Large-scale satellite network routing simulation and optimization method based on digital twin

Abstract

This invention discloses a large-scale satellite network routing simulation and optimization method based on digital twins, primarily addressing the problem that existing large-scale satellite network routing methods rely solely on predicting future traffic based on the current network state, leading to delayed routing decisions. The implementation scheme involves: measuring the topology and link load information of a real low-Earth orbit (LEO) satellite network using a medium-Earth orbit (MEO) satellite; collecting the measurement information and using it to construct state information in a twin space; adopting a deep Digital Quad-Quality Network (DQN) model in the twin space, using inter-satellite topology and link load as the state space, and link load, latency, and hop count as optimization objectives, calculating and verifying the optimal path for all node pairs in the satellite network; synchronizing the verified DQN network model to the real network, and then using the MEO satellite to output the optimal path for the current real LEO satellite network traffic based on this synchronized network model. This invention effectively avoids the impact of erroneous paths on the real network, improves the overall network throughput, and can be used for routing decisions.

Description

Technical Field

[0001] This invention belongs to the field of satellite communication technology, and specifically relates to a method for large-scale satellite network routing simulation and optimization, which can be used for routing decisions. Background Technology

[0002] Large-scale low-Earth orbit (LEO) satellite networks represent a new generation of internet revolution technology. LEO satellites operate at lower altitudes, resulting in lower communication latency and higher data transmission rates, bringing more possibilities to the future space-based information industry.

[0003] Routing technology has always been a key technology for realizing satellite network communication. For low-Earth orbit (LEO) satellite networks, existing inter-satellite routing algorithms face challenges: on the one hand, due to the high density and dynamic topology changes of satellite networks, frequent inter-satellite link disconnections and handovers occur; on the other hand, due to the uneven distribution of global traffic, some inter-satellite links become congested, leading to resource utilization failing to meet service demands. For these problems caused by traditional satellites, the DRL algorithm, due to its decentralization and strong adaptability, has been widely used in areas including network routing. However, the DRL method relies on collected data to perceive the environmental state; therefore, missing, outdated, and inaccurate state information data can make the DRL method difficult to use effectively.

[0004] Digital twin (DT) technology is a novel approach that simulates real networks by constructing a virtual model of the physical world. It is a promising technology with wide applications across various fields. By creating a virtual copy of the real network environment, DT can accurately simulate the operation of algorithms in actual networks, helping to optimize algorithm design and parameter selection. Currently, research combining digital twin technology with routing algorithms is still in its early stages, with relatively few relevant references, and the algorithm models used are relatively simple, making it difficult to handle the routing requirements of complex network conditions such as low-Earth orbit satellites.

[0005] Patent document with application number "CN202310520847.1" discloses "An SDN Routing Optimization Method Based on Digital Twins," which uses an SDN controller as a proxy for the routing algorithm. Based on historical network state information collected in the twin space, it asynchronously updates the neural network parameters of the routing model in the real network, thereby planning a suitable forwarding path for each traffic. However, since the digital twin construction in this method is mainly aimed at terrestrial communication networks with fixed topologies, it is not suitable for low-Earth orbit satellite networks with frequent topology updates. Furthermore, the algorithm design lacks verification of the path correctness in the twin system, which may lead to incorrect paths affecting the real network.

[0006] Patent document with application number "CN202310344143.3" discloses "A Satellite QoS Routing Method and Apparatus Based on Digital Twin". It obtains the network topology map at the current moment through digital twin technology, avoids inter-satellite links with handover and congestion risks, and uses an ant colony algorithm to calculate the optimal path in the twin system and add it to the routing table. Finally, the routing table is uploaded to the real satellite network, improving the transmission stability and routing efficiency of the satellite network for service transmission. However, the ant colony algorithm, as a heuristic algorithm, suffers from long solution time and is prone to getting trapped in local optima when solving routing problems. It cannot fully meet the needs of predicting and modeling future network environments, as well as the real-time interaction between the twin system and the real network. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of the prior art by providing a large-scale satellite network routing simulation and optimization method based on digital twins, so as to effectively avoid the impact of erroneous paths on the real network and improve the overall network throughput.

[0008] The technical approach to achieving the objective of this invention is to combine digital twin technology with low-Earth orbit satellite network scenarios, create a replica of the real satellite network in the twin space, use the DQN algorithm to select the optimal route and verify it in the twin space, and then use the real low-Earth orbit satellite network to select the optimal path for traffic transmission based on the verified DQN model, so as to effectively avoid the impact of incorrect paths on the real network and improve the throughput of the entire network.

[0009] Based on the above ideas, the technical solution of the present invention includes the following steps:

[0010] 1) Obtain the topology and link load status information of the real low-Earth orbit satellite network through medium-Earth orbit satellites, and use this network status information to construct a twin space of the real low-Earth orbit satellite network;

[0011] 2) In the twin space, a network model using the DQN algorithm is adopted, and the optimal path is selected for all source and destination nodes of the low-Earth orbit satellite network based on the three indicators of hop count, latency, and load.

[0012] 3) The selected optimal path is verified in twin space based on the comprehensive evaluation index CRE, and the network model of the DQN algorithm is updated. Then the verified DQN network model is synchronized to the medium-Earth orbit satellite.

[0013] 4) The medium-Earth orbit satellite receives the network model from the twin space, outputs the optimal path for the current traffic based on the network model and generates a routing table, and sends the generated routing table to the real low-Earth orbit satellite network, and performs routing through the optimal path recorded in the routing table.

[0014] Compared with the prior art, the present invention has the following advantages:

[0015] Firstly, this invention adopts a network model based on the DQN algorithm in twin space, using hop count, latency, and load as the three indicators for selecting the optimal path for all source and destination nodes of the low-Earth orbit satellite network. This avoids the problem of traditional heuristic algorithms taking a long time to solve routing problems and easily getting trapped in local optima.

[0016] Secondly, by adopting the comprehensive evaluation index CRE to verify the selected optimal path in twin space, this invention can avoid the impact of erroneous paths on the real network and improve the overall network throughput. Attached Figure Description

[0017] Figure 1 This is a usage scenario diagram of the present invention;

[0018] Figure 2 This is a flowchart illustrating the implementation of the present invention. Detailed Implementation

[0019] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] Reference Figure 1 This example demonstrates a use case involving both a physical network and a twin space. The physical network comprises a network of medium-Earth orbit (MEO) satellites and a network of real low-Earth orbit (LEO) satellites. The twin space consists of a Deep Q-learning network model (DQN) and a twin LEO satellite network. The MEO satellite acts as the network control center, responsible for collecting state information from the real LEO satellite network, updating routing tables, and distributing the data. The real LEO satellite network is responsible for routing and forwarding incoming service flows. The DQN network model is responsible for training and decision-making regarding routes. The twin LEO satellite network, composed of network state information collected from the MEO satellites, serves as the traffic forwarding network within the twin space, verifying the routing paths output by the DQN network model and providing feedback.

[0021] Reference Figure 2 In the above scenario, this example performs large-scale satellite network routing simulation and optimization based on digital twins. The implementation steps are as follows:

[0022] Step 1: Medium-Earth Orbit (MEO) satellites acquire status information of the actual Low-Earth Orbit (LEO) satellite network.

[0023] 1.1) Based on the visibility relationship with the real low-Earth orbit satellite network, the medium-Earth orbit satellite obtains the topology information of the real low-Earth orbit satellite network. This topology information is constructed in the form of an adjacency matrix with a size of ncnt×4. Here, ncnt represents the number of real low-Earth orbit satellite nodes that have a visibility relationship with the medium-Earth orbit satellite, and 4 represents the number of links for each real low-Earth orbit satellite node. Each bit of the adjacency matrix stores the peer node number of the corresponding link of the current low-Earth orbit satellite node. If the peer node of the corresponding link does not exist, -1 is stored.

[0024] 1.2) Real LEO satellites collect link load information of their direct links. This link load information is constructed in the form of an adjacency matrix with a size of ncnt×4. Here, ncnt represents the number of real LEO satellite nodes that have a visibility relationship with the current MEO satellite, and 4 represents the number of links of each real LEO satellite node. Each bit of the adjacency matrix stores the number of existing service data packets on the link corresponding to the current LEO satellite node. If the peer node of the corresponding link does not exist, -1 is stored, and the link load information is sent to the cluster head satellite.

[0025] 1.3) The cluster head satellite uses Dijkstra's algorithm to select the low-Earth orbit satellite node with the minimum total hop count among all low-Earth orbit satellite nodes in the real low-Earth orbit satellite network, and which has a visibility relationship with the current medium-Earth orbit satellite, as the cluster head satellite node. It also integrates the link load information sent by the other real low-Earth orbit satellites and reports it to the medium-Earth orbit satellite as a relay node.

[0026] The topology information and link load information are the status information of the actual low-Earth orbit satellite network obtained by the medium-Earth orbit satellite.

[0027] Step 2: Construct inter-satellite links and traffic distribution in the twin space of a real low-Earth orbit satellite network using the status information obtained from medium-Earth orbit satellites.

[0028] 2.1) The Deep Q-Learning Network (DQN) model is selected, which includes a training network and a target network. It is used to select agent nodes as agents to interact with the twin space through a centralized training mode.

[0029] 2.2) Using a medium-Earth orbit satellite as the agent in the DQN network model in twin space, responsible for centralized training of the routing algorithm;

[0030] 2.3) Construct inter-satellite links in twin space based on topology information obtained from medium-Earth orbit satellites:

[0031]

[0032] Where conn ijThis represents the topological connection between current node i and node j. The values ​​of i and j range from 0 to n, where n represents the actual number of low-Earth orbit satellite nodes. If node i and node j are connected, then conn... ij =1, otherwise conn ij =-1;

[0033] 2.4) Construct the traffic distribution flow in twin space based on the link load information obtained from medium-Earth orbit satellites:

[0034]

[0035] Among them l ij This represents the link load information between the current node i and node j. If node i and node j are connected, then l ij It equals the current link load size; otherwise, l ij =-1.

[0036] Step 3: In the twin space, adopt the DQN network model and select the optimal path for all source and destination nodes of the low-Earth orbit satellite network based on the three indicators of hop count, latency, and load.

[0037] 3.1) Initialize the parameters of the training and target networks of the DQN network model in the twin space, including the state space s of the network environment at the current time step. t Action set a t Rewards r t The exploration rate ε is set, and the number of training rounds is set to episode = 1, 2, ..., n. In this example, n is 1000.

[0038] 3.1.1) Initialize the state space s of the network environment at the current moment. t for:

[0039]

[0040] Among them: hc ij Represents the number of hops in the shortest path from source node i to destination node j; load ij This represents the link load between the source node i and the destination node. L represents the transmission delay from source node i to destination node j. ij λ represents the path from node i to node j; λ represents the arrival rate of the node, i.e., the number of data packets received by the current node per unit time; μ represents the service rate of the node, i.e., the number of data packets processed by the current node per unit time; L represents the data packet size, which is a fixed value of 1500; W represents the service queue length, which is a fixed value of 1024.

[0041] 3.1.2) Initialize the action set a of the network environment at the current moment.t for:

[0042]

[0043] Where: path ij This represents the path from source node i to destination node j, which uses the K-shortest path algorithm (KSP) to calculate k pre-selected paths [a0...a ......a......a......a...... i ...a k And randomly select a pre-selected path a from k paths. i as path ij The initialization result;

[0044] 3.1.3) Initialize the reward r of the network environment at the current moment. t for:

[0045] r t =αr p +βr hc +ηr load

[0046] in: The normalized reward value representing the delay p, p cd The transmission delay from node c to node d is represented by mp. cd p represents the maximum transmission delay from node c to node d. ca d represents the transmission delay from node c to node a, where c represents the source node, a represents the destination node, and d represents the next-hop node from the source node to the destination node.

[0047] The normalized reward value representing the number of jumps hc, hc ca len represents the number of hops from node c to node a. ca This represents the total length of the link from node c to node a;

[0048] The normalized reward value representing the link load (load). ca This represents the sum of the link loads along the path from node c to node a; the coefficients α, β, and η are all 1 / 3.

[0049] 3.1.4) Initialize the network exploration rate ε at the current moment:

[0050] ε=ε decay ×ε

[0051] Where ε decay This represents the attenuation factor, which is a fixed value of 0.995 in this example, and the initial value of ε is set to 1.

[0052] 3.2) Obtain the state space s at the current time. t The data is input into the training network of DQN, and the medium-Earth orbit satellite acts as the agent, using a set random number as the state space s. t Select action set a t :

[0053] If the set random number is less than the exploration rate ε, the agent, according to the K-shortest path algorithm (KSP), randomly selects a path from k pre-selected paths for each node combination (i,j) to form an action set a. t And obtain the action set a t The corresponding reward r t ;

[0054] Otherwise, calculate the Q-value of the DQN network output in each pre-selected path, and select the pre-selected path that maximizes the Q-value of the DQN network output from the k pre-selected paths as the action set a. t The action set a is obtained from the output of the trained network. t The corresponding reward r t The formula for calculating the Q value is:

[0055] Q(s t ,a t )←Q(s t ,a t )+α[r t +γmaxQ-Q(s t ,a t )]

[0056] In the formula, α represents the learning rate, used to control the rate at which the Q-value is updated; in this example, it is taken as a fixed value of 0.1. γ represents the discount factor, used to control the degree of influence of the reward; in this example, it is taken as a fixed value of 0.9. maxQ represents the maximum Q-value in the historical record, and r t Represents the current moment based on the corresponding action set a t The rewards received;

[0057] 3.3) Based on the action set a output by the training network t Select paths in the twin network for all node pairs consisting of source and destination nodes;

[0058] 3.4) In the twin network, the selected path is routed to obtain the state space s at the next time step. t ':

[0059] 3.4.1) Obtain the set of all current business flows [(src1,dest1)...(src1,dest1)] based on the flow distribution in the twin space. i ,dest i )…(src m,dest m )], where (src i ,dest i ) represents the source node and destination node of the i-th service flow, and m represents the total number of service flows in the current network;

[0060] 3.4.2) Based on action set a t The corresponding path For each node in the business flow set, combine (src) i ,dest i Perform routing;

[0061] 3.4.3) After routing is completed, collect the transmission delay, hop count, and link load information of all node combinations (i,j) to construct the state space s for the next time step. t+1 The set (s) generated in this round is obtained. t ,a t ,r t ,s t+1 );

[0062] 3.5) The set (s) generated in this round t ,a t ,r t ,s t+1 Add the experience replay pool to the DQN network and update the exploration rate ε = ε decay ×ε;

[0063] 3.6) The training network randomly selects one sample from the experience replay pool to update the target network parameters, in order to provide a relatively stable Q value and avoid overestimation in the DQN algorithm;

[0064] 3.7) Repeat steps 3.2)-3.6) until the training rounds end (episode = 1, 2, ..., n). In this example, n is 1000. Then execute step 3.8).

[0065] 3.8) Collect the transmission delay, hop count, and link load of all node combinations (i,j) in the Siamese network at the end of each training round, and use them as the state space s t The input is fed into the DQN target network, and the target network outputs a set of actions, a. t That is, the optimal path that consists of all source and destination nodes forming a node pair.

[0066] Step 4: Validate the selected optimal path in twin space using the comprehensive evaluation index CRE as a benchmark, update the network model of the DQN algorithm, and then synchronize the validated DQN network model to the medium-Earth orbit satellite.

[0067] 4.1) Based on the state set s corresponding to the currently selected optimal patht Calculate the overall routing performance (CRE):

[0068]

[0069] Where, d ij 'Represents the number of hops from node i to node j calculated using the selected optimal path, l ij 'Represents the sum of link loads calculated from node i to node j along the selected optimal path, p ij 'Represents the transmission delay calculated from node i to node j using the selected optimal path, where i represents the source node and j represents the destination node;

[0070] 4.2) Based on the current network state set s t Obtain the next time-step state set s from the experience replay pool of the DQN network. t+1 Calculate the overall routing efficiency (CRE).

[0071] 4.3) s t+1 As a new set of states s t Repeat step 4.2) N times, where N is a fixed value of 10 in this example;

[0072] 4.4) Calculate the average of the CRE results in N tests and compare it with the CRE value obtained in step 4.1) to verify the selected optimal path;

[0073] If the CRE value obtained in step 4.1) is greater than the average value, then the path passes the verification and step 4.6) is executed;

[0074] Otherwise, based on the current network state set s t The target network re-outputs the action set a. t 'and the corresponding reward r t Then, the action set a is run in the twin network. t 'The corresponding path, obtain the next time-space state set s' t+1 ', and the set (s) generated this time t ',a t ',r t ',s t+1 ') Add it to the experience replay pool of the DQN network and execute step 4.5);

[0075] 4.5) s t+1 'The set of states corresponding to the optimal path selected in the next round is used to repeat steps 4.1)-4.4);

[0076] 4.6) Update the DQN network model to the validated network model and synchronize it to the medium-Earth orbit satellite.

[0077] Step 5: Based on the DQN network model validated by the comprehensive evaluation index CRE, generate the optimal path for traffic output of the current real low-Earth orbit satellite network and generate a routing table.

[0078] 5.1) The real low-Earth orbit satellite network uses Dijkstra's algorithm of the static routing protocol to calculate the initial routing table for each low-Earth orbit satellite node;

[0079] 5.2) After the service flow is injected into the real low-Earth orbit (LEO) satellite network, the LEO satellite starts a timed thread to monitor the link load level of its own satellite node. This link load level is determined according to the ratio of the number of existing service data packets in the inter-satellite link to the maximum capacity of the inter-satellite link. A ratio in the range [0, 0.25] is designated as Level 1, [0.25, 0.5] as Level 2, [0.5, 0.75] as Level 3, and so on.

[0080] [0.75,1] is set to level 4, and the maximum number of inter-satellite links to be carried in this example is 1024;

[0081] If the link load level does not change, the low-Earth orbit satellites make routing decisions based on the initial routing table;

[0082] Otherwise, proceed to step 5.3).

[0083] 5.3) The medium-Earth orbit satellite obtains the state space of the current real low-Earth orbit satellite network, inputs it into the DQN network model verified by the comprehensive evaluation index CRE, outputs a set of actions, and selects the optimal path for all node combinations based on the set of actions to form the whole network routing table of the real low-Earth orbit satellite network.

[0084] Step 6: The routing table composed of medium-Earth orbit satellites is sent to the actual low-Earth orbit satellite network to realize the routing update.

[0085] 6.1) Based on Dijkstra's algorithm, select the node with the smallest total hop count among all low-Earth orbit satellite nodes in the real low-Earth orbit satellite network as the relay satellite node;

[0086] 6.2) The medium-Earth orbit satellites send the entire network routing table of the actual low-Earth orbit satellite network to the relay satellite nodes;

[0087] 6.3) After receiving the routing table, the relay satellite node transmits it to other real low-Earth orbit satellite nodes in a flooding manner, thereby updating the routing table across the entire real low-Earth orbit satellite network.

[0088] The above description is merely a specific example of the present invention and does not constitute any limitation on the present invention. Obviously, those skilled in the art, after understanding the content and principles of the present invention, may make various modifications and changes in form and details without departing from the principles and structure of the present invention. However, these modifications and changes based on the ideas of the present invention are still within the scope of protection of the claims of the present invention.

[0089] It should be noted that the step numbers in the specification and claims of this invention are only for the purpose of clearly describing the embodiments of this invention and facilitating understanding, and their order is not limited.

Claims

1. A method for large-scale satellite network routing simulation and optimization based on digital twins, characterized in that, Includes the following steps: 1) Obtain the topology and link load status information of the real low-Earth orbit satellite network through medium-Earth orbit satellites, and use the status information of this network to construct the inter-satellite links and traffic distribution of the twin space of the real low-Earth orbit satellite network; 2) In the twin space, a network model using the DQN algorithm is adopted. The optimal path is selected for all source and destination nodes of the low-Earth orbit satellite network based on three metrics: hop count, latency, and load. The implementation steps include the following: 2a) Initialize the parameters of the training network and the target network in the deep Q-learning model DQN in the twin space, including the set of network environment states s at the current time. t Action set a t Rewards r t Let ε be the exploration rate, and let episode = 1, 2, ..., n be the number of training iterations. t This includes transmission latency, hop count, and link load, where n is the maximum number of training rounds set. 2b) Set the state set s t The exploration rate ε is input into the DQN model, and the medium-Earth orbit satellite, acting as the agent of the DQN model, generates the current input state set s based on a set random number. t Select action set a t : If the set random number is less than the exploration rate ε, the agent selects a random action as the action set a. t ; Otherwise, select the action with the largest Q-value among all outputs of the DQN network as the action set a. t And output by the training network; 2c) Based on the action set a output by the trained network t Select paths in the twin network for all node pairs consisting of source and destination nodes; 2d) Run the selected path in the Siamese network and obtain the next time-step state set s. t ', and the set (s) generated in this round t ,a t ,r t ,s t ') Add it to the experience replay pool in the DQN network; 2e) The training network randomly selects one sample from the experience replay pool to update the target network parameters and provides a relatively stable Q value to avoid overestimation in the DQN algorithm; 2f) Repeat steps 2b)-2e) until the training rounds end (episode = 1, 2, ..., n), then execute step 2g); 2g) Collect the transmission delay, hop count, and link load of the twin network at the end of each training round, and use them as the state set s t The input is fed into the DQN target network, and the target network outputs a set of actions, a. t That is, the optimal path consisting of all source and destination nodes forming a node pair; 3) The selected optimal path is validated in twin space using the Comprehensive Evaluation Representation (CRE) as a benchmark, and the network model of the DQN algorithm is updated simultaneously. The network model is then synchronized to the medium-Earth orbit satellite. The CRE is calculated using the following formula: Where, d ij 'Represents the number of hops from node i to node j calculated using the selected optimal path, l ij 'Represents the sum of link loads calculated from node i to node j along the selected optimal path, p ij 'Represents the transmission delay calculated from node i to node j using the selected optimal path, where i represents the source node and j represents the destination node; 4) The medium-Earth orbit satellite receives the network model from the twin space, outputs the optimal path for the current traffic based on the network model and generates a routing table, and sends the generated routing table to the real low-Earth orbit satellite network, and performs routing through the optimal path recorded in the routing table.

2. The method according to claim 1, characterized in that, Step 1) involves obtaining the topology and link load status information of a real low-Earth orbit satellite network using medium-Earth orbit satellites. The steps include the following: 1a) Medium-Earth orbit satellites obtain the topology information of real low-Earth orbit satellite networks based on their visibility relationship with the real low-Earth orbit satellite networks; 1b) The real low-Earth orbit satellites collect the link load information of their own direct links, that is, the number of service data packets in the current link, and send it to the cluster head satellite. The cluster head satellite integrates the link load information sent by the other real low-Earth orbit satellites and reports it to the medium-Earth orbit satellite as a relay node.

3. The method according to claim 1, characterized in that, Step 1) Construct inter-satellite links and traffic distribution in the twin space of a real low-Earth orbit satellite network using network state information, as follows: A deep Q-learning (DQN) network model is selected, which includes a training network and a target network, to select agent nodes as agents to interact with the twin space through a centralized training mode. Using medium-Earth orbit satellites as agents in the DQN network model within the twin space, and using topology information as inter-satellite links in the twin space, and link load information as traffic distribution in the twin space, a twin space for the low-Earth orbit satellite network is obtained.

4. The method according to claim 1, characterized in that, Step 2) describes the network model in twin space that adopts the DQN algorithm, using hop count, latency, and load as the three indicators for selecting the optimal path for all source and destination nodes of the low-Earth orbit satellite network.

5. The method according to claim 1, characterized in that, Step 3) uses the comprehensive evaluation index CRE as a benchmark to verify the selected optimal path in the twin space. The implementation steps include the following: 3a) Based on the state set s corresponding to the currently selected optimal path t Calculate the comprehensive evaluation index CRE; 3b) Based on the current network state set s t Obtain the next time-step state set s from the experience replay pool of the DQN network. t 'Calculate the comprehensive evaluation index CRE and s t 'As a new set of states s t Repeat step 3b) N times; 3c) Compare the average of the CRE results from the N tests with the CRE value obtained in step 3a); If the CRE value obtained in step 3a) is greater than the average value, then the path passes the validation. Otherwise, based on the current network state set s t The target network re-outputs the action set a. t Then run the action set a in the twin network. t The corresponding path to obtain feedback reward r t and the set of states s at the next moment t ', and the set (s) generated this time t ,a t ,r t ,s t ')Put it into the experience replay pool in the DQN network and execute 3d); 3d) will s t 'The set of states s corresponding to the optimal path selected in the next round' t Repeat steps 3a)-3c).

6. The method according to claim 1, characterized in that, The network model updated in step 3) is the DQN network model verified by the medium-Earth orbit satellite in twin space through the comprehensive evaluation index CRE, which serves as the network model for inputting service flows into the real low-Earth orbit satellite.

7. The method according to claim 1, characterized in that, Step 4) Based on the DQN network model validated through the comprehensive evaluation metric CRE, output the optimal path for the current traffic and generate a routing table. The steps include the following: 4a) Real low-Earth orbit satellite networks use Dijkstra's algorithm of static routing protocols to calculate the initial routing table for each low-Earth orbit satellite node; 4b) After the service flow is injected into the real low-Earth orbit satellite network, the low-Earth orbit satellite monitors the link load level of its own satellite node in real time; If the link load level does not change, the low-Earth orbit satellites make routing decisions based on the initial routing table; Otherwise, proceed to step 4c); 4c) Medium-Earth orbit satellites use the topology and link load information currently collected from the actual low-Earth orbit satellite network as a state set s t The input is fed into the DQN network model validated by the comprehensive evaluation metric CRE, and the output action set a is generated. t And based on the action set a t Generate the complete routing table for a real low-Earth orbit satellite network.

8. The method according to claim 1, characterized in that, Step 4) involves sending the generated routing table to the actual low-Earth orbit satellite network. The steps include the following: 4e) According to Dijkstra's algorithm, the node with the smallest total hop count among all low-Earth orbit satellite nodes in the real low-Earth orbit satellite network is selected as the relay satellite node. 4f) The medium-Earth orbit satellites send the entire network routing table of the actual low-Earth orbit satellite network to the relay satellite nodes; After receiving the routing table, the 4g relay satellite node transmits it to other real low-Earth orbit satellite nodes in a flooding manner, thereby updating the routing table across the entire real low-Earth orbit satellite network.

Images