Low earth orbit satellite network mac layer scheduling system based on large language model multi-agent

Abstract

This invention provides a low-Earth orbit (LEO) satellite network MAC layer scheduling system based on a large language model and multiple agents. The system includes a policy selection agent deployed on the satellite side and scenario analysis, policy generation, and policy verification agents deployed on the ground station side. The satellite-side agent is responsible for selecting and invoking scheduling policies based on the current communication state, while the three ground-side agents are responsible for analyzing new communication scenarios, generating scheduling policies, and performing policy simulation verification, thus forming a collaborative reasoning system driven by a large language model. This invention offers the following advantages: by constructing a satellite-ground collaborative multi-agent system, it fully utilizes the semantic understanding and code generation capabilities of the large language model, enabling continuous evolution and adaptive updating of scheduling policies, and significantly improving the resource scheduling efficiency and operational performance of the LEO satellite network MAC layer in complex communication scenarios.

Description

Technical Field

[0001] This invention relates to the fields of communication networks and artificial intelligence, and in particular to a MAC layer scheduling system for low-Earth orbit satellite networks based on a large language model multi-agent system. Background Technology

[0002] Low Earth Orbit (LEO) satellite communication systems are gradually becoming an important component of next-generation global communication networks due to their advantages such as low latency, high bandwidth, and wide coverage. However, compared with geostationary orbit (GEO) satellite systems, LEO satellites exhibit characteristics such as high-speed motion, short satellite-to-ground visibility windows, frequent changes in network topology, and highly dynamic channel conditions. These characteristics present the Media Access Control (MAC) layer with more complex technical challenges in resource scheduling, access control, and quality of service (QoS) assurance. Especially in scenarios with multiple users accessing concurrently and dynamically changing service demands, the MAC layer scheduling strategy's responsiveness to environmental changes directly impacts the overall communication performance of the system.

[0003] In early satellite communication systems, MAC layer scheduling mechanisms often employed rule-based static or semi-static access methods, such as channel access methods based on random access or reservation mechanisms, fixed time slot allocation mechanisms, and priority-based round-robin scheduling schemes. These methods were suitable for scenarios with small network scales, relatively stable topologies, or relatively simple service types. However, when applied to low-Earth orbit satellite constellation systems, due to the high-speed orbit of satellites, frequent switching of serving satellites, and dense access by ground terminals, traditional scheduling mechanisms struggled to adjust scheduling parameters in a timely manner to adapt to rapidly changing network conditions. This was particularly problematic in scenarios requiring guaranteed quality of service, concurrent access by multiple users, and significant fluctuations in channel resources, often exhibiting issues such as delayed response, insufficient robustness, and low scheduling efficiency.

[0004] To overcome the insufficient adaptability of traditional rule-driven MAC protocols in dynamic environments, the field of communication systems has begun to explore intelligent optimization methods with environmental awareness and adaptive capabilities. Among these, intelligent MAC design, represented by deep reinforcement learning, has gradually gained attention. These methods typically model the MAC layer scheduling process as an interaction between an agent and the communication environment, achieving dynamic adjustment of scheduling behavior through policy learning. Although related research has achieved certain performance improvements in some scenarios, reinforcement learning-based methods generally suffer from limitations such as complex training processes, sensitivity to environmental feedback, insufficient model interpretability, and difficulty in stable operation in highly dynamic, strongly coupled network environments. Therefore, they cannot directly meet the comprehensive requirements of low-Earth orbit satellite communication systems for real-time performance, reliability, and engineering feasibility.

[0005] In recent years, with the rapid development of large language models and multi-agent systems, intelligent decision-making frameworks based on semantic understanding, task decomposition, and collaborative reasoning have gradually become a research hotspot. Multi-agent systems, by dividing complex decision-making tasks into multiple functional units and processing them collaboratively, have shown new potential in dealing with dynamic environments and complex decision-making problems. This trend provides new technical possibilities for the intelligent generation and dynamic updating of MAC layer scheduling strategies for low-Earth orbit satellite networks. However, how to effectively introduce the large language model multi-agent mechanism into satellite communication scheduling scenarios while meeting onboard resource constraints and system real-time requirements still requires further research. Summary of the Invention

[0006] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a low-Earth orbit satellite network MAC layer scheduling system based on a large language model and multiple agents, which solves the problems of poor adaptability of fixed strategies, rigid generation methods, and difficulty in deploying intelligent models in limited platforms in the existing low-Earth orbit satellite MAC layer scheduling.

[0007] To achieve the above and other related objectives, the present invention provides the following technical solution:

[0008] A multi-agent scheduling system for low-Earth orbit satellite networks based on a large language model includes a policy selection agent deployed on the satellite side and a scenario analysis agent, a policy generation agent, and a policy verification agent deployed on the ground station side. The policy selection agent selects the scheduling policy that best matches the current communication scenario from a local policy library based on the characteristics of the current communication state and executes the MAC layer scheduling by calling the corresponding policy code. The scenario analysis agent analyzes the operation logs uploaded from the satellite side, identifies new communication scenarios not yet covered by the existing policy library, and extracts and structures the key features of the new communication scenarios.

[0009] The strategy generation agent is used to generate candidate scheduling strategy code that meets the requirements of the target communication scenario by calling a large language model based on the scenario description data output by the scenario analysis agent. The strategy verification agent is used to perform functional verification and basic performance verification on the candidate scheduling strategy code output by the strategy generation agent, and determine whether the candidate scheduling strategy code is executable and effective in the target communication scenario based on the verification results.

[0010] In one embodiment of the present invention, the strategy selection agent adopts a lightweight language model, which is used to match, sort and select the current communication state features and the applicable scenario descriptions of each candidate scheduling strategy in the strategy library.

[0011] In one embodiment of the present invention, the strategy selection agent is used to select the scheduling strategy that best matches the current communication scenario from the local strategy library based on the current communication state characteristics, and call the corresponding strategy code to execute MAC layer scheduling, including: obtaining the scenario feature vector and candidate scheduling strategy set of the current communication window, and inputting the scenario feature vector and candidate scheduling strategy set of the current communication window into a lightweight language model; calculating the matching score of each candidate scheduling strategy through the lightweight language model, obtaining the target scheduling strategy identifier that best matches the current communication scenario based on the calculation result, and calling the corresponding strategy code to execute MAC layer scheduling.

[0012] In one embodiment of the present invention, before obtaining the scene feature vector and candidate scheduling strategy set of the current communication window, and inputting the scene feature vector and candidate scheduling strategy set of the current communication window into the lightweight language model, the method includes: continuously collecting running data for multiple scheduling cycles within the current communication window, and performing in-window statistical processing on the running data to obtain the scene feature vector of the current communication window; constructing the scene feature vector of the current communication window into a scene observation description of the current communication window, and obtaining each candidate scheduling strategy from the local strategy library according to the scene observation description to obtain the candidate scheduling strategy set.

[0013] In one embodiment of the present invention, the scene analysis agent is used to analyze the operation logs uploaded by the satellite side, identify new communication scenarios not yet covered by the existing policy library, and extract and structure the key features of the new communication scenarios, including: segmenting the operation logs based on preset scene change detection rules, obtaining candidate scene segments with stable communication states based on the segmentation results; extracting scene features that can characterize the current communication state from each candidate scene segment; comparing the scene features with the applicable scene descriptions corresponding to each stored policy in the existing policy library for coverage, and confirming whether the candidate scene segments have been covered by the existing policy library based on the comparison results.

[0014] In one embodiment of the present invention, the step of comparing the scene features with the applicable scene descriptions corresponding to each stored strategy in the existing strategy library for coverage, and determining whether the candidate scene segment has been covered by the existing strategy library based on the comparison result, includes: if the candidate scene segment does not meet the preset coverage conditions with any scenario description corresponding to any strategy in the existing strategy library, or if the performance index of the currently executed strategy under the candidate scene segment is lower than the preset requirement, then the candidate scene segment is determined to be an uncovered new communication scene.

[0015] In one embodiment of the present invention, the strategy generation agent is used to generate candidate scheduling strategy code that meets the requirements of the target communication scenario by calling a large language model based on the scenario description data output by the scenario analysis agent. The process includes: formatting the scenario description data and constructing a strategy generation request for the large language model; inputting the strategy generation request into the large language model, and having the large language model generate candidate scheduling strategy code based on the scenario description data, code template, and interface constraints of the target communication scenario.

[0016] In one embodiment of the present invention, the policy verification agent is used to perform functional verification and basic performance verification on the candidate scheduling policy code output by the policy generation agent, and determine whether the candidate scheduling policy code is executable and effective in the target communication scenario based on the verification results. This includes: deploying a typical scenario corresponding to the target communication scenario in a simulation platform, and loading the candidate scheduling policy code into a preset MAC layer scheduling interface to form a verification environment; performing functional verification and basic performance verification on the candidate scheduling policy code, and determining whether the candidate scheduling policy code is executable and effective in the target communication scenario based on the verification results.

[0017] In one embodiment of the present invention, the step of performing functional verification and basic performance verification on the candidate scheduling strategy code, and determining whether the candidate scheduling strategy code is executable and effective in the target communication scenario based on the verification results, includes: performing functional verification on the candidate scheduling strategy code; after the candidate scheduling strategy code passes the functional verification, performing basic performance verification on the candidate scheduling strategy code; wherein, the basic performance verification includes evaluating one or more of the following indicators of the candidate scheduling strategy code in the target communication scenario: throughput, latency, packet delivery rate, deadline fulfillment rate, queue backlog, and high-priority service guarantee capability.

[0018] When the candidate scheduling strategy code passes both functional verification and basic performance verification, it is determined that the candidate scheduling strategy code has executability and scheduling validity, and is submitted to the subsequent strategy storage or synchronization module; when the candidate scheduling strategy code fails verification, a verification result report is output and fed back to the strategy generation agent, so as to be used for subsequent strategy correction and iterative generation.

[0019] As described above, the MAC layer scheduling system for low-Earth orbit satellite networks based on a large language model multi-agent system of the present invention has the following beneficial effects:

[0020] First, this invention introduces the generation and reasoning capabilities of a large language model based on a multi-agent collaborative mechanism, and constructs a scheduling policy update system with self-evolution capabilities. This system can continuously expand the coverage of the policy library to new communication scenarios, thereby significantly improving the system's adaptability to complex and ever-changing scheduling environments and overcoming the problem of insufficient generalization capability of traditional fixed policies.

[0021] Secondly, by deploying the scheduling strategy generation and verification tasks on nodes with sufficient ground computing resources, this invention effectively avoids the limitation of limited spaceborne platform resources, enabling the system to achieve dynamic strategy updates without relying on complex satellite-side models, while ensuring a coordinated balance between strategy generation quality and execution efficiency.

[0022] Furthermore, this invention adopts a satellite-ground collaborative structural design, utilizing the visible window within the satellite's operational cycle to complete data interaction and strategy synchronization, forming a closed-loop generation mechanism for scheduling strategies driven by operational data. It possesses good system scalability and operational stability, making it suitable for deployment in multi-satellite and heterogeneous network environments.

[0023] Furthermore, by constructing a clearly structured and functionally independent intelligent agent system, this invention achieves task decoupling, clear semantic hierarchy, and explicit information flow during the policy generation process, thereby improving the interpretability and engineering maintainability of the scheduling system and facilitating rapid adaptation under different scheduling models and communication protocols.

[0024] In summary, this invention fully leverages the semantic understanding and code generation capabilities of large language models, combines the modular advantages of multi-agent systems with the complementary and collaborative mechanism of satellite and ground resources, and realizes the intelligent evolution of low-Earth orbit satellite MAC layer scheduling systems while ensuring real-time performance, reliability, and engineering feasibility. This provides a new technical path for building highly adaptive satellite communication networks. Attached Figure Description

[0025] Figure 1 The diagram shows the overall architecture of the low-orbit satellite network MAC layer scheduling system based on a large language model multi-agent disclosed in this embodiment of the invention.

[0026] Figure 2 The diagram shown is a flowchart of the policy switching process of satellite-side agents in the MAC layer scheduling system of a low-orbit satellite network based on a large language model multi-agent disclosed in an embodiment of the present invention.

[0027] Figure 3 The diagram shows the generation strategy flowchart of ground multi-agents in the MAC layer scheduling system of low-orbit satellite network based on large language model multi-agents disclosed in the embodiments of the present invention.

[0028] Figure 4The diagram shows a throughput comparison of the MAC layer scheduling system for low-orbit satellite networks based on a large language model multi-agent disclosed in an embodiment of the present invention.

[0029] Figure 5 The diagram shows a time delay comparison of the low-orbit satellite network MAC layer scheduling system based on a large language model multi-agent disclosed in an embodiment of the present invention.

[0030] Figure 6 The chart shown is a comparison of delivery rates for the low-Earth orbit satellite network MAC layer scheduling system based on a large language model multi-agent disclosed in this embodiment of the invention. Detailed Implementation

[0031] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. It should be noted that, unless otherwise specified, the following embodiments and features described herein can be combined with each other.

[0032] Please see Figure 1 This invention provides a low-Earth orbit satellite network MAC layer scheduling system based on a large language model and multiple agents, and particularly relates to a media access control (MAC) layer scheduling system suitable for low-Earth orbit satellite communication systems. Specifically, it is a technical solution that uses a large language model to construct a multi-agent collaborative system and achieves intelligent generation and adaptive updating of MAC layer scheduling strategies through satellite-ground joint operation. This system integrates multiple technologies such as satellite network scheduling, natural language processing, agent systems and automatic code generation, and is suitable for the evolution and deployment of scheduling algorithms in dynamic communication environments.

[0033] The MAC layer scheduling system of this low-Earth orbit satellite network includes a policy selection agent deployed on the satellite side and a scenario analysis agent, a policy generation agent, and a policy verification agent deployed on the ground station side. The three ground agents are deployed on the ground station or nearby nodes with computing power. Among them, the satellite-side agent is responsible for selecting and invoking the scheduling policy according to the current communication status, while the three ground station-side agents are responsible for completing the analysis of new communication scenarios, the generation of scheduling policies, and the simulation and verification of policies, thus forming a collaborative reasoning system driven by a large language model.

[0034] The strategy selection agent is set up on the satellite side. It is used to select the scheduling strategy that best matches the current communication scenario from the local policy library based on the current communication state characteristics, and call the corresponding policy code to execute MAC layer scheduling. The strategy selection agent is constructed using a lightweight language model. The lightweight language model is a small pre-trained language model with limited parameters. It is used to match, sort and select the current communication state characteristics and the applicable scenario descriptions of each candidate scheduling strategy in the policy library, without participating in the generation of new scheduling strategy code.

[0035] Specifically, the satellite first collects operational status data within the current communication window and statistically summarizes the raw monitoring data over several consecutive scheduling cycles to form current communication status characteristics. These characteristics include at least link throughput characteristics, service priority characteristics, and queue head latency characteristics, and may also include one or more of the following: number of users, service queue backlog, historical execution strategy identifiers, and corresponding performance indicators. Subsequently, the satellite constructs the current communication status characteristics into a scenario observation description of the current communication window and organizes the applicable scenario descriptions and strategy identifiers corresponding to each candidate scheduling strategy in the local strategy library into a candidate strategy set, which is then input into a lightweight language model. Based on preset selection instructions, the lightweight language model compares and infers the scenario observation description of the current communication window with the applicable scenario descriptions of multiple candidate scheduling strategies in a unified input context, and directly outputs the target scheduling strategy identifier that best matches the current communication scenario.

[0036] After determining the target scheduling strategy, the satellite side calls the strategy code corresponding to the target scheduling strategy and outputs the resource allocation results to each user terminal or each service flow through the preset MAC layer scheduling interface. In this way, the online decision-making process of the satellite side is limited to state feature extraction, candidate strategy matching and target strategy invocation, thereby achieving faster response and stable execution under limited computing resources.

[0037] More specifically, the satellite continuously collects operational data for multiple scheduling cycles within the current communication window. This operational data includes the instantaneous throughput, service priority, and head-of-line delay for each user terminal or service flow. Subsequently, the operational data undergoes statistical processing within the window to generate a scene feature vector for the current communication window. This scene feature vector can be composed of one or more of the following: average throughput, throughput fluctuation, priority distribution, average head-of-line delay, head-of-line delay percentile, number of users, and queue backlog.

[0038] The local policy library pre-stores multiple verified scheduling policies. Each scheduling policy corresponds to a set of applicable scenario description information and a unique policy identifier. When selecting a policy, the scenario feature vector and the applicable scenario description information corresponding to each candidate scheduling policy are input into the lightweight language model. The lightweight language model outputs the matching score or ranking result of each candidate scheduling policy and selects the policy identifier with the highest matching score or the highest ranking as the output result.

[0039] After obtaining the policy identifier, the satellite locates the corresponding policy code in the local policy library based on the policy identifier, loads the policy code into the preset MAC layer scheduling interface to perform resource allocation calculation, and outputs the scheduling result for the current scheduling period. If the scene characteristics change in subsequent communication windows, the above process of feature extraction, feature input, policy matching, and policy invocation is repeated to achieve dynamic policy selection under changing communication scenarios. Please refer to the above process. Figure 2 .

[0040] The scenario analysis agent is used to analyze the operation logs uploaded from the satellite side, identify new communication scenarios not yet covered by the existing policy library, and extract and structure the key features of the new communication scenarios to provide input for subsequent policy generation and policy verification. The operation logs include at least the link throughput information, service priority information, queue head delay information, current execution policy identifier, and corresponding scheduling result information for multiple consecutive scheduling cycles.

[0041] Specifically, the scenario analysis agent first segments the operation log according to the preset scenario change detection rules to obtain candidate scenario segments with relatively stable communication states. Then, it extracts a set of structured features that can characterize the current communication state from each candidate scenario segment. Next, it compares the coverage of the structured feature set with the applicable scenario descriptions corresponding to each stored policy in the existing policy library to determine whether the candidate scenario segment has been covered by the existing policy library. If the candidate scenario segment does not meet the preset coverage conditions with any policy corresponding to any policy in the existing policy library, or if the performance index of the currently executed policy under the candidate scenario segment is lower than the preset requirements, then the candidate scenario segment is determined to be an uncovered new communication scenario.

[0042] For newly identified communication scenarios, the scenario analysis agent further generates standardized scenario description data, which includes scenario summary, key constraints, target optimization direction and simulation parameter suggestions, for use by subsequent policy generation agents and policy verification agents.

[0043] More specifically, the scenario analysis agent receives the operation logs uploaded from the satellite side and reads the throughput information, priority information, queue head delay information, current execution strategy identifier, and scheduling result information for multiple consecutive scheduling cycles in chronological order. Then, the operation logs are segmented according to preset thresholds and a continuous window: when the statistical change of at least one state feature between adjacent time periods continuously exceeds the corresponding threshold, it is determined that the communication scenario has switched, and a new candidate scenario segment is formed; when the change does not continuously meet the preset conditions, the original candidate scenario segment remains unchanged.

[0044] For each candidate scenario segment, the scenario analysis agent performs statistical processing on the runtime logs within the segment to extract scenario features. These features include one or more of the following: average throughput, throughput fluctuation, priority distribution, average latency at the head of the queue, latency percentile at the head of the queue, business backlog, and number of users. Subsequently, the scenario features are compared with the applicable scenario descriptions corresponding to existing scheduling policies in the policy library using similarity comparison or rule matching. When the similarity between the scenario features and an existing scenario description is higher than a preset threshold, and the throughput, latency, delivery rate, or high-priority business assurance indicators of the current execution policy in that segment meet preset requirements, it is determined that the candidate scenario segment has been covered by the existing policy library.

[0045] When neither the candidate scenario fragment nor the existing scenario description meets the coverage condition, or when a similar scenario exists but the current execution strategy exhibits performance anomalies, strategy failures, or reduced scheduling effectiveness under that fragment, the candidate scenario fragment is determined to be an uncovered new communication scenario. Performance anomalies or strategy failures can manifest as one or more of the following: a significant decrease in throughput, a significant increase in latency, a decrease in packet delivery rate, a decrease in deadline fulfillment rate, or insufficient high-priority service guarantee capability. For this uncovered new communication scenario, the scenario analysis agent outputs standardized scenario description data, which includes a scenario summary, key constraints, target optimization direction, and simulation parameter suggestions, and passes this data to the subsequent strategy generation agent and strategy verification agent.

[0046] The strategy generation agent is used to generate candidate scheduling strategy code that meets the requirements of the target communication scenario by calling a large language model based on the scenario description data output by the scenario analysis agent. The scenario description data includes at least link state characteristics, service load characteristics, service priority distribution characteristics, latency constraint characteristics, user number characteristics, currently executed strategies and their performance information, as well as optimization objectives, constraints and simulation environment parameter suggestions for the target scenario.

[0047] Specifically, the policy generation agent first formats the scene description data to construct a policy generation request for the large language model. In addition to the structured feature information of the target scene, the policy generation request also includes a preset scheduling policy code template, a fixed MAC layer scheduling interface definition, input and output parameter descriptions, and computational complexity constraints, which are used to limit the generation range of the large language model. Subsequently, the policy generation request is input into the large language model, which generates candidate scheduling policy codes based on the scene description data, code template, and interface constraints of the target communication scene.

[0048] After obtaining the candidate scheduling strategy code, the strategy generation agent performs post-processing to make it meet the established interface specifications and operational requirements, and forms a standard strategy file that can be called and verified by the subsequent strategy verification agent. In this way, the strategy generation agent does not generate unconstrained free code, but rather runnable scheduling strategy code that is adapted to the target scenario and compatible with the existing scheduling interface.

[0049] More specifically, the policy generation agent first receives standardized scenario description data output by the scenario analysis agent. This standardized scenario description data includes link state characteristics, service load characteristics, service priority distribution characteristics, latency constraint characteristics, user number characteristics, currently executed policies and their performance information, as well as optimization objectives, constraints, and simulation environment parameter suggestions. Subsequently, the policy generation agent organizes the above scenario description data together with a preset code template, a fixed MAC layer scheduling interface definition, input and output variable descriptions, and complexity limits into a policy generation request, and inputs the policy generation request into the large language model.

[0050] The large language model outputs a candidate scheduling strategy code based on the strategy generation request. The candidate scheduling strategy code describes the resource allocation logic under the current input scenario. The resource allocation logic includes at least the state reading rules, business priority processing rules, resource allocation calculation rules, and result output rules. Subsequently, the strategy generation agent performs post-processing on the candidate scheduling strategy code. The post-processing includes syntax integrity checking, function name and parameter name standardization, input and output variable mapping with template interface, incompatible code snippets pruning or correction, and strategy file encapsulation.

[0051] After post-processing, a scheduling strategy code that meets the preset interface requirements is generated and output to the strategy verification agent for functional verification and performance evaluation. When the candidate scheduling strategy code fails the verification process, the strategy generation agent can also receive the failure reason information and correction suggestion information, and reorganize them with the original scene description data and input them into the large language model again to generate a new candidate scheduling strategy code.

[0052] The policy verification agent is used in the NS3 simulation platform to perform functional and basic performance verification on the candidate scheduling policy code output by the policy generation agent, and to determine whether the candidate scheduling policy code is executable and effective in the target communication scenario based on the verification results. The policy verification agent receives the candidate scheduling policy code, target scenario description data, baseline policy identifier, and historical operation log data corresponding to the target communication scenario, and suggests the construction of the corresponding simulation verification scenario in the NS3 simulation platform based on the target scenario description data and simulation parameters.

[0053] Specifically, the policy verification agent first deploys a typical scenario corresponding to the target communication scenario in the simulation platform, and loads the candidate scheduling policy code into the preset MAC layer scheduling interface to form a verification environment. Then, the candidate scheduling policy code is functionally verified to confirm whether it can be correctly loaded by the simulation platform, whether it meets the preset interface definition, and whether it can output a valid scheduling result. After the candidate scheduling policy code passes the functional verification, it is then subjected to basic performance verification to evaluate one or more of the following indicators in the target communication scenario: throughput, latency, packet delivery rate, deadline fulfillment rate, queue backlog, and high-priority service guarantee capability.

[0054] When a candidate scheduling strategy code passes both functional verification and basic performance verification, it is determined to be executable and effective in scheduling, and is submitted to the subsequent strategy storage or synchronization module; when it fails verification, a verification result report is output and fed back to the strategy generation agent for subsequent strategy correction and iterative generation.

[0055] Once verified, the new strategy is uploaded to the satellite as a "strategy supplement for this new scenario." The onboard agent incorporates it into the strategy library for management. Strategy updates utilize an LRU mechanism to manage the elimination and replacement of scheduling strategies in the strategy library. It's important to note that the satellite's local strategy library stores multiple verified scheduling strategies and records the most recent usage time for each strategy. When the satellite invokes a scheduling strategy, it updates the most recent usage time of that strategy. When a new scheduling strategy needs to be written to the strategy library, it first checks if the library has reached its capacity limit. If not, the new strategy is written directly. If it has reached its capacity limit, the longest-unused scheduling strategy is selected for elimination, removed from the strategy library, and the new strategy is written. This approach prioritizes recently invoked scheduling strategies and eliminates long-unused ones, maintaining the strategy library's adaptability to current communication scenarios within limited storage space. The satellite-to-ground data interaction design is synchronized with the satellite transit cycle, ensuring the agent system completes data download and new strategy upload within the physical link reach window.

[0056] More specifically, the policy verification agent first receives candidate scheduling policy code, target scenario description data, baseline policy identifier, and historical operation log data. Based on the link state characteristics, service load characteristics, service priority distribution characteristics, latency constraint characteristics, and simulation parameter suggestions in the target scenario description data, it constructs a typical simulation scenario corresponding to the target communication scenario in the NS3 simulation platform. Subsequently, it connects the candidate scheduling policy code to the preset MAC layer scheduling interface and runs it under the typical simulation scenario.

[0057] Before formally evaluating the candidate scheduling strategy code, the strategy verification agent can first load the baseline strategy into the simulation platform, run it under the same simulation parameters, and compare the simulation output with the corresponding performance results in the historical running log. When the difference between the two exceeds the preset range, the simulation parameters are adjusted to make the simulation scenario as close as possible to the actual running scenario. After completing the above simulation scenario deployment, the strategy verification agent performs functional verification on the candidate scheduling strategy code. The functional verification includes: whether the candidate scheduling strategy code can be loaded correctly, whether it conforms to the preset MAC layer scheduling interface definition, and whether it can run continuously without syntax errors, interface call errors, abnormal interruptions, or illegal resource allocation results.

[0058] After the candidate scheduling strategy code passes functional verification, the strategy verification agent further performs a basic performance evaluation on it. The basic performance evaluation includes obtaining the throughput, latency, packet delivery rate, deadline fulfillment rate, queue backlog, and high-priority service guarantee indicators of the candidate scheduling strategy code under typical simulation scenarios, and comparing them with the corresponding performance results of the baseline strategy under the same conditions. When the candidate scheduling strategy code reaches or exceeds the baseline strategy in the preset core indicators, or meets the preset minimum performance threshold requirements of the target scenario, it is determined to have passed the basic performance evaluation; otherwise, it is determined to have failed the basic performance evaluation. After verification, the strategy verification agent outputs a verification result report and submits the candidate scheduling strategy code that has passed the verification to the subsequent strategy storage module, and feeds back the candidate scheduling strategy code that has failed the verification along with the reasons for failure to the strategy generation agent.

[0059] Once verified, the strategies are uploaded to the satellite via the space-to-ground communication link, where the onboard agent incorporates them into the strategy library. The strategy storage adopts an LRU (Least Recently Used) update mechanism, dynamically replacing low-frequency strategies based on the strategy call frequency to ensure that the strategy library maintains the optimal combination within its limited capacity. To adapt to the periodic visibility of the space-to-ground link, the MAC layer scheduling system of this low-Earth orbit satellite network is designed to align the strategy synchronization window with the satellite transit time, completing the bidirectional interaction of downlinking operational data and uplinking new strategies within the ground visibility window.

[0060] The entire system is designed to implement a closed-loop process of policy deployment, from "scene awareness—semantic generation—simulation verification—policy deployment," enabling the scheduling policy to continuously expand and adapt as the communication environment evolves. The onboard portion uses a lightweight inference module to quickly complete policy selection, while the ground portion uses a highly complex model to complete policy generation and verification tasks. This approach improves the overall intelligence level and communication performance assurance capabilities of the system while meeting satellite resource constraints. Please refer to the above process description. Figure 3 .

[0061] Furthermore, regarding the simulation / experiment setup: In one implementation, to verify the effectiveness of the scheduling algorithm generated by the system of this invention, a multi-user concurrent access scenario is constructed in a network simulation environment. The scenario includes five user terminals (UEs): UE1-UE2 carry high-priority services (corresponding to DC-GBR type services with 5QI=87), using a small packet service model with a medium packet arrival rate (pps); UE3-UE5 carry low-priority services (corresponding to Non-GBR type services with 5QI=80), using a large packet service model with a relatively high packet arrival rate (pps), thus forming a mixed service competition environment of "high-priority light load + low-priority heavy load" to cover typical congestion and resource contention conditions. To reflect the expired removal mechanism for latency-constrained services, a PDCP layer discarding strategy is enabled: when the cumulative dwell time of a data packet in the queue / protocol stack exceeds a preset threshold, the data packet will be discarded in advance to prevent expired data from continuing to occupy radio resources and further deteriorating the timeliness of high-priority services. Based on this scenario, the scheduling strategy automatically generated and distributed by the architecture of this invention (hereinafter referred to as "the scheduling strategy generated by this invention") and a variety of typical scheduling strategies, including PF, MR and RR, were deployed as a comparison to verify the performance differences and guarantee capability differences in mixed priority business scenarios.

[0062] II. Evaluation Metrics: Throughput, latency, and delivery rate (Rx / Tx) are used as evaluation metrics. Throughput measures the effective transmission capacity of the system; latency measures the degree to which service timeliness is met; and delivery rate Rx / Tx represents the ratio of successfully received data (or packets) to transmitted data (or packets), reflecting the impact of packet loss, congestion dropping, and premature PDCP dropping on service completion. The above metrics are calculated as a total average, as well as high-priority average and low-priority average by service level, to reflect the scheduling strategy's ability to differentiate and guarantee different service levels and the efficiency of resource utilization.

[0063] III. Experimental Results and Analysis: (e.g.) Figures 4 to 6 As shown, in a mixed service scenario where the PDCP dropping mechanism is enabled and low-priority services are heavy-loaded with large packets at high PPS, the scheduling strategy generated by this invention can achieve a better trade-off between "prioritizing high-priority services" and "maintaining the available throughput of low-priority services", thereby obtaining a more stable overall service quality.

[0064] Specifically, Figure 4For throughput comparison, the scheduling strategy generated by this invention achieves an average total throughput of approximately 2.55 Mbps, which is close to the overall level of PF (approximately 2.69 Mbps), significantly better than RR (approximately 1.30 Mbps), and better than MR (approximately 2.38 Mbps). This indicates that under this highly competitive load condition, the scheduling strategy generated by this invention can maintain a high overall resource utilization rate. Regarding high-priority service throughput, the average high-priority throughput of the scheduling strategy generated by this invention is approximately 1.35 Mbps, higher than PF (approximately 1.27 Mbps) and RR (approximately 1.28 Mbps), and significantly higher than MR (approximately 0.679 Mbps). This demonstrates that the scheduling strategy generated by this invention... Even under heavy load interference in low-priority scenarios, it can still provide more stable and effective transmission for latency-sensitive services. In terms of low-priority service throughput, the average throughput of low-priority services generated by the scheduling strategy of this invention is approximately 3.35 Mbps, which is significantly higher than the approximately 1.32 Mbps of RR, and is on the same order of magnitude as the approximately 3.64 Mbps of PF and the approximately 3.51 Mbps of MR. This demonstrates that the scheduling strategy generated by this invention can still make full use of the remaining resources to maintain the available transmission capacity of low-priority services while ensuring high-priority services. Therefore, the scheduling strategy generated by this invention can maintain the overall throughput level while enhancing high-priority throughput, and avoid the problem of significantly limited throughput of RR under heavy load.

[0065] Figure 5 To compare latency results, given the significant differences in latency between different methods for high and low priority services, a logarithmic scale is used in the figure. It can be seen that the scheduling strategy generated by this invention has an average latency of approximately 135ms for high-priority services, significantly lower than PF's approximately 191ms and better than RR's approximately 139ms. This indicates that the scheduling strategy generated by this invention can effectively suppress queuing backlog in high-priority services under congestion, making it closer to the timeliness requirements of latency-sensitive services. Regarding the total average latency, the scheduling strategy generated by this invention is approximately 988ms, lower than PF's approximately 1349ms and RR's approximately 1514ms. This shows that it not only has stronger latency control capabilities for high-priority services but also alleviates queuing and delays caused by congestion at the system level, thereby improving the overall throughput-latency tradeoff. It should be noted that MR exhibits a lower average latency in this set of experiments (e.g., a total average of approximately 608ms), but this phenomenon should be considered in conjunction with… Figure 6 Common understanding of delivery rate results: When a large number of data packets are dropped prematurely or fail to be delivered successfully under congestion conditions, the latency of the statistical arrival sample may be lower. Therefore, "lower latency" alone is not enough to prove that the business experience is better. Delivery completion rate still needs to be examined at the same time.

[0066] Figure 6The delivery rate comparison results are used to characterize the deliverability and completion rate of various services under congestion contention and PDCP early discarding mechanisms; from Figure 6 As can be seen, the scheduling strategy generated by this invention achieves a delivery rate of approximately 0.742 for high-priority services, which is higher than 0.695 for PF and approximately 0.702 for RR, and significantly higher than approximately 0.372 for MR. This indicates that the scheduling strategy generated by this invention can more effectively guarantee the successful delivery of high-priority services under strong competitive load conditions, reducing drops due to congestion and expiration, thereby making high-priority services more stable in terms of "deliverability". Regarding the delivery rate of low-priority services, the scheduling strategy generated by this invention achieves a delivery rate of approximately 0.069, significantly higher than approximately 0.02 for RR. 7, and close to PF's approximately 0.075 and MR's approximately 0.072, indicating that the scheduling strategy generated by this invention can maintain a certain delivery level for low-priority services while ensuring high-priority services, and will not cause low-priority services to become unavailable due to excessive skewness; in terms of total average delivery rate, the scheduling strategy generated by this invention is approximately 0.085, close to PF's approximately 0.090, and higher than MR's approximately 0.079 and RR's approximately 0.044, further demonstrating that it can still achieve a high overall delivery completion rate under mixed load and drop mechanism enabled;

[0067] comprehensive Figures 4 to 6 As can be seen, the scheduling strategy generated by the architecture of this invention can simultaneously enhance throughput, reduce latency and improve delivery rate for high-priority services in mixed priority service scenarios, and improve the overall system balance without significantly sacrificing overall throughput and low-priority availability, thereby demonstrating a more reasonable resource allocation and differentiated guarantee effect.

[0068] In summary, this invention achieves dynamic generation, verification, and deployment of scheduling strategies by constructing a space-ground collaborative multi-agent system. The onboard agents are deployed in low-Earth orbit satellites with regenerative forwarding capabilities. Their built-in strategy library stores multiple communication resource scheduling strategy codes and can select the optimal strategy based on the characteristics of the current communication scenario. Multiple agents are deployed on the ground station side, each responsible for scenario feature extraction and analysis, scheduling strategy generation and encapsulation, and strategy testing and simulation verification, forming a closed loop of scheduling strategy evolution based on a large language model. Each agent on the ground side automatically generates runnable scheduling strategy code through multi-round collaborative reasoning and information interaction, and performs performance evaluation in a network simulation environment. Verified new strategies are uploaded to the satellite via the space-ground link to update the strategy library. This invention fully utilizes the capabilities of large language models in semantic understanding and code generation, achieving continuous evolution and adaptive updating of scheduling strategies, significantly improving the resource scheduling efficiency and operational performance of the MAC layer of low-Earth orbit satellite networks in complex communication scenarios.

[0069] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. All equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this invention should still be covered by the claims of this invention.

Claims

1. A MAC layer scheduling system for low-Earth orbit satellite networks based on a large language model and multiple agents, characterized in that: This includes a policy selection agent deployed on the satellite side, and scenario analysis agents, policy generation agents, and policy verification agents deployed on the ground station side. The strategy selection agent is used to select the scheduling strategy that best matches the current communication scenario from the local policy library based on the characteristics of the current communication state, and call the corresponding strategy code to execute MAC layer scheduling. The scenario analysis agent is used to analyze the operation logs uploaded by the satellite side, identify new communication scenarios not yet covered by the existing policy library, and output scenario description data about the new communication scenarios. The scenario description data includes link status characteristics, service load characteristics, service priority distribution characteristics, latency constraint characteristics, user number characteristics, currently executed policies and their performance information, as well as optimization objectives, constraints and simulation environment parameter suggestions for the target communication scenario. The strategy generation agent is used to generate candidate scheduling strategy code that meets the requirements of the target communication scenario by calling a large language model based on the scenario description data output by the scenario analysis agent. The policy verification agent is used to perform functional verification and basic performance verification on the candidate scheduling policy code output by the policy generation agent, and to determine whether the candidate scheduling policy code is executable and effective in the target communication scenario based on the verification results; wherein, the basic performance verification includes evaluating one or more of the following indicators of the candidate scheduling policy code in the target communication scenario: throughput, latency, packet delivery rate, deadline satisfaction rate, queue backlog, and high-priority service guarantee capability. The scenario analysis agent analyzes the operational logs uploaded from the satellite, identifies new communication scenarios not yet covered by the existing policy library, and outputs scenario description data for the new communication scenarios, including: The operation log is segmented based on a preset scene change detection rule, and candidate scene segments with stable communication status are obtained based on the segmentation results. Extract scene features that can characterize the current communication state from each candidate scene fragment; The scenario features are compared with the applicable scenario descriptions corresponding to each stored strategy in the existing strategy library for coverage. Based on the comparison results, it is determined whether the candidate scenario fragment has been covered by the existing strategy library. If the candidate scenario fragment does not meet the preset coverage conditions with any scenario description corresponding to any strategy in the existing strategy library, or if the performance index of the currently executed strategy under the candidate scenario fragment is lower than the preset requirements, then the candidate scenario fragment is determined to be an uncovered new communication scenario.

2. The low-Earth orbit satellite network MAC layer scheduling system based on a large language model multi-agent according to claim 1, characterized in that: The strategy selection agent employs a lightweight language model, which is used to match, sort, and select the current communication state features from the applicable scenario descriptions of each candidate scheduling strategy in the strategy library.

3. The low-Earth orbit satellite network MAC layer scheduling system based on a large language model multi-agent according to claim 2, characterized in that: The strategy selection agent is used to select the scheduling strategy that best matches the current communication scenario from the local policy library based on the characteristics of the current communication state, and to call the corresponding policy code to execute MAC layer scheduling, including: Obtain the scene feature vector and candidate scheduling strategy set of the current communication window, and input the scene feature vector and candidate scheduling strategy set of the current communication window into the lightweight language model; The matching score of each candidate scheduling strategy is calculated using a lightweight language model. Based on the calculation results, the target scheduling strategy identifier that best matches the current communication scenario is obtained, and the corresponding strategy code is called to execute MAC layer scheduling.

4. The low-Earth orbit satellite network MAC layer scheduling system based on a large language model multi-agent according to claim 3, characterized in that: Before obtaining the scene feature vector and candidate scheduling strategy set of the current communication window, and inputting the scene feature vector and candidate scheduling strategy set of the current communication window into the lightweight language model, the following steps are included: Within the current communication window, running data for multiple scheduling cycles is continuously collected, and the running data is statistically processed within the window to obtain the scene feature vector of the current communication window. The scene feature vector of the current communication window is used to construct a scene observation description of the current communication window, and each candidate scheduling strategy is obtained from the local policy library based on the scene observation description, thereby obtaining a set of candidate scheduling strategies.

5. The low-Earth orbit satellite network MAC layer scheduling system based on a large language model multi-agent according to claim 1, characterized in that: The strategy generation agent is used to generate candidate scheduling strategy code that meets the requirements of the target communication scenario by calling a large language model based on the scenario description data output by the scenario analysis agent, including: The scene description data is formatted, and a strategy generation request for a large language model is constructed. The strategy generation request is input into the large language model, which then generates candidate scheduling strategy code based on the scenario description data, code template, and interface constraints of the target communication scenario.

6. The low-Earth orbit satellite network MAC layer scheduling system based on a large language model multi-agent according to claim 1, characterized in that: The policy verification agent is used to perform functional and basic performance verification on the candidate scheduling policy code output by the policy generation agent, and to determine whether the candidate scheduling policy code is executable and effective in the target communication scenario based on the verification results, including: Deploy a typical scenario corresponding to the target communication scenario in the simulation platform, and load the candidate scheduling strategy code into the preset MAC layer scheduling interface to form a verification environment; The candidate scheduling strategy code is functionally and fundamentally verified, and the results are used to determine whether the candidate scheduling strategy code is executable and effective in the target communication scenario.

7. The low-Earth orbit satellite network MAC layer scheduling system based on a large language model multi-agent according to claim 6, characterized in that: The process of performing functional and basic performance verification on the candidate scheduling strategy code, and determining whether the candidate scheduling strategy code is executable and effective in the target communication scenario based on the verification results, includes: The candidate scheduling strategy code is functionally verified. After the candidate scheduling strategy code passes the functional verification, the candidate scheduling strategy code is then subjected to basic performance verification. When the candidate scheduling strategy code passes both functional verification and basic performance verification, it is determined that the candidate scheduling strategy code has executability and scheduling effectiveness, and is submitted to the subsequent strategy storage or synchronization module. If the candidate scheduling strategy code fails the verification, a verification result report is output and fed back to the strategy generation agent for subsequent strategy correction and iterative generation.

Images