transmission costs and latency, and improves the efficiency of satellite internet traffic prediction.
[0065] 3. This invention allows for the addition or removal of participating satellites in a satellite network, exhibiting high scalability. New satellites can join federated learning and be trained using local data, thereby expanding the scale and capabilities of the entire system. Furthermore, by training models and updating parameters on real-time time-series traffic data on each satellite, real-time traffic prediction and performance optimization can be achieved. Changes in satellite internet traffic can be responded to and adapted to quickly, providing timely routing decision support and services.
[0066] 4. The satellite of this invention uses differential privacy technology to add noise to local data, and then uploads the noise-added data to the ground station server for model parameter aggregation. The ground station server does not directly access the user's original data, but only the noise-added model parameters. This ensures the privacy and security of user data. Even if the ground station server is attacked or leaked, the user's sensitive information will not be exposed. While protecting user data privacy, it can achieve accurate prediction and optimization of traffic. Users can use satellite internet services with confidence without worrying about the leakage of personal information.
[0067] In summary, the federated learning framework enables individual satellites in a satellite internet to train models using local traffic data and upload the trained model parameters to a ground station server for aggregation. Through multiple iterations of model updates and parameter aggregation, a global traffic prediction model can be obtained, exhibiting good accuracy and robustness. This avoids the problems of centralized data storage and processing while also protecting the privacy and security of user data. Attached Figure Description
[0068] Figure 1 This is a usage scenario diagram of the present invention.
[0069] Figure 2 This is a flowchart of the satellite internet traffic transmission process of the present invention.
[0070] Figure 3 This is a flowchart of the method of the present invention.
[0071] Figure 4 This is a flowchart of the process of training the traffic prediction model of the present invention.
[0072] Figure 5 This is a comparison chart of the flow prediction method of this invention and the centralized flow prediction method. Detailed Implementation
[0073] The present invention will now be described in further detail with reference to the accompanying drawings.
[0074] Reference Figure 1The application scenario of this invention is an orbital satellite network comprising C satellites, N orbits, and S ground stations. With the widespread adoption and increasing use of the internet, satellite internet traffic exhibits high growth and significant fluctuations. Traffic variations across different regions, time periods, and users are highly complex, requiring accurate prediction to meet network resource and service demands. Furthermore, satellite internet resources are limited, necessitating rational planning and management to satisfy user requirements for bandwidth and service quality. Current research on prediction models for satellite internet largely employs centralized models, which can easily lead to wasted network resources and data leaks. Therefore, this invention utilizes a federated model framework to develop a distributed prediction model for satellite internet.
[0075] Reference Figure 2 , Figure 3 Satellite internet traffic prediction methods based on a federated learning framework include:
[0076] Step 1: Construct an orbital satellite internet model, which includes ground stations, satellites, ground station control centers, and users;
[0077] The STK satellite toolkit generates satellite and ground station trajectory files. The satellite and ground station read the location information from the trajectory files to obtain their current location. Users are responsible for generating traffic services, which are then propagated to the destination user through inter-satellite links and satellite-to-ground links. The ground station control center is responsible for establishing the entire network topology and calculating the corresponding routing table based on the network status and source and destination satellite information, and sending it to the source satellite.
[0078] Step 2: Determine the satellite internet traffic transmission process;
[0079] Step 2.1: Calculate the visibility relationship between C satellites and S ground stations, generating a satellite-to-ground visibility matrix. Row information represents satellite numbers, and column information represents ground station numbers, with 1 indicating visibility and 0 indicating invisibility. The formula for calculating the visibility correlation is as follows:
[0080] F1 = cos(w) s cos(w) g cos(j) s -j g )+sin(w s sin(w) g )
[0081]
[0082] Among them, w s ,w g These are the latitudes of the satellite and the ground station, respectively. s ,j gThese represent the longitudes of the satellite and the ground station, respectively; R is the Earth's radius; h represents the satellite's orbital altitude; and θ represents the satellite's elevation angle.
[0083] When F1≥F2, the satellite and ground station are visible to each other, and the flag is set to 1.
[0084] When F1 < F2, the satellite and ground station are not visible to each other, and the flag is 0;
[0085] Step 2.2: The ground station control center collects satellite trajectory information, queue and neighbor information to generate network topology, and uses the shortest path algorithm to calculate the static routing table. Due to the variability of satellite topology, the satellite-to-ground visibility relationship and routes are updated every minute.
[0086] Step 2.3: The satellite periodically sends HELLO packets to neighboring satellites to detect link status, propagation delay, and queue length information in the satellite, and uses this information to generate a delay matrix. The delay matrix is then transmitted to the ground station control center through a static routing table. After generating the delay matrix, the ground station control center uses the Floyd algorithm to generate a QoS routing table for the transmission of traffic data packets generated by the initial user.
[0087] Step 3: Establish a federated learning framework;
[0088] Using satellites as clients of federated learning and ground stations as central servers, each satellite can train on locally processed historical traffic data without uploading the traffic data to the ground station for retraining. Only the model parameters trained on the satellites need to be uploaded to the ground station. The global objective function of the federated learning framework is:
[0089]
[0090] Where, m c =|D c |, Represents the entire dataset.
[0091] The local objective function is as follows:
[0092]
[0093] Where loss(w;x) represents the training loss of model parameters w and data points x;
[0094] Step 4: Train the prediction model using the federated learning framework;
[0095] See Figure 4 After determining the overall framework of federated learning, satellites and ground stations need to collaboratively train the prediction model without sharing historical traffic data. The specific steps are as follows:
[0096] Step 4.1: The satellite client trains the traffic prediction model parameters based on the local objective function:
[0097] When the satellite connects to the ground station, the ground station sends the model parameters to the satellite. The satellite uses the model parameters to initialize the local model and uses the LSTM algorithm to train the local dataset locally. After the model converges, the updated local parameters are uploaded to the central server ground station.
[0098] Step 4.2: The ground station server aggregates all traffic prediction model parameters from all participating satellites:
[0099] After all participating satellite clients upload their updated local parameters in a round, the ground station server uses the FedAvg algorithm to aggregate the model parameters for this round, ultimately obtaining a global model parameter.
[0100] Step 4.3: The ground station server transmits the new global model parameters to the participating satellites:
[0101] After all participating satellites receive the new global model parameters from the ground station, they use the validation traffic dataset to verify the performance of the current model. If the current model performs better than the best model in the previous round, the distributed global model parameters are used as the initial model for the next round of training, and the model parameters are updated to adapt to the characteristics and changes of the validation traffic dataset. Otherwise, the next step is determined.
[0102] Step 4.4: Train the prediction model iteratively until the required number of iterations is reached or the model converges:
[0103] Determine whether the prediction model iteration has reached the target or whether the optimal model trained in the current round has converged. If the global target training conditions are met, end the training; otherwise, continue the iteration.
[0104] Step 4.5: Evaluate and improve the prediction model;
[0105] Step 4.5.1: Evaluate the aggregation effect of ground station parameters
[0106] The aggregation effect of the ground station is judged based on whether the global model parameters sent by the ground station server in step 4 are consistent with the training effect of most local satellites. When there are many problems with the aggregation model, an aggregation algorithm with dynamic weight allocation is adopted to adjust the weights according to the contribution of each satellite in order to obtain a better global model.
[0107] Step 4.5.2: Introduce data sampling and label sharing mechanisms in the federated learning framework.
[0108] Sampling data from different regions increases data diversity and improves the model's generalization ability. Simultaneously, public, non-sensitive tag information can be shared among satellites, reducing the problem of imbalanced tag data.
[0109] Step 4.5.3: In terms of data protection, the application of differential privacy can be further optimized.
[0110] Researching different differential privacy algorithms and parameter settings to minimize the impact on model prediction accuracy while protecting data privacy requires balancing and adjusting between privacy protection and model accuracy.
[0111] Step 5: Implement real-time satellite internet traffic prediction based on the trained traffic prediction model;
[0112] Real-time satellite internet traffic prediction is achieved using a prediction model trained based on a federated learning framework. The model takes the traffic data of all participating satellites at the current moment as input and predicts the traffic data that satellites will need to forward at the next moment. Based on the satellite traffic forwarding situation at the next moment, the satellite QoS routing table is updated in a timely manner to achieve satellite network load balancing and improve the overall QoS performance of the satellite internet. The specific steps are as follows:
[0113] Step 5.1: Traffic Monitoring and Analysis
[0114] Real-time monitoring of traffic load on each satellite in the satellite internet, including data transmission volume and rate. Based on the traffic monitoring data, analyze the load of each satellite to identify those with high and low loads.
[0115] Step 5.2: Real-time data migration
[0116] In real time, a portion of the data from satellites with high loads is migrated to satellites with lower loads to achieve load balancing. This data migration is achieved by adjusting the data transmission path and controlling latency.
[0117] Step 5.3: Real-time monitoring and adjustment
[0118] Over time, satellite traffic load is continuously monitored, and routing strategies are constantly updated based on new traffic forecasts and network topology information. Transmission paths are adjusted in real time to adapt to dynamic changes in the network environment, ensuring optimal routing and stability.
[0119] Step 5.4: Expand the satellite network according to needs. After the overall function is realized, expand the satellite Internet according to the data transmission scale. The expanded satellites participate in federated learning in a timely manner and exchange model parameters with the ground station server.
[0120] See Figure 5In satellite communication systems, when the system service rate increases dramatically, i.e., the demand for traffic services expands rapidly, the packet loss rate and latency will rise significantly, affecting the QoS capabilities of satellite internet. Data is transmitted to a queue of 60 satellites at an average service rate of 10 Mbps. The traffic distribution for the next time slot is predicted using both traditional centralized prediction and the federated learning framework of this invention. In the federated learning framework, five ground stations act as ground station servers to calculate the satellite-to-ground visibility relationship for different time slots. The satellite client trains the dataset using an LSTM model with a learning rate of 0.05. Each satellite is trained according to a local objective function and uploads updated parameters to the ground station servers. The ground station servers aggregate the model parameters using a federated averaging algorithm and distribute them. This process is iterated 10 times until the model converges, yielding the final overall prediction result.
[0121] The comparison shows that the determination coefficient R of this invention is... 2 The results are higher than those of centralized prediction, proving that the model of this invention has a better fit. The root mean square error (RMSE) and mean absolute error (MAE) of this invention are smaller than those of traditional centralized prediction, proving that the model of this invention has a better fit. It can be seen that this invention increases the accuracy of the prediction model. Based on the peak and fluctuation of the predicted traffic data, the satellite can update the routing strategy in advance, thereby maximizing the use of network resources and improving bandwidth utilization.
[0122] Satellite internet covers multiple regions, each with its own satellites. Network connections between these regions may vary in bandwidth, latency, and signal quality. To provide high-quality communication services, service quality optimization is needed tailored to the characteristics of each region. First, each region's satellites use local datasets for traffic prediction. Through a federated learning framework, the prediction model parameters for each region are uploaded to a ground station server for aggregation. The ground station server then generates a global traffic prediction model based on the prediction models for each region. Based on this global traffic prediction model, network operators can rationally plan and adjust network resources for each region. For regions predicted to experience high traffic during peak hours, bandwidth and storage resources can be increased to improve service quality and ensure stable connections and high-speed transmission for users even during peak times. For regions with poor network quality, measures can be taken to optimize routing and scheduling strategies to avoid network congestion and excessive latency. Guided by the global model, routing paths and transmission strategies can be adjusted in real time to improve network throughput and response speed. Satellite internet traffic prediction technology based on the federated learning framework can optimize and guarantee service quality according to the network characteristics and traffic demands of different regions. By rationally planning network resources and optimizing network strategies, it can provide more stable and efficient communication services, thereby improving user experience and satisfaction.
[0123] Consider a satellite internet system where multiple users connect to the internet via satellite. Each user generates a large amount of traffic data, containing their browsing habits, personal preferences, and sensitive information. Traditional centralized traffic prediction methods may require storing all users' raw data on a central server, leading to risks of user privacy leaks and misuse. Using a federated learning framework, each user's data is processed and trained locally without sharing the raw data. The satellite uses differential privacy technology to add noise to the local data before uploading the noisy data to the central server for model parameter aggregation. Through the federated learning framework, the central server does not directly access the users' raw data, only the noisy model parameters. This ensures user data privacy and security; even if the ground station server is attacked or compromised, users' sensitive information will not be exposed. Satellite internet traffic prediction technology based on the federated learning framework can accurately predict and optimize traffic while protecting user data privacy. Users can use satellite internet services with confidence, without worrying about the leakage of personal information.
[0124] A satellite internet traffic prediction system based on a federated learning framework includes:
[0125] Satellite Internet Model Module: Constructs an orbital satellite internet model, which includes ground stations, satellites, ground station control centers, and users;
[0126] Traffic transmission module: Based on the orbital satellite internet model constructed by the satellite internet model construction module, determine the satellite internet traffic transmission process;
[0127] Federated Learning Framework Module: Establishes a federated learning framework by using satellites as clients and ground stations as servers;
[0128] Model Training Module: The federated learning framework module is used to train the traffic prediction model.
[0129] Real-time traffic prediction module: Real-time satellite internet traffic prediction is achieved based on the traffic prediction model obtained from the model training module.
[0130] A satellite internet traffic prediction device based on a federated learning framework includes:
[0131] Memory: A computer program for storing the satellite internet traffic prediction method based on the federated learning framework;
[0132] Processor: Used to implement a satellite internet traffic prediction method based on a federated learning framework when executing the computer program.
[0133] A satellite internet traffic prediction medium based on a federated learning framework is provided, wherein the computer-readable storage medium stores a computer program that, when executed by a processor, enables a satellite internet traffic prediction method based on a federated learning framework.
Claims
1. A satellite internet traffic prediction method based on a federated learning framework, characterized in that, Includes the following steps: Step 1: Construct an orbital satellite internet model, which includes ground stations, satellites, ground station control centers, and users; Step 2: Based on the orbital satellite internet model constructed in Step 1, determine the satellite internet traffic transmission process; Step 3: Establish a federated learning framework using satellites as the client and ground stations as the server. Step 4: Train the traffic prediction model using the satellite internet traffic transmission process from Step 2 and the federated learning framework established in Step 3. Step 5: Real-time satellite internet traffic prediction is achieved based on the traffic prediction model trained in Step 4. The specific process of training the traffic prediction model in step 4 is as follows: Step 4.1: The satellite client trains the traffic prediction model parameters based on the local objective function: When the satellite connects to the ground station, the ground station server sends the traffic prediction model parameters to the satellite client. The satellite client uses the model parameters to initialize the local model and uses the Long Short-Term Memory (LSTM) deep learning algorithm to train the local dataset locally. After the model converges, the updated local parameters are uploaded to the ground station server. Step 4.2: The ground station server aggregates all traffic prediction model parameters from all participating satellites: The ground station server uses a federated averaging algorithm to aggregate the model parameters to obtain a global set of model parameters. The formula is as follows: in, Representatives of the satellite sample size Indicates the first Model parameters for each satellite; Step 4.3: The ground station server transmits the new global model parameters to the participating satellite clients: The ground station server sends the new global model parameters aggregated at the current moment back to each participating satellite that uploaded the model parameters. It uses the validation dataset to verify the performance of the current model. If the current model performance is lower than the error of the best model in the previous round, the distributed global model parameters are used as the initial model parameters for the next round of training. The model parameters are then updated to adapt to the features and changes of the validation dataset. Otherwise, the model parameters from the previous round are used for the next step of judgment. Step 4.4: Train the prediction model iteratively until the required number of iterations is reached or the model converges: Through periodic aggregation and distribution of model parameters, the satellite continuously iterates through local dataset training and model updates until the model converges or reaches the preset number of training rounds. Step 5 uses the traffic data of all participating satellites at the current moment as input to the model to predict the traffic data that the satellites will need to forward at the next moment. Based on the traffic forwarding situation of the satellites at the next moment, the satellite service quality routing table is updated in a timely manner to achieve load balancing of the orbital satellite network and improve the overall service quality of satellite internet. The specific steps are as follows: Step 5.1: Traffic monitoring and analysis. Monitor the traffic load of each satellite in the satellite internet in real time, including data transmission volume and transmission rate. Based on the traffic monitoring data, analyze the load of each satellite and identify the satellites with high load and low load. Step 5.2: Real-time data migration is achieved by migrating a portion of the data from satellites with high loads to satellites with lower loads in real time to achieve load balancing. Data migration is achieved by adjusting the data transmission path and delay control methods. Step 5.3: Real-time monitoring and adjustment. Monitor satellite traffic load in real time, continuously update routing strategies based on new traffic forecasts and network topology information, adjust transmission paths in real time to adapt to dynamic changes in the network environment, and ensure optimal routing and stability. Step 5.4: Expand the satellite network according to needs. After the overall function is realized, expand the satellite Internet according to the data transmission scale. The expanded satellites participate in federated learning in a timely manner and exchange model parameters with the ground station server.
2. The satellite internet traffic prediction method based on a federated learning framework according to claim 1, characterized in that, In step 1, the satellite constellation orbit parameters are used To indicate, Indicates the number of orbits. This indicates the number of satellites in each orbit. Indicates the orbital inclination angle. Indicates the satellite's orbital altitude. Indicates the number of ground stations; The orbital period of the satellite is Each track has There are [number] satellites, with a total of [number] satellites. The cycle of satellite internet topology changes is: Satellite latitude and longitude are used This indicates that the latitude and longitude of the ground station are expressed in... express.
3. The satellite internet traffic prediction method based on a federated learning framework according to claim 1, characterized in that, The specific process of step 2 is as follows: Step 2.1: Calculation satellites and The visibility relationships of each ground station are used to generate a satellite-to-ground visibility matrix. The row information is the satellite number, and the column information is the ground station number. 1 represents visible, and 0 represents invisible. The formula for calculating visibility-related judgments is as follows: in, These are the latitudes of the satellite and the ground station, respectively. These are the longitudes of the satellite and the ground station, respectively. For the Earth's radius, Indicates the satellite's orbital altitude. Indicates the satellite's elevation angle; when At that time, the satellite and ground station are visible to each other, and the flag is 1; when At this time, the satellite and ground station are not visible to each other, and the flag is 0; Step 2.2: The ground station control center collects satellite trajectory information, queue and neighbor information to generate network topology, and uses the shortest path algorithm to calculate the static routing table, setting the satellite and ground station visibility relationship and route to be updated once in each time slot; Step 2.3: The satellite periodically sends HELLO packets to neighboring satellites to detect link status, propagation delay, and queue length information in neighboring satellites, generates a delay matrix, and transmits the delay matrix to the ground station control center through a static routing table. After generating the delay matrix, the ground station control center uses the interpolation method (Floyd algorithm) to generate a Quality of Service (QoS) routing table for the transmission of traffic data packets generated by the initial user.
4. The satellite internet traffic prediction method based on a federated learning framework according to claim 1, characterized in that, In step 3, satellites are used as clients for federated learning. There are 1 client, and the ground station acts as the server. There are two servers. When the satellite and ground station are visible, the satellite client uploads its trained model parameters to the ground station server. The satellite client enables the ground station server and satellite client to collaboratively train the traffic prediction model without sharing training data by minimizing a global objective function. The global objective function is as follows: in, , Represents the entire dataset. The local objective function is as follows: in, Represents model parameters and data points Training loss.
5. The satellite internet traffic prediction method based on a federated learning framework according to claim 1, characterized in that, The specific training traffic prediction model in step 4 is evaluated and improved as follows; In step 4.1, data sampling and label sharing are introduced for the local dataset. For public, non-sensitive label information, it can be shared between satellites to reduce the imbalance of label data. The satellite client uses differential privacy technology to add noise to the local data and uploads the noise-added data to the ground station server. The aggregation effect of the ground station is judged based on whether the global model parameters issued by the ground station server in step 4.3 are consistent with the training effect of most local satellites. When there are many problems with the aggregation model, an aggregation algorithm with dynamic weight allocation is adopted to adjust the weights according to the contribution of each satellite in order to obtain a better global model.
6. A satellite internet traffic prediction system based on a federated learning framework, used to implement the method of claim 1, characterized in that, include: Satellite Internet Model Module: Constructs an orbital satellite internet model, which includes ground stations, satellites, ground control centers, and users; Traffic transmission module: Based on the orbital satellite internet model constructed by the satellite internet model construction module, determine the satellite internet traffic transmission process; Federated Learning Framework Module: Establishes a federated learning framework by using satellites as clients and ground stations as servers; Model Training Module: The federated learning framework module is used to train the traffic prediction model. Real-time traffic prediction module: Real-time satellite internet traffic prediction is achieved based on the traffic prediction model obtained from the model training module.
7. A satellite internet traffic prediction device based on a federated learning framework, characterized in that, include: Memory: for storing the computer program of the satellite internet traffic prediction method based on the federated learning framework as described in any one of claims 1-5; Processor: Used to implement the satellite internet traffic prediction method based on the federated learning framework as described in any one of claims 1-5 when executing the computer program.
8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the satellite internet traffic prediction method based on a federated learning framework as described in any one of claims 1-5.