Method for accelerating learning of low earth orbit satellite network traffic scheduling based on elephant flow perception

Abstract

The application discloses an elephant flow perception-based learning acceleration low-orbit satellite network flow scheduling method, and belongs to the field of low-orbit satellite network flow scheduling; the method comprises the following steps: firstly, a dynamic threshold is calculated based on a global flow demand matrix to identify an elephant flow, and a mouse flow is statically unloaded to obtain a remaining inter-satellite link capacity, so that flow dimensionality reduction is realized; secondly, the translational invariance of a satellite topology is utilized, and a candidate path set of the elephant flow is rapidly instantiated through a translation operator; subsequently, a competition relationship between the remaining link and the path is modeled as a time-varying bipartite graph, and a graph neural network and a demand coordination layer are utilized to extract high-order flow feature embedding satisfying a flow conservation constraint; finally, the elephant flow is modeled as an independent agent, and an anti-fact multi-agent reinforcement learning algorithm is utilized to output an optimal flow splitting ratio in parallel. Through the fusion of flow stripping and topology perception depth, the dimension disaster and decision-making time delay problems in a giant constellation are effectively solved, and the global load balancing capability is significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of low-Earth orbit satellite network traffic scheduling, specifically involving a learning-accelerated low-Earth orbit satellite network traffic scheduling method based on elephant flow perception. Background Technology

[0002] In recent years, large-scale low-Earth orbit (LEO) satellite constellations have expanded rapidly, aiming to provide low-latency, high-bandwidth communication services globally and bridge the digital divide in remote areas. Within these mega-constellations, traffic engineering (TE) plays a crucial role in minimizing end-to-end latency and preventing network link congestion by optimizing network resource utilization. However, the highly dynamic topology and massive concurrent traffic of LEO satellite networks pose significant challenges to traditional traffic engineering methods. In actual satellite network traffic distribution, a significant "heavy-tailed" characteristic exists. Specifically, the network is filled with a massive number of lightweight "miceflows," which, while having a negligible impact on total bandwidth consumption, greatly increase the state dimension and solution space of global routing optimization; while a small number of high-bandwidth "elephantflows," though fewer in number, are key to causing link congestion and determining the overall load balancing performance of the network. Therefore, how to fundamentally reduce the optimization dimensionality of the system by stripping away the "rat stream" in a massive concurrent request environment, thereby focusing computing resources on the critical "elephant stream" for fast and accurate scheduling, is a key technical challenge that urgently needs to be addressed to achieve efficient congestion control in giant satellite networks.

[0003] Existing technologies have proposed implementing traffic scheduling or offloading in specific scenarios; for example: Patent application CN106453130A discloses a traffic scheduling system and method based on accurate identification of "elephant flows." This technical solution proposes a mechanism to distinguish between large and small flows in a software-defined network (SDN). Its core lies in identifying elephant flows by monitoring the amount of data in the TCP queue during terminal data transmission. Then, based on existing routing protocols, it calculates the path with the largest available bandwidth for the identified elephant flow as the optimal forwarding path, and the SDN controller issues flow tables for scheduling. However, while this technology achieves elephant flow identification, its subsequent scheduling heavily relies on traditional bandwidth calculation logic (such as heuristic search or shortest path algorithms). In giant satellite networks with thousands of nodes, calculating and comparing remaining bandwidth one by one among a massive number of candidate paths requires enormous computational overhead. This serial calculation mode results in decision latency often reaching several seconds, far lagging behind the millisecond-level topology change speed of low-Earth orbit satellite networks, easily leading to routing policy failure and severe congestion in local links.

[0004] Patent application CN115150335A discloses a method and system for optimal traffic segmentation based on deep reinforcement learning. This technical solution involves an artificial intelligence-based routing allocation strategy. It obtains the current state of all routers requiring traffic segmentation, establishes a router state matrix, and constructs a training dataset based on this matrix. By specifying the average throughput within a traffic cycle as the reward, a deep reinforcement learning model is trained, enabling the model to directly output the optimal traffic segmentation ratio for each router in the next time step. However, this technique directly concatenates the router states into a matrix and inputs it to the reinforcement learning model. First, this scheme does not utilize the "heavy-tail effect" of satellite traffic for preprocessing (i.e., it does not remove the massive amount of "rat flows"), causing the reinforcement learning model to be overwhelmed by countless irrelevant background flows, leading to a severe "curse of dimensionality" and making model convergence difficult. Second, the simple state matrix cannot effectively capture the complex interference relationships between satellite nodes, links, and massive concurrent paths (such as congestion caused by multiple "elephant flows" simultaneously competing for the same inter-satellite link), preventing the agent from making globally optimal decisions that satisfy strict flow conservation constraints.

[0005] Patent application CN112954385A discloses an adaptive traffic splitting decision method based on cybernetics and data-driven approaches. This technical solution focuses on multi-path transmission of application-layer data, defining the network environment as the state input of the intelligent agent module. It adaptively determines the optimal splitting ratio for multiple paths through a machine learning model to distribute video and other data. Simultaneously, it incorporates cybernetics, adjusting the video bitrate based on dynamic changes in the receiver buffer to further optimize the accuracy of the machine learning decision model in practical systems. However, this technology is a typical end-to-end traffic splitting control, with its input state and reward mechanism primarily focused on the application experience of a single client (such as video buffer occupancy and bitrate). This mechanism lacks a macroscopic understanding of the global network topology and cannot assess and optimize the "maximum link utilization (MLU)" of the core backbone links. Forcibly applying it to satellite network scheduling would lead to each data source operating independently, easily causing uncontrollable traffic collisions on core inter-satellite links.

[0006] Patent application CN114567365A discloses a routing method and system for load balancing in low-Earth orbit (LEO) satellite networks. This technical solution discloses a LEO satellite network routing method based on deep reinforcement learning (DQN). Its core lies in the fact that the current satellite node obtains the state information (including spatial location, queuing delay, available frequency resources, and signal-to-noise ratio) of neighboring satellite nodes through communication interaction, and combines this information with its own state, inputting it into a trained Deep Q-Network (DQN) model. This model directly determines the best neighboring satellite node for the next hop and forwards data, repeating this process hop-by-hop until the target node is reached. This solution aims to improve operational reliability under high load without relying on inter-satellite link state prediction. Although this technology utilizes DQN to achieve state-aware intelligent routing, it still has significant limitations in giant LEO constellations: First, the method employs a hop-by-hop local decision-making mechanism, which is essentially a single-path forwarding mechanism. It cannot perform reasonable optimal proportioning of high-bandwidth "elephant flows" like end-to-end multi-path routing, making it highly susceptible to micro-burst congestion on the selected "next hop" link. Secondly, this scheme fails to consider the "heavy-tail" characteristic of satellite traffic and does not remove the massive background "rat streams," resulting in a large amount of redundant noise in the input state space of the DQN model. This leads to severe dimensionality curse and convergence difficulties at the thousand-node scale. Finally, relying solely on simple concatenation of adjacent node states as model input lacks the structured feature extraction of complex global "link-path" interference relationships by graph neural networks (GNNs), making it difficult to fundamentally reduce the maximum link utilization of the entire network.

[0007] In summary, existing technologies are either limited by the high latency of traditional algorithms or by the failure of intelligent algorithms in high-dimensional representations. This invention aims to solve the technical problem of how to accurately detect network congestion bottlenecks within a very short time (milliseconds) in large-scale low-Earth orbit satellite networks with massive numbers of nodes, and to perform dimensionality reduction and globally optimal allocation of massive concurrent traffic, thereby minimizing maximum link utilization (MLU).

[0008] To address this, this invention proposes a learning-accelerated low-Earth orbit satellite network traffic scheduling method based on elephant flow perception, which deeply integrates traffic distribution patterns (separation and dimensionality reduction of mouse flow / elephant flow), graph neural networks (capturing topology and path interference), and multi-agent reinforcement learning (millisecond-level parallel inference). Summary of the Invention

[0009] To address the shortcomings of existing technologies, the present invention aims to provide a learning-accelerated low-Earth orbit satellite network traffic scheduling method based on elephant flow perception, thereby solving the problems in existing technologies.

[0010] The objective of this invention can be achieved through the following technical solutions: Learning-based methods for accelerating traffic scheduling in low-Earth orbit satellite networks, including: Obtain the global traffic demand matrix of the low-Earth orbit satellite network at the current moment and calculate the dynamic identification threshold; determine the node pairs with traffic demand greater than or equal to the dynamic identification threshold as elephant flow, otherwise determine them as mouse flow; unload the mouse flow along the static shortest path, and deduct the bandwidth capacity occupied by the mouse flow from the inter-satellite links of the low-Earth orbit satellite network to obtain the remaining inter-satellite link capacity. For each of the elephant flows, the relative offset vector between its source node and destination node is calculated; based on the translation invariance of the low-orbit satellite + grid topology, the reference path sequence corresponding to the relative offset vector in the pre-computation cache is mapped to the actual candidate path using the translation operator to obtain the elephant flow candidate path set; The competitive relationship between the inter-satellite links with the remaining inter-satellite link capacity and the candidate path set of elephant stream is modeled as a time-varying bipartite graph; a graph neural network is used to perform message passing on the time-varying bipartite graph, and a demand coordination layer is introduced to pool and coordinate the features of candidate paths serving the same elephant stream communication demand, and output high-order stream feature embeddings. Each elephant flow is modeled as an independent agent, and the high-order flow features are embedded into the policy network corresponding to the independent agent. A multi-agent reinforcement learning algorithm is used to output the flow splitting ratio of each elephant flow on the corresponding actual candidate path in parallel.

[0011] Furthermore, the dynamic recognition threshold The calculation formula is: in, Let be the average traffic demand of all nodes in the global traffic demand matrix. The standard deviation of the traffic demand of all nodes in the global traffic demand matrix is ​​given. Preset hyperparameters for controlling the sparsity of elephant flow.

[0012] Furthermore, through the translation operator The formula for mapping the actual candidate path is: in, The coordinates of the source node corresponding to the elephant flow are two-dimensional coordinates. The reference path sequence, This refers to the relative coordinate offset of each node in the reference path sequence. These refer to the scale of the low-orbit satellite network in both longitude and latitude.

[0013] Furthermore, the time-varying bipartite graph is defined as: in, This is the set of active inter-satellite links at the current moment. This is the candidate path set for the elephant stream; if inter-satellite links Included in actual candidate paths In the middle, the adjacency relationship between links and paths is established. .

[0014] Furthermore, the update rule of the graph neural network is as follows: in, This is an augmented double adjacency matrix that includes the adjacency relationships between the inter-satellite links and the actual candidate paths. For graph neural networks The feature matrix of the layer, It is a non-linear activation function. For the first Multilayer perceptron.

[0015] Furthermore, the demand coordination layer updates the features using the following formula: in, These are the updated candidate path features. To serve the same elephant stream communication needs The set of candidate paths, where POOL represents the feature pooling operation, and q is the set. Any candidate path in the list, For the first Deep neural networks.

[0016] Furthermore, the multi-agent reinforcement learning algorithm adopts a centralized training and distributed execution architecture, and introduces a counterfactual baseline to calculate the advantage function of each independent agent. Its formula is: in, For the global action value function, This is the global state. For the joint action of all intelligent agents, For the local flow features of the k-th agent, For the policy network, For the joint action of the other intelligent agents excluding the k-th intelligent agent, Let k be the action space of the k-th agent. In the action space Action variables for inner traversal.

[0017] Learning-based acceleration of low-Earth orbit satellite network traffic scheduling systems using elephant flow perception, implementing the above methods, includes: Flow identification and offloading module: Obtain the global traffic demand matrix of the low-Earth orbit satellite network at the current moment and calculate the dynamic identification threshold; identify node pairs with traffic demand greater than or equal to the dynamic identification threshold as elephant flows, otherwise identify them as mouse flows; offload the mouse flows along the static shortest path, and deduct the bandwidth capacity occupied by the mouse flows from the inter-satellite links of the low-Earth orbit satellite network to obtain the remaining inter-satellite link capacity. Path instantiation module: For each of the elephant flows, calculate the relative offset vector between its source node and destination node; based on the translation invariance of low-orbit satellite + grid topology, use the translation operator to map the reference path sequence corresponding to the relative offset vector in the pre-computation cache to the actual candidate path, and obtain the elephant flow candidate path set; State embedding module: Models the competition relationship between the inter-satellite links with the remaining inter-satellite link capacity and the elephant flow candidate path set as a time-varying bipartite graph; performs message passing on the time-varying bipartite graph using a graph neural network, and introduces a demand coordination layer to pool and concatenate the features of candidate paths serving the same elephant flow communication demand, and outputs high-order flow feature embedding. Allocation Module: Each elephant flow is modeled as an independent agent, and the high-order flow features are embedded into the policy network corresponding to the independent agent. The multi-agent reinforcement learning algorithm is used to output the traffic splitting ratio of each elephant flow on the corresponding actual candidate path in parallel.

[0018] A computer storage medium storing a readable program that, when executed, instructs a computing device to perform the above-described learning-accelerated low-Earth orbit satellite network traffic scheduling method based on elephant flow perception.

[0019] An electronic device includes: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform operations corresponding to the above-described learning-accelerated low-Earth orbit satellite network traffic scheduling method based on elephant flow perception.

[0020] The beneficial effects of this invention are: 1. This invention eliminates the interference of massive background small flows on network optimization by using an elephant flow / mouse flow separation mechanism and a fast path mapping technique based on relative coordinates, thereby reducing the search complexity of candidate paths to... This is beneficial for providing a low-dimensional, high signal-to-noise ratio optimization space for subsequent intelligent algorithms.

[0021] 2. This invention is based on the message passing and demand coordination layer (DCL) of the "link-path" time-varying bipartite graph. It can perceive the changes in the remaining capacity of inter-satellite links and the competition and interference of elephant flows on different paths in real time. This is conducive to improving the model's control over strict flow conservation constraints and its ability to accurately represent global link congestion bottlenecks.

[0022] 3. This invention combines Counterfactual Baseline with multi-agent reinforcement learning, and drives the flow allocation ratio with a centralized training and decentralized execution (CTDE) strategy to output extremely fast (20 milliseconds). This not only achieves a 24%~30% reduction in maximum link utilization (MLU), but also helps to improve the system's millisecond-level non-retraining resilience and recovery capability in the event of sudden link failures. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of the learning-accelerated low-Earth orbit satellite network traffic scheduling method based on elephant flow perception, as described in this invention. Figure 2 This is a comparison chart of the cumulative distribution of maximum link utilization (MLU) under different elephant flow identification thresholds of the present invention; Figure 3 This is a box plot comparing the normalized maximum link utilization performance of the method of this invention with that of existing mainstream routing algorithms. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1 The global optimization objective of the Traffic Engineering (TE) problem in this invention is to minimize the maximum link utilization (MLU) of the network, denoted as . Set up nodes. To the node The traffic demand is , The set of all shortest paths, decision variables Indicates allocation to the first The traffic proportion of each path. Its global mathematical model is defined as: min Received in, Indicates whether the traffic path passes through a link This refers to the link capacity.

[0027] To efficiently solve the above problems, this invention proposes a learning-accelerated low-Earth orbit satellite network traffic scheduling method based on elephant flow perception, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain the global traffic demand matrix of the low-Earth orbit satellite network at the current moment and calculate the dynamic identification threshold; determine the node pairs with traffic demand greater than or equal to the dynamic identification threshold as elephant flow, otherwise determine them as mouse flow; unload the mouse flow along the static shortest path, and deduct the bandwidth capacity occupied by the mouse flow from the inter-satellite links of the low-Earth orbit satellite network to obtain the remaining inter-satellite link capacity. Low-Earth orbit (LEO) satellite network traffic exhibits significant heavy-tail characteristics. To fundamentally reduce the state dimension and solution space of subsequent intelligent routing optimization, this invention first extracts the global traffic demand matrix at the current moment. (Total number of satellite nodes), calculate its average. and standard deviation Subsequently, a dynamic recognition threshold was constructed. : in, To control the sparsity of elephant flow hyperparameters, a reasonable recommended value range is 3 to 10 (preferably 5). This is because the flow of node pairs... If the result is positive, it's classified as an elephant flow; otherwise, it's classified as a mouse flow. The magnitude of this parameter directly determines the system's sensitivity to elephant flow feature capture, significantly impacting the system's dimensionality reduction effect and global load balancing performance. Taking values ​​that are too small (such as 3 or 4) will result in an excessively low recognition threshold, causing a large amount of ordinary traffic to be overclassified as "elephant traffic." This not only introduces redundant traffic noise, interfering with the reinforcement learning agent's accurate capture of core network bottlenecks, but also leads to a rapid expansion of the state space, directly negating the dimensionality reduction acceleration effect brought by the large / small flow separation mechanism. Values ​​that are too large (such as 6 or 10) can lead to overly stringent identification criteria, causing many truly high-bandwidth-demand traffic flows to fail to be detected (i.e., too many elephant flows are misclassified as mouse flows). These hidden heavy-load flows that are not optimized by multi-agent systems will be forcibly allocated to the static shortest path, which can easily cause uncontrollable local congestion on core inter-satellite links.

[0028] Verified through system simulation (e.g.) Figure 2 As shown), whether it's the GDP traffic matrix scenario on the left or the POP traffic matrix scenario on the right, when When the value is 5, the system can most accurately strip away redundant traffic and lock onto the core elephant flow, achieving the best balance between reducing the solution space and ensuring low network congestion.

[0029] Logical stripping mechanism: All mouse flows are directly offloaded along the static shortest path, and their occupied bandwidth capacity is deducted from the inter-satellite links. This stripping operation isolates massive amounts of background traffic that do not cause congestion, drastically reducing the problem dimension, thus allowing the subsequent graph neural network to accurately focus on the key elephant flows that cause congestion.

[0030] Step 2: For each of the elephant flows, calculate the relative offset vector between its source node and destination node; based on the translation invariance of the low-orbit satellite + grid topology, use the translation operator to map the reference path sequence corresponding to the relative offset vector in the pre-computation cache to the actual candidate path, and obtain the elephant flow candidate path set; For the identified elephant flow, a candidate path set is instantiated using the translation invariance of low-Earth orbit satellites and the Grid topology. The source node is then computed. With the target node relative offset vector in, This represents the orbital index offset of the destination node relative to the source node. This represents the satellite index offset within the orbit of the destination node relative to the source node. The orbital index of the target satellite node. This is the orbit index of the source satellite node. This is the index of the location of the target satellite node within its orbit. is the position index of the source satellite node within its orbit, and is the total number of orbital planes in the low-Earth orbit satellite network. The total number of satellites in each orbital plane; The initialization steps of the pre-computation cache library are as follows: During the system offline phase, utilizing the translational symmetry of the low-Earth orbit satellite network topology, a relative coordinate system is constructed using any satellite node as the virtual reference origin (0,0); based on the minimum hop count as a metric, the top K shortest paths from the virtual reference origin to all possible relative coordinate offsets in the network are pre-calculated using the shortest path algorithm, and this set of paths is persistently saved as the pre-computation cache library. During the online execution phase, the corresponding reference path sequence in the cache library will be pre-calculated. By translation operator Directly mapped to actual candidate paths in, This refers to the actual candidate path generated after the translation operator mapping (i.e., a sequence consisting of the absolute coordinates of a series of actual satellite nodes). Reference path for offline caching The virtual relative node coordinates corresponding to the k-th jump; thus, with The candidate path set for elephant flow is completed with extremely low complexity (i.e., computational complexity that is linearly related to the total number of satellite nodes in the network). This mechanism enables rapid instantiation and completely avoids the massive node traversal required by traditional online global pathfinding, greatly reducing scheduling latency.

[0031] Step 3: Model the competition relationship between the inter-satellite links with the remaining inter-satellite link capacity and the elephant flow candidate path set as a time-varying bipartite graph; use a graph neural network to perform message passing on the time-varying bipartite graph, and introduce a demand coordination layer to pool and concatenate the features of candidate paths serving the same elephant flow communication demand, and output high-order flow feature embeddings. Will The competitive relationship of satellite network resources at any given time is modeled as a time-dependent bipartite graph. ,in For active inter-satellite link sets, This is the set of candidate paths for the elephant flow. If the link... Belongs to path Then an adjacency relationship is established. Initialize the feature matrix containing link capacity and path requirements: in, Let be the initial global feature matrix of the time-varying bipartite graph at time t. To encode the initial feature matrix of the link nodes that contains the real-time remaining capacity and latency status of currently active inter-satellite links, This is the initial feature matrix of path nodes that encodes the volume and priority of the elephant flow traffic demand. Execute on a bipartite graph Layer message passing, update rules are as follows: in, It is a non-linear activation function. For graph neural networks The feature matrix of the layer, For the first Multilayer perceptron; To augment the biadjacency matrix; Among them, the augmented double adjacency matrix Defined as: in, This is an identity matrix corresponding to the dimension and the number of active inter-satellite links. A binary path-link indicator matrix is ​​used to indicate whether a candidate path passes through a specific inter-satellite link. This is the transpose of the path-link indicator matrix. This is the identity matrix corresponding to the dimension and the number of candidate paths in the elephant flow; This mechanism is achieved through The bottleneck information is propagated back to the path.

[0032] To ensure that multiple alternative paths for the same elephant flow satisfy the "flow conservation" constraint, a demand coordination layer (DCL) is introduced after each GNN update. This is for serving the same demand. Feature pooling and splicing coordination are performed: in, These are the updated candidate path features. For the first Deep neural networks with multiple layers To represent the feature pooling operation, The candidate path features before the update. For the set Any candidate path in the list. To serve the same elephant stream communication needs The set of candidate paths; The final output encodes a high-order flow feature embedding that incorporates dynamic topology and path interference relationships. .

[0033] Step 4: Model each elephant flow as an independent agent, embed the high-order flow features into the policy network corresponding to the independent agent, and use a multi-agent reinforcement learning algorithm to output the flow split ratio of each elephant flow on the corresponding actual candidate path in parallel.

[0034] Model each elephant flow as an independent intelligent agent. At time t, the agent acquires its local flow characteristics. , by policy network Output continuous action This refers to the traffic splitting ratio on the alternative paths. The system adopts a centralized training and distributed execution (CTDE) architecture. To address the credit assignment challenge caused by massive concurrent traffic, a counterfactual baseline is introduced. Calculate the advantage function. The baseline is obtained by marginalizing the agent's action space and calculating its expectation: in For intelligent agents The counterfactual baseline; For intelligent agents Discrete action space; For intelligent agents Any alternative action; In a given local state Next action The probability of the strategy; In the global state Other intelligent agent actions Under fixed conditions, intelligent agents Execute action The value function of the centralized action at that time.

[0035] intelligent agent Advantage function The calculation is as follows: in, For intelligent agents The advantage function is used to quantify the independent contribution of its actions to the global reward; In the current global state Execute joint actions The actual global action value.

[0036] In a multi-agent reinforcement learning environment, to guide all agents to collaboratively minimize system-level congestion, this invention uses a global reward function shared by all elephant stream agents. Defined as the negative of the current maximum link utilization (MLU) of the network: in, for The global maximum link utilization (MLU) is calculated through forward traffic allocation at every moment. This reward mechanism ensures that the local traffic splitting actions of each agent directly serve the global optimization goal of alleviating the most congested links in the entire network. Based on this global reward... The system uses a centralized criterion network to calculate the aforementioned action value function. .

[0037] Finally, the independent contributions of all elephant flow agents to the global MLU optimization are accumulated through a multi-agent policy gradient algorithm to update the policy network parameters. : in, The gradient represents the expected return of the policy network; Indicates the current joint strategy Expectation calculation; The total number of elephant-flow agents in the system; These are neural network parameters shared by all agents, including parameters from the policy network and the graph neural network.

[0038] Once trained, the Elephant Flow agent can make decentralized adaptive traffic allocation decisions in milliseconds using a single network forward propagation.

[0039] Example 2 In this embodiment, a learning-accelerated low-Earth orbit satellite network traffic scheduling system based on elephant flow perception is proposed, including: Flow identification and offloading module: Obtain the global traffic demand matrix of the low-Earth orbit satellite network at the current moment and calculate the dynamic identification threshold; identify node pairs with traffic demand greater than or equal to the dynamic identification threshold as elephant flows, otherwise identify them as mouse flows; offload the mouse flows along the static shortest path, and deduct the bandwidth capacity occupied by the mouse flows from the inter-satellite links of the low-Earth orbit satellite network to obtain the remaining inter-satellite link capacity. Path instantiation module: For each of the elephant flows, calculate the relative offset vector between its source node and destination node; based on the translation invariance of low-orbit satellite + grid topology, use the translation operator to map the reference path sequence corresponding to the relative offset vector in the pre-computation cache to the actual candidate path, and obtain the elephant flow candidate path set; State embedding module: Models the competition relationship between the inter-satellite links with the remaining inter-satellite link capacity and the elephant flow candidate path set as a time-varying bipartite graph; performs message passing on the time-varying bipartite graph using a graph neural network, and introduces a demand coordination layer to pool and concatenate the features of candidate paths serving the same elephant flow communication demand, and outputs high-order flow feature embedding. Allocation Module: Each elephant flow is modeled as an independent agent, and the high-order flow features are embedded into the policy network corresponding to the independent agent. The multi-agent reinforcement learning algorithm is used to output the traffic splitting ratio of each elephant flow on the corresponding actual candidate path in parallel.

[0040] Example 3 In this embodiment, the method proposed in Embodiment 1 is verified through a specific example; the details are as follows: The simulation environment constructed a large-scale mega-constellation containing 1584 low-Earth orbit satellites (referencing the first-stage constellation configuration of Starlink Phase 1), with satellites distributed across 72 orbital planes, each containing 22 satellites. To simulate real-world global network demands, the traffic generation model employed a gravity model, randomly sampling cities globally based on their GDP or population as weights. The sampled city pairs were then connected to the nearest satellite nodes, generating 100 sets of burst traffic matrices based on GDP distribution and 100 sets based on population distribution, respectively. Simulation hardware platform: The algorithm ran on a high-performance computing platform equipped with an NVIDIA RTX 4090 GPU.

[0041] Baseline Algorithm Comparison: To verify the effectiveness of this invention (ELATE framework), it is compared and verified with the following five existing mainstream traffic scheduling algorithms: SP (Shortest Path Algorithm): Without splitting traffic, it directly allocates all traffic to a single shortest path.

[0042] SE (Equivalent Multi-Path Shortest Path Algorithm): The mouse flow takes a single shortest path, while the elephant flow splits multiple equivalent candidate shortest paths into equal proportions on an absolute average basis.

[0043] LS (Local Search Algorithm): Iteratively reroutes traffic on congested links to other alternative paths.

[0044] LP (Global Linear Programming Algorithm): Uses traditional commercial solvers (such as Gurobi) to perform accurate linear programming solutions on global mathematical models.

[0045] CESLP (Cluster-based Linear Programming Splitting Algorithm): Divides the entire network into multiple clusters and optimizes the splitting of elephant flows using solvers within local clusters.

[0046] Through testing under the aforementioned high-load traffic matrix, this invention demonstrates a clear advantage in network congestion prevention and execution efficiency. Maximum link utilization (MLU) decreases significantly: simulation results show (e.g.) Figure 3 The left figure shows the GDP traffic matrix scenario, and the right figure shows the POP traffic matrix scenario. Because the SP and SE algorithms cannot effectively avoid congested links, their normalized median MLU is as high as approximately 0.75 and 0.67, respectively. The median MLU of traditional LP and LS algorithms is between 0.60 and 0.65. However, the learning acceleration framework of this invention can accurately perceive topology bottlenecks and perform adaptive allocation. Its MLU curve remains consistently at the lowest level under various traffic loads, with the median MLU dropping to approximately 0.53. Compared to the traditional SP algorithm, it achieves a maximum MLU reduction of up to 29%. Even compared to the best-performing CESLP algorithm, this invention achieves approximately 8% further optimization. This data fully demonstrates the superior global load balancing capability of this invention when handling heavy-tail traffic.

[0047] Furthermore, it achieves a significant speed improvement in computational decision-making: In the massive state space of 1584 satellites, traditional solvers face severe dimensionality curses. The LP algorithm takes 34.76 seconds per solution, and the CESLP algorithm, due to subproblem iteration limitations, even takes over 60 seconds. This invention, however, drastically reduces computational dimensionality through mouse-flow decomposition and completely eliminates serial iteration, relying entirely on GPUs for matrix-parallel forward inference. Its actual average time per inference is only 10.57 milliseconds (strictly guaranteed to be within 20 milliseconds). Compared to traditional baseline algorithms, this invention achieves a computational speed improvement of 11 to 3288 times.

[0048] This embodiment fully demonstrates through specific data that the present invention can achieve better network congestion control than traditional global optimization algorithms with millisecond-level decision-making speed, while avoiding the high latency defect of the solver. It perfectly meets the real-time scheduling requirements of high-frequency topology changes and massive connections in low-orbit satellite networks.

[0049] Based on a similar inventive concept, embodiments of the present invention also provide a computer storage medium storing a readable program that, when run by a processor, can execute the above-described learning-accelerated low-Earth orbit satellite network traffic scheduling method based on elephant flow perception.

[0050] Based on a similar inventive concept, this invention provides an electronic device, including: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described learning-accelerated low-Earth orbit satellite network traffic scheduling method based on elephant flow perception.

[0051] Based on a similar inventive concept, embodiments of the present invention also provide a computer program product, including computer instructions, which instruct a computing device to perform the operations corresponding to the above-described learning-accelerated low-Earth orbit satellite network traffic scheduling method based on elephant flow perception.

[0052] The methods of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code that, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses the code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for performing the methods shown herein.

[0053] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A learning-accelerated low-Earth orbit satellite network traffic scheduling method based on elephant flow perception, characterized in that, include: Obtain the global traffic demand matrix of the low-Earth orbit satellite network at the current moment, and calculate the dynamic identification threshold; Node pairs with traffic demand greater than or equal to the dynamic identification threshold are identified as elephant flows, otherwise they are identified as mouse flows; the mouse flows are unloaded along the static shortest path, and the bandwidth capacity occupied by the mouse flows is deducted from the inter-satellite links of the low-Earth orbit satellite network to obtain the remaining inter-satellite link capacity. For each of the elephant flows, the relative offset vector between its source node and destination node is calculated; based on the translation invariance of the low-orbit satellite + grid topology, the reference path sequence corresponding to the relative offset vector in the pre-computation cache is mapped to the actual candidate path using the translation operator to obtain the elephant flow candidate path set; The competitive relationship between the inter-satellite links with the remaining inter-satellite link capacity and the candidate path set of elephant stream is modeled as a time-varying bipartite graph; a graph neural network is used to perform message passing on the time-varying bipartite graph, and a demand coordination layer is introduced to pool and coordinate the features of candidate paths serving the same elephant stream communication demand, and output high-order stream feature embeddings. Each elephant flow is modeled as an independent agent, and the high-order flow features are embedded into the policy network corresponding to the independent agent. A multi-agent reinforcement learning algorithm is used to output the flow splitting ratio of each elephant flow on the corresponding actual candidate path in parallel.

2. The learning-accelerated low-Earth orbit satellite network traffic scheduling method based on elephant flow perception as described in claim 1, characterized in that, The dynamic recognition threshold The calculation formula is: in, Let be the average traffic demand of all nodes in the global traffic demand matrix. The standard deviation of the traffic demand of all nodes in the global traffic demand matrix is ​​given. Preset hyperparameters for controlling the sparsity of elephant flow.

3. The learning-accelerated low-Earth orbit satellite network traffic scheduling method based on elephant flow perception as described in claim 1, characterized in that, By translation operator The formula for mapping the actual candidate path is: in, The coordinates of the source node corresponding to the elephant flow are two-dimensional coordinates. The reference path sequence, This refers to the relative coordinate offset of each node in the reference path sequence. These refer to the scale of the low-orbit satellite network in both longitude and latitude.

4. The learning-accelerated low-Earth orbit satellite network traffic scheduling method based on elephant flow perception according to claim 1, characterized in that, The time-varying bipartite graph is defined as follows: in, This is the set of active inter-satellite links at the current moment. This is the candidate path set for the elephant stream; if inter-satellite links Included in actual candidate paths In the middle, the adjacency relationship between links and paths is established. .

5. The learning-accelerated low-Earth orbit satellite network traffic scheduling method based on elephant flow perception according to claim 4, characterized in that, The update rule of the graph neural network is as follows: in, This is an augmented double adjacency matrix that includes the adjacency relationships between the inter-satellite links and the actual candidate paths. For graph neural networks The feature matrix of the layer, It is a non-linear activation function. For the first Multilayer perceptron.

6. The learning-accelerated low-Earth orbit satellite network traffic scheduling method based on elephant flow perception according to claim 1, characterized in that, The demand coordination layer updates features using the following formula: in, These are the updated candidate path features. To serve the same elephant stream communication needs The set of candidate paths, where POOL represents the feature pooling operation, and q is the set. Any candidate path in the list, For the first Deep neural networks.

7. The learning-accelerated low-Earth orbit satellite network traffic scheduling method based on elephant flow perception according to claim 1, characterized in that, The multi-agent reinforcement learning algorithm employs a centralized training and distributed execution architecture, and introduces a counterfactual baseline to calculate the advantage function of each independent agent. Its formula is: in, For the global action value function, This is the global state. For the joint action of all intelligent agents, For the local flow features of the k-th agent, For the policy network, For the joint action of the other intelligent agents excluding the k-th intelligent agent, Let k be the action space of the k-th agent. In the action space Action variables for inner traversal.

8. A learning-accelerated low-Earth orbit satellite network traffic scheduling system based on elephant flow perception, comprising the method described in any one of claims 1-7, characterized in that, include: Flow identification and offloading module: Obtain the global traffic demand matrix of the low-Earth orbit satellite network at the current moment and calculate the dynamic identification threshold; Node pairs with traffic demand greater than or equal to the dynamic identification threshold are identified as elephant flows, otherwise they are identified as mouse flows; the mouse flows are unloaded along the static shortest path, and the bandwidth capacity occupied by the mouse flows is deducted from the inter-satellite links of the low-Earth orbit satellite network to obtain the remaining inter-satellite link capacity. Path instantiation module: For each of the elephant flows, calculate the relative offset vector between its source node and destination node; based on the translation invariance of low-orbit satellite + grid topology, use the translation operator to map the reference path sequence corresponding to the relative offset vector in the pre-computation cache to the actual candidate path, and obtain the elephant flow candidate path set; State embedding module: Models the competition relationship between the inter-satellite links with the remaining inter-satellite link capacity and the elephant flow candidate path set as a time-varying bipartite graph; performs message passing on the time-varying bipartite graph using a graph neural network, and introduces a demand coordination layer to pool and concatenate the features of candidate paths serving the same elephant flow communication demand, and outputs high-order flow feature embedding. Allocation Module: Each elephant flow is modeled as an independent agent, and the high-order flow features are embedded into the policy network corresponding to the independent agent. The multi-agent reinforcement learning algorithm is used to output the traffic splitting ratio of each elephant flow on the corresponding actual candidate path in parallel.

9. A computer storage medium storing a readable program, characterized in that, When the program runs, it can instruct the computing device to execute the learning-accelerated low-Earth orbit satellite network traffic scheduling method based on elephant flow perception as described in any one of claims 1-7.

10. An electronic device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the learning-accelerated low-Earth orbit satellite network traffic scheduling method based on elephant flow perception as described in any one of claims 1-7.

Images