A method and system for adaptive slicing of a supernova constellation cloud network resource and SLA guarantee

Abstract

The present application relates to a kind of based on spatiotemporal diagram convolution Mega constellation cloud network resource adaptive slicing and SLA guarantee method and system, belong to satellite communication and network technical field, the present application method includes: construct Mega constellation cloud network fusion spatiotemporal diagram model, the computing, storage resource of satellite node and the bandwidth resource of inter-satellite link are uniformly represented;Spatiotemporal diagram convolution network is constructed, and cloud network resource state is forecasted in multiple steps in advance;Based on the prediction result and business SLA demand, construct SLA-aware cloud network joint slicing optimization model;SLA violation prediction and closed-loop adaptive slicing reconstruction mechanism are constructed, and incremental resource reconfiguration is triggered before SLA violation actually occurs;Orbit domain-network domain double-layer slicing coordination mechanism is constructed, scalable slicing management under Mega constellation is realized;The present application realizes the joint adaptive slicing of cloud network resource and forward-looking SLA guarantee, can support the real-time slicing management of large-scale satellite constellation.

Description

Technical Field

[0001] This invention relates to a method and system for adaptive slicing and SLA guarantee of resources in the Giants Cone cloud network based on spatiotemporal graph convolution, belonging to the field of satellite communication and network technology. Background Technology

[0002] With the rapid development of low-orbit satellite internet, the giant constellation network, represented by Starlink and OneWeb, has reached a scale of thousands or even tens of thousands of satellites, providing global users with diversified services such as broadband access, Internet of Things, and emergency communication. The giant constellation network has the following significant characteristics: (1) huge network scale, large number of nodes, and complex topology; (2) highly dynamic topology changes, with high-speed movement of low-orbit satellites causing frequent connection and disconnection of inter-satellite links; (3) significant differences in business needs, with different industry scenarios having large differences in requirements for indicators such as latency, bandwidth, and reliability; (4) limited on-board resources, with computing, storage, and energy resources limited by the satellite's payload capacity. Network slicing technology is a technology that virtualizes physical network resources into multiple logically isolated dedicated networks, which can provide customized service guarantees for differentiated services. SLA (Service Level Agreement) is a technical contract between operators and users that agrees on service quality indicators, including key indicators such as latency, bandwidth, reliability, and jitter. In the giant constellation scenario, how to achieve adaptive slicing of cloud network resources and guarantee the SLA of each slice is the core problem that urgently needs to be solved. Existing technologies have the following shortcomings: Most existing satellite network resource allocation schemes focus only on the allocation of single-dimensional resources such as link bandwidth or power. For example, the "Satellite Communication Link Resource Allocation Method Based on AI Dynamic Planning" (CN202511879081) only allocates power and bandwidth jointly, without incorporating computing and storage resources into a unified slicing framework. However, actual services (such as edge computing and data caching) require the coordinated supply of computing, storage, and bandwidth resources, and single-dimensional allocation cannot meet the needs of cloud-network convergence scenarios. Most existing slicing management schemes only trigger resource adjustments after SLAs have been violated. For example, the "Satellite Network Slice Migration Method and System Based on Graph Token and Large Language Model Collaboration" (CN202511906937) only passively triggers migration decisions after link failures, failing to perform preventative resource reconfiguration before violations occur. The high dynamism of the Giants constellation results in a significant time delay from violation detection to adjustment completion, and passive responses cannot effectively guarantee SLAs.

[0003] Existing solutions mostly employ centralized optimization, which results in extremely high computational complexity for satellite constellations with thousands of nodes, making real-time solutions difficult. For example, the "Data Transmission Method and Related Equipment for Differentiated Quality of Service Assurance in Satellite Networks" (CN202511777565) calculates utility values ​​link by link across the entire network, but the computational overhead is unacceptable at the scale of satellite constellations. Existing solutions model the spatiotemporal dynamics of satellite networks in a relatively coarse-grained manner. For instance, the "Dynamic Network Representation and Resource Optimization Method and Device Based on Spatiotemporal Attributes" (CN202511094370) introduces spatiotemporal attributes, but only uses them for the joint objective function of path optimization, failing to fully utilize spatiotemporal correlations for resource status prediction and thus unable to support predictive slicing decisions. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and propose a method and system for adaptive slicing and SLA assurance of Giants cloud network resources. By using a spatiotemporal graph convolutional network to predict the resource status of Giants cloud network, adaptive allocation of cloud network resources is achieved through prediction-driven SLA-aware slicing optimization. Furthermore, through an SLA violation prediction and closed-loop reconstruction mechanism and a two-layer coordination mechanism between the orbital domain and the network domain, efficient and scalable slice management and SLA assurance at the Giants scale are achieved.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for adaptive slicing and SLA guarantee of resources in the Giant Star constellation cloud network includes the following steps: S1. Construct a spatiotemporal graph model for cloud-network convergence in the Giants constellation, including: representing the computing resources, storage resources, and inter-satellite link bandwidth resources of satellite nodes in the Giants constellation as a resource state graph that dynamically changes with time and spatial location, wherein each node is associated with a multi-dimensional resource feature vector and each edge is associated with a link quality feature vector; S2. Based on the aforementioned Giant Star constellation cloud-network fusion spatiotemporal graph model, a spatiotemporal graph convolutional network is constructed. This network includes a spatial graph convolutional module and a temporally gated convolutional module. The spatial graph convolutional module extracts spatial dependency features of the satellite network topology based on a dynamic adjacency matrix, while the temporally gated convolutional module extracts temporal evolution features of resource states based on a sliding time window. The spatial graph convolutional module and the temporally gated convolutional module are executed alternately to obtain cloud-network resource state prediction results for future multi-time slots. S3. Based on the cloud network resource status prediction results and service SLA requirements, construct an SLA-aware cloud network joint slice optimization model. With the maximization of slice SLA satisfaction rate and the maximization of cloud network resource utilization as joint optimization objectives, solve the resource configuration scheme of each slice under the capacity constraints of computing resources, storage resources, and bandwidth resources and the SLA index constraints. S4. Construct an SLA violation prediction and closed-loop adaptive slice reconstruction mechanism, including: real-time monitoring of the SLA index of each slice; when the spatiotemporal graph convolutional network predicts that a slice has an SLA violation risk in a future time slot, triggering resource reconfiguration of the slice, and migrating the slice with the predicted violation to a resource-sufficient orbital domain through incremental resource adjustment, proceeding to step S5; if the spatiotemporal graph convolutional network does not predict that any slice has an SLA violation risk in a future time slot, proceeding to step S5; S5. Construct a two-layer slicing coordination mechanism between the orbital domain and the network domain, including: dividing orbital clusters and allocating cluster-level resource quotas at the orbital domain level according to the satellite orbital motion law; and performing fine-grained resource scheduling and slice instantiation within the clusters at the network domain level according to real-time service requirements. The two layers work together to achieve scalable slice management at the scale of the Giants constellation.

[0006] In the above-mentioned adaptive slicing and SLA guarantee method for the Giant Star Cloud Network resources, step S1, the method for constructing the Giant Star Cloud Network fusion spatiotemporal graph model includes: The Giants constellation network is modeled as the following time-varying graph: G(t) = (V, E(t), X V(t) X E(t) ) in: V is the union of the set of satellite nodes and the set of ground station nodes; This is the set of inter-satellite links and satellite-to-ground links active in time slot t; X V(t) This is a multidimensional resource feature matrix for nodes. The feature vector of each node contains available computing resources, available storage resources, current load rate, and track position coordinates. X E(t) The feature matrix is ​​the edge feature matrix. The feature vector of each link includes available bandwidth, propagation delay, bit error rate, and remaining link lifetime.

[0007] In the above-mentioned adaptive slicing and SLA guarantee method for resources of the Giant Star Cloud Network, in step S2, the spatial graph convolution module adopts a graph convolution network to perform spatial feature aggregation on the node feature matrix and dynamic adjacency matrix of time slot t. The temporally gated convolution module performs temporal dimension convolution on the temporal feature sequence output by the spatial convolution, and uses gated linear units to capture the temporal periodicity and trend of resource status.

[0008] In the aforementioned adaptive slicing and SLA guarantee method for the Giant Star Cloud Network resources, the node representation update method of the spatial graph convolution module is as follows: H s (l+1) = sigma(D tilde(t) (-1 / 2) A tilde(t) D tilde(t) (-1 / 2) H s (l) W s (l) ), where A tilde(t) = A(t) + I is the adjacency matrix with self-loops, A(t) is the initial adjacency matrix, I is the identity matrix, D tilde(t) Let W be the corresponding degree matrix. s (l) H is the trainable weight matrix for the l-th spatial convolution layer. s (l) H represents the node state before the update. s (l+1) The updated node state; In the time-gated convolution module, H t = sigma(W1 H s (L) )x(W2 H s (L) ),in This is a one-dimensional convolution operation, where sigma is the sigmoid function; W2 is the weight matrix for the second layer, and H... s (L) For node states, W1 is a layer weight matrix, and H... t This is a time-gated convolution function.

[0009] In the above-mentioned adaptive slicing and SLA guarantee method for cloud network resources in the Giant Star constellation, step S3, the method for constructing the SLA-aware cloud network joint slicing optimization model includes: Define the SLA requirement for each network slice s as a quadruple SLA. s = (D smax B smin , R smin J smax ), where D smax B smin , R smin J smax These represent the upper limit of end-to-end latency, the lower limit of guaranteed bandwidth, the lower limit of reliability, and the upper limit of latency jitter, respectively. The SLA satisfaction level of slice s is defined as the proportion of all service flows under that slice that meet the SLA metric, i.e., eta s =(1 / |F s |) sumf (I(QoS indicators meet SLA requirements)), where F s Let s be the set of business flows carried by slice s, I be the indicator function, and sum f The sum of indicators required to meet SLA requirements; the joint optimization objective is: max alpha (1 / |S|) sum s(eta_s) +beta U resource U resource The total cloud resource utilization rate is represented by alpha and beta, with |S| being the sum of SLA requirements for all slices. s(eta_s) This is the sum of SLA satisfaction for all slices.

[0010] In the above-mentioned adaptive slicing and SLA guarantee method for the Giant Star Cloud Network resources, in step S3, the capacity constraints and SLA index constraints of the computing resources, storage resources, and bandwidth resources include: the total amount of computing resources allocated to each satellite node does not exceed its available computing resources, the total amount of storage resources allocated does not exceed its available storage resources, the total amount of bandwidth allocated to each link does not exceed its available bandwidth, and the SLA index of each slice meets its agreed threshold.

[0011] In the aforementioned Giant Star Cloud Network resource adaptive slicing and SLA guarantee method, step S4 involves constructing an SLA violation prediction and closed-loop adaptive slicing reconstruction mechanism, including: Based on the future K-step resource state prediction output by the spatiotemporal graph convolutional network, calculate the SLA satisfaction prediction value of each slice in the future time slot; If the predicted value is lower than the SLA safety threshold of the slice, it is determined that the slice has an SLA violation risk in future time slots, and a violation warning is generated; otherwise, proceed to step S5. Upon receiving a violation warning, an incremental slice reconstruction is initiated: based on the predicted resource availability, alternative satellite nodes with sufficient resources are found within the orbital cluster to which the current slice belongs, and only the affected virtual function instances and link mappings are migrated, while the unaffected parts of the slice remain unchanged; If there are insufficient resources in the current orbital cluster, a request is made to the orbital domain coordination layer for cross-cluster resource allocation to migrate some slice instances to a neighboring orbital cluster, ensuring business continuity during the migration process; After the slice reconstruction is completed, the resource allocation status in the spatiotemporal graph model is updated, and the SLA index continues to be monitored in real time, forming a closed-loop control of prediction-early warning-reconstruction-verification.

[0012] In the above-mentioned adaptive slicing and SLA guarantee method for the Giants constellation cloud network resources, in step S5, at the orbital domain level, the satellites in the Giants constellation are divided into several orbital clusters according to the satellite orbital inclination and right ascension of the ascending node. Each orbital cluster consists of continuous satellites in the same orbital plane and satellites in adjacent orbital planes. Resource quotas are allocated to each orbital cluster. The resource quotas are determined based on the historical service load statistics and predicted service demand of the cluster, and are periodically adjusted with the orbital period as the granularity. At the network domain level, within each orbital cluster, the cluster head satellite or ground control station performs fine-grained slice instantiation within the cluster-level resource quota based on real-time service requests and SLA requirements. This includes selecting specific satellite nodes for the slice, allocating computing and storage resources, and reserving bandwidth for inter-satellite links.

[0013] In the above-mentioned adaptive slicing and SLA guarantee method for resources in the Giant Star constellation cloud network, the orbital domain layer and the network domain layer are coordinated through a resource quota feedback mechanism. When the actual resource consumption of the network domain layer continuously deviates from the quota, the network domain layer reports the deviation information to the orbital domain layer. The orbital domain layer reallocates the quota according to the resource distribution of the entire network to achieve cross-cluster resource balance.

[0014] In the aforementioned adaptive slicing and SLA guarantee method for the Giant Star Cloud network resources, the method employs a deep reinforcement learning agent for solution, specifically including: The state space consists of the resource state prediction results of the spatiotemporal graph convolutional network, the SLA satisfaction of the current slice, and the remaining resources of each node. The action space selects a resource allocation scheme for each slice, including the selection of satellite nodes for virtual function deployment, the allocation of computing and storage resources, and the reserved amount of link bandwidth; Reward function, r t =alpha (1 / |S|) sum s(eta_s) + beta U resource - lambda C migration C migration The migration overhead incurred by slice reconstruction, lambda is the migration penalty coefficient, |S| is the sum of SLA requirements for all slices, and sum is the migration cost. s(eta_s) U is the sum of the SLA satisfaction scores for all slices. resource The overall cloud network resource utilization rate is represented by alpha and beta, which are weighting coefficients. The agent adopts an Actor-Critic architecture, where the actor network outputs the probability distribution of resource allocation actions, the critic network evaluates the long-term value of actions, and iterative optimization is achieved through the policy gradient method.

[0015] A Giant Star Cloud Network Resource Adaptive Slicing and SLA Guarantee System includes: The spatiotemporal graph model construction module constructs a spatiotemporal graph model for the Giants constellation cloud-network convergence, including: representing the computing resources, storage resources, and inter-satellite link bandwidth resources of satellite nodes in the Giants constellation as a resource state graph that dynamically changes with time and spatial location, wherein each node is associated with a multi-dimensional resource feature vector, and each edge is associated with a link quality feature vector; The convolutional network construction module, based on the aforementioned Giant Star constellation cloud-network fusion spatiotemporal graph model, constructs a spatiotemporal graph convolutional network. This network includes a spatial graph convolutional module and a temporally gated convolutional module. The spatial graph convolutional module extracts spatial dependency features of the satellite network topology based on a dynamic adjacency matrix, while the temporally gated convolutional module extracts temporal evolution features of resource states based on a sliding time window. The spatial graph convolutional module and the temporally gated convolutional module are executed alternately to obtain cloud-network resource state prediction results for future multi-time slots. The resource allocation scheme calculation module, based on the cloud network resource status prediction results and service SLA requirements, constructs an SLA-aware cloud network joint slice optimization model. With the maximization of slice SLA satisfaction rate and the maximization of cloud network resource utilization as joint optimization objectives, it solves the resource allocation scheme for each slice under the capacity constraints of computing resources, storage resources, and bandwidth resources and the SLA index constraints. The violation prediction and reconstruction module constructs an SLA violation prediction and closed-loop adaptive slice reconstruction mechanism, including: real-time monitoring of the SLA index of each slice; when the spatiotemporal graph convolutional network predicts that a slice has an SLA violation risk in a future time slot, it triggers the resource reconfiguration of the slice and migrates the slice with predicted violations to a resource-sufficient orbital domain through incremental resource adjustment. The scalable slice management module constructs a two-layer slice coordination mechanism between the orbital domain and the network domain. This includes: dividing orbital clusters and allocating cluster-level resource quotas at the orbital domain level based on satellite orbital motion patterns; and performing fine-grained resource scheduling and slice instantiation within clusters at the network domain level based on real-time service requirements. The two layers work together to achieve scalable slice management at the scale of the Giants constellation.

[0016] A Giant Star Cloud Network Resource Adaptive Slicing and SLA Guarantee System includes: The spatiotemporal graph modeling module is used to construct the spatiotemporal graph model of the Giant Star Cloud Network Convergence, which uniformly represents computing, storage, and bandwidth resources as a time-varying graph structure. The spatiotemporal graph convolutional prediction module is used to perform multi-step prediction of cloud network resource status based on spatiotemporal graph convolutional networks, and output predictions of topology, resource availability and link quality for future time slots; The SLA-aware slice optimization module is used to build a joint optimization model based on resource status prediction and business SLA requirements to solve the cloud network resource allocation scheme for each slice. The SLA violation prediction and closed-loop reconstruction module is used to predict the risk of SLA violations in slices, trigger incremental slice reconstruction, and verify the reconstruction effect to form closed-loop control. The orbital domain-network domain two-layer coordination module is used to allocate cluster-level resource quotas at the orbital domain level and perform fine-grained slice instantiation at the network domain level. The two layers work together to achieve scalable slice management.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects: (1) This invention provides a method for adaptive slicing and SLA guarantee of cloud network resources in the Giants constellation, including: constructing a spatiotemporal graph model of cloud network fusion in the Giants constellation to uniformly represent the computing, storage resources and inter-satellite link bandwidth resources of satellite nodes; constructing a spatiotemporal graph convolutional network to perform multi-step forward prediction of cloud network resource status; constructing an SLA-aware cloud network joint slicing optimization model based on the prediction results and service SLA requirements; constructing an SLA violation prediction and closed-loop adaptive slicing reconstruction mechanism to trigger incremental resource reconfiguration before the actual occurrence of SLA violation; constructing a two-layer slicing coordination mechanism of orbital domain and network domain to realize scalable slicing management at the Giants constellation scale; This invention realizes joint adaptive slicing and forward-looking SLA guarantee of cloud network resources, and can support real-time slicing management of large-scale satellite constellations.

[0018] (2) Cloud-network joint slicing significantly improves resource utilization. This invention incorporates computing, storage, and bandwidth resources into a unified slicing optimization framework. Compared with existing solutions that only optimize bandwidth, it can simultaneously meet the dual needs of services for computing and transmission, avoiding resource waste caused by mismatch between computing and bandwidth resource allocation.

[0019] (3) Prediction-driven SLA assurance significantly reduces the violation rate. This invention utilizes the look-ahead prediction capability of spatiotemporal graph convolution to perform preventative slice reconstruction before SLA violations actually occur.

[0020] (4) Closed-loop adaptive reconstruction to ensure robustness. This invention constructs a "prediction-early warning-reconstruction-verification" closed loop. After reconstruction, the SLA index is continuously verified to ensure the reconstruction effect. If necessary, the reconstruction intensity is upgraded to form an adaptive convergence guarantee mechanism.

[0021] (5) Two-layer coordination mechanism, which can be extended to large-scale nodes. This invention decouples the management of the Giants Confluence slice into two levels: cluster-level quota management and intra-cluster instantiation, through two-layer coordination of the orbital domain and the network domain, which significantly reduces the computational complexity.

[0022] (6) Incremental reconstruction minimizes migration overhead. This invention adopts an incremental strategy during slice reconstruction, migrating only the affected parts while keeping the unaffected parts unchanged, thus significantly reducing service interruption time. Attached Figure Description

[0023] Figure 1 This is an overall flowchart of the adaptive slicing and SLA guarantee method for Giant Star Cloud Network resources in an embodiment of the present invention; Figure 2 This is a schematic diagram of the spatiotemporal graph model of the Giant Star Cloud Network Fusion in an embodiment of the present invention; Figure 3 This is a schematic diagram of the spatiotemporal graph convolutional network structure in an embodiment of the present invention; Figure 4 This is a flowchart of SLA violation prediction and closed-loop adaptive slice reconstruction in an embodiment of the present invention; Figure 5 This is a schematic diagram of the orbital domain-network domain two-layer slicing coordination mechanism in an embodiment of the present invention; Figure 6 This is a module structure diagram of the Giant Star Cloud Network resource adaptive slicing and SLA guarantee system in an embodiment of the present invention. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments: like Figure 1 As shown, this invention provides a method for adaptive slicing and SLA guarantee of resources in the Giant Star Cloud Network based on spatiotemporal graph convolution, including the following steps: I. Constructing a spatiotemporal graph model for the convergence of the Giant Star constellation cloud and network This includes representing the computing resources, storage resources, and inter-satellite link bandwidth resources of satellite nodes in the Giant Pivot constellation as a resource state diagram that dynamically changes with time and spatial location. Each node is associated with a multi-dimensional resource feature vector, and each edge is associated with a link quality feature vector.

[0025] The Giant Star Network is modeled as a time-varying graph G(t) = (V, E(t), X). V(t) X E(t) This provides a unified representation of the computing, storage, and inter-satellite link bandwidth resources of satellite nodes. Node set V: Contains all satellite nodes and ground station nodes. Satellite nodes simultaneously handle communication forwarding, on-board computing, and data storage functions, while ground station nodes provide centralized cloud computing capabilities and network management functions.

[0026] The dynamic link set E(t) is determined by satellite orbital dynamics and communication geometry, and includes three types: inter-satellite links within the same orbit, inter-satellite links between different orbits, and satellite-to-ground links. Due to the high-speed motion of low-Earth orbit satellites, the link set changes continuously over time.

[0027] Node multidimensional resource feature matrix X V(t) Each node's feature vector contains available computing resources, available storage resources, current load rate, and orbital position coordinates.

[0028] Edge feature matrix X E(t) The feature vector of each link includes available bandwidth, propagation delay, bit error rate, and remaining link lifetime.

[0029] This model represents cloud (computing + storage) and network (bandwidth + link) resources in a unified time-varying graph framework, providing a data foundation for subsequent spatiotemporal graph convolution prediction and joint slicing optimization.

[0030] II. Constructing a spatiotemporal graph convolutional network for cloud network resource status prediction A cloud-network joint slicing optimization model with SLA awareness is constructed. The joint optimization objectives are to maximize the slice SLA satisfaction rate and maximize the cloud-network resource utilization rate. Under the capacity constraints of computing resources, storage resources, and bandwidth resources and the SLA index constraints, the resource configuration scheme of each slice is solved.

[0031] This includes: Based on the spatiotemporal graph model from step one, constructing a spatiotemporal graph convolutional network (STGCN), which consists of alternating stacks of spatial graph convolutional modules and temporally gated convolutional modules. Spatial graph convolution module: For each time slot, spatial features are aggregated based on a dynamic adjacency matrix to capture the spatial topological dependencies between satellite nodes within the same time slot. Node representation updates employ graph convolution formulas, utilizing the time-varying adjacency matrix of the satellite network to achieve dynamic topology awareness.

[0032] The spatial graph convolutional module employs a graph convolutional network to aggregate spatial features from the node feature matrix and dynamic adjacency matrix at time slot t. The node representation update method of the spatial graph convolutional module is: H s (l+1) = sigma(D tilde(t) (-1 / 2) A tilde(t) D tilde(t) (-1 / 2) H s (l) W s (l) ), where A tilde(t) = A(t) + I is the adjacency matrix with self-loops, A(t) is the initial adjacency matrix, I is the identity matrix, D tilde(t) Let W be the corresponding degree matrix. s (l) H is the trainable weight matrix for the l-th spatial convolution layer. s (l) H represents the node state before the update. s (l+1) The updated node state; Temporally Gated Convolution Module: This module performs a one-dimensional convolution on the temporal feature sequence output by the spatial convolution, and uses gated linear units (GLUs) to extract temporal periodicity and trends. The GLU gating mechanism enables the model to adaptively focus on the temporal features most important for future predictions.

[0033] The temporally gated convolution module performs temporal convolution on the temporal feature sequence output by the spatial convolution, employing gated linear units to capture the temporal periodicity and trend of resource states. In the temporally gated convolution module, H... t = sigma(W1 H s (L) )x(W2 H s (L) ),in This is a one-dimensional convolution operation, where sigma is the sigmoid function; W2 is the weight matrix for the second layer, and H... s (L) For node states, W1 is a layer weight matrix, and H... t This is a time-gated convolution function.

[0034] Multi-layer stacking and prediction output: After multiple layers of spatial-temporal convolutional modules are stacked alternately, the cloud network resource status prediction for the next K time slots is output through a fully connected prediction layer, including the predicted values ​​of available computing resources and storage resources of each node, and the predicted values ​​of available bandwidth and latency of each link.

[0035] The core advantage of this module lies in the fact that, by leveraging the periodicity and determinism of satellite orbital motion, spatiotemporal graph convolution can accurately predict future topological changes and resource status, providing forward-looking information for predictive slicing decisions.

[0036] III. Constructing an SLA-Aware Cloud-Network Joint Slicing Optimization Model Based on the cloud network resource status prediction results and service SLA requirements from step two, an SLA-aware cloud network joint slicing optimization model is constructed. With the joint optimization objectives of maximizing slice SLA satisfaction rate and maximizing cloud network resource utilization, the resource configuration scheme of each slice is solved under the capacity constraints of computing resources, storage resources, and bandwidth resources and the SLA index constraints.

[0037] SLA Requirement Definition: The SLA requirement for each network slice is defined as a four-tuple, representing the upper limit of end-to-end latency, the lower limit of guaranteed bandwidth, the lower limit of reliability, and the upper limit of latency jitter. SLA requirements differ significantly between different types of slices.

[0038] SLA Satisfaction Calculation: The SLA satisfaction of a slice is defined as the proportion of all business flows in that slice that satisfy all SLA metrics.

[0039] Joint optimization objective: to maximize the slice SLA satisfaction rate and the cloud network resource utilization rate.

[0040] Constraints include computing resource constraints, storage resource constraints, bandwidth constraints, and SLA (Service Level Agreement) constraints.

[0041] In this embodiment, the method for constructing the SLA-aware cloud-network joint slicing optimization model specifically includes: Define the SLA requirement for each network slice s as a quadruple SLA. s = (D smax B smin , R smin J smax ), where D smax B smin , R smin J smax These represent the upper limit of end-to-end latency, the lower limit of guaranteed bandwidth, the lower limit of reliability, and the upper limit of latency jitter, respectively. The SLA satisfaction level of slice s is defined as the proportion of all service flows under that slice that meet the SLA metric, i.e., eta s =(1 / |F s |) sum f (I(QoS indicators meet SLA requirements)), where F s Let s be the set of business flows carried by slice s, I be the indicator function, and sum f The sum of indicators required to meet SLA requirements; the joint optimization objective is: max alpha (1 / |S|) sum s(eta_s) +beta U resource U resource The total cloud resource utilization rate is represented by alpha and beta, with |S| being the sum of SLA requirements for all slices. s(eta_s) This is the sum of SLA satisfaction rates for all slices. Constraints include: the total computing resources allocated to each satellite node do not exceed its available computing resources; the total storage resources allocated do not exceed its available storage resources; the total bandwidth allocated to each link does not exceed its available bandwidth; and the SLA metrics for each slice meet their agreed-upon thresholds.

[0042] The core innovation of this joint optimization model lies in using the predicted future resource state as the optimization input, rather than the current instantaneous state, so that the slicing decision is forward-looking; at the same time, it unifies the optimization of computing, storage and bandwidth resources to achieve true cloud-network joint slicing.

[0043] IV. Constructing an SLA violation prediction and closed-loop adaptive slice reconstruction mechanism This includes real-time monitoring of the SLA metrics of each slice. When the spatial-graphical convolutional network predicts that a slice has an SLA violation risk in a future time slot, it triggers resource reconfiguration for that slice. Through incremental resource adjustment, the slice predicted to have a violation is migrated to a resource-sufficient orbital domain, and then proceeds to step five. If the spatial-graphical convolutional network does not predict that any slice has an SLA violation risk in a future time slot, then proceeds to step five.

[0044] SLA violation prospective prediction: Based on the prediction results of the spatiotemporal graph convolutional network, the predicted SLA satisfaction value of each slice in future time slots is calculated prospectively. When the predicted value is lower than the safety threshold, a violation risk is determined and an early warning is generated.

[0045] Incremental slice reconstruction: Incremental reconstruction is initiated upon receiving an alert. The core principle is to minimize migration overhead: priority is given to finding alternative nodes within the current orbital cluster, migrating only the affected virtual function instances, and keeping the unaffected parts of the slice unchanged; if resources within the cluster are insufficient, cross-cluster allocation is requested.

[0046] Closed-loop verification: After reconstruction, update the spatiotemporal graph model status and continue monitoring SLA indicators to form a closed loop of "prediction-early warning-reconstruction-verification". If verification fails, escalate the reconstruction efforts.

[0047] The innovation of this mechanism lies in its use of the forward-looking predictive capabilities of spatiotemporal graph convolution to perform preventative reconstruction before SLA violations actually occur, thus avoiding the lag of passive response.

[0048] In this embodiment, the SLA violation prediction and closed-loop adaptive slice reconstruction mechanism is constructed, which specifically includes the following steps: Based on the future K-step resource state prediction output by the spatiotemporal graph convolutional network, calculate the SLA satisfaction prediction value of each slice in the future time slot; If the predicted value is lower than the SLA safety threshold of the slice, it is determined that the slice has an SLA violation risk in future time slots, and a violation warning is generated; otherwise, proceed to step five. Upon receiving a violation warning, an incremental slice reconstruction is initiated: based on the predicted resource availability, alternative satellite nodes with sufficient resources are found within the orbital cluster to which the current slice belongs, and only the affected virtual function instances and link mappings are migrated, while the unaffected parts of the slice remain unchanged; If there are insufficient resources in the current orbital cluster, a request is made to the orbital domain coordination layer for cross-cluster resource allocation to migrate some slice instances to a neighboring orbital cluster, ensuring business continuity during the migration process; After the slice reconstruction is completed, the resource allocation status in the spatiotemporal graph model is updated, and the SLA index continues to be monitored in real time, forming a closed-loop control of prediction-early warning-reconstruction-verification.

[0049] V. Constructing a two-layer slicing coordination mechanism between the track domain and the network domain This includes dividing orbital clusters and allocating cluster-level resource quotas at the orbital domain level based on satellite orbital motion patterns, and performing fine-grained resource scheduling and slice instantiation within clusters at the network domain level based on real-time service requirements. The two layers work together to achieve scalable slice management at the scale of the Giants constellation.

[0050] Orbital domain layer: Based on the satellite orbital inclination and right ascension of the ascending node, the satellites in the constellation Giants are divided into several orbital clusters. Each orbital cluster consists of continuous satellites in the same orbital plane and satellites in adjacent orbital planes. Resource quotas are allocated to each orbital cluster. The quotas are determined based on the historical service load statistics of the cluster and AI-predicted service demand, and are periodically adjusted with the orbital period as the granularity.

[0051] Network Domain Layer: Within each orbital cluster, the cluster head satellite or ground control station performs fine-grained slice instantiation within the cluster-level resource quota based on real-time service requests and SLA requirements. This includes selecting specific satellite nodes for the slice, allocating computing and storage resources, and reserving bandwidth for inter-satellite links. Two-layer coordination: The two layers coordinate through a resource quota feedback mechanism. When the actual resource consumption of the network domain layer continuously deviates from the quota, it reports the deviation information. The network domain layer reports the deviation information to the track domain layer, and the track domain layer reallocates the quota according to the resource distribution of the entire network to achieve cross-cluster resource balance.

[0052] The innovation of this mechanism lies in the fact that by using coarse-grained quota management in the orbital domain and fine-grained instantiation division of labor in the network domain, the complexity of slice management at the giant constellation scale is significantly reduced, thus achieving scalable slice management.

[0053] The proposed method for adaptive resource slicing and SLA guarantee in the Giant Star Cloud Network employs a deep reinforcement learning agent for solution, specifically including: The state space consists of the resource state prediction results of the spatiotemporal graph convolutional network, the SLA satisfaction of the current slice, and the remaining resources of each node. The action space selects a resource allocation scheme for each slice, including the selection of satellite nodes for virtual function deployment, the allocation of computing and storage resources, and the reserved amount of link bandwidth; Reward function, r t =alpha (1 / |S|) sum s(eta_s) + beta U resource - lambda C migration C migrationThe migration overhead incurred by slice reconstruction, lambda is the migration penalty coefficient, |S| is the sum of SLA requirements for all slices, and sum is the migration cost. s(eta_s) U is the sum of the SLA satisfaction scores for all slices. resource The overall cloud network resource utilization rate is represented by alpha and beta, which are weighting coefficients. The agent adopts an Actor-Critic architecture, where the actor network outputs the probability distribution of resource allocation actions, the critic network evaluates the long-term value of actions, and iterative optimization is achieved through the policy gradient method.

[0054] This invention provides a Giant Star Cloud Network resource adaptive slicing and SLA guarantee system, comprising the following modules: The spatiotemporal graph model construction module constructs a spatiotemporal graph model for the Giants constellation cloud-network convergence, including: representing the computing resources, storage resources, and inter-satellite link bandwidth resources of satellite nodes in the Giants constellation as a resource state graph that dynamically changes with time and spatial location, wherein each node is associated with a multi-dimensional resource feature vector, and each edge is associated with a link quality feature vector; The convolutional network construction module, based on the aforementioned Giant Star constellation cloud-network fusion spatiotemporal graph model, constructs a spatiotemporal graph convolutional network. This network includes a spatial graph convolutional module and a temporally gated convolutional module. The spatial graph convolutional module extracts spatial dependency features of the satellite network topology based on a dynamic adjacency matrix, while the temporally gated convolutional module extracts temporal evolution features of resource states based on a sliding time window. The spatial graph convolutional module and the temporally gated convolutional module are executed alternately to obtain cloud-network resource state prediction results for future multi-time slots. The resource allocation scheme calculation module, based on the cloud network resource status prediction results and service SLA requirements, constructs an SLA-aware cloud network joint slice optimization model. With the maximization of slice SLA satisfaction rate and the maximization of cloud network resource utilization as joint optimization objectives, it solves the resource allocation scheme for each slice under the capacity constraints of computing resources, storage resources, and bandwidth resources and the SLA index constraints. The violation prediction and reconstruction module constructs an SLA violation prediction and closed-loop adaptive slice reconstruction mechanism, including: real-time monitoring of the SLA index of each slice; when the spatiotemporal graph convolutional network predicts that a slice has an SLA violation risk in a future time slot, it triggers the resource reconfiguration of the slice and migrates the slice with predicted violations to a resource-sufficient orbital domain through incremental resource adjustment. The scalable slice management module constructs a two-layer slice coordination mechanism between the orbital domain and the network domain. This includes: dividing orbital clusters and allocating cluster-level resource quotas at the orbital domain level based on satellite orbital motion patterns; and performing fine-grained resource scheduling and slice instantiation within clusters at the network domain level based on real-time service requirements. The two layers work together to achieve scalable slice management at the scale of the Giants constellation.

[0055] This invention provides a Giant Star Cloud Network resource adaptive slicing and SLA guarantee system, comprising the following modules: The spatiotemporal graph modeling module is used to construct the spatiotemporal graph model of the Giant Star Cloud Network Convergence, which uniformly represents computing, storage, and bandwidth resources as a time-varying graph structure. The spatiotemporal graph convolutional prediction module is used to perform multi-step prediction of cloud network resource status based on spatiotemporal graph convolutional networks, and output predictions of topology, resource availability and link quality for future time slots; The SLA-aware slice optimization module is used to build a joint optimization model based on resource status prediction and business SLA requirements to solve the cloud network resource allocation scheme for each slice. The SLA violation prediction and closed-loop reconstruction module is used to predict the risk of SLA violations in slices, trigger incremental slice reconstruction, and verify the reconstruction effect to form closed-loop control. The orbital domain-network domain two-layer coordination module is used to allocate cluster-level resource quotas at the orbital domain level and perform fine-grained slice instantiation at the network domain level. The two layers work together to achieve scalable slice management.

[0056] Example 1 like Figure 1 As shown, the adaptive slicing and SLA guarantee method for Giants Cloud Network resources based on spatiotemporal graph convolution in this embodiment includes the following steps: Step S1: Construct the Giant Star Cloud-Network Fusion Spatiotemporal Graph Model like Figure 2 As shown, taking the Walker constellation, which contains 1,584 low-Earth orbit satellites, as an example, the satellites are distributed across 72 orbital planes, with 22 satellites in each orbital plane, an orbital altitude of 550 km, and an orbital inclination of 53 degrees.

[0057] The network is modeled as a time-varying graph G(t) = (V, E(t), X). V(t) X E(t) ): The node set V contains 1584 satellite nodes and 12 ground station nodes, for a total of 1596 nodes.

[0058] Each satellite node has computing resources (such as onboard FPGA / DSP, equivalent to 2-8 vCPUs), storage resources (4-16GB solid-state storage), and communication forwarding capabilities. Ground station nodes have even richer computing and storage resources.

[0059] The dynamic link set E(t) is determined by orbital dynamics: inter-satellite links within the same orbital plane maintain stable connections between adjacent satellites; inter-satellite links between satellites in adjacent orbital planes are established and periodically connected and disconnected as the satellites move; satellite-to-ground links are established between a satellite and a ground station and are constrained by elevation angle and distance. Using one orbital period (approximately 95 minutes) as the time granularity, the period is divided into T=24 time slots, each approximately 4 minutes long.

[0060] The node feature vector contains available computing resources, available storage resources, current load rate, orbital longitude, and orbital latitude. The computing and storage resources are time-varying functions (dynamically changing with the business load), and the orbital position is accurately calculated by ephemeris.

[0061] The link feature vector includes available bandwidth, propagation delay, bit error rate, and remaining link lifetime. Bandwidth and delay vary with the geometric distance of the link, and the remaining link lifetime is calculated based on orbit prediction.

[0062] Step S2: Construct a spatiotemporal graph convolutional network for cloud network resource state prediction. like Figure 3 As shown, the spatiotemporal graph convolutional network consists of two stacked spatial-temporal convolutional blocks, each containing a spatial graph convolutional layer and a temporally gated convolutional layer.

[0063] Spatial graph convolutional layer: Input is a sliding window T w =8 time slot node feature matrix sequences and corresponding dynamic adjacency matrix sequences. For each time slot, neighbor node information is aggregated through graph convolution. The first spatial convolutional layer maps 5-dimensional node features to a 64-dimensional hidden space, and the second layer maintains the 64-dimensionality.

[0064] Temporally Gated Convolutional Layer: A one-dimensional convolution (kernel size 3) is performed on the temporal sequence output from the spatial convolution, and temporal features are extracted through the GLU gating mechanism. The GLU's sigmoid gate adaptively controls the information flow, enabling the model to focus on the temporal features most relevant to the prediction.

[0065] Prediction output layer: Two spatiotemporal convolutional blocks followed by a fully connected layer, outputting the prediction of the future K=6 time slots (approximately 24 minutes), including the predicted values ​​of available computing resources and storage resources for each node, and the predicted values ​​of available bandwidth and latency for each link.

[0066] Training method: Spatiotemporal map snapshot data collected in historical orbital cycles are used as the training set, with the real future state as the label. The MSE loss function is used for training, the Adam optimizer is used, the learning rate is 0.001, and the training converges after 200 rounds.

[0067] Step S3: Construct an SLA-aware cloud-network joint slicing optimization model Three types of network slices are defined: ultra-low latency slices (D...max =15ms, R min =99.999%, J max =2ms, used for real-time interaction and remote control), large bandwidth slicing (B min =100Mbps, D max =80ms, BER<10 -3 (For high-definition video and data synchronization) and massive connection slices (connection success rate >99.5%, aggregate packet loss rate <1%, for IoT and sensor networks).

[0068] For each arriving slice request, a resource allocation scheme is solved using a deep reinforcement learning agent based on the predicted resource state from step S2. The agent employs an Actor-Critic architecture, with the actor network being a two-layer fully connected network (128-64 neurons) that outputs the action probability distribution; the critic network evaluates the long-term value of the actions. Iterative optimization is performed using the policy gradient method, and the system is trained offline and then deployed online at a ground control center.

[0069] Step S4: Construct an SLA violation prediction and closed-loop adaptive slice reconstruction mechanism like Figure 4 As shown, the closed-loop reconfiguration process is as follows: (1) Prediction: Run the space-time graph convolutional network in each time slot to predict the SLA satisfaction of each slice in the next 6 steps.

[0070] (2) Warning: Set the safety threshold to 0.95 (i.e., if the SLA satisfaction rate is lower than 95%, a warning will be issued). If the predicted value is lower than the threshold, a violation warning will be generated.

[0071] (3) Reconstruction: After receiving the warning, the incremental reconstruction strategy is as follows: Level 1 Reconfiguration (Local Adjustment): Within the current orbital cluster, affected virtual functions are migrated to nearby satellite nodes with sufficient resources, adjusting only the relevant link mappings. Migration overhead is low, making it suitable for scenarios with localized resource shortages.

[0072] Secondary refactoring (cross-cluster migration): If there are insufficient resources within a cluster, a cross-cluster allocation is requested to migrate some slice instances to a neighboring orbital cluster. This has a higher overhead and is suitable for scenarios with insufficient cluster-level resources.

[0073] Level 3 Refactoring (Overall Remapping): If the first two levels of refactoring fail to meet the SLA, the slice is remapped as a whole, and the deployment locations and link mappings of all virtual functions are replanned. This has the highest overhead and is used as a fallback solution.

[0074] (4) Verification: After reconstruction, update the spatiotemporal graph model state and re-predict the SLA satisfaction. If the safety threshold is met, the reconstruction is successful; otherwise, upgrade to a higher level of reconstruction.

[0075] (5) Closed loop: Continuous monitoring to form an adaptive convergence guarantee loop.

[0076] Step S5: Construct a two-layer slicing coordination mechanism between the orbital domain and the network domain. like Figure 5 As shown, the two-tier coordination mechanism is implemented as follows: Orbital Domain Layer: The 72 orbital planes are divided into 12 orbital clusters, grouped into sets of 6 adjacent planes. Each cluster contains approximately 132 satellites. Based on the service demand forecast from step S2, computing, storage, and bandwidth resource quotas are allocated to each cluster. The quotas are adjusted at the granularity of the orbital period (95 minutes) and remain stable within each adjustment interval.

[0077] Network Domain Layer: Each orbital cluster designates a cluster head satellite (selected as the satellite with the lowest current load and most abundant resources) responsible for instantiating slices within the cluster. The cluster head allocates specific satellite nodes, computing / storage resources, and link bandwidth within the cluster-level quota based on real-time service requests and SLA requirements.

[0078] Two-layer coordination: At the end of each orbital cycle, each cluster head reports resource usage deviation information. If a cluster continues to exceed its quota, the orbital domain layer allocates quotas from adjacent clusters with sufficient resources to achieve resource balance across the entire network. The granularity of cross-cluster allocation is based on orbital planes, avoiding the signaling overhead caused by frequent fine-grained adjustments.

[0079] Example 2 like Figure 6 As shown, the Giant Star Cloud Network Resource Adaptive Slicing and SLA Guarantee System based on Spatiotemporal Graph Convolution in this embodiment includes: Spatiotemporal diagram modeling module: Deployed at the ground control center, it receives ephemeris data, telemetry data and network monitoring data, and builds and continuously updates the Giants constellation cloud-network fusion spatiotemporal diagram model.

[0080] Spatiotemporal graph convolutional prediction module: Deployed on a GPU server in the ground control center, it runs a trained STGCN model, takes the current spatiotemporal graph snapshot as input, and outputs a resource state prediction for the next 6 steps.

[0081] SLA Perception Slice Optimization Module: Deployed at the ground control center, it runs a deep reinforcement learning agent to solve the cloud network resource allocation scheme for each slice based on prediction results and business SLA requirements.

[0082] SLA violation prediction and closed-loop reconfiguration module: Deployed at the ground control center, it continuously evaluates the predicted SLA satisfaction value of each slice, triggers incremental reconfiguration, and verifies the reconfiguration effect.

[0083] The orbital domain-network domain two-layer coordination module is deployed at the ground control center and is responsible for cluster-level quota management. The cluster head management function of the network domain layer is deployed at the cluster head satellites or ground area control stations of each orbital cluster and is responsible for intra-cluster slice instantiation.

[0084] The modules work together to form a complete closed loop of "modeling-prediction-optimization-assistance-coordination", enabling adaptive slicing and SLA assurance of the Giant Star Cloud Network resources.

[0085] The above description is only the best specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes 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 protection scope of the present invention.

[0086] The contents not described in detail in this specification are common knowledge to those skilled in the art.

Claims

1. A method for adaptive slicing and SLA guarantee of resources in the Giant Star constellation cloud network, characterized in that, Includes the following steps: S1. Construct a spatiotemporal graph model for cloud-network convergence in the Giants constellation, including: representing the computing resources, storage resources, and inter-satellite link bandwidth resources of satellite nodes in the Giants constellation as a resource state graph that dynamically changes with time and spatial location, wherein each node is associated with a multi-dimensional resource feature vector and each edge is associated with a link quality feature vector; S2. Based on the aforementioned Giant Star constellation cloud-network fusion spatiotemporal graph model, a spatiotemporal graph convolutional network is constructed. This network includes a spatial graph convolutional module and a temporally gated convolutional module. The spatial graph convolutional module extracts spatial dependency features of the satellite network topology based on a dynamic adjacency matrix, while the temporally gated convolutional module extracts temporal evolution features of resource states based on a sliding time window. The spatial graph convolutional module and the temporally gated convolutional module are executed alternately to obtain cloud-network resource state prediction results for future multi-time slots. S3. Based on the cloud network resource status prediction results and service SLA requirements, construct an SLA-aware cloud network joint slice optimization model. With the maximization of slice SLA satisfaction rate and the maximization of cloud network resource utilization as joint optimization objectives, solve the resource configuration scheme of each slice under the capacity constraints of computing resources, storage resources, and bandwidth resources and the SLA index constraints. S4. Construct an SLA violation prediction and closed-loop adaptive slice reconstruction mechanism, including: real-time monitoring of the SLA index of each slice; when the spatiotemporal graph convolutional network predicts that a slice has an SLA violation risk in a future time slot, triggering resource reconfiguration of the slice, and migrating the slice with the predicted violation to a resource-sufficient orbital domain through incremental resource adjustment, proceeding to step S5; if the spatiotemporal graph convolutional network does not predict that any slice has an SLA violation risk in a future time slot, proceeding to step S5; S5. Construct a two-layer slicing coordination mechanism between the orbital domain and the network domain, including: dividing orbital clusters and allocating cluster-level resource quotas at the orbital domain level according to the satellite orbital motion law; and performing fine-grained resource scheduling and slice instantiation within the clusters at the network domain level according to real-time service requirements. The two layers work together to achieve scalable slice management at the scale of the Giants constellation.

2. The method for adaptive slicing and SLA guarantee of resources in the Giant Star Cloud Network according to claim 1, characterized in that, In step S1, the method for constructing the Giant Star Cloud-Network Fusion Spatiotemporal Graph Model includes: The Giants constellation network is modeled as the following time-varying graph: G(t) = (V, E(t), X V(t) , X E(t) ) in: V is the union of the set of satellite nodes and the set of ground station nodes; This is the set of inter-satellite links and satellite-to-ground links active in time slot t; X V(t) This is a multidimensional resource feature matrix for nodes. The feature vector of each node contains available computing resources, available storage resources, current load rate, and track position coordinates. X E(t) The feature matrix is ​​the edge feature matrix. The feature vector of each link includes available bandwidth, propagation delay, bit error rate, and remaining link lifetime.

3. The method for adaptive slicing and SLA guarantee of resources in the Giant Star Cloud Network according to claim 1, characterized in that, In step S2, the spatial graph convolution module uses a graph convolutional network to perform spatial feature aggregation on the node feature matrix and dynamic adjacency matrix of time slot t; The temporally gated convolution module performs temporal dimension convolution on the temporal feature sequence output by the spatial convolution, and uses gated linear units to capture the temporal periodicity and trend of resource status.

4. The method for adaptive slicing and SLA guarantee of resources in the Giant Star Cloud Network according to claim 3, characterized in that, The node representation update method of the spatial graph convolution module is as follows: H s (l+1) = sigma(D tilde(t) (-1 / 2) A tilde(t) D tilde(t) (-1 / 2) H s (l) W s (l) ), where A tilde(t) = A(t) + I is the adjacency matrix with self-loops, A(t) is the initial adjacency matrix, I is the identity matrix, D tilde(t) Let W be the corresponding degree matrix. s (l) H is the trainable weight matrix for the l-th spatial convolution layer. s (l) H represents the node state before the update. s (l+1) The updated node state; In the time-gated convolution module, H t = sigma(W1 H s (L) )x(W2 H s (L) ),in This is a one-dimensional convolution operation, where sigma is the sigmoid function; W2 is the weight matrix for the second layer, and H... s (L) For node states, W1 is a layer weight matrix, and H... t This is a time-gated convolution function.

5. The method for adaptive slicing and SLA guarantee of resources in the Giant Star Cloud Network according to claim 1, characterized in that, In step S3, the method for constructing the SLA-aware cloud-network joint slicing optimization model includes: Define the SLA requirement for each network slice s as a quadruple SLA. s = (D smax B smin , R smin J smax ), where D smax B smin , R smin J smax These represent the upper limit of end-to-end latency, the lower limit of guaranteed bandwidth, the lower limit of reliability, and the upper limit of latency jitter, respectively. The SLA satisfaction level of slice s is defined as the proportion of all service flows under that slice that meet the SLA metric, i.e., eta s = (1 / |F s |) sum f (I(QoS indicators meet SLA requirements)), where F s Let s be the set of business flows carried by slice s, I be the indicator function, and sum f The sum of indicators required to meet SLA requirements; the joint optimization objective is: max alpha (1 / |S|) sum s(eta_s) + beta U resource U resource The total cloud resource utilization rate is represented by alpha and beta, with |S| being the sum of SLA requirements for all slices. s(eta_s) This is the sum of SLA satisfaction for all slices.

6. The method for adaptive slicing and SLA guarantee of resources in the Giant Star Cloud Network according to claim 1, characterized in that, In step S3, the capacity constraints and SLA index constraints of the computing resources, storage resources, and bandwidth resources include: the total amount of computing resources allocated to each satellite node does not exceed its available computing resources, the total amount of storage resources allocated does not exceed its available storage resources, the total amount of bandwidth allocated to each link does not exceed its available bandwidth, and the SLA index of each slice meets its agreed threshold.

7. The method for adaptive slicing and SLA guarantee of resources in the Giant Star Cloud Network according to claim 1, characterized in that, In step S4, an SLA violation prediction and closed-loop adaptive slice reconstruction mechanism is constructed, including: Based on the future K-step resource state prediction output by the spatiotemporal graph convolutional network, calculate the SLA satisfaction prediction value of each slice in the future time slot; If the predicted value is lower than the SLA safety threshold of the slice, it is determined that the slice has an SLA violation risk in future time slots, and a violation warning is generated; otherwise, proceed to step S5. Upon receiving a violation warning, an incremental slice reconstruction is initiated: based on the predicted resource availability, alternative satellite nodes with sufficient resources are found within the orbital cluster to which the current slice belongs, and only the affected virtual function instances and link mappings are migrated, while the unaffected parts of the slice remain unchanged; If there are insufficient resources in the current orbital cluster, a request is made to the orbital domain coordination layer for cross-cluster resource allocation to migrate some slice instances to a neighboring orbital cluster, ensuring business continuity during the migration process; After the slice reconstruction is completed, the resource allocation status in the spatiotemporal graph model is updated, and the SLA index continues to be monitored in real time, forming a closed-loop control of prediction-early warning-reconstruction-verification.

8. The method for adaptive slicing and SLA guarantee of resources in the Giant Star Cloud Network according to claim 1, characterized in that, In step S5, at the orbital domain level, the satellites in the constellation Giants are divided into several orbital clusters based on their orbital inclination and right ascension of the ascending node. Each orbital cluster consists of consecutive satellites in the same orbital plane and satellites in adjacent orbital planes. Resource quotas are allocated to each orbital cluster. These resource quotas are determined based on the cluster's historical service load statistics and predicted service demand, and are periodically adjusted with the orbital period as the granularity. At the network domain level, within each orbital cluster, the cluster head satellite or ground control station performs fine-grained slice instantiation within the cluster-level resource quota based on real-time service requests and SLA requirements. This includes selecting specific satellite nodes for the slice, allocating computing and storage resources, and reserving bandwidth for inter-satellite links.

9. The method for adaptive slicing and SLA guarantee of resources in the Giant Star Cloud Network according to claim 8, characterized in that, The orbital domain layer and the network domain layer are coordinated through a resource quota feedback mechanism. When the actual resource consumption of the network domain layer continuously deviates from the quota, the network domain layer reports the deviation information to the orbital domain layer. The orbital domain layer then reallocates the quota according to the resource distribution of the entire network to achieve cross-cluster resource balance.

10. The method for adaptive slicing and SLA guarantee of resources in the Giant Star Cloud Network according to claim 1, characterized in that, The method employs a deep reinforcement learning agent to solve the problem, specifically including: The state space consists of the resource state prediction results of the spatiotemporal graph convolutional network, the SLA satisfaction of the current slice, and the remaining resources of each node. The action space selects a resource allocation scheme for each slice, including the selection of satellite nodes for virtual function deployment, the allocation of computing and storage resources, and the reserved amount of link bandwidth; Reward function, r t =alpha (1 / |S|) sum s(eta_s) + beta U resource - lambda C migration C migration The migration overhead incurred by slice reconstruction, lambda is the migration penalty coefficient, |S| is the sum of SLA requirements for all slices, and sum is the migration cost. s(eta_s) U is the sum of the SLA satisfaction scores for all slices. resource The overall cloud network resource utilization rate is represented by alpha and beta, which are weighting coefficients. The agent adopts an Actor-Critic architecture, where the actor network outputs the probability distribution of resource allocation actions, the critic network evaluates the long-term value of actions, and iterative optimization is achieved through the policy gradient method.

11. A Giant Star Cloud Network Resource Adaptive Slicing and SLA Guarantee System, characterized in that, include: The spatiotemporal graph model construction module constructs a spatiotemporal graph model for the Giants constellation cloud-network convergence, including: representing the computing resources, storage resources, and inter-satellite link bandwidth resources of satellite nodes in the Giants constellation as a resource state graph that dynamically changes with time and spatial location, where each node is associated with a multi-dimensional resource feature vector and each edge is associated with a link quality feature vector; The convolutional network construction module, based on the aforementioned Giant Star constellation cloud-network fusion spatiotemporal graph model, constructs a spatiotemporal graph convolutional network. This network includes a spatial graph convolutional module and a temporally gated convolutional module. The spatial graph convolutional module extracts spatial dependency features of the satellite network topology based on a dynamic adjacency matrix, while the temporally gated convolutional module extracts temporal evolution features of resource states based on a sliding time window. The spatial graph convolutional module and the temporally gated convolutional module are executed alternately to obtain cloud-network resource state prediction results for future multi-time slots. The resource allocation scheme calculation module, based on the cloud network resource status prediction results and service SLA requirements, constructs an SLA-aware cloud network joint slice optimization model. With the maximization of slice SLA satisfaction rate and the maximization of cloud network resource utilization as joint optimization objectives, it solves the resource allocation scheme for each slice under the capacity constraints of computing resources, storage resources, and bandwidth resources and the SLA index constraints. The violation prediction and reconstruction module constructs an SLA violation prediction and closed-loop adaptive slice reconstruction mechanism, including: real-time monitoring of the SLA index of each slice; when the spatiotemporal graph convolutional network predicts that a slice has an SLA violation risk in a future time slot, it triggers the resource reconfiguration of the slice and migrates the slice with the predicted violation to the resource-sufficient orbital domain through incremental resource adjustment. The scalable slice management module constructs a two-layer slice coordination mechanism between the orbital domain and the network domain. This includes: dividing orbital clusters and allocating cluster-level resource quotas at the orbital domain level based on satellite orbital motion patterns; and performing fine-grained resource scheduling and slice instantiation within clusters at the network domain level based on real-time service requirements. The two layers work together to achieve scalable slice management at the scale of the Giants constellation.

12. A resource adaptive slicing and SLA guarantee system for the Giant Star constellation cloud network, characterized in that, include: The spatiotemporal graph modeling module is used to construct the spatiotemporal graph model of the Giant Star Cloud Network Convergence, which uniformly represents computing, storage, and bandwidth resources as a time-varying graph structure. The spatiotemporal graph convolutional prediction module is used to perform multi-step prediction of cloud network resource status based on spatiotemporal graph convolutional networks, and output predictions of topology, resource availability and link quality for future time slots; The SLA-aware slice optimization module is used to build a joint optimization model based on resource status prediction and business SLA requirements to solve the cloud network resource allocation scheme for each slice. The SLA violation prediction and closed-loop reconstruction module is used to predict the risk of SLA violations in slices, trigger incremental slice reconstruction, and verify the reconstruction effect to form closed-loop control. The orbital domain-network domain two-layer coordination module is used to allocate cluster-level resource quotas at the orbital domain level and perform fine-grained slice instantiation at the network domain level. The two layers work together to achieve scalable slice management.

Images