Low-orbit giant constellation network routing method based on graph neural network and reinforcement learning

Abstract

The invention discloses a low-orbit giant constellation network routing method based on a graph neural network and reinforcement learning, and is applied to the technical field of wireless communication. The method aims at solving the problem that the complex service quality requirement cannot be guaranteed due to the characteristics of large network topology scale and strong network dynamic property of a giant constellation network in the existing routing technology. According to the method, firstly, a multi-domain routing management architecture aiming at the giant constellation network is adopted, a virtual domain special for cross-domain routing is constructed, a global path is calculated based on abstract topology, and then a model combining a graph neural network and reinforcement learning is adopted in each domain, so that the global path of the giant constellation network is obtained. Through accurate prediction of path quality by the graph neural network, assisting the reinforcement learning agent in selecting an optimal path in the domain, and finally obtaining a cross-domain complete path through splicing the optimal path in the domain; according to the invention, under the high dynamic environment of the giant constellation network, routing with low computational complexity and sufficient guarantee of complex service quality requirements can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communications, and in particular relates to a low-orbit giant constellation network routing method based on graph neural network and reinforcement learning. Background Art

[0002] With the advent of the 6G era, new scenarios such as the Internet of Vehicles (IoV), smart cities, and extended reality (XR) are constantly emerging, and the data traffic in the network is showing explosive growth. The service quality requirements of different businesses are becoming more and more complex and sophisticated. Due to the high communication costs and limited infrastructure, the ground communication network cannot meet the global communication needs. Satellite networks have become an important supplement to ground networks because they can provide full coverage, all-weather, and low-latency communication services. At the same time, in order to cope with the explosive growth of network traffic, giant constellation networks represented by Starlink have gradually become the development paradigm of future satellite networks. However, giant constellation networks have the characteristics of large topological scale and strong network dynamics, which leads to complex cross-domain routing calculations and makes it difficult to guarantee complex service quality requirements including multiple indicators such as latency, packet loss rate, and throughput.

[0003] Existing satellite routing mechanisms mainly focus on small satellite networks and can be divided into two categories: distributed and centralized. The literature "Y.Lyu, H.Hu, R.Fan, Z.Liu, J.An, and S.Mao, "Dynamic routing for integrated satellite-terrestrial networks: A constrained multi-agent reinforcement learning approach," IEEE Journal on Selected Areas in Communications, vol.42, no.5, pp.1204–1218, 2024." uses a distributed multi-agent reinforcement learning mechanism to select the next-hop satellite that best meets the energy constraints and service quality requirements for each satellite as an independent agent. However, this type of distributed method mainly uses local information to calculate routing and relies on frequent information exchange, which will further increase network congestion and cannot guarantee that the complete path can meet the complex service quality requirements of diverse services. The literature "Han Z, Zhao G, Xing Y, et al. Dynamicrouting for software-defined LEO satellite networks based on ISLattributes[C] / / 2021IEEE Global Communications Conference(GLOBECOM).IEEE, 2021: 1-6." uses a centralized multi-attribute intersatellite link utility mechanism, establishes a path utility model by dynamically adjusting the weights of the attributes, and selects the optimal path that best meets the service quality requirements. However, this type of centralized method relies on collecting global information to calculate the route, which is difficult to cope with the massive links and highly dynamic topology of giant constellation networks, and cannot effectively guarantee complex service quality requirements. Therefore, it is of great significance to study routing methods suitable for low-orbit giant constellation networks. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a low-orbit giant constellation network routing method based on graph neural network and reinforcement learning, which can achieve a balance between computational complexity and service quality requirements in the highly dynamic environment of the giant constellation network.

[0005] The technical solution adopted by the present invention is: a low-orbit giant constellation network routing method based on graph neural network and reinforcement learning, comprising:

[0006] Deployment of multi-domain routing management architecture: deploy a root controller and multiple local controllers; divide the domains to which each satellite in the giant constellation network belongs; the root controller is used to manage all local controllers; each local controller can monitor all satellites in the domain;

[0007] Constructing virtual domains: The root controller adds border satellites of neighboring domains connected to each domain to construct virtual domains for cross-domain routing;

[0008] Obtaining the global path: The root controller obtains the abstract topology based on the virtual domains where the source satellite node and the destination satellite node are located and the local controllers of the virtual domains passed through, and then obtains the feasible global path based on the abstract topology; the abstract topology consists of the source satellite node, the destination satellite node, and the boundary satellite nodes of the virtual domains passed through;

[0009] Determine the optimal route within the virtual domain: The local controller corresponding to each virtual domain involved in the feasible global path uses the route selection model that has been trained, inputs the network status information within the domain into the selection model, and selects the model to output the optimal intra-domain path; when training, the selection model aims to output the intra-domain path with the highest service quality index within the domain;

[0010] Determine the optimal cross-domain route: The root controller seamlessly splices the optimal path in each virtual domain involved in the feasible global path to obtain a cross-domain complete path, and selects the cross-domain complete path with the highest global service quality index as the cross-domain optimal route output.

[0011] Beneficial effects of the present invention: A low-orbit giant constellation network routing method based on graph neural network and reinforcement learning of the present invention can efficiently realize cross-domain routing calculation in a dynamic environment and fully guarantee complex service quality requirements; the multi-domain routing management architecture adopted by the present invention can reduce the complexity of routing calculation and reduce routing planning time; the model combining graph neural network and reinforcement learning proposed by the present invention can adaptively adjust routing decisions by predicting network status, avoid network congestion, and improve resource utilization. Therefore, the present invention has the advantages of short routing planning time and high routing decision quality, and can be directly applied to giant constellation networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is the algorithm workflow diagram of the present invention;

[0013] Figure 2 A model that combines graph neural networks and reinforcement learning;

[0014] Figure 3 The model training process for combining graph neural networks and reinforcement learning;

[0015] Figure 4The average delay performance of different routing algorithms under different traffic intensities and constellation networks of different sizes;

[0016] in, Figure 4 (a) The corresponding constellation contains 256 satellites, Figure 4 (b) The corresponding constellation contains 768 satellites, Figure 4 (c) The corresponding constellation contains 1584 satellites; DETAILED DESCRIPTION

[0017] To facilitate those skilled in the art to understand the technical content of the present invention, the present invention is further explained below with reference to the accompanying drawings.

[0018] The algorithm workflow of the present invention is as follows Figure 1 As shown, the specific implementation steps are:

[0019] Step 1: Adopting a multi-domain routing management architecture for a mega-constellation network

[0020] The multi-domain routing management architecture realizes the decoupling of the control plane and the data plane based on the SDN architecture. The control plane consists of a root controller and multiple local controllers deployed on the ground. The satellite domains are divided based on the latitude and longitude grid and geographic population density. The root controller is used to manage multiple local controllers. Each local controller corresponds to a domain and can monitor all satellites in the domain. The data plane consists of a large number of low-orbit satellites and only undertakes basic data forwarding tasks.

[0021] The root controller communicates and exchanges data with local controllers to dynamically adjust service quality indicators and routing decisions within or across domains to improve the utilization of giant constellation network resources.

[0022] Step 2: Construct a virtual domain dedicated to cross-domain routing

[0023] A virtual domain is constructed by adding neighboring boundary satellites connected to the original domain. If there are multiple boundary satellites between two adjacent domains, the boundary satellite with the largest link available bandwidth is selected first.

[0024] Step 3: Calculate the global path based on the abstract topology

[0025] From a global perspective, the virtual domains where the source satellite nodes and the destination satellite nodes are located are used to estimate the virtual domains that need to be passed through. The abstract topology is obtained using the local controllers involved. The abstract topology only includes the source satellite nodes, the destination satellite nodes, and the boundary satellite nodes. On this basis, the feasible global path is calculated according to the source and destination pairs of each traffic demand:

[0026] R' m =[R' m,1 ,R' m,2 ,...,R' m,K']

[0027] Among them, R' m represents the set of all feasible global paths for the mth source-destination pair, which has a total of K', each of which can be expressed as:

[0028]

[0029] Among them, R' m,k' represents the k'th feasible global path of the mth source-destination pair, VD represents the virtual domain, and J' represents the total number of virtual domains that need to be passed under the current path. Represents the j'th virtual domain that the traffic of the m'th source-destination pair passes through when taking the k'th feasible global path.

[0030] Step 4: The local controller of each domain uses a model combining graph neural network and reinforcement learning to calculate the optimal routing in the domain. The routing model combining graph neural network and reinforcement learning is as follows: Figure 2 As shown in the figure, the local controller of each domain acts as an intelligent agent, collects network status information from the environment, and uses this information to learn the dependencies between links and paths based on the graph neural network, obtains network performance prediction results such as delay, packet loss rate, and throughput of different routing schemes, and calculates the Q value that measures the quality of the path. On this basis, the intelligent agent learns routing decisions with the goal of obtaining the maximum reward, and obtains the optimal route in the domain under the current environmental state. Among them, each local controller acts as an intelligent agent, and collects the virtual domain topology, traffic matrix, and available bandwidth matrix of links in the domain at the current moment from the giant constellation network environment to form a state space:

[0031]

[0032] Among them, t represents the current time sequence number, s t represents the state space at time t, represents the virtual domain topology at time t, TM t represents the flow matrix at time t, C t represents the available bandwidth matrix of intra-domain links at time t;

[0033] The agent uses Yen's algorithm to obtain K shortest paths in the domain according to the source nodes and boundary nodes in the virtual domain as the discrete action space:

[0034] A t =[p 1 ,p 2 ...,p K ]

[0035] Among them, A t represents the discrete action space at time t, p represents the path within the domain, and K represents the total number of shortest paths within the virtual domain;

[0036] The agent uses the ∈-greedy strategy to select actions, randomly selecting a path in the domain from the action space with probability ∈, or selecting the path in the domain with the maximum Q value according to the path quality Q value estimated by the graph neural network with probability 1-∈. The process of selecting an action based on the maximum Q value can be expressed as:

[0037]

[0038] Among them, a' represents the action space A at time t t All actions in a t Represents the selected action, from the action space A t The action selected in a t is the corresponding path in the domain, argmax is the maximum index function, Q(s t ,a') represents the current state space s t The Q value calculation process of the next action a'.

[0039] The Q value represents the path quality, which is expressed as a weighted combination of three key service quality indicators: delay, packet loss rate, and throughput. The intelligent agent uses a graph neural network model based on the existing satellite network status information s t For the action space A t All actions in are predicted in turn, and the specific Q value calculation process is as follows:

[0040] The graph neural network model of the intelligent agent adopts the message passing neural network architecture, which is divided into three stages: initialization, message passing, and readout. First, the state s at the current time t collected in the satellite network environment is t It is parsed into three levels of features: link l, path p, and domain d, where the initial link feature x l Including available bandwidth, betweenness centrality, path initial characteristics x p Including traffic matrix, topology matrix, domain initial feature x d is the average available bandwidth of links in the domain. In the initialization phase, the initial features of links, paths, and domains are used to construct the hidden states of links and paths, and filled to a fixed dimension:

[0041] h l ←[x l ,x d ,0,...,0]

[0042] h p ←[x p ,0,...,0]

[0043] Among them, x l represents the initial characteristics of the link, x p represents the initial features of the path, xd represents the initial features of the domain, uses 0 to fill the dimension, ← represents the update assignment, h l 、h p They represent the link hiding state and path hiding state respectively.

[0044] The message passing phase learns the potential cyclic dependencies between links and paths through the three operations of message function Message(·), aggregation function Aggregation(·), and update function Update(·). The message aggregation and update are performed interactively at the path and link level and repeated T' times. First, for each path, its hidden state is related to the states of all the links it contains. The specific operations are as follows:

[0045]

[0046] in, They represent the hidden states of the paths and links in the current t'th cycle, li represents the i-th link l, represents the attention coefficient of the i-th link l on path p, is the path message vector of the t'th cycle. Message(·) means encoding the hidden state to generate the message vector needed for subsequent use. Aggregation(·) means using the attention coefficient for each link on the path in the t'th cycle Perform weighted message aggregation. The attention coefficient is obtained by concatenating the two feature vectors of betweenness centrality and average available bandwidth of links in the domain after softmax operation. Update(·) indicates that after the message aggregation on the link is completed, the path state is recursively updated using the gated recurrent unit and saved as the intermediate state before the link state is updated. After the message aggregation operation is performed on all links on the path, the t'th loop ends and the final state is updated to the initial state of the next round. After the path information is updated, the operation of all links in the network topology is entered. Since the links have similar dependencies on the paths, the same operation is used:

[0047]

[0048] in, are the link message vector and path message vector of the t'+1th cycle respectively, is the updated link hidden state after the t'th cycle ends.

[0049] After completing T' rounds of message transmission iterations, a regression model is established in the readout phase through the readout function Readout(·) to obtain the Q value representing the path service quality indicator:

[0050]

[0051] in, Represents the final path hidden state after completing T' cycles, The three-dimensional indicators of delay, packet loss rate and throughput are output separately. 1 represents the first layer readout function, which is composed of a fully connected network and a nonlinear activation function selu. Q represents the weighted combination of the three-dimensional indicators of latency, packet loss rate, and throughput. R 2 represents the second-layer readout function, which is composed of a fully connected network. Since the current state space s t Represents the structural information of the topology. Selecting actions based on this state space will change the distribution of traffic in the topology. Therefore, the final path hidden state can be obtained by combining the specific state and action through cyclic spatial convolution. The embedding of traffic distribution and topological structure is realized. On this basis, the readout function Readout(·) is used to establish a regression model to learn the final path hidden state. The relationship between the delay, packet loss rate and throughput and the path quality indicators (such as delay, packet loss rate and throughput) is used to predict the quality of each action in the discrete action space, and the weighted combination of the delay, packet loss rate and throughput three-dimensional indicators is expressed as the Q value. Finally, by finding the largest Q value, the path that best meets the satellite network service quality requirements is selected.

[0052] The agent starts from the action space A t The action selected in a t That is, the corresponding intra-domain path will be obtained from the network environment, and the service quality index after the path change will be calculated based on it. First, the utility function U of the routing optimization problem is defined as:

[0053] U(QoS)=log(QoS)

[0054] Among them, QoS represents the service quality indicator, which can be replaced by the specific values ​​of average delay, packet loss rate, and throughput. Since reinforcement learning usually takes maximizing rewards as the goal of learning decisions, the service quality indicator of the intra-domain path after weighted combination is defined as follows:

[0055]

[0056] Among them, m,k ,lo m,k and m,k They represent the normalized average delay, packet loss rate, and throughput of the mth source-destination pair when the traffic takes the kth path. represents the service quality index after weighted combination when the traffic of the mth source-destination pair adopts the kth path, β 1 , β 2 and β3 De m,k ,lo m,k and m,k The corresponding weight parameters and satisfy β 1 +β 2 +β 3 =1, by changing the relative size, it can adapt to the service quality requirements of diversified businesses. By uniformly quantifying the path quality, the path with the largest utility function value is the optimal path in the domain. The reward r obtained by the agent after executing the action t is the sum of the utility functions of all traffic requests, defined as:

[0057]

[0058] in, Indicates that the traffic of the mth source-destination pair uses the optimal path k * The agent then takes the current state, the current action, the state after the action, and the reward as a tuple (s t ,a t ,r t ,s t+1 ) is stored in the experience replay pool. When the number of tuples stored in the experience replay pool reaches a threshold, random small batches are sampled from it for training the graph neural network. The training process adopts the existing dual network architecture consisting of a main Q network and a target Q network. The main Q network Q(s t ,a t ; θ) is a graph neural network used to predict the Q value of each action in the current state, s t represents the network status at the current time t, a t represents the action selected at the current time t, and θ is the network parameter of the main Q network at the current time. In each training iteration, the parameters of the main Q network are updated according to the experience playback samples; the target Q network Q(s t ,a t θ - ) has the same structure as the main Q network, but its network parameters θ - It is updated by periodic replication from the main Q network, which reduces the frequency of parameter updates compared to the main Q network, and can effectively avoid turbulence in the training process. During training, the loss function L that needs to be minimized is defined as the mean square error E between the target Q value and the Q value predicted by the main Q network:

[0059]

[0060] Among them, y t represents the prediction result of the target Q network. The update of the main Q network parameters is achieved through the gradient descent method, and the gradient of the loss function is expressed as:

[0061]

[0062] Among them, γ t It is used to measure the importance of current rewards and future rewards. The parameter update process is expressed as:

[0063]

[0064] in, It represents the gradient of the parameter θ of the loss function L(θ), and λ represents the learning rate. The trained model can be directly deployed on the local controller of each domain. By inputting the topology of the virtual domain, the traffic matrix, the available bandwidth of the intra-domain links and other information, the optimal intra-domain path can be output.

[0065] Step 5: Connect the optimal routes within the domain to obtain a complete cross-domain path that guarantees service quality requirements

[0066] Based on the virtual domain, the optimal path in the domain is seamlessly spliced, so that the service quality indicators of the combined cross-domain paths can be compared, and then the cross-domain complete path that guarantees the service quality requirements can be obtained. m,k' For example, the specific global path service quality indicator merging process is as follows:

[0067]

[0068] in, represents the j'th virtual domain through which the k'th feasible global path of the m'th source-destination pair passes, p * Indicates the optimal path selected in the virtual domain. and They represent the average delay, packet loss rate and throughput of the optimal path in the virtual domain predicted by the graph neural network model, del m,k' 、los m,k' and thr m,k' They represent the average delay, packet loss rate, and throughput of the cross-domain complete path respectively. The delay of the cross-domain complete path is the sum of the intra-domain path delays, the packet loss rate is related to the product of the intra-domain path packet loss rate, and the throughput is the minimum value of all intra-domain path throughputs. Finally, the utility function defined above is used to define the quality of the cross-domain complete path as the delay del normalized by the maximum value. m,k' , Packet loss rate los m,k' ,Thr m,k' A weighted combination of three indicators:

[0069] U' m,k' =-β 1 log(del m,k' )-β 2log(los m,k' )+β 3 log(thr m,k' )

[0070] Among them, U' m,k' Represents the global path R' m,k' The path quality, β 1 , β 2 and β 3 respectively m,k' 、los m,k' and thr m,k' The corresponding weight parameters and satisfy β 1 +β 2 +β 3 =1, the specific value is consistent with the parameter setting of the intra-domain routing. Cross-domain complete path quality value U' m,k' The larger the value is, the better the quality of the merged path is, and thus a cross-domain complete path that can best guarantee complex service quality requirements can be obtained.

[0071] The effects of the present invention are further illustrated by the following simulation and experimental verification:

[0072] Experimental results:

[0073] The present invention builds a low-orbit giant constellation network simulation platform through NS3, and establishes three different-sized Walker-Delta constellations, which contain 256, 768, and 1584 satellites respectively, with an orbital altitude of 550km and an orbital inclination of 55°. The forwarding rate of the intersatellite link is set to 300Mbps, and the total traffic intensity range of the network service is set to 1200Mbps to 1800Mbps. The Adam optimizer with a learning rate of 0.01 is used to train the routing model combining graph neural network and reinforcement learning, in which the link state dimension and path state dimension of the graph neural network are both set to 32, the number of readout units is set to 8, and the number of iterations of the message passing process is set to 3.

[0074] In addition, four advanced solutions are selected for comparison: the shortest path algorithm (SPF), the dynamic routing algorithm based on intersatellite link attributes (IADR), the deep Q network algorithm (DQN), and the collaborative multi-agent reinforcement learning algorithm (CoMARL). Among them, SPF uses the Dijkstra algorithm to obtain the shortest path, IADR solves the multi-attribute routing optimization problem by quantifying the link status, DQN uses the traditional convolutional neural network to predict the path quality, and CoMARL uses distributed collaborative multi-agents to obtain paths for different service quality requirements.

[0075] First, the routing model based on graph neural network and reinforcement learning was trained. The experimental results are as follows: Figure 3As the number of iterations increases, the reward obtained by the model through learning routing decisions gradually increases, and begins to converge when the number of iterations increases to about 120. At the same time, among all parameter combinations, setting the link state dimension and path state dimension of the graph neural network to 32 and the number of output units to 8 obtains the highest reward.

[0076] Next, the performance of the method proposed in the present invention was verified by various service quality indicators, among which the experimental results of average delay are as follows: Figure 4 As shown in Figure 2, the average delay performance of different routing algorithms under different traffic intensities and constellation networks of different sizes is demonstrated. Figure 4 Taking (a) as an example, as the total traffic intensity increases, the delays of all algorithms increase to varying degrees, among which SPF's delay always remains the highest, and then the delay performance of the IADR, DQN and CoMARL algorithms is significantly better than SPF, while the delay of the method proposed in the present invention always remains the lowest, and the increase is also the smallest. At the highest total traffic intensity (1800Mbps), the delay is reduced by 22.56% compared with SPF. Figure 4 (a)-(c) show that as the constellation network scale increases, the latency of all algorithms decreases, among which the algorithm proposed in the present invention always maintains the lowest latency. At the largest constellation scale (1584 satellites) and the highest total traffic intensity (1800Mbps), the latency is reduced by 38.43% compared with the SPF algorithm.

[0077] The above experimental results show that the present invention realizes routing that can efficiently guarantee complex service quality indicators in a giant constellation network by constructing a routing method based on graph neural network and reinforcement learning, and can fully cope with the challenges of large network scale and strong topology dynamics.

[0078] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. For those skilled in the art, the present invention may have various changes and variations. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A low-orbit giant constellation network routing method based on graph neural network and reinforcement learning, characterized in that: The following steps are involved: Multi-domain routing management architecture deployment: deploy a root controller and multiple local controllers; Divide the domains to which each satellite in the giant constellation network belongs; the root controller is used to control all local controllers; each local controller can monitor all satellites in the domain; Constructing virtual domains: The root controller adds border satellites of neighboring domains connected to each domain to construct virtual domains for cross-domain routing; Obtaining the global path: The root controller obtains the abstract topology based on the virtual domains where the source satellite node and the destination satellite node are located and the local controllers of the virtual domains passed through, and then obtains the feasible global path based on the abstract topology; The abstract topology consists of source satellite nodes, destination satellite nodes, and boundary satellite nodes of the virtual domains passed through; Determine the optimal route within the virtual domain: The local controller corresponding to each virtual domain involved in the feasible global path uses the route selection model that has been trained, inputs the network status information within the domain into the selection model, and selects the model to output the optimal intra-domain path; when training, the selection model aims to output the intra-domain path with the highest service quality index within the domain; Determine the optimal cross-domain route: The root controller seamlessly splices the optimal path in each virtual domain involved in the feasible global path to obtain a cross-domain complete path, and selects the cross-domain complete path with the highest global service quality index as the cross-domain optimal route output.

2. The method according to claim 1, characterized in that: The routing selection model is a model based on graph neural network and reinforcement learning; the intra-domain network status information includes status features at three levels: link, path, and domain to represent the virtual domain topology, traffic matrix, and available bandwidth of intra-domain links.

3. The method according to claim 2, characterized in that: Link state features include the available bandwidth and betweenness centrality of a single link; path state features include the traffic matrix and topology matrix; domain state features include the average available bandwidth of links within the domain.

4. The method according to claim 2 or 3, characterized in that: When training the routing model, spatial convolution is used to learn the cyclic dependencies between links and paths, and messages are aggregated, updated, and iterated interactively at the path and link levels to improve the model's adaptability to dynamic network environments.

5. The method according to claim 1, characterized in that The service quality index is obtained by a weighted combination of average delay, packet loss rate and throughput.

6. The method according to claim 5, characterized in that The specific method of weighted combination of the service quality indicator U through average delay, packet loss rate and throughput is as follows: U=-β1log(de)-β2log(lo)+β3log(th) Among them, de, lo and th are the average delay, packet loss rate and throughput respectively, β1, β2 and β3 are the weight parameters corresponding to de, lo and th respectively and satisfy β1+β2+β3=1.

7. The method according to claim 6, characterized in that The calculation method of average delay, packet loss rate and throughput in the global service quality indicators is as follows: Among them, del m,k' 、los m,k' and thr m,k' They represent the k'th cross-domain complete path R' of the mth source-destination pair. m,k' Average latency, packet loss rate, and throughput; represents the j'th virtual domain through which the k'th feasible global path of the m'th source-destination pair passes, p * Indicates the optimal path selected in the virtual domain. and Virtual domains The optimal path within p * The average latency, packet loss rate and throughput of 8. The method according to claim 1, characterized in that The domains to which each satellite in the giant constellation network belongs are divided based on latitude and longitude grids and geographic population density.

9. The method according to claim 1, characterized in that: If there are multiple boundary satellites in the neighborhood, the boundary satellite with the largest available bandwidth is preferentially added to construct the virtual domain.

10. The method according to claim 1, characterized in that The root controller communicates and exchanges data with local controllers to dynamically adjust service quality indicators and routing decisions within or across domains to improve the utilization of giant constellation network resources.

Images