A space-ground integrated network resource allocation method based on a graph attention network
Patent Information
- Application Number
- CN202611217043.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-12
- Publication Date
- 2026-10-16
AI Technical Summary
[0005]3.非独立同分布数据：不同地理区域（如海洋、无人区、城市）的业务请求分布极不均匀，传统集中式优化算法计算复杂度过高且难以实时收敛
1、动态注意力机制提升资源分配精度：通过图注意力网络的动态权重分配机制，模型能够自动识别异构网络环境中的关键节点和链路，在聚合邻域信息时为重要邻居分配更高权重，提升资源分配精度。
Smart Images
Figure CN122933567A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wireless communication network technology, and in particular to a resource allocation method for a Space-Ground Integrated Network (SGIN) based on graph attention networks, which is applicable to dynamic resource scheduling in scenarios where low-Earth orbit satellite constellations and terrestrial networks are integrated. Background Technology
[0002] The integrated space-ground network achieves seamless global coverage by integrating terrestrial networks (4G / 5G / 6G), high-altitude platforms, and low-Earth orbit / medium-Earth orbit / geostationary orbit satellites. Compared with traditional terrestrial networks, the integrated space-ground network has characteristics such as highly dynamic topology, resource heterogeneity, and non-independent and co-distributed data.
[0003] 1. High dynamic topology: Low-Earth orbit satellites move at speeds of several kilometers per second, resulting in frequent switching of satellite-to-ground links (on a minute-by-minute basis), and the network topology changes drastically over time.
[0004] 2. Resource heterogeneity: It involves the joint optimization of computing resources (onboard processors), spectrum resources (multi-band), power resources and storage resources, and the capabilities of different nodes vary greatly.
[0005] 3. Non-independent and identically distributed data: Business requests are distributed very unevenly across different geographical regions (such as oceans, uninhabited areas, and cities). Traditional centralized optimization algorithms have excessively high computational complexity and are difficult to converge in real time.
[0006] With the rapid development of integrated space-ground networks, network scale continues to expand, and service types are becoming increasingly diversified, posing serious challenges to network resource management. Existing network resource scheduling technologies mainly face the following technical problems: First, traditional heuristic algorithms struggle to adapt to dynamic network environments. Resource scheduling methods based on rules or simple optimization models are ill-equipped to make accurate real-time scheduling decisions when faced with rapidly changing network conditions and dynamic fluctuations in business demands, resulting in low resource utilization efficiency.
[0007] Second, although deep reinforcement learning has been applied to the field of resource scheduling, existing methods usually input the network state as a whole into the model, failing to distinguish the differences in the importance of different features to the decision, resulting in low learning efficiency and slow convergence speed when the model faces a large number of network nodes.
[0008] Third, business requests are usually time-dependent, and the characteristics of the current task are related to the characteristics of the historical task. Existing resource scheduling methods often ignore the impact of this temporal correlation on service caching and resource pre-configuration.
[0009] To address these issues, some studies have attempted to introduce attention mechanisms to enhance feature extraction capabilities or employ multi-agent reinforcement learning to solve large-scale decision-making problems. However, current technologies have not yet formed a complete and comprehensive solution capable of collaboratively optimizing service caching, task scheduling, and resource allocation, while also possessing time-series awareness and differentiated feature extraction capabilities. Summary of the Invention
[0010] To overcome the shortcomings of existing technologies, this invention provides a method for allocating resources in a space-ground integrated network based on graph attention networks.
[0011] To achieve the aforementioned objectives of the invention, the technical solution adopted to solve its technical problems is as follows: A method for allocating resources in a space-ground integrated network based on graph attention networks includes the following steps: Step 1: Obtain real-time status information of the integrated space-ground network, including network topology, node resource status, and link quality; Step 2: Construct the network topology graph structure data. The nodes in the network are the vertices of the graph, the communication links between the nodes are the edges of the graph, and the resource status of the nodes and the quality of the links are the feature attributes of the corresponding vertices and edges. Step 3: Input the graph structure data carrying feature attributes into the pre-trained multi-head graph attention network model; Step 4: The graph attention network model dynamically calculates the weight coefficients between adjacent nodes through the attention mechanism and outputs resource allocation decisions; Step 5: Optimize and adjust the communication and computing resources in the integrated space-ground network.
[0012] Furthermore, in step 1, the real-time status information includes: network topology, node resource status, and link quality; The network topology includes the connection relationships between nodes and the visible time window; The node resource status includes remaining computing power, storage space, energy status, and available bandwidth; The link quality includes propagation delay, available bandwidth, bit error rate, and link lifetime; The data is automatically collected by sensors deployed on satellites, drones, and ground base stations and transmitted to the data processing center in real time.
[0013] Furthermore, step 2 includes the following: Mapping the network topology as a dynamic graph sequence Each time snapshot This includes the current network status; For each graph The following methods are used to extract the features of nodes and edges: Node feature vectors This includes the node's computing resources, storage resources, energy status, and geographical location. Edge feature vector This includes the link's propagation delay, available bandwidth, and link stability. To capture the spatiotemporal correlation characteristics of network resources, attention representations are calculated in two dimensions: structural neighborhood and temporal neighborhood. Structural neighborhood attention is used to aggregate information of neighboring nodes at the current time, while temporal neighborhood attention is used to capture the state evolution pattern of the same node at different time steps.
[0014] Furthermore, step 3 includes the following: The graph structure data carrying spatiotemporal features is input into a pre-trained multi-head graph attention network model. The core computation process of the model is as follows: For nodes and its neighboring nodes , No. The attention coefficient of the layer is calculated as follows:
[0015] in, for T Attention weight vector at any time, The weight matrix is a learnable matrix. and Divided into the first The feature vectors of two adjacent nodes in a layer. This represents a vector concatenation operation. For nodes The attention coefficients mentioned above reflect the neighbor set of the neighbor nodes. Relative to the central node The degree of importance; The attention coefficients are normalized using the Softmax function:
[0016] in, The normalized attention coefficient. For the current target node The unnormalized raw attention coefficients For all traversed nodes The unnormalized original attention coefficients; node The new feature representation is obtained by weighted aggregation of neighbor features:
[0017] in, For nodes After the update (the first one) Weighted aggregated neighbor features of layer output It is a ReLU nonlinear activation function; To enhance the model's expressive power, a multi-head attention mechanism is adopted. The outputs of the individual attention heads are concatenated:
[0018] in, For the first Node in the head To the neighbors Normalized attention weights For the first Learnable linear matrices of a certain size.
[0019] Furthermore, step 4 includes the following: The hidden state of the nodes after being encoded by the graph attention network is input into the decoder module. The decoder adopts a sequence-to-sequence architecture and generates a resource allocation action sequence element by element. The resource allocation decision includes at least one of the following: bandwidth allocation decision, power control decision, computation offloading decision, service function chain deployment decision, and routing path decision. Specifically, for service function chain deployment scenarios, the decoder generates decisions as follows: At the current time step, based on the already deployed virtual network function sequence and combined with the node state output by the graph attention network, the next virtual network function deployment node is selected. This process can be formalized as follows:
[0020] in, For time Selected deployment node, For time The complete state, The current focal point Embedded representation, For time The context vector.
[0021] Furthermore, step 5 includes the following: Based on all generated resource allocation decisions, specific resource reconfiguration operations are executed through the software-defined network controller or the local agent of each node, including bandwidth reallocation, transmit power adjustment, and task offloading path switching. The optimization and adjustment process of communication and computing resources in the integrated space-ground network includes: Collect state trajectory data of the integrated space-ground network under different operational scenarios to construct a training dataset:
[0022] in, This is a network state diagram. For time Selected deployment node, The reward function obtained; Combining resource allocation accuracy and fairness metrics, an end-to-end training algorithm based on a near-end policy optimization is employed, and a loss function is designed as follows:
[0023] in, The importance sampling ratio, This is the estimated value of the dominance function. To trim hyperparameters, For the loss of value, The coefficient representing the value loss. For the entropy of the strategy, The entropy regularization coefficient; During the training phase, the global state information model parameters are updated using a proximal policy optimization algorithm. During the execution phase, each node relies only on local observations for distributed decision-making. The reward function during the training phase is designed as follows:
[0024] Here, `service_acceptance` indicates whether the service request was successfully deployed (+1 for success, -1 for failure), `service_cost` is the total resources consumed during deployment, and `latency_penalty` is the penalty for end-to-end latency exceeding a threshold. This is the discount factor.
[0025] By employing the above technical solutions, this invention has the following advantages and positive effects compared with the prior art: 1. Dynamic attention mechanism improves resource allocation accuracy: Through the dynamic weight allocation mechanism of graph attention network, the model can automatically identify key nodes and links in heterogeneous network environment, and assign higher weights to important neighbors when aggregating neighborhood information, thereby improving the accuracy of resource allocation.
[0026] 2. Enhanced dynamic adaptability through spatiotemporal joint modeling: By calculating attention representations in both structural and temporal neighborhoods, accurate modeling of the spatiotemporal relationships of network resources is achieved. This enables the model to effectively predict topology change trends and adjust resource strategies in advance, reducing losses caused by frequent link switching. Attached Figure Description
[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is an overall flowchart of the integrated space-ground network resource allocation method based on graph attention networks of the present invention; Figure 2 This is a schematic diagram of the structure of the multi-head graph attention network in an embodiment of the present invention; Figure 3 This illustrates the performance improvement of different algorithms during training in the embodiments of the present invention; Figure 4 This is a diagram showing the relationship between system performance and the number of users in an embodiment of the present invention; Figure 5 This is a comparison chart of the convergence speeds of different algorithms in the embodiments of the present invention. Detailed Implementation
[0028] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.
[0029] like Figure 1 As shown, this embodiment discloses a method for allocating resources in a space-ground integrated network based on graph attention networks, including the following steps: Step 1: Obtain real-time status information of the integrated space-ground network, including network topology, node resource status, and link quality.
[0030] Specifically, in step 1, the real-time status information includes: network topology, node resource status, and link quality; The network topology includes the connection relationships between nodes and the visible time window; The node resource status includes remaining computing power, storage space, energy status, and available bandwidth; The link quality includes propagation delay, available bandwidth, bit error rate, and link lifetime; The data is automatically collected by sensors deployed on satellites, drones, and ground base stations and transmitted to the data processing center in real time.
[0031] In this embodiment, a space-ground integrated network comprising 66 low-Earth orbit satellites, 10 ground gateways, and 50 edge servers is deployed in the simulation scenario. Network status data is collected using the OMNeT++ network simulator within a 10-second time window. The data collected at each time step includes: remaining computing resources of satellite nodes (0-100 GFLOPS), energy status (0-100%), connection status with ground gateways (binary), and link propagation latency (ms-level).
[0032] Step 2: Construct the network topology graph structure data. The nodes in the network are the vertices of the graph, the communication links between the nodes are the edges of the graph, and the resource status of the nodes and the quality of the links are the feature attributes of the corresponding vertices and edges.
[0033] Specifically, step 2 includes the following: Mapping the network topology as a dynamic graph sequence Each time snapshot This includes the current network status; For each graph The following methods are used to extract the features of nodes and edges: Node feature vectors This includes the node's computing resources, storage resources, energy status, and geographical location. Edge feature vector This includes the link's propagation delay, available bandwidth, and link stability. To capture the spatiotemporal correlation characteristics of network resources, attention representations are calculated in two dimensions: structural neighborhood and temporal neighborhood. Structural neighborhood attention is used to aggregate information of neighboring nodes at the current time, while temporal neighborhood attention is used to capture the state evolution pattern of the same node at different time steps.
[0034] In this embodiment, the network state at each time step is constructed as a graph sequence. The node set contains all satellites and ground equipment, and edges represent communication links with visible relationships. Each node's feature vector... The dimension is 8, and the feature vector of each edge is... The dimension is 4.
[0035] Step 3: Input the graph structure data carrying feature attributes into the pre-trained multi-head graph attention network model.
[0036] Specifically, step 3 includes the following: The graph structure data carrying spatiotemporal features is input into a pre-trained multi-head graph attention network model. The core computation process of the model is as follows: For nodes and its neighboring nodes , No. The attention coefficient of the layer is calculated as follows:
[0037] in, for T Attention weight vector at any time, The weight matrix is a learnable matrix. and Divided into the first The feature vectors of two adjacent nodes in a layer. This represents a vector concatenation operation. For nodes The attention coefficients mentioned above reflect the neighbor set of the neighbor nodes. Relative to the central node The degree of importance; The attention coefficients are normalized using the Softmax function:
[0038] in, The normalized attention coefficient. For the current target node The unnormalized raw attention coefficients For all traversed nodes The unnormalized original attention coefficients; node The new feature representation is obtained by weighted aggregation of neighbor features:
[0039] in, For nodes After the update (the first one) Weighted aggregated neighbor features of layer output It is a ReLU nonlinear activation function; To enhance the model's expressive power, a multi-head attention mechanism is adopted. The outputs of the individual attention heads are concatenated:
[0040] in, For the first Node in the head To the neighbors Normalized attention weights For the first Learnable linear matrices of a certain size.
[0041] In this embodiment, a two-layer multi-head graph attention network is constructed, with four attention heads in each layer and a hidden layer dimension of 128. The input layer maps the original features to the latent space; the first layer GAT aggregates information within the one-hop neighbor range, and the second layer GAT expands the receptive field to two-hop neighbors. The processing flow is as follows: For a service function chain request to be deployed, the GAT model takes the current network state graph as input and outputs an encoded vector for each node. This encoded vector aggregates the node's own state and the influence weights of its neighboring nodes.
[0042] Step 4: The graph attention network model dynamically calculates the weight coefficients between adjacent nodes through the attention mechanism and outputs resource allocation decisions.
[0043] Specifically, step 4 includes the following: The hidden state of the nodes after being encoded by the graph attention network is input into the decoder module. The decoder adopts a sequence-to-sequence architecture and generates a resource allocation action sequence element by element. The resource allocation decision includes at least one of the following: bandwidth allocation decision, power control decision, computation offloading decision, service function chain deployment decision, and routing path decision. Specifically, for service function chain deployment scenarios, the decoder generates decisions as follows: At the current time step, based on the already deployed virtual network function sequence and combined with the node state output by the graph attention network, the next virtual network function deployment node is selected. This process can be formalized as follows:
[0044] in, For time Selected deployment node, For time The complete state, The current focal point Embedded representation, For time The context vector.
[0045] In this embodiment, a Seq2Seq decoder is used to generate the VNF deployment sequence in an autoregressive manner. At each decoding time step, the decoder selects the optimal deployment node through a softmax layer based on the node state encoded by GAT and the context vector of the deployed sequence.
[0046] To address link reliability requirements, a link mapping algorithm was designed: after selecting deployment nodes, the shortest path algorithm is used to calculate the transmission path between nodes that meets bandwidth constraints.
[0047] Step 5: Optimize and adjust the communication and computing resources in the integrated space-ground network.
[0048] Specifically, step 5 includes the following: Based on all generated resource allocation decisions, specific resource reconfiguration operations are executed through the software-defined network controller or the local agent of each node, including bandwidth reallocation, transmit power adjustment, and task offloading path switching. The optimization and adjustment process of communication and computing resources in the integrated space-ground network includes: Collect state trajectory data of the integrated space-ground network under different operational scenarios to construct a training dataset:
[0049] in, This is a network state diagram. For time Selected deployment node, The reward function obtained; Combining resource allocation accuracy and fairness metrics, an end-to-end training algorithm based on a near-end policy optimization is employed, and a loss function is designed as follows:
[0050] in, The importance sampling ratio, This is the estimated value of the dominance function. To trim hyperparameters, For the loss of value, The coefficient representing the value loss. For the entropy of the strategy, The entropy regularization coefficient; During the training phase, the global state information model parameters are updated using a proximal policy optimization algorithm. During the execution phase, each node relies only on local observations for distributed decision-making. The reward function during the training phase is designed as follows:
[0051] Here, `service_acceptance` indicates whether the service request was successfully deployed (+1 for success, -1 for failure), `service_cost` is the total resources consumed during deployment, and `latency_penalty` is the penalty for end-to-end latency exceeding a threshold. This is the discount factor.
[0052] In this embodiment, a proximal policy optimization algorithm is used during the training phase, and the discount factor in the reward function is set to .
[0053] Figure 3 This is a comparison graph showing the performance improvement of different algorithms during training. The horizontal axis represents the number of training rounds, and the vertical axis represents the average cumulative reward. The larger the value, the better the algorithm performance. This patent has ±1 The curve represents the uncertainty of results from multiple experiments; the narrower the boundary shadow, the more stable the algorithm. Compared with graph convolutional networks, the performance of this patented method is improved by 22.5%.
[0054] Figure 4 The graph illustrates the relationship between system performance and the number of users across different algorithms. The horizontal axis represents the number of users, and the vertical axis represents bandwidth utilization, signifying the total system throughput. A higher value indicates a larger system capacity and more efficient resource utilization. Under all user modes, the method presented in this patent achieves optimal performance, representing an improvement of approximately 10% compared to graph convolutional networks. The performance advantage of this patent becomes more significant as the number of users increases, demonstrating its resource allocation capabilities in complex multi-user environments.
[0055] Figure 5 This graph compares the convergence speeds of different algorithms. The horizontal axis represents different algorithms, and the vertical axis represents the number of training steps required to achieve the target performance. Fewer steps mean better performance. The error bars represent the standard deviation of multiple experiments; the smaller the value, the more stable the algorithm. This patented algorithm can achieve the same performance with fewer training steps, reducing training time and computational cost. Its convergence speed is 35.70% faster than graph convolutional networks.
[0056] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for allocating resources in a space-ground integrated network based on graph attention networks, characterized in that, Includes the following steps: Step 1: Obtain real-time status information of the integrated space-ground network, including network topology, node resource status, and link quality; Step 2: Construct the network topology graph structure data. The nodes in the network are the vertices of the graph, the communication links between the nodes are the edges of the graph, and the resource status of the nodes and the quality of the links are the feature attributes of the corresponding vertices and edges. Step 3: Input the graph structure data carrying feature attributes into the pre-trained multi-head graph attention network model; Step 4: The graph attention network model dynamically calculates the weight coefficients between adjacent nodes through the attention mechanism and outputs resource allocation decisions; Step 5: Optimize and adjust the communication and computing resources in the integrated space-ground network.
2. The method for allocating integrated space-ground network resources based on graph attention networks according to claim 1, characterized in that, In step 1, the real-time status information includes: network topology, node resource status, and link quality; The network topology includes the connection relationships between nodes and the visible time window; The node resource status includes remaining computing power, storage space, energy status, and available bandwidth; The link quality includes propagation delay, available bandwidth, bit error rate, and link lifetime; The data is automatically collected by sensors deployed on satellites, drones, and ground base stations and transmitted to the data processing center in real time.
3. The method for allocating space-ground integrated network resources based on graph attention networks according to claim 1, characterized in that, Step 2 includes the following: Mapping the network topology as a dynamic graph sequence Each time snapshot This includes the current network status; For each graph The following methods are used to extract the features of nodes and edges: Node feature vectors This includes the node's computing resources, storage resources, energy status, and geographical location. Edge feature vector This includes the link's propagation delay, available bandwidth, and link stability. To capture the spatiotemporal correlation characteristics of network resources, attention representations are calculated in two dimensions: structural neighborhood and temporal neighborhood. Structural neighborhood attention is used to aggregate information of neighboring nodes at the current time, while temporal neighborhood attention is used to capture the state evolution pattern of the same node at different time steps.
4. The method for allocating space-ground integrated network resources based on graph attention networks according to claim 1, characterized in that, Step 3 includes the following: The graph structure data carrying spatiotemporal features is input into a pre-trained multi-head graph attention network model. The core computation process of the model is as follows: For nodes and its neighboring nodes , No. The attention coefficient of the layer is calculated as follows: in, for T Attention weight vector at any time, The weight matrix is a learnable matrix. and Divided into the first The feature vectors of two adjacent nodes in a layer This represents a vector concatenation operation. For nodes The attention coefficients mentioned above reflect the neighbor set of the neighbor nodes. Relative to the central node The degree of importance; The attention coefficients are normalized using the Softmax function: in, Attention coefficients after normalization For the current target node The unnormalized raw attention coefficients For all traversed nodes The unnormalized raw attention coefficients; node The new feature representation is obtained by weighted aggregation of neighbor features: in, For nodes After the update (the first one) Weighted aggregated neighbor features of layer output It is a ReLU nonlinear activation function; To enhance the model's expressive power, a multi-head attention mechanism is adopted. The outputs of the individual attention heads are concatenated: in, For the first Nodes in the head To the neighbors Normalized attention weights For the first A learnable linear matrix of size.
5. A method for allocating space-ground integrated network resources based on graph attention networks according to claim 1, characterized in that, In step 4, Includes the following: The hidden state of the nodes after being encoded by the graph attention network is input into the decoder module. The decoder adopts a sequence-to-sequence architecture and generates a resource allocation action sequence element by element. The resource allocation decision includes at least one of the following: bandwidth allocation decision, power control decision, computation offloading decision, service function chain deployment decision, and routing path decision. Specifically, for service function chain deployment scenarios, the decoder generates decisions as follows: At the current time step, based on the already deployed virtual network function sequence and combined with the node state output by the graph attention network, the next virtual network function deployment node is selected. This process can be formalized as follows: in, For time Selected deployment node, For time The complete state, The current focal node Embedded representation, For time The context vector.
6. A method for allocating space-ground integrated network resources based on graph attention networks according to claim 1, characterized in that, Step 5 includes the following: Based on all generated resource allocation decisions, specific resource reconfiguration operations are executed through the software-defined network controller or the local agent of each node, including bandwidth reallocation, transmit power adjustment, and task offloading path switching. The optimization and adjustment process of communication and computing resources in the integrated space-ground network includes: Collect state trajectory data of the integrated space-ground network under different operational scenarios to construct a training dataset: in, This is a network state diagram. For time Selected deployment node, The reward function obtained; Combining resource allocation accuracy and fairness metrics, an end-to-end training algorithm based on a near-end policy optimization is employed, and a loss function is designed as follows: in, The importance sampling ratio, This is the estimated value of the dominance function. To trim hyperparameters, For the loss of value, The coefficient representing the value loss. For the entropy of the strategy, The entropy regularity coefficient; During the training phase, the global state information model parameters are updated using a proximal policy optimization algorithm. During the execution phase, each node relies only on local observations for distributed decision-making. The reward function during the training phase is designed as follows: Here, `service_acceptance` indicates whether the service request was successfully deployed (+1 for success, -1 for failure), `service_cost` is the total resources consumed during deployment, and `latency_penalty` is the penalty for end-to-end latency exceeding a threshold. This is the discount factor.