A Domain-Specific Intelligent Routing Method for Large-Scale Low-Earth Orbit Satellite Networks

Abstract

This invention discloses a domain-specific intelligent routing method for large-scale low-Earth orbit (LEO) satellite networks, belonging to the field of satellite communication technology. This method constructs a network model based on multi-dimensional resource characteristics and combines an improved Transformer architecture regional traffic prediction algorithm to provide resource constraints and spatiotemporal correlation characteristics for routing decisions. Based on geographical region division into virtual nodes and autonomous domains, the LEO satellite network routing process is decoupled into intra-domain and inter-domain parts. Intra-domain routing employs a multi-agent deep Q-network, while inter-domain routing uses a summation tree-based deep Q-network. Routing and resource scheduling algorithms are designed separately for each domain. Through hierarchical routing and decoupled resource scheduling, a domain-specific intelligent routing system with both flexibility and scalability is constructed. This invention significantly enhances the service capabilities of LEO satellite networks under extreme scenarios such as high-frequency topology changes, strictly limited resources, and sudden spatiotemporal changes in traffic, providing a highly reliable, low-latency routing solution for integrated space-ground communication, emergency disaster early warning, and global real-time data transmission.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of satellite communication, and particularly relates to a large-scale low-orbit satellite network sub-domain intelligent routing method. BACKGROUND

[0002] With the rapid development of low-orbit satellite network technology, it shows great potential in space-ground integrated communication, global coverage and real-time data transmission. However, low-orbit satellite network faces technical challenges such as high dynamic topology, strict satellite-borne resource constraints and multi-dimensional space-time traffic distribution imbalance. Due to the low orbit height, the network topology structure of low-orbit satellite changes frequently with the satellite operation, which makes it difficult for traditional routing mechanism to adapt to the demand of rapid topology reconstruction. At the same time, low-orbit satellite network faces multi-dimensional resource constraints such as limited on-board computing power, limited storage resources and dynamic energy supply, which makes the routing optimization problem more complex. The traffic distribution of low-orbit satellite network has significant space-time imbalance characteristics. The geographical distribution of ground user demand, the dynamic coverage of satellite constellation and the time-varying characteristics of inter-satellite link lead to the multi-scale space-time correlation and nonlinear evolution law of network traffic, and traditional traffic prediction methods are difficult to achieve high-precision prediction.

[0003] In practical applications, the routing optimization technology of low-orbit satellite network faces the following key problems: traditional routing algorithms are mainly designed based on static network model, which is difficult to effectively cope with the high-frequency changes of low-orbit satellite network topology and dynamic resource constraints, resulting in slow routing convergence speed and low resource utilization. Existing traffic prediction methods are mostly based on static space-time model of independent data source, which is difficult to effectively fuse the space-time correlation of multi-dimensional data, resulting in insufficient filtering of redundant information, deviation of key feature extraction, and difficulty in adapting to real-time analysis requirements of constellation dynamic reconstruction and environmental mutation. Traditional routing optimization methods ignore the strong coupling characteristics of multi-dimensional resources in satellite network, and the dynamic correlation of orbit configuration and energy load, the chain feedback of communication energy consumption and battery life, and other cross-domain coupling effects, making it difficult to achieve global optimization of routing decision.

[0004] In view of the above problems, the present application provides a large-scale low-orbit satellite network sub-domain intelligent routing method, which has the following innovation points and advantages:

[0005] The low-orbit satellite network simulation modeling is realized based on multi-dimensional resource feature extraction and representation, and the influence weight of multi-factor coupling on network performance is quantified to provide resource constraint conditions and space-time correlation features for routing algorithm, including space environment model, energy consumption model, illumination model, on-orbit life model, and multi-dimensional resource coupling characteristics and dynamic interaction model. The space environment model designs constellation configuration, inter-satellite link and communication link model based on orbital parameters such as altitude, inclination, number of orbital planes, number of satellites in each orbit, and phase factor of adjacent orbital planes. The energy consumption model quantifies satellite antenna receiving power, transmitting power and data transmission energy consumption, and establishes the mathematical relationship between satellite communication energy consumption and transmission distance, data volume and channel state. The illumination model predicts the solar panel power generation efficiency and the influence of the eclipse period on satellite power supply through the real-time interaction between satellite orbital position and earth shadow area. The on-orbit life model constructs the relationship between battery reliability and lost life according to the discharge depth of lithium-ion battery. The multi-dimensional resource coupling characteristics dynamic interaction model constructs the coupling mechanism and dynamic correlation of multi-dimensional resources by analyzing the space environment, energy consumption, illumination and on-orbit life models, and carries out collaborative modeling and system optimization.

[0006] The low-orbit satellite network regional traffic prediction algorithm is realized based on the improved Transformer architecture, a global traffic tidal effect model is constructed, and the prediction results are mapped to the routing decision state space to drive resource allocation and congestion avoidance strategy, including data traffic modeling and low-orbit satellite network traffic prediction algorithm based on patch time series transformer (PatchTST). The earth surface is divided into 288 traffic blocks according to the latitude and longitude every 15°, and the traffic intensity is converted according to the real statistical data to complete the low-orbit satellite network traffic representation of each traffic block. The original traffic sequence is decomposed into multiple single frequency and independent sub-sequences using empirical mode decomposition (EMD) technology, and the PatchTST model is used to predict each sub-sequence, capture the complex patterns and long-term dependencies in time series, and finally improve the prediction accuracy and computational efficiency.

[0007] A large-scale low-Earth orbit (LEO) satellite network model is constructed based on the division of virtual nodes and autonomous regions in geographic regions. Specifically, this includes a hexagonal grid division method for geographic regions and a LEO satellite network routing method based on the fusion of virtual nodes and autonomous regions. The Earth's surface is divided into equivalent coverage areas using a hexagonal grid, with each grid cell defined as a virtual node and autonomous region. By spatially associating the satellite's ground projection position with the Earth's surface, the hexagonal grid is mapped onto the constellation. Satellites in the same block belong to the same autonomous region and are aggregated into an abstract virtual node. Thus, the routing process is decoupled into intra-domain and inter-domain parts.

[0008] An improved multi-agent deep Q-network (MADQN) is used to implement a routing and resource scheduling algorithm within the low-Earth orbit (LEO) satellite network domain. Specifically, it includes a real-time Markov decision process (RTMDP) and a self-attention mechanism. The routing and resource scheduling problem within the autonomous system is modeled as a real-time Markov decision process. Each satellite acts as an agent, receiving inputs including multi-dimensional resource feature information from neighboring nodes and information from its previous action. The output action is the transmission direction for the next hop. The self-attention mechanism is used to learn the correlation matrix information of the satellite network state and assign dynamic weights to key matrices, thereby enhancing the multi-agent's understanding and response capabilities to the environmental state.

[0009] A Deep Q-Network with SumTree (DQN-ST) algorithm for inter-domain routing and resource scheduling in low-Earth orbit satellite networks is implemented. Specifically, it includes a dynamically adaptive domain boundary node mapping mechanism and a DQN-ST algorithm architecture optimized for routing tasks. The algorithm updates the domain boundary node and topology connection status in real time through a weighted score of node validity and resource availability, establishing a dynamic mapping relationship for cross-domain routing. An improved priority experience replay mechanism based on SumTree is employed to determine the priority of satellite routing task samples according to the absolute value of the temporal differential error. Priority samples are then sampled and reinforced learning is performed on these priority samples. The optimal next-hop satellite node is selected based on the Q-value, and the aggregated results output as an intelligent decision for inter-domain routing.

[0010] Through these innovations, this invention can solve problems such as insufficient multi-dimensional resource coupling modeling, limited accuracy of spatiotemporal traffic prediction, and difficulty in global optimization of cross-domain routing in low-Earth orbit satellite networks. It realizes routing decision-making and resource scheduling driven by dynamic resource representation and spatiotemporal collaborative optimization, significantly enhancing the service capabilities of low-Earth orbit satellite networks in extreme scenarios such as high-frequency topology changes, strictly limited resources, and spatiotemporal traffic mutations. It provides a highly reliable, low-latency constellation-level routing solution for integrated space-ground communication, emergency disaster early warning, and global real-time data transmission. Summary of the Invention

[0011] In view of this, the present invention provides a domain-specific intelligent routing method for large-scale low-Earth orbit satellite networks to eliminate or improve the defects existing in the prior art, such as insufficient multi-dimensional resource coupling modeling, limited spatiotemporal traffic prediction accuracy, and difficulty in global optimization of cross-domain routing in low-Earth orbit satellite networks. It realizes routing decision-making and resource scheduling driven by dynamic resource representation and spatiotemporal collaborative optimization, thereby enhancing the service capabilities of low-Earth orbit satellite networks in scenarios such as high-frequency topology changes, strictly limited resources, and spatiotemporal traffic mutations.

[0012] This invention provides a simulation modeling method for low-Earth orbit satellite networks based on multi-dimensional resource feature extraction and representation. By defining specific parameters for each orbital layer, designing the connection logic between satellites, and considering factors such as space environment, illumination, and satellite lifetime, a systematic coupled modeling framework is proposed, providing theoretical and technical support for building efficient and stable space-based internet services. Specifically, it includes the following technical features:

[0013] The aforementioned space environment model designs constellation configurations, inter-satellite links, and communication link models based on orbital parameters such as altitude, inclination, number of orbital planes, number of satellites in each orbit, and phase factor of the phase difference between adjacent orbital planes. The neighbor satellite connection relationship of a satellite with a logical address (m, n) is defined by the following formula:

[0014]

[0015] The energy consumption model quantifies the satellite antenna's received power, transmitted power, and data transmission energy consumption, establishing a mathematical relationship between satellite communication energy consumption and transmission distance, data volume, and channel status; and derives the inter-satellite transmission rate using Shannon's formula. Among them, w (i，j) σ is the bandwidth of the wireless channel. 2 Noise power; free space propagation loss during transmission. The expression is Where, d (i，j) Let λ be the distance between satellites, λ be the wavelength of the electromagnetic wave, c be the speed of light, and f be the operating frequency; the antenna receiving power is... Given the satellite's receiving power, its transmitting power can be deduced. The expression is

[0016] The aforementioned illumination model predicts the power generation efficiency of solar panels and the impact of the Earth's shadow period on satellite power supply through real-time interaction between the satellite's orbital position and the Earth's shadow region. A cylindrical shadow model is used to predict the shadow period, with the following constraints: If this formula is not satisfied, it means that the satellite is in a lit area;

[0017] The aforementioned on-orbit lifetime model constructs a relationship between battery reliability and lost lifetime based on the impact of lithium-ion battery discharge depth on battery cycle life consumption; it is based on the lithium-ion battery discharge depth D(t) and cycle life consumption L. t1,t2 The functional relationship is The fitted expression for battery reliability R(t) is as follows:

[0018] The aforementioned dynamic interaction model of multidimensional resource coupling characteristics, by revealing the constraints of orbital configuration on communication link distance, the cumulative impact of energy mode switching caused by ground shadow events on battery life, and the interaction mechanism between dynamic tracking strategies and energy consumption distribution, breaks through the limitations of traditional models that independently model single-dimensional resources. It constructs the coupling mechanism and dynamic correlation of multidimensional resources and carries out collaborative modeling and system optimization. Through Monte Carlo simulation, it quantifies the joint impact of orbital altitude on communication distance and energy load, providing a data and model foundation for subsequent research on low-Earth orbit satellite network traffic prediction and routing technology.

[0019] This invention provides a low-Earth orbit satellite network regional traffic prediction algorithm based on an improved Transformer architecture. It divides the Earth's surface into 288 traffic blocks according to latitude and longitude, and utilizes EMD technology to decompose the complex original traffic sequence into multiple single-frequency and independent IMF subsequences, thereby effectively removing noise and simplifying the model training process. Subsequently, the PatchTST model is used to predict each IMF subsequence. Combining the advantages of the Transformer architecture, it captures complex patterns and long-term dependencies in the time series, ultimately improving prediction accuracy and computational efficiency. Specifically, it includes the following technical features:

[0020] The aforementioned data service traffic modeling divides the Earth's surface into regions every 15° of latitude and longitude, and calculates the traffic intensity based on the number of internet users in each region, thus obtaining the global data service traffic intensity L. n The temporal distribution of traffic is represented by a discount factor γ. Within each traffic block, the longitude of its center is selected as a reference to calculate the time zone of that block. It is assumed that the discount factor is 1 when users are most active and 0 when there is almost no activity. The low-Earth orbit satellite network traffic of any traffic block is represented as C.n =αγL n n = 1, 2, ..., 288, in Mbps;

[0021] The low-Earth orbit satellite network traffic prediction algorithm based on PatchTST employs EMD technology to decompose the original traffic sequence into multiple single-frequency and independent IMF subsequences and a residual signal, effectively separating different frequency components and making each subsequence more stable and easier to model. For each IMF subsequence and residual signal, a PatchTST model is introduced for prediction. The PatchTST model combines the advantages of the Transformer architecture, maintaining high efficiency when processing long sequences and capturing complex patterns and dependencies in the time series. The output results are summed to obtain the prediction output.

[0022] The EMD implementation method assumes the original sequence is X(t), which can be decomposed into a finite number of IMF subsequences c. j and a remaining term r n , The formula for calculating its standard deviation is: In the formula h i-1，j h i，j Let i represent the intrinsic mode function, j represent the number of iterations, and N represent the total number of iterations.

[0023] The method for constructing a traffic prediction model based on PatchTST, for a set of multivariate time series samples, assuming that the length of the review window (i.e., the time data length of the input model) is L, yields samples (x1,...,x...). t ,...,x L ), where, at time slot t, x t It is a one-dimensional column vector containing M elements, predicting the data values ​​(x) for the next T time slots. L+1 ,...,x L+T The model structure includes a forward process, a block process, and a Transformer encoder.

[0024] The aforementioned forward process represents the i-th one-dimensional sequence as follows: Each sequence is independently input into the Transformer backbone network, and the prediction result for that sequence is calculated as follows:

[0025] The block-based process described above involves each input one-dimensional time series x. (i) The data is divided into blocks, with the block length denoted as P and the non-overlapping region between two consecutive blocks denoted as S. After the block division process, a block sequence will be generated. Where N represents the number of blocks generated. The memory usage and computational complexity of attention graphs are expressed in S. 2 The speed is reduced. Therefore, under conditions of limited training time and GPU memory resources, block operations allow the model to understand longer historical sequences, thereby significantly improving the model's predictive performance;

[0026] The Transformer encoder described above uses a trainable linear projection W P ∈R D×P And learnable addition position encoding W pos ∈R D×N Mapping the blocks to a latent representation space of dimension D, the data input to the Transformer encoder is then represented as follows: in Each attention head in a multi-head attention model, h = 1, ..., H, is transformed into a query matrix. Key matrix and value matrix in Attention output is obtained using scaling product. The attention output is processed through a normalization layer and a feedforward network layer with residual connections, resulting in z. (i) ∈R D×N The prediction results are obtained using a fully connected layer with a linear head.

[0027] This invention provides a large-scale low-Earth orbit satellite network model based on the division of virtual nodes and autonomous regions within a geographical region. By performing a simplified cylindrical projection of the Earth's surface and dividing it into 112 regular hexagonal grid cells, each grid cell is defined as a virtual node and autonomous region, thus shielding the impact of high-speed satellite motion on topology dynamics. Subsequently, the surface grid is associated with the constellation through satellite nadir point position mapping, aggregating satellites within the same block into virtual nodes. Global routing is decoupled from intra-domain real-time decision-making and inter-domain cross-domain scheduling, significantly reducing the size of the routing table and computational complexity. Finally, by combining a hybrid strategy of flooding updates and local updates, routing convergence speed is improved and on-board storage resource requirements are reduced in high-frequency topology reconstruction scenarios. Specifically, it includes the following technical features:

[0028] The described hexagonal grid partitioning method for geographic regions involves performing a simplified cylindrical projection of the Earth's surface and converting it into a rectangle with an aspect ratio of 2:1. This rectangle is then divided into 112 regular hexagonal grids, each covering a 15°×15° latitude and longitude area, representing a virtual node and an Autonomous System (AS). The grid mapping rules ensure seamless connection between adjacent virtual nodes, reducing the number of routing hops.

[0029] The low-Earth orbit satellite network routing method based on the fusion of virtual nodes and autonomous systems (AS) maps the hexagonal grid structure of the Earth's surface onto the satellite constellation by spatially associating the satellite's ground projection position with the surface grid. Satellites in the same block belong to the same AS and are aggregated into an abstract virtual node. The routing and resource information update strategy combines flooding and local updates. Flooding updates inter-domain information, spreading the boundary satellite node information of the block to the entire satellite constellation. Local updates update intra-domain information, sharing and updating information such as the status of each satellite and link within the AS.

[0030] This invention provides an improved MADQN-based routing and resource scheduling algorithm for low-Earth orbit satellite networks. By modeling the problem as a real-time Markov decision process, it addresses the real-time performance limitations of traditional MDP frameworks in dynamic environments. Subsequently, the MADQN algorithm is employed, with each satellite acting as an agent. Inputs include satellite state information and link state information from neighboring nodes, and the output action is one of four possible transmission directions. To more effectively handle complex and dynamically changing state information, a self-attention mechanism is introduced to learn the relevant key matrices of the satellite network state and dynamically allocate weights, thereby enhancing the understanding and response capabilities to environmental states. Specifically, it includes the following technical features:

[0031] The real-time Markov decision process described above has a state space s. t This includes satellite power consumption, link quality, battery reliability, neighbor node status, and the result of the satellite node's last action decision, action space a. t For the transmission directions of "up, down, left, and right", the reward function r(s) t a t This includes comprehensive latency, energy consumption, and link stability. In the RTMDP framework, actions and states change simultaneously. The agent can observe the next state immediately after selecting an action and directly infer the next action. This real-time decision-making mechanism allows the agent to track rapidly changing environmental conditions more closely, ensuring that routing and resource scheduling can respond promptly to changes in the topology and state of the low-Earth orbit satellite network.

[0032] The expression for the action value function is shown below:

[0033]

[0034] In the formula, Indicates that in a given state s t and action a t Transition to a specific state s t+1 The expected probability. Represents an action pair (s) in a given state. t+1 ,a tUnder the condition of π, execute action a according to the current strategy π. t+1 Subsequently, the expected cumulative reward for the sequence of states that the agent may encounter;

[0035] The state value function expression is shown in the following equation:

[0036]

[0037] In the formula, Represents an action pair (s) in a given state. t ,a t The expected cumulative reward for the sequence of states that the agent may encounter under the given conditions.

[0038] The self-attention mechanism described above maps the input neighbor node state information into query matrices by constructing a multi-head attention structure. Key matrix and value matrix Attention weights are calculated according to the formula Where d k The dimension scaling factor of the key vector is used to avoid the gradient saturation problem of the Softmax function. By computing the outputs of multiple attention heads in parallel and concatenating them into a joint representation, the algorithm can dynamically allocate weights, highlight the states of neighboring nodes with high link stability and high resource redundancy, and suppress noise interference. The attention output is then input into the fully connected network after residual connection and layer normalization to generate routing decision actions.

[0039] This invention provides an inter-domain routing and resource scheduling algorithm for low-Earth orbit satellite networks based on DQN-ST. By designing a dynamically adaptive domain boundary node mapping mechanism, it comprehensively considers node validity and resource availability to solve the boundary node failure problem caused by high-speed satellite motion. Subsequently, it proposes a priority experience replay deep Q-network based on an improved Sumtree, achieving global collaborative optimization of cross-domain routing and resource scheduling through task-aware state-space modeling and dynamic bandwidth allocation mechanisms. Specifically, it includes the following technical features:

[0040] The aforementioned dynamic adaptive domain boundary node mapping mechanism addresses the issue of intra-domain nodes forwarding requests to domain boundary nodes when cross-autonomous domain communication is required. Due to the continuous high-speed movement of satellites, satellite nodes may move from one region to another, causing the previously selected domain boundary nodes and inter-domain communication links to become invalid. The mechanism comprehensively considers node effectiveness and available resources to evaluate intra-domain satellite nodes and select the mapping satellites for the domain boundary nodes. The calculation of node effectiveness considers not only the geometric distance to adjacent domains at the current moment but also the time sensitivity of specific routing tasks. The evaluation of node resource indicators requires dynamic weighting based on actual business needs. The node effectiveness E(s) is defined as follows:i ) = dis(s i )s i ∈{s1, s2, ..., s n}, node available resources R(s) i ) = Cal(s i )+Cache(s i )s i ∈{s1, s2, ..., s n If the weighting coefficients of the two indicators are set to 0.7 and 0.3 respectively, then the standardized weighted composite score will be... Mapped to actual routing decisions, the system can automatically select boundary nodes that meet both the Quality of Service (QoS) requirements and have the best resource utilization.

[0041] The DQN-ST algorithm architecture optimized for routing tasks described above applies a priority experience replay mechanism based on Sumtree to the deep Q-network algorithm, with the state space s t = {l(t), u(t), τ(t), γ(t)}, where l(t) represents the satellite node where the task in time slot t is located, u(t) represents the link connection status of the satellite node where the task in time slot t is located, τ(t) represents the data size of the task in time slot t, and γ(t) represents the network status information of the virtual constellation in time slot t; action space a t = {h(t), B(t)}, where h(t) represents the satellite node for the agent to perform the next forwarding in time slot t; B(t) represents the bandwidth resources allocated to the task in time slot t; the priority experience replay mechanism based on Sumtree considers the spatiotemporal correlation of routing tasks in actual deployment, and assigns higher TD error weights to task samples with high latency sensitivity or high bandwidth requirements, with priority p i =|δ(i)|+∈, where ∈ is a minimal constant, and the sampling probability P(i) is calculated according to the formula Calculation; Based on the above calculation and according to the Q value calculation results, select the optimal next-hop satellite node h(t) to ensure the efficiency and stability of cross-domain communication.

[0042] The beneficial effects of this invention are as follows:

[0043] This invention provides a domain-based intelligent routing method for large-scale low-Earth orbit (LEO) satellite networks, belonging to the field of satellite communication technology. This method is based on LEO satellite network simulation modeling using multi-dimensional resource feature extraction and characterization, providing resource constraints and spatiotemporal correlation characteristics for subsequent routing algorithms. An improved Transformer architecture-based regional traffic prediction algorithm for LEO satellite networks constructs a global traffic tidal effect model to drive routing decisions and resource scheduling. A large-scale LEO satellite network model based on geographically divided virtual nodes and autonomous systems (ASIs) is used, with hexagonal networks effectively covering the Earth's surface to obtain ASIs, decoupling the routing process into intra-domain and inter-domain components. An improved MADQN-based intra-domain routing and resource scheduling algorithm for LEO satellite networks achieves an efficient real-time decision-making mechanism, ensuring that routing and resource scheduling can respond promptly to changes in network topology and state. An DQN-ST-based inter-domain routing and resource scheduling algorithm for LEO satellite networks solves the problem of boundary node failure caused by high-speed satellite movement, achieving global collaborative optimization of cross-domain routing and resource scheduling.

[0044] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0045] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0046] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings:

[0047] Figure 1 This is a schematic diagram illustrating the steps of a domain-based intelligent routing method for a large-scale low-Earth orbit satellite network according to an embodiment of the present invention.

[0048] Figure 2 This is a flowchart illustrating a domain-based intelligent routing method for large-scale low-Earth orbit satellite networks according to an embodiment of the present invention.

[0049] Figure 3 This is a schematic diagram of a low-orbit satellite network simulation modeling method and a multi-dimensional resource coupling model in one embodiment of the present invention.

[0050] Figure 4 This is a flowchart illustrating a low-Earth orbit satellite network traffic prediction method according to an embodiment of the present invention.

[0051] Figure 5 This is a flowchart illustrating an empirical mode decomposition method according to an embodiment of the present invention.

[0052] Figure 6 This is a framework diagram of an intra-domain routing and resource scheduling algorithm based on MADQN in one embodiment of the present invention.

[0053] Figure 7 This is a diagram of an algorithm architecture for inter-domain routing optimization based on DQN-ST in one embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0055] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0056] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0057] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0058] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0059] To address the challenges of complex multi-factor coupling and dynamic characteristic characterization in large-scale low-Earth orbit (LEO) satellite networks, the difficulty in predicting traffic in spatiotemporally unbalanced LEO networks, and the inability of LEO network routing technologies to cope with on-orbit changes and environmental impacts, this invention provides a domain-based intelligent routing method for large-scale LEO satellite networks. This method is based on a multi-dimensional resource feature extraction and characterization-based simulation modeling approach for LEO satellite networks and an improved Transformer architecture-based regional traffic prediction algorithm for LEO satellite networks. This provides resource constraints and spatiotemporal correlation characteristics for routing decisions. Based on a large-scale LEO satellite network model that divides virtual nodes and autonomous domains into geographical regions, the routing process is decoupled into intra-domain and inter-domain parts. An improved MADQN-based intra-domain routing and resource scheduling algorithm and a DQN-ST-based inter-domain routing and resource scheduling algorithm for LEO satellite networks are used to achieve hierarchical routing and decoupled resource scheduling. This breaks through the traditional single-layer routing framework and constructs a domain-based intelligent routing system that balances flexibility and scalability. Figure 1 As shown, the method includes the following steps S101 to S105:

[0060] Step S101: Construct a multi-dimensional dynamic characterization model of low-Earth orbit satellite networks to quantify resource constraints and spatiotemporal correlation characteristics. By receiving satellite constellation configuration parameters and real-time operating status, establish space environment models, energy consumption models, illumination models, and on-orbit lifetime models to reveal the coupling mechanism and dynamic interaction characteristics of multi-dimensional resources.

[0061] Step S102: Implement low-Earth orbit satellite network traffic prediction to support efficient and accurate resource management and scheduling decisions, avoiding network congestion and resource waste. Divide the Earth's surface into 288 latitude and longitude blocks, separate the noise and core components of the traffic sequence using empirical mode decomposition technology, and capture long-range spatiotemporal dependencies using a block time series transformer. Map the prediction results to a spatial state that can be parsed for routing decisions.

[0062] Step S103: Construct a virtual node topology to decouple the global routing optimization problem. A hexagonal grid is used to divide the Earth's surface into equivalent coverage areas. The constellation is mapped to 112 autonomous regions based on the satellite's sub-satellite point positions, forming an intra-regional satellite cluster topology. A hybrid update strategy is designed to reduce on-board storage resource requirements.

[0063] Step S104: Perform real-time routing decisions and resource scheduling in a dynamic environment within the autonomous system to optimize local network transmission efficiency and resource utilization. Using a multi-agent collaborative learning framework, each satellite is modeled as an independent decision-making unit, taking into account the link status, remaining energy, and historical action sequences of neighboring nodes, and outputting the next-hop transmission direction. A self-attention mechanism is used to dynamically evaluate the weights of critical links, and a multi-dimensional reward function drives policy gradient updates, enabling rapid convergence and load balancing of satellite nodes in scenarios with drastic topology changes.

[0064] Step S105: Implement global routing optimization and priority resource collaborative scheduling in cross-domain scenarios to ensure the service quality of high-value services. By comprehensively considering indicators such as node effectiveness and available node resources, evaluate and select the mapping satellites for domain boundary nodes. Utilize a priority experience replay mechanism based on a Sumtree structure to globally perceive the network status of the virtual constellation composed of virtual nodes, achieving more efficient and accurate routing selection and resource scheduling.

[0065] In this invention, the low-Earth orbit satellite network parameters consist of satellite constellation configuration parameters and real-time operating status data.

[0066] like Figure 2 The diagram shows a flowchart of a domain-based intelligent routing method for large-scale low-Earth orbit satellite networks.

[0067] In step S101, a multi-dimensional resource coupling characterization framework is established to reveal the influence of the interaction between orbital parameters, energy load, communication requirements and environmental constraints on the dynamic characteristics of the network.

[0068] Specifically, the publicly available datasets of major international large-scale operational constellations are incomplete, involve complex multi-factor coupling, and are difficult to simulate dynamic characteristics. Therefore, it is necessary to construct network structure and communication resource models, energy and environmental dynamic models, and multi-dimensional resource coupling characteristics and dynamic interaction models. In practical implementation, by defining specific parameters for each orbital layer, designing the connection logic between satellites, and considering factors such as space environment, illumination, and satellite lifetime, a systematic coupling modeling framework is proposed, providing theoretical and technical support for building efficient and stable space-based internet services.

[0069] In step S102, a traffic tidal variation model is constructed to capture complex patterns and dependencies in the time series, improve the accuracy and efficiency of traffic prediction, and support efficient and accurate low-orbit satellite network resource management and scheduling decisions.

[0070] Specifically, traditional satellite network routing algorithms and flow control strategies suffer from serious lag issues. By dividing the Earth's surface into 288 flow blocks according to latitude and longitude, and using EMD technology to decompose the complex original flow sequence into multiple single-frequency and independent IMF subsequences, noise is effectively removed and the model training process is simplified. Subsequently, the PatchTST model is used to predict each IMF subsequence. Combining the advantages of the Transformer architecture, long-term series data is processed efficiently, ultimately improving prediction accuracy and computational efficiency.

[0071] In step S103, autonomous domains and virtual nodes are divided, and the complex global routing problem is decomposed into two levels that are easier to handle: local and global. This makes routing selection and resource scheduling more efficient and flexible, and enables hierarchical routing within and between domains.

[0072] Specifically, the virtual node concept is an effective solution to shield against the influence of satellite motion. In traditional virtual node-based routing algorithms, each virtual node is assigned a fixed geographical location. During constellation operation, only one satellite is providing services to ground users at any given time, failing to fully utilize satellite resources. By performing a simplified cylindrical projection of the Earth's surface and dividing it into 112 regular hexagonal grid cells, each grid is defined as a virtual node and autonomous region, thus shielding the impact of high-speed satellite motion on topology dynamics. Subsequently, the surface grid is associated with the constellation through satellite sub-satellite point location mapping, aggregating satellites within the same block into virtual nodes, and decoupling global routing into intra-domain real-time decision-making and inter-domain cross-domain scheduling, significantly reducing the size of the routing table and computational complexity. Finally, by combining a hybrid strategy of flooding updates and local updates, routing convergence speed is improved and on-board storage resource requirements are reduced in high-frequency topology reconstruction scenarios.

[0073] In step S104, MADQN is introduced for intra-domain routing and resource scheduling, realizing an efficient real-time decision-making mechanism and ensuring that routing and resource scheduling can respond promptly to changes in network topology and status.

[0074] Specifically, traditional MDP frameworks have limitations for highly dynamic and real-time-critical systems like low-Earth orbit satellite networks. This paper addresses this by modeling the problem as a real-time Markov decision process using the MADQN algorithm. Each satellite acts as an agent, receiving inputs including satellite state information from neighboring nodes and link state information, and outputting one of four possible transmission directions. To more effectively handle complex and dynamically changing state information, a self-attention mechanism is introduced to learn the relevant key matrices of the satellite network state and dynamically assign weights, thereby enhancing the understanding and response capabilities to environmental conditions.

[0075] In step S105, the DQN-ST algorithm is introduced to perceive the network state of the virtual constellation composed of virtual nodes from a global perspective, improve the generalization ability of the model, avoid overfitting, and ensure the efficiency and stability of cross-domain communication.

[0076] Specifically, for satellite networks, due to the continuous high-speed movement of satellites, satellite nodes constantly move from one region to the next. This causes the originally selected domain boundary nodes and inter-domain communication links to become invalid, requiring the satellite network to recalculate and select new boundary nodes. By designing a dynamically adaptive domain boundary node mapping mechanism that comprehensively considers node validity and resource availability, the problem of boundary node failure caused by high-speed satellite movement is solved. Subsequently, a priority experience replay deep Q-network based on an improved Sumtree is proposed. Through task-aware state-space modeling and dynamic bandwidth allocation mechanisms, global collaborative optimization of cross-domain routing and resource scheduling is achieved.

[0077] In some embodiments, multi-dimensional resource feature extraction and characterization are performed based on the low-Earth orbit satellite network simulation modeling method provided by the present invention to obtain a multi-dimensional resource coupling model, such as... Figure 3 As shown, the specific steps include S301 to S305:

[0078] Step S301: Construct a space environment model. Based on orbital parameters such as altitude, inclination, number of orbital planes, number of satellites in each orbit, and phase factor of adjacent orbital planes, design the constellation configuration, inter-satellite links, and communication link models. The neighbor satellite connection relationship of satellite logical address (m,n) is defined by the following formula:

[0079]

[0080] Step S302: Construct an energy consumption model to quantify the satellite antenna's received power, transmitted power, and data transmission energy consumption; establish the mathematical relationship between satellite communication energy consumption and transmission distance, data volume, and channel status; derive the inter-satellite transmission rate using Shannon's formula. Among them, w (i,j) σ is the bandwidth of the wireless channel. 2 Noise power; free space propagation loss during transmission. The expression is Where, d (i,j) Let λ be the distance between satellites, λ be the wavelength of the electromagnetic wave, c be the speed of light, and f be the operating frequency; the antenna receiving power is... Given the satellite's receiving power, its transmitting power can be deduced. The expression is

[0081] Step S303: Construct an illumination model to predict the solar panel power generation efficiency and the impact of the Earth's shadow period on satellite power supply through real-time interaction between the satellite's orbital position and the Earth's shadow region; use a cylindrical shadow model to predict the shadow period, with the following constraints: If the equation is not satisfied, it means that the satellite is in a lit area.

[0082] Step S304: Construct an on-orbit lifetime model. Based on the effect of lithium-ion battery discharge depth on battery cycle life consumption, establish the relationship between battery reliability and lost lifetime; based on the lithium-ion battery discharge depth D(t) and cycle life consumption L... t1,t2 The functional relationship is The fitted expression for battery reliability R(t) is as follows:

[0083] Step S305: Construct a dynamic interaction model of multidimensional resource coupling characteristics. By revealing the constraints of orbital configuration on communication link distance, the cumulative impact of energy mode switching caused by ground shadow events on battery life, and the interaction mechanism between dynamic tracking strategies and energy consumption distribution, this model breaks through the limitations of traditional models that independently model single-dimensional resources. It constructs the coupling mechanism and dynamic correlation of multidimensional resources and conducts collaborative modeling and system optimization. Monte Carlo simulation is used to quantify the joint impact of orbital altitude on communication distance and energy load, providing data and model foundation for subsequent research on low-Earth orbit satellite network traffic prediction and routing technology.

[0084] like Figure 4 As shown, the low-Earth orbit satellite network regional traffic prediction algorithm based on the improved Transformer architecture of the present invention specifically includes the following steps S401 to S405:

[0085] Step S401: Data service traffic modeling. The Earth's surface is divided into regions every 15° of latitude and longitude, resulting in 288 traffic blocks. The data service traffic intensity is calculated based on the number of internet users within each block, treating each person as a node. 6 The traffic intensity of each node is equivalent to 1Mbps. The global data service traffic intensity is obtained based on statistics of the actual number of global internet users, and the traffic intensity of each traffic block is L. n Since most traffic will be distributed via terrestrial communication networks, we can assume that the proportion of traffic distributed via LEO satellite networks, α, is set to 5%. The temporal distribution of traffic is represented by a discount factor γ. Within each traffic block, the longitude of its center is selected as a reference to calculate the time zone of that block. Assuming a discount factor of 1 for peak user activity and 0 for near-inactive periods, a tidal variation in traffic can be constructed. The LEO satellite network traffic for any traffic block is represented as C. n=ɑγL n n = 1, 2, ..., 288, in Mbps.

[0086] Step S402: Empirical Mode Decomposition (EMD) is used to decompose the complex original flow sequence into multiple single-frequency and independent IMF subsequences. Each IMF subsequence is not only easier to model, but also retains the key features of the original signal, enabling subsequent prediction models to better capture short-term fluctuations and long-term trends. Assuming the original sequence is X(t), this sequence can be decomposed into a finite number of IMF subsequences c. j and a remaining term r n ,Right now The formula for calculating its standard deviation is: In the formula h i-1，j h i，j Let represent the intrinsic mode function, i represent the iteration number, j represent the function order, and N represent the total number of iterations. The specific implementation flow of the EMD algorithm is as follows: Figure 5 As shown.

[0087] Step S403: Implement the forward process of the PatchTST model, representing the i-th one-dimensional sequence as... Each sequence is independently input into the Transformer backbone network, and the prediction result for that sequence is calculated as follows:

[0088] Step S404: Implement the block segmentation process of the PatchTST model, where each input is a one-dimensional time series x. (i) The data is divided into blocks, which can be overlapping or non-overlapping. Let P be the block length and S be the non-overlapping region between two consecutive blocks. The resulting block sequence is then generated. Where N represents the number of blocks generated. After performing block partitioning, the input token size can be reduced from L to approximately L / S, and the memory usage and computational complexity of the attention graph are reduced by S. 2 The speed is reduced. Therefore, under conditions of limited training time and GPU memory resources, block operations allow the model to understand longer historical sequences, thereby significantly improving the model's predictive performance.

[0089] Step S405: Implement the Transformer encoder using a trainable linear projection W P ∈R D×P And learnable addition position encoding W pos ∈R D×NMapping the blocks to a latent representation space of dimension D, the data input to the Transformer encoder is then represented as follows: in Each attention head in a multi-head attention model, h = 1, ..., H, is transformed into a query matrix. Key matrix and value matrix in Attention output is obtained using scaling product. The attention output is processed through a normalization layer and a feedforward network layer with residual connections, resulting in z. (i) ∈R D×N The prediction results are obtained using a fully connected layer with a linear head.

[0090] The large-scale low-Earth orbit satellite network model based on geographical region division of virtual nodes and autonomous domains provided by this invention specifically includes the following steps S601 to S602:

[0091] Step S601: Divide the Earth's surface into a hexagonal grid based on the geographic region. However, since it is impossible to strictly divide the sphere into a hexagonal grid, the Earth's surface is subjected to a simplified cylindrical projection and converted into a rectangle with an aspect ratio of 2:1. This rectangle is then divided into 112 regular hexagonal grids, each grid representing a virtual node and an autonomous region.

[0092] Step S602: Based on the decoupling of virtual nodes and autonomous systems, the routing method for low-Earth orbit satellite networks maps the hexagonal grid structure of the Earth's surface onto the satellite constellation by spatially associating the satellite's ground projection position with the surface grid. Satellites in the same block will belong to the same autonomous system and aggregate into an abstract virtual node. A routing and resource information update strategy combining flooding and local updates is adopted. Inter-domain information is updated through flooding, flooding the boundary satellite node information of the block to the entire satellite constellation. Intra-domain information is updated through local updates, sharing and updating information such as the status of each satellite and link within the autonomous system.

[0093] This invention also provides an improved MADQN-based intra-domain routing and resource scheduling algorithm for low-Earth orbit satellite networks, the framework of which is as follows: Figure 6 As shown, the specific steps include S701 to S702:

[0094] Step S701: Model a real-time Markov decision process, with state space s t This includes satellite power consumption, link quality, battery reliability, neighbor node status, and the result of the satellite node's last action decision, action space a. t For the transmission directions of "up, down, left, and right", the reward function r(s)t a t This includes comprehensive latency, energy consumption, and link stability. In the RTMDP framework, actions and states change simultaneously. After selecting an action, the agent can immediately observe the next state and directly infer the next action. This real-time decision-making mechanism allows the agent to more closely track rapidly changing environmental conditions, ensuring that routing and resource scheduling can respond promptly to changes in the topology and state of the low-Earth orbit satellite network.

[0095] The expression for the action value function is shown below:

[0096]

[0097] In the formula, Indicates that in a given state s t and action a t Transition to a specific state s t+1 The expected probability. Represents an action pair (s) in a given state. t+1 ,a t Under the condition of π, execute action a according to the current strategy π. t+1 Subsequently, the expected cumulative reward for the sequence of states that the agent may encounter;

[0098] The state value function expression is shown in the following equation:

[0099]

[0100] In the formula, Represents an action pair (s) in a given state. t a t The expected cumulative reward for the sequence of states that the agent may encounter under the given conditions.

[0101] Step S702: Introduce a self-attention mechanism. By constructing a multi-head attention structure, the input neighbor node state information is mapped to query matrices respectively. Key matrix and value matrix Attention weights are calculated according to the formula Where d k The dimension scaling factor of the key vector is used to avoid the gradient saturation problem of the Softmax function. By computing the outputs of multiple attention heads in parallel and concatenating them into a joint representation, the algorithm can dynamically allocate weights, highlight the states of neighboring nodes with high link stability and high resource redundancy, and suppress noise interference. The attention output is then input into the fully connected network after residual connection and layer normalization to generate routing decision actions.

[0102] The DQN-ST-based inter-domain routing and resource scheduling algorithm for low-Earth orbit satellite networks provided by this invention has the following framework: Figure 7 As shown, the specific steps include S801 to S802:

[0103] Step S801: Implement dynamic adaptive domain boundary node mapping. Considering indicators such as node validity and available resources, evaluate satellite nodes within the domain and select the mapping satellites for the domain boundary nodes. In actual satellite network cross-domain routing scenarios, the selection of domain boundary nodes needs to be deeply coupled with the characteristics of dynamic routing tasks. When a satellite node detects the need to forward a cross-domain service request, it predicts a candidate node set {s1, s2, ..., s...} based on ephemeris data. n The orbital position within a future time interval Δt, where the value of Δt is closely related to the lifecycle of a typical routing task. Node effectiveness metric E(s) i ) = dis(s i The calculation of the node resource metric R(s) considers not only the geometric distance to the adjacent domain at the current moment, but also the time sensitivity of the specific routing task. For low-latency services such as real-time video streaming, nodes closer to the next domain are selected to shorten propagation latency. For batch data transmission tasks, nodes with longer remaining service time are preferred to ensure connection continuity. i ) = Cal(s i )+CacHe(s i The evaluation needs to be dynamically weighted to meet actual business needs. By mapping the standardized score Score = 0.7E′ + 0.3R′ to the actual routing decision, the system can automatically select the boundary node that can meet the business service quality requirements and has the best resource utilization.

[0104] The validity of the node is represented by the following formula:

[0105] E(s i ) = dis(s i )s i ∈{s1,s2,…,s n}

[0106] The available resources of the node are represented by the following formula:

[0107] R(s i ) = Cal(s i )+Cache(s i )s i ∈{s1,s2,…,s n}

[0108] The weight coefficients of the two indicators are set to 0.7 and 0.3 respectively, and the standardized weighted comprehensive score is shown in the following formula.

[0109]

[0110] Step S802: Implement the task-oriented optimized DQN-ST algorithm architecture, applying the priority experience replay mechanism based on Sumtree to the deep Q-network algorithm, with state space s t = {l(t), u(t), τ(t), γ(t)}, where l(t) represents the satellite node where the task in time slot t is located, u(t) represents the link connection status of the satellite node where the task in time slot t is located, τ(t) represents the data size of the task in time slot t, and γ(t) represents the network status information of the virtual constellation in time slot t; action space a t = {h(t), B(t)}, where h(t) represents the satellite node for the agent to perform the next forwarding in time slot t; B(t) represents the bandwidth resources allocated to the task in time slot t.

[0111] In the state space design, l(t) not only includes the current satellite node's location information but also integrates the routing task queue status undertaken by the node (such as the number of cross-domain connections being processed and the priority of packets to be forwarded), enabling the agent to perceive the impact of local load on global routing. The link connection status in u(t) constructs a link stability matrix using parameters such as actual measured bit error rate and latency jitter, providing physical layer support for action selection. The bandwidth allocation B(t) in the action space needs to be linked with the service type of the specific routing task. For example, to meet the sudden large data transmission needs in low-orbit satellite Earth observation missions, the algorithm expands the dynamic adjustment range of B(t) and establishes a nonlinear relationship model between bandwidth allocation and throughput and packet loss rate through online learning. The priority experience replay mechanism based on Sumtree needs to consider the spatiotemporal correlation of routing tasks in actual deployment. By assigning higher TD error weights to task samples with high latency sensitivity or high bandwidth requirements, priority p i =|δ(i)|+∈, where ∈ is a minimal constant, and the sampling probability P(i) is calculated according to the formula Calculation; Based on the above calculations and Q-value results, select the optimal next-hop satellite node h(t) to ensure the efficiency and stability of cross-domain communication.

[0112] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a domain-based intelligent routing method for large-scale low-Earth orbit satellite networks.

[0113] Corresponding to the above method, the present invention also provides an apparatus comprising a computer device, the computer device including a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the apparatus performs the steps of the method as described above.

[0114] This invention also provides a computer-readable storage medium having a computer program stored thereon. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium known in the art.

[0115] In summary, this invention provides a domain-based intelligent routing method for large-scale low-Earth orbit (LEO) satellite networks, belonging to the field of satellite communication technology. This method is based on a LEO satellite network simulation modeling method using multi-dimensional resource feature extraction and characterization, and a regional traffic prediction algorithm for LEO satellite networks based on an improved Transformer architecture. It provides resource constraints and spatiotemporal correlation characteristics for routing decisions. Based on a large-scale LEO satellite network model that divides virtual nodes and autonomous domains into geographical regions, the routing process is decoupled into intra-domain and inter-domain parts. Based on an improved MADQN intra-domain routing and resource scheduling algorithm for LEO satellite networks and a DQN-ST inter-domain routing and resource scheduling algorithm, it achieves hierarchical routing and decoupled resource scheduling, breaking through the traditional single-layer routing framework and constructing a domain-based intelligent routing system that balances flexibility and scalability. This invention significantly enhances the service capabilities of LEO satellite networks under extreme scenarios such as high-frequency topology changes, strictly limited resources, and sudden spatiotemporal changes in traffic, providing a highly reliable and low-latency routing solution for integrated space-ground communication, emergency disaster early warning, and global real-time data transmission.

[0116] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0117] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0118] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0119] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A large-scale low-orbit satellite network sub-domain intelligent routing method, characterized by: Step one, based on multi-dimensional resource feature extraction and representation, realize low-orbit satellite network simulation modeling, quantize the influence weight of multi-factor coupling on network performance, provide resource constraint conditions and space-time correlation characteristics for routing algorithm, support real-time and adaptability of routing decision; Step two, based on improved Transformer architecture, realize low-orbit satellite network regional traffic prediction algorithm, build global traffic tidal effect model, and map the prediction results to routing decision state space to drive resource allocation and congestion avoidance strategy; Step three, construct a large-scale low-orbit satellite network model based on geographical area division virtual nodes and autonomous domain, decouple the routing process into intra-domain and inter-domain parts; Step four, based on improved multi-agent deep Q network MADQN, realize low-orbit satellite network intra-domain routing selection and resource scheduling algorithm, model each satellite as an agent, based on real-time Markov decision process RTMDP, input multi-dimensional resource features of neighbor nodes and last action information, output next hop transmission direction, combine self-attention mechanism to learn matrix information of satellite network state, enhance the understanding and response ability of multi-agent to environment state; Step five, based on deep Q network DQN-ST combined with sum tree mechanism, realize low-orbit satellite network inter-domain routing selection and resource scheduling algorithm, establish dynamic mapping relationship of cross-domain routing, use SumTree improved priority experience replay mechanism, determine satellite routing task sample priority according to timing difference error, select optimal next hop satellite node based on Q value, aggregate and output inter-domain routing intelligent decision. 2.The method of claim 1, wherein, In step one, the low-orbit satellite network simulation modeling based on multi-dimensional resource feature extraction and representation includes space environment model, energy consumption model, illumination model, on-orbit life model, and multi-dimensional resource coupling characteristics and dynamic interaction model; The space environment model is based on constellation configuration design, inter-satellite link and communication link model, and the orbit parameters include altitude, inclination, number of orbit planes, number of satellites in each orbit and phase factor of adjacent orbit planes The neighbor satellite connection relationship of the logical address of the satellite is defined by the following formula: ， The energy consumption model quantifies satellite antenna receiving power, transmitting power and data transmission energy consumption, and establishes the mathematical relationship between satellite communication energy consumption and transmission distance, data volume and channel state; Derivation of inter-satellite transmission rate by Shannon formula wherein, B is the bandwidth of the wireless channel, N is the noise power; free space propagation loss during transmission The expression is wherein, D is the distance between satellites, λ is the wavelength of the electromagnetic wave, c is the speed of light, f is the operating frequency; the antenna receiving power is and when the receiving power of a given satellite is given, the satellite transmitting power can be deduced The expression is ; The light model is characterized in that, through real-time interaction between a satellite orbit position and an earth shadow area, the influence of solar panel power generation efficiency and a satellite power supply during a shadow period is predicted; a cylindrical shadow model is used to predict the shadow period, and the restriction condition is If the formula is not satisfied, it is indicated that the satellite is in a light area. The in-orbit life model is characterized in that a function relationship between battery reliability and lost life is constructed according to the discharge depth of the lithium ion battery and the cycle life consumption of the battery; and the function relationship is based on the discharge depth of the lithium ion battery and the cycle life consumption of the lithium ion battery , ; and a fitting expression of the battery reliability is . The multi-dimensional resource coupling characteristics dynamic and interaction model analyzes the space environment, energy consumption, illumination and on-orbit life model representation, constructs the coupling mechanism and dynamic correlation of multi-dimensional resources, and carries out collaborative modeling and system optimization; Through Monte Carlo simulation, the joint influence of orbital height on communication distance and energy load is quantified. 3.The method of claim 1, wherein, In step two, the low-orbit satellite network regional traffic prediction algorithm based on improved Transformer architecture includes data traffic modeling and PatchTST-based traffic prediction algorithm; The data service traffic modeling divides the earth surface according to longitude and latitude every 15°, and converts the traffic intensity according to the number of Internet users in each block, and obtains the global data service traffic intensity The distribution of the traffic in time is represented by a discount coefficient ; in each traffic block, the longitude of the center position is selected as a reference to calculate the time zone of the block, and it is assumed that the discount coefficient of the most active time of the user is 1 and the discount coefficient of the almost inactive time is 0; the low-orbit satellite network traffic of any traffic block is represented as , and the unit is Mbps; The PatchTST-based low-orbit satellite network traffic prediction algorithm adopts EMD technology to decompose the original traffic sequence into multiple single-frequency and mutually independent intrinsic mode function (IMF) sub-sequences and a residual signal, effectively separates different frequency components, so that each sub-sequence is more stable and easier to model; the PatchTST model is introduced for each IMF sub-sequence and residual signal for prediction; the PatchTST model combines the advantages of the Transformer architecture, can maintain high efficiency when processing long sequences, and capture complex patterns and dependencies in time series; the output results are added as the prediction output; The implementation method of the EMD assumes that the original sequence is , which is decomposed into a finite number of IMF subsequences and a residual , ; the standard deviation is calculated by the formula , wherein , represents an intrinsic mode function, represents the number of iterations, represents the order of the function, represents the total number of iterations; The construction method of the PatchTST-based traffic prediction model, for a group of multivariate time series samples, assumes that the review window, i.e. the time data length of inputting the model, is , and samples are obtained, wherein at time slot , is a one-dimensional column vector containing elements, and the data values predicted for the next time slots; the model structure includes a forward process, a block process and a Transformer encoder; The forward process represents the first one-dimensional sequence as each sequence is independently input into a Transformer backbone network, and the prediction result of the sequence is calculated ; The block process, each input one-dimensional time series Firstly, the block process is performed, and the block length is denoted as The non-overlapping region between two consecutive blocks is denoted as After the block process, a block sequence is generated ; wherein, The number of blocks generated is denoted as The memory usage and computational complexity of the attention diagram are reduced at a rate of Therefore, under the condition of limited GPU memory resources and training time, the block operation can enable the model to understand longer historical sequences, thereby significantly improving the prediction performance of the model; The Transformer encoder, through a trainable linear projection , and learnable additive positional encoding , maps the patches into a latent representation space of dimension , the data representation input to the Transformer encoder is , where ; each attention head of the multi-head attention converts it into a query matrix , a key matrix , and a value matrix , where , ; the attention output is obtained using scaled product ; the result generated after the attention output is processed through a normalization layer and a feed-forward network layer with a residual connection is represented as , and finally a fully connected layer with a linear head is used to obtain the prediction result .

4. The method of claim 1, wherein, In step three, the large-scale low-orbit satellite network model based on geographic area division of virtual nodes and autonomous domains includes a geographic area hexagonal grid division method and a low-orbit satellite network routing method based on fusion of virtual nodes and autonomous domains. The geographic area hexagonal grid division method performs a simple cylindrical projection on the earth's surface and converts it into a rectangle with an aspect ratio of 2:1, and then divides the rectangle into 112 regular hexagonal grids, each covering a 15°x15° area in latitude and longitude; the grid mapping rule ensures seamless connection between adjacent virtual nodes, reducing the number of routing hops; The low-orbit satellite network routing method based on fusion of virtual nodes and autonomous domains maps the hexagonal grid structure on the ground to the satellite constellation by associating the satellite ground projection position with the ground grid space, and the satellite group in the same block belongs to the same autonomous domain and is aggregated into an abstract virtual node; The routing and resource information updating strategy combines flooding update and local update, updates inter-domain information by flooding method, and floods the boundary satellite node information of the block to the entire satellite constellation; The intra-domain information is updated by local update method, and the state information of each satellite and link is shared and updated within the autonomous domain.

5. The method of claim 1, wherein, In step four, the low-orbit satellite network intra-domain routing selection and resource scheduling algorithm based on the improved MADQN includes a real-time Markov decision process and a self-attention mechanism. The real-time Markov decision process is characterized in that the state space The action space includes satellite energy consumption, link quality, battery reliability, neighbor node state and last action decision result of the satellite node The transmission direction is "up, down, left and right", and the reward function The reward function includes comprehensive delay, energy consumption and link stability; in the RTMDP framework, the action and the state change simultaneously, the agent can observe the next state immediately after selecting the action, and directly infer the next action, the real-time decision mechanism allows the agent to track the fast-changing environmental conditions more closely, and ensures that the routing selection and resource scheduling can respond to the changes of the low-orbit satellite network topology and state in time; The action value function expression is as follows: ， where denotes the expected probability of transitioning to a particular state given a state and action ; denotes the expected cumulative reward of a sequence of states that the agent can encounter given a state-action pair and following the current policy after performing action ; The state value function expression is as follows: ， wherein represents the expected cumulative reward for a sequence of states that the agent can encounter given the state-action pair condition. The self-attention mechanism described above maps the input neighbor node state information into query matrices by constructing a multi-head attention structure. Key matrix and value matrix Attention weights are calculated according to the formula ,in The dimension scaling factor of the key vector is used to avoid the gradient saturation problem of the Softmax function. By computing the outputs of multiple attention heads in parallel and concatenating them into a joint representation, the algorithm can dynamically allocate weights, highlight the states of neighboring nodes with high link stability and high resource redundancy, and suppress noise interference. The attention output is then input into the fully connected network after residual connection and layer normalization to generate routing decision actions.

6. The method of claim 1, wherein, In step five, the low-orbit satellite network inter-domain routing selection and resource scheduling algorithm based on DQN-ST includes a dynamic adaptive domain boundary node mapping mechanism and a DQN-ST algorithm architecture for routing task optimization. The dynamic adaptive domain boundary node mapping mechanism, when a node needs to communicate across the autonomous domain, the node in the domain forwards the request information to the domain boundary node, due to the continuous high-speed movement of the satellite, the satellite node will continuously move from one area to the next area, at this time the originally selected domain boundary node and the inter-domain communication link will be invalid, considering the node effectiveness and the node available resources, evaluating the satellite nodes in the domain, selecting the mapping satellite of the domain boundary node; the calculation of the node effectiveness index not only considers the geometric distance to the adjacent domain at the current time, but also needs to combine the time sensitivity of the specific routing task, the evaluation of the node resource index needs to be dynamically weighted according to the actual business demand, so the node effectiveness , the node available resources , the weight coefficients of the two indexes are set to 0.7 and 0.3 respectively, then the weighted comprehensive score after standardization , is mapped to the actual routing decision, the system can automatically select the boundary node which can meet the quality of service (QoS) requirements and has the optimal resource utilization rate. The DQN-ST algorithm architecture for route task optimization applies the priority experience replay mechanism improved based on Sumtree to the deep Q network algorithm, the state space wherein represents a satellite node where the time slot task is located, represents a link connection condition of the satellite node where the time slot task is located, represents a data size of the time slot task, represents network state information of the time slot virtual constellation; the action space wherein, represents a satellite node where the time slot intelligent agent performs the next forwarding; represents a bandwidth resource allocated to the task by the time slot; the priority experience replay mechanism based on Sumtree considers the space-time correlation of the routing task in actual deployment, and by giving higher TD error weights to task samples with high time delay sensitivity or high bandwidth demand, the priority wherein is a minimum constant, and the sampling probability is calculated according to the formula Based on the above calculation, and according to the Q value calculation result, the optimal next hop satellite node is selected to ensure the efficiency and stability of cross-domain communication.

7. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to realize the steps of the method in any one of claims 1 to 6.

Images