Content caching method considering dynamic network topology in satellite edge calculation

Abstract

According to the content caching method considering the dynamic network topology in the satellite edge calculation, when modeling is conducted on the content caching problem in the satellite edge calculation, the dynamic nature of the satellite network topology is considered, and dual optimization of transmission time delay and transmission energy consumption is achieved under the constraint condition of the satellite caching capacity and the repeated caching number; according to the method, the content caching problem is solved by utilizing multi-agent graph attention reinforcement learning, and the purpose of solving a caching strategy in a dynamic environment is achieved. And the low-orbit satellite selects to cache the content to the local, so that the transmission delay and transmission energy consumption of the user for acquiring the content are reduced, the increase of the transmission delay and the transmission energy consumption caused by the topology change of the satellite network is avoided, and the cache resources of the satellite are effectively utilized.

Description

Technical Field

[0001] This invention belongs to the field of computer application technology, specifically relating to a content caching method that considers dynamic network topology in satellite edge computing. Background Technology

[0002] With the rapid development of mobile internet and multimedia applications, users' demand for low-latency, highly reliable data services is constantly growing. Edge computing technology has emerged to address this need. This technology deploys computing and storage resources at base stations or edge servers close to the user, offloading some content and computing tasks from the central cloud to the network edge, thereby effectively reducing content transmission latency and improving service response speed. This method of deploying caching servers at terrestrial base stations to provide low-latency content distribution to users within the coverage area is called terrestrial edge caching technology.

[0003] However, terrestrial edge caching relies on robust terrestrial infrastructure. In remote areas, marine regions, or disaster-prone areas where terrestrial communication facilities are scarce or lacking, this technology is difficult to apply, resulting in the inability to achieve continuous global coverage. To address this issue, a new type of converged computing and communication system—satellite computing—has emerged in recent years. As a crucial component of an integrated space-air-terrestrial network, this system combines satellite communication with edge computing technology, enabling service providers to cache content on LEO satellites, thereby providing efficient content distribution services to users in remote areas or regions with weak infrastructure. This satellite-based content caching method is known as satellite caching technology.

[0004] Unlike ground-based edge caches, which rely on fixed infrastructure, satellite caches depend on a constellation of high-speed moving low-Earth orbit (LEO) satellites, resulting in a highly dynamic network topology. LEO satellites constantly change their relative positions to other satellites and ground terminals during their orbital operation, leading to frequent changes in inter-satellite link connections. This dynamic topology presents satellite caches with several challenges:

[0005] First, frequent changes in satellite network topology lead to constant shifts in direct connections between satellites. Satellites that could communicate directly before may lose connection in the next time slot, requiring content to be relayed through more satellites before reaching users, thus increasing transmission latency and energy consumption. Second, caching decisions lag behind network state changes. When satellites make caching decisions based on outdated topology information, cached content may be difficult to access quickly under the new network topology, reducing cache hit rate and increasing energy consumption. Third, satellite resources are limited, and user needs are diverse. LEO satellites have limited computing and storage capabilities, making it impossible to cache all potentially popular content; furthermore, satellites move across different regions, and cached content moves with them, potentially becoming irrelevant to users in the current coverage area. Additionally, user interests change dynamically over time, making static caching strategies ill-suited to these changes, leading to cache invalidation and increased content access latency. Summary of the Invention

[0006] This invention addresses the shortcomings of the aforementioned background technology by providing a content caching method for satellite edge computing that considers dynamic network topology. When modeling the content caching problem in satellite edge computing, the dynamic nature of the satellite network topology is taken into account, achieving dual optimization of transmission latency and transmission energy consumption under constraints of satellite cache capacity and the number of duplicate caches. This method utilizes multi-agent graph attention reinforcement learning to solve the content caching problem, achieving the goal of solving caching strategies in a dynamic environment.

[0007] A content caching method considering dynamic network topology in satellite edge computing includes the following steps:

[0008] Step S1: Model the satellite network topology using a time-varying network, obtain the link stability based on the volatility in the time-varying network, obtain the link centrality based on the degree centrality in the time-varying network, and construct the weights for the impact of topology changes based on link stability and centrality.

[0009] Step S2: Build a system model for content caching in satellite edge computing, and then formalize the content caching problem considering dynamic network topology with the goal of minimizing the average transmission cost and the constraints of satellite cache capacity and duplicate cache number.

[0010] Step S3: Establish a content caching action selection model based on multi-agent graph attention reinforcement learning, and train its network parameters;

[0011] Step S4: Input the state of the local satellite into the trained multi-agent graph attention reinforcement learning-based content caching action selection model, output the action of the local satellite, and cache the corresponding content for each satellite according to the action.

[0012] The beneficial effects of this invention are:

[0013] 1) When modeling the content caching problem in satellite edge computing, the dynamic nature of the satellite network topology is considered, and dual optimization of transmission latency and transmission energy consumption is achieved under the constraints of satellite cache capacity and the number of duplicate caches.

[0014] 2) This method uses multi-agent graph attention reinforcement learning to solve the content caching problem, thereby achieving the goal of solving caching strategies in a dynamic environment.

[0015] 3) Low-Earth orbit satellites choose to cache content locally, reducing the transmission latency and energy consumption for users to access content, avoiding the increase in transmission latency and energy consumption due to changes in satellite network topology, and effectively utilizing the satellite's caching resources. Attached Figure Description

[0016] Figure 1 This is a diagram of the satellite edge computing architecture in a specific embodiment of the present invention.

[0017] Figure 2 This is a specific flowchart of a particular embodiment of the present invention.

[0018] Figure 3 This is a schematic diagram of a content caching action selection model based on multi-agent graph attention reinforcement learning in a specific embodiment of the present invention.

[0019] Figure 4 This is a schematic diagram of the training algorithm flow for the content caching action selection model based on multi-agent graph attention reinforcement learning in a specific embodiment of the present invention.

[0020] Figure 5 The figure shows the average time delay experimental results in a specific embodiment of the present invention.

[0021] Figure 6 The figure shows the average energy consumption experimental results in a specific embodiment of the present invention. Detailed Implementation

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0023] like Figure 1 As shown in the diagram, the satellite edge computing architecture of this invention is described in detail below: When a user requests content in the LEO satellite coverage area, the requested content can be obtained from the access LEO satellite. If the access LEO satellite does not cache the content, the content can be obtained from the neighboring LEO satellites of the access LEO satellite. If neither the access LEO satellite nor the neighboring LEO satellites cache the content, the content can be obtained from the GEO satellite.

[0024] like Figure 2 As shown, the specific process of this embodiment of the invention includes:

[0025] Step S1: Construct the weights for the impact of topology changes.

[0026] The time-varying topology of Low Earth Orbit (LEO) satellite networks is modeled based on time-varying network theory. Since the orbits of LEO satellites can be accurately predicted using ephemeris data, the satellite positions and link states in each time slice can be precisely determined. The inter-satellite topology can be represented as a time-varying graph. Where V = {1, ..., S} represents the set of LEO satellites. This represents the set of links that change over time. Within time slot t, the local time-varying map of LEO satellite i consists of subgraphs. Indicated. Wherein, V i (t) includes all neighboring LEO satellites within the N-hop range, when e i,j When (t) = 1, there is an edge between LEO satellite i and LEO satellite j.

[0027] Based on time-varying networks, two key metrics are introduced: link stability and link centrality, to characterize link properties. The former reflects the link's ability to maintain continuous connectivity, while the latter describes the link's importance in data forwarding and network coverage. Since these two metrics directly affect content transmission performance, they are considered together in caching strategies to improve system reliability and efficiency.

[0028] The volatility of LEO satellite i is used to measure its ability to maintain connectivity over a period of time. Given a local time-varying plot G of LEO satellite i... i (t), the volatility of LEO satellite i is calculated based on the volatility in the time-varying network as follows:

[0029]

[0030] Among them, U i,j The following definition indicates whether two LEO satellites have established a connection during their orbital period:

[0031]

[0032] It is important to note the indicator SN i This value is used to measure the link maintenance capability of LEO satellite i. The smaller the value, the lower the satellite's volatility, and the higher the link's sustainability.

[0033] Let SE i,j The stability of link (i,j) is represented by the average volatility of LEO satellite i and LEO satellite j, and is calculated as follows:

[0034]

[0035] Let CN i,j The degree centrality of LEO satellite i is calculated according to the degree centrality in time-varying networks as follows:

[0036]

[0037] This metric measures the importance of LEO satellites in the topology; the higher the value, the greater their influence and the more critical relay role in the network.

[0038] Let CE i,j The centrality of link (i,j) is represented by the reciprocal of the average degree centrality of LEO satellite i and LEO satellite j, and is calculated as follows:

[0039]

[0040] First, the stability and centrality indices are normalized to the interval [0,1] using the min-max normalization method, and the normalized values ​​are denoted as SE'. i,i and CE' i,j Subsequently, the weight of the topology change affecting link (i,j) is represented by a weighted sum of these two normalized metrics:

[0041] W i,j =μSE′ i,j +(1-μ)CE′ i,j

[0042] Where μ is the weighting coefficient, used to balance the importance of stability and centrality.

[0043] Step S2: Build a system model for content caching in satellite edge computing, and formalize the content caching problem that considers dynamic network topology.

[0044] Step S2-1: Build a system model for content caching in satellite edge computing.

[0045] The system model for content caching in satellite edge computing includes one geostationary Earth Orbit (GEO) satellite as a cloud server and S LEO satellites as edge servers. The set of all LEO satellites is denoted as […]. Each LEO satellite has a storage capacity ξ s The LEO satellite closest to the user is called the access satellite, and LEO satellites within N hops of the access satellite are called neighbor satellites. Assume there are F distinct content items in the system, denoted as... Each piece of content can be cached at most B times across all satellites. Given a set of users... Each user u∈U will generate a content request, denoted as R. u (t)={c u,1 (t),...,c u,f (t),...c u,F (t)}. Where, when user u requests content f, c u,f (t) = 1, otherwise c u,f (t) = 0. Time is divided into several equally spaced time slices, denoted as . Where T represents the satellite's orbital period. Let... This is for content caching decisions. Where x... s,f (t) indicates whether content f is cached within time slot t and in satellite s (x s,f (t) = 1 indicates a cache, x s,f (t) = 0 indicates that it is not cached).

[0046] Step S2-2: Formalize the content caching problem in satellites that considers dynamic network topology:

[0047] exist Figure 1 In the communication model shown, user u first connects to its nearest LEO satellite m via a user-to-LEO (UL) link (u,m), which is called the access satellite. The access satellite can forward requests to neighboring satellites in a multi-hop manner via LEO-to-LEO (LL) links. The forwarding path contains at most N hops along the shortest path. Alternatively, the access satellite can directly connect to a GEO satellite via an LEO-to-GEO (LG) link. Therefore, any content retrieval path always includes a UL link and may additionally include LL or LG links. Based on this, this embodiment defines Path. u,s Let be the set of links contained in the shortest path from user u to satellite s of the cached content f, represented as:

[0048] Path u,s ={(u,m),(m,1),...,(ω,s)}

[0049] Where ω represents the path u,s The satellite preceding satellite s.

[0050] For path within time slot t u,s For each link (i,j) on the network, let r i,j (t) represents its transmission rate. According to Shannon's theorem, r i,j (t) is calculated as follows:

[0051]

[0052] Among them, b i,j (t) represents the link bandwidth, pi,j (t) represents the transmission power. σ represents the channel gain. i,j Indicates noise power.

[0053] set up Let f represent the transmission delay caused by transmitting content f through link (i,j) within time slot t. Then we have:

[0054]

[0055] Where |f| represents the size of the content f.

[0056] Similarly, let This represents the transmission energy consumed in transmitting content f through link (i,j) within time slot t, expressed as:

[0057]

[0058] Where, p i,j (t) is the transmission power of link (i,j).

[0059] Based on the stability and centrality of the link, the transmission delay of content f requested by user u is defined as:

[0060]

[0061] The corresponding transmission energy consumption is:

[0062]

[0063] The total transmission delay of all requests served by LEO satellites within time slice t. Total transmission energy consumption Defined as:

[0064]

[0065]

[0066] Where F is the content set, c u,f (t) represents the variable indicating user u's request for content f, U s (t) represents the set of all users within the coverage area of ​​LEO satellites.

[0067] Within time slot t, the transmission cost of LEO satellites s is defined as:

[0068]

[0069] in, and Let represent the total transmission delay and total transmission energy consumption after max-min normalization, respectively, and λ be the weighting coefficient of the two.

[0070] This embodiment models the content caching problem considering dynamic network topology as an optimization problem, with the goal of minimizing the average transmission cost under constraints of satellite cache capacity and duplicate cache count. The problem is expressed as follows:

[0071]

[0072] Among them, constraints C1 and C2 are used to ensure that the cache placement decision variables are in binary form; constraint C2 ensures the storage capacity ξ of each LEO satellite. s Not exceeded; constraint C2 limits the number of duplicate caches of the same content. Where B is the maximum number of duplicate caches allowed by the content provider.

[0073] Step S3: Establish a content caching action selection model based on multi-agent graph attention reinforcement learning, and train its network parameters.

[0074] Step S3-1: For the cache X(t) of all satellites, solve the problem using a content caching action selection model based on multi-agent graph attention reinforcement learning; this invention models the content caching problem considering dynamic network topology as a Markov decision process. Each LEO satellite is treated as an independent agent. This invention uses triples. To describe this Markov decision process. Among them, This is represented as the global state. Specifically, the state of satellite i in time slot t is represented as... in, This represents the set of all currently cached content for satellite i. Let A(t) represent the set of user requests received by satellite i; A(t) = {A1(t), ..., A...} i (t),…,A S (t)} represents the joint action of all users. The action of satellite i in time slot t is represented as A. i (t)={a i,1 (t),…,a i,f (t),…,a i,F (t)}, where a i,f (t) is a binary variable used to indicate whether content f is cached in satellite i. These actions must satisfy the following constraints: (Satellite cache capacity constraint) (Constraint on the number of duplicate cache entries); Indicates when the satellite is in a state And the joint reward obtained after performing action A(t). Each time a low-Earth orbit satellite performs a corresponding action, it receives a corresponding reward based on its behavior. The learning objective for each low-Earth orbit satellite is to learn the optimal content caching strategy to reduce the average transmission cost. Since the goal of reinforcement learning is to maximize the reward value, the immediate reward is defined as... The cost minimization problem is transformed into the reward maximization problem in reinforcement learning.

[0075] Step S3-2: Construct a content caching action selection model based on multi-agent graph attention reinforcement learning. Each agent in the multi-agent graph attention reinforcement learning model contains a multi-head graph attention network (Multi-head GAT) layer, a gated recurrent unit (GRU) layer, and an actor-critic layer.

[0076] The multi-head GAT layer captures spatial features between neighboring satellites, enabling the agent to model complex and time-varying topologies. The GRU layer captures temporal patterns in user requests, allowing the agent to learn the temporal evolution of user preferences. Finally, the Actor-Critic layer generates caching decision actions based on the extracted spatiotemporal features. Through reinforcement learning, this module utilizes spatiotemporal feature representations and adaptive policy updates to optimize long-term caching performance.

[0077] The multi-head GAT layer consists of K attention heads, each independently computing spatial features, which are then concatenated. For the k-th attention head, the attention coefficients between LEO satellites i and j within time slot t are... The calculation formula is:

[0078]

[0079] in, and This represents the node features after a linear transformation. This represents the link characteristics after a linear transformation. VM k and EM k The weight matrix is ​​a learnable matrix. Let be a trainable attention vector, and LeakyReLU(·) be the activation function.

[0080] Subsequently, satellite i in time slot t has spatial characteristics h i (t) is obtained by aggregating the features of its neighbors:

[0081]

[0082] Here, || represents the concatenation operation. This represents the normalized attention coefficient.

[0083] After extracting the spatial features, h i (t) Input to the GRU layer to capture temporal correlations. Let h' i (t) represents the spatiotemporal characteristics of LEO satellite i within time slice t. The GRU layer first calculates the reset gate y(t) and the update gate q(t), as follows:

[0084] y(t)=sigmoid(W y [h′ i (t-1),h i (t)]+b y )

[0085] q(t)=sigmoid(W q [h′ i (t-1),h i (t)]+b q )

[0086] Among them, W y W q U y and U q Let b be the weight matrix. y and b q is the bias term, and sigmoid(·) is the sigmoid activation function.

[0087] Subsequently, the candidate hidden states are calculated using the tanh(·) activation function, i.e.:

[0088]

[0089] Where ⊙ represents element-wise multiplication, W h and b h These are the weight matrix and the bias term, respectively.

[0090] Finally, the GRU layer updates the spatiotemporal features as follows:

[0091]

[0092] By combining the Multi-head GAT layer and the GRU layer, it is possible to effectively capture both the structural dependencies between LEO satellites and the temporal evolution characteristics of user needs.

[0093] In the Actor-Critic layer, a centralized training and distributed execution framework is used, with each LEO satellite i containing an Actor network. Where, θ i (t) represents the network parameters of the Actor network on the i-th satellite. This network uses the local spatiotemporal characteristics h' i (t) is the input, and the output is the local action. Simultaneously maintain a target Actor network The core objective is to improve the stability of Actor network training. During initialization, the network parameters of the target Actor network are... The network parameters are subsequently updated continuously via soft updates to improve the stability of the Actor network training. Centralized Critic network Q φ(t) Based on global state Estimate the global cumulative reward with joint action A(t) Where φ(t) represents the network parameters of the Critic network. A synchronized target Critic network is also maintained. To improve the stability of Critic network training. During initialization, the network parameters of the target Critic network are... Subsequently, the network parameters are continuously updated via soft updates to improve the stability of Critic network training.

[0094] Step S3-3: Train the parameters in the content caching action selection model based on multi-agent graph attention reinforcement learning. The specific steps are as follows:

[0095] Step S3-3-1: Initialize the Actor network for all LEO satellites and centralized Critic network Q φ(t) and target network and

[0096] Step S3-3-2: Each satellite inputs its local status. Spatial features h are obtained in the Multi-head GAT layer i (t). Then, input its local spatial features h. i (t) is used to obtain the spatiotemporal features h' in the GRU layer. i (t). Finally, input its local spatiotemporal features h' i (t) Obtain the local action A from the Actor network. i (t). The local actions of all satellites constitute the joint action A(t).

[0097] Step S3-3-3: Execute the joint action A(t) to obtain the joint reward. The global state transitions to the state of the next time slot. Storage Quadruple To the replay buffer P.

[0098] Step S3-3-4: All Actor network parameters θ i Both (t) and the Critic network parameter φ(t) are obtained by sampling mini-batch empirical tuples randomly from the replay buffer P. Update the program to reduce the relevance of experience and stabilize the learning process.

[0099] Step S3-3-5: Randomly sample a mini-batch of experience from the replay buffer P to train the Critic network parameters φ(t). The target value is calculated using the following formula:

[0100]

[0101] Where γ is the discount factor. A'(t+1) represents the next target action from all target Actor networks. The Critic network parameters φ(t) are optimized by minimizing the mean squared error (MSE) loss:

[0102]

[0103] Step S3-3-6: Randomly sample a mini-batch of experience from the replay buffer P to train the Actor network parameters θ. i (t). Updated based on policy gradient:

[0104]

[0105] in, The dominance function is used to represent action A. i (t) Relative gain compared to the expected action outcome. Benchmark value. Defined as:

[0106]

[0107] Step S3-3-7: After updating the Actor network and Critic network, the target network will undergo a soft update to improve training stability:

[0108]

[0109] Where τ∈(0,1) is a smoothing factor used to control the update rate of the target network.

[0110] Step S3-3-6: Repeat S3-3-2 to S3-3-7 until the reward value converges, completing the training.

[0111] Step S4: Input the observed system state The action selection model outputs the action A(t) of the satellite using the trained multi-agent graph attention reinforcement learning content cache model, and each LEO satellite caches the corresponding content based on this action.

[0112] To demonstrate the effectiveness of this embodiment, preliminary experiments were conducted. The method proposed in this embodiment is compared with Random, LRU, CDCD, and Cooperative methods. In Random, the method randomly generates a caching strategy for each satellite; that is, each LEO satellite does not consider user request characteristics, content popularity, or network status when caching content, relying entirely on random selection for cache allocation. In LRU, the method caches user-requested content on the LEO satellites it connects to. When the satellite cache space is full, a Least Recently Used (LRU) mechanism is used to evict the least recently requested content to make room for new content. In CDCD, the method uses a time-slicing model and snapshot merging mechanism to represent the dynamic topology of the satellite network and makes caching decisions based on the community partitioning concept. First, the entire satellite network is divided into multiple satellite communities; then, nodes with higher importance in the community are selected as caching nodes; finally, popular content is cached on these nodes. This method improves caching efficiency to some extent, but it fails to fully consider the time-varying nature of user preferences, thus limiting its effectiveness in real-world scenarios where content popularity changes dynamically. In Cooperative, this method divides the caching space of each satellite into a fixed cache area and an optional cache area. The content in the fixed cache area is optimized through theoretical analysis to fully utilize content popularity; the optional cache area uses an iterative algorithm to calculate the optimal content distribution scheme across the entire network. While this method improves content utilization, it ignores the dynamic characteristics of the satellite network topology, limiting its applicability in real-world, complex scenarios.

[0113] The comparison results are as follows Figure 5 and Figure 6 As shown in the figure, the experimental results demonstrate that the method proposed in this embodiment, which considers dynamic network topology, exhibits the best performance in terms of both latency and energy consumption. The method proposed in this embodiment is particularly suitable for use in content caching for satellite edge computing.

[0114] The above description is only a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. Any equivalent modifications or changes made by those skilled in the art based on the content disclosed in the present invention should be included within the scope of protection set forth in the claims.

Claims

1. A content caching method considering dynamic network topology in satellite edge computing, characterized by: The method includes the following steps: Step S1: Model the satellite network topology using a time-varying network, obtain the link stability based on the volatility in the time-varying network, obtain the link centrality based on the degree centrality in the time-varying network, and construct the weights for the impact of topology changes based on link stability and centrality. Step S2: Build a system model for content caching in satellite edge computing, and then formalize the content caching problem considering dynamic network topology with the goal of minimizing the average transmission cost and the constraints of satellite cache capacity and duplicate cache number. Step S3: Establish a content caching action selection model based on multi-agent graph attention reinforcement learning, and train its network parameters; Step S4: Input the state of the local satellite into the trained multi-agent graph attention reinforcement learning-based content caching action selection model, output the action of the local satellite, and cache the corresponding content for each satellite according to the action.

2. The content caching method for satellite edge computing considering dynamic network topology according to claim 1, characterized in that: In step S1, constructing the weights for the impact of topology changes includes the following specific steps: The time-varying topology of the low-Earth orbit (LEO) satellite network is modeled based on time-varying network theory; the inter-satellite topology is represented as a time-varying graph. Where V = {1, ..., S} represents the set of LEO satellites. Represents a set of links that change over time; in time slot t, the local time-varying map of LEO satellite i consists of subgraphs. Represents; where i(t) contains all neighboring LEO satellites within the N-hop range, when e i,j When (t) = 1, there is an edge between LEO satellite i and LEO satellite j. Time is divided into several equally spaced time slices, denoted as Where T represents the satellite's orbital period; The volatility of LEO satellite i is used to measure its ability to maintain connectivity over a period of time; given a local time-varying graph G of LEO satellite i i (t), the volatility of LEO satellite i is calculated based on the volatility in the time-varying network as follows: Among them, U i,j The following definition indicates whether two LEO satellites have established a connection during their orbital period: Indicator SN i This value is used to measure the link maintenance capability of LEO satellite i. The smaller the value, the lower the satellite's volatility and the higher the link's continuity. Let SE i,j The stability of link (i,j) is represented by the average volatility of LEO satellite i and LEO satellite j, and is calculated as follows: Let CN i,j The degree centrality of LEO satellite i is calculated according to the degree centrality in time-varying networks as follows: Indicator CN i,j Used to measure the importance of LEO satellite i in the topology; the larger the value, the stronger its influence and the more critical relay role in the network. Let CE i,j The centrality of link (i,j) is represented by the reciprocal of the average degree centrality of LEO satellite i and LEO satellite j, and is calculated as follows: First, the stability and centrality indices are normalized to the interval [0,1] using the min-max normalization method, and the normalized values ​​are denoted as SE'. i,j and CE' i,j Subsequently, the weight of the topology change affecting link (i,j) is represented by the weighted sum of these two normalized indices: W i,j =μSE′ i,j +(1-μ)CE′ i,j Where μ is the weighting coefficient, used to balance the importance of stability and centrality.

3. The content caching method for satellite edge computing considering dynamic network topology according to claim 2, characterized in that: In step S2, the specific steps are as follows: Step S2-1: Build a system model for content caching in satellite edge computing: The system model for content caching in satellite edge computing includes one geostationary GEO satellite as a cloud server and S LEO satellites as edge servers; the set of all LEO satellites is denoted as . Each LEO satellite has a storage capacity ξ s ; The LEO satellite closest to the user is called the access satellite, and the LEO satellites that are no more than N hops away from the access satellite are called neighboring satellites. Suppose that there are F distinct content items in the system, denoted as F0. Each piece of content can be cached at most B times across all satellites; given a set of users Each user u∈U will generate a content request, denoted as R. u (t)={c u,1 (t),...,c u,f (t),...c u,F (t)}; where, when user u requests content f, c u,f (t) = 1, otherwise c u,f (t) = 0; time is divided into several equally spaced time slices, denoted as _(t) = 0; Where T represents the satellite's orbital period; let... For content caching decisions; where x s,f (t) indicates whether content f is cached in satellite s within time slot t, x s,f (t) = 1 indicates a cache, x s,f (t) = 0 indicates that it is not cached; Step S2-2: Formalize the content caching problem in satellites that considers dynamic network topology: In the communication model, user u first connects to its nearest LEO satellite m via the user-to-LEO UL link (u,m), which is called the access satellite. The access satellite forwards the request to neighboring satellites in a multi-hop manner via the LEO-to-LEO LL link. The forwarding path contains at most N hops along the shortest path. Alternatively, the access satellite can directly connect to the GEO satellite via the LEO-to-GEO LG link. Therefore, any content retrieval path always includes a UL link and may additionally include LL or LG links. Based on this, Path is defined. u,s Let be the set of links contained in the shortest path from user u to satellite s of the cached content f, represented as: Path u,s ={(u,m),(m,1),...,(ω,s)} Where ω represents the path u,s The satellite preceding satellite s; For path within time slot t u,s For each link (i,j) on the network, let r i,j (t) represents its transmission rate; according to Shannon's theorem, r i,j (t) is calculated as follows: Among them, b i,j (t) represents the link bandwidth, p i,j (t) represents the transmission power. σ represents the channel gain. i,j Indicates noise power; set up Let f represent the transmission delay caused by transmitting content f through link (i,j) within time slot t. Then we have: Where |f| represents the size of the content f; set up This represents the transmission energy consumed in transmitting content f through link (i,j) within time slot t, expressed as: Based on the stability and centrality of the link, the transmission delay of content f requested by user u is defined as: The corresponding transmission energy consumption is: The total transmission delay of all requests served by LEO satellites within time slice t. Total transmission energy consumption Defined as: Where F is the content set, c u,f (t) represents the variable indicating user u's request for content f, U s (t) represents the set of all users within the coverage area of ​​LEO satellites s; Within time slot t, the transmission cost of LEO satellites s is defined as: in, and Let represent the total transmission delay and total transmission energy consumption after max-min normalization, respectively, and λ be the weighting coefficient of the two. The content caching problem considering dynamic network topology is modeled as an optimization problem. The objective is to minimize the average transmission cost for all satellite caches X(t) under constraints of satellite cache capacity and the number of duplicate caches. The problem is expressed as follows: Among them, constraints C1 and C2 are used to ensure that the cache placement decision variables are in binary form; constraint C2 ensures the storage capacity ξ of each LEO satellite. s Not exceeded; Constraint C2 limits the number of duplicate caches of the same content, where B is the maximum number of duplicate caches allowed by the content provider.

4. The content caching method for satellite edge computing considering dynamic network topology according to claim 3, characterized in that: In step S3, the specific steps are as follows: Step S3-1: For the cache X(t) of all satellites, solve the content caching action selection model based on multi-agent graph attention reinforcement learning; model the content caching problem of dynamic network topology as a Markov decision process, where each LEO satellite is an independent agent; use triples To describe this Markov decision process; where, Represented as the global state, the state of satellite i in time slot t is represented as... This represents the set of all currently cached content for satellite i. Let A(t) represent the set of user requests received by satellite i; A(t) = {A1(t), ..., A...} i (t),…,A S (t)} represents the joint action of all users, and the action of satellite i in time slot t is represented as A. i (t)={a i,1 (t),…,a i,f (t),…,a i,F (t)}, where a i,f (t) is a binary variable used to indicate whether content f is cached in satellite i; these actions must satisfy the satellite cache capacity constraint. and duplicate cache quantity constraints Indicates when the satellite is in a state Furthermore, the joint reward obtained after performing action A(t) is provided, and each time a low-Earth orbit satellite completes a corresponding action, it receives a reward based on its behavior; the learning objective for each low-Earth orbit satellite is to learn the optimal content caching strategy to reduce the average transmission cost; the immediate reward is defined as... The cost minimization problem is transformed into the reward maximization problem in reinforcement learning; Step S3-2: Construct a content caching action selection model based on multi-agent graph attention reinforcement learning. Each agent in the multi-agent graph attention reinforcement learning model contains a multi-head graph attention network layer (Multi-head GAT), a gated recurrent unit layer (GRU), and an actor-critic layer. The Multi-head GAT layer is used to capture spatial features between neighboring satellites, enabling the agent to model complex and time-varying topologies; the GRU layer is used to capture temporal patterns of user requests, enabling the agent to learn the temporal evolution of user preferences; finally, the Actor-Critic layer generates cache decision actions based on the extracted spatiotemporal features; through reinforcement learning, this module utilizes spatiotemporal feature representations and adaptive policy updates to optimize long-term cache performance. The multi-head GAT layer consists of K attention heads, each independently calculating spatial features, which are then concatenated. For the k-th attention head, the attention coefficients between LEO satellites i and j within time slot t are calculated. The calculation formula is: in, and This represents the node features after a linear transformation. Represents the link characteristics after linear transformation; VM k and EM k The weight matrix is ​​a learnable matrix. Let be a trainable attention vector, and LeakyReLU(·) be the activation function; Subsequently, satellite i in time slot t has spatial characteristics h i (t) is obtained by aggregating the features of its neighbors: Here, || represents the concatenation operation. This represents the normalized attention coefficient; After extracting the spatial features, h i (t) Input the GRU layer to capture temporal correlations; let h' i (t) represents the spatiotemporal characteristics of LEO satellite i within time slice t; the GRU layer first calculates the reset gate y(t) and the update gate q(t), as follows: y(t)=sigmoid(W y [h′ i (t-1),h i (t)]+b y ) q(t)=sigmoid(W q [h′ i (t-1),h i (t)]+b q ) Among them, W y W q U y and U q Let b be the weight matrix. y and b q The bias term is sigmoid(·), and sigmoid(·) is the sigmoid activation function. Subsequently, the candidate hidden states are calculated using the tanh(·) activation function, i.e.: Where ⊙ represents element-wise multiplication, W h and b h These are the weight matrix and the bias term, respectively. Finally, the GRU layer updates the spatiotemporal features as follows: By combining the Multi-head GAT layer and the GRU layer, it is possible to effectively capture the structural dependencies between LEO satellites and the temporal evolution characteristics of user needs simultaneously. In the Actor-Critic layer, a centralized training and distributed execution framework is used, with each LEO satellite i containing an Actor network. θ i (t) represents the network parameters of the Actor network on the i-th satellite at time slot t; the network uses the local spatiotemporal characteristics h' i (t) is the input, and the output is the local action. Simultaneously maintain a target Actor network The core objective is to improve the stability of Actor network training; during initialization, the network parameters of the target Actor network are... Subsequently, its network parameters are continuously updated via soft updates to improve the stability of Actor network training; centralized Critic network Q φ(t) Based on global state Estimate the global cumulative reward with joint action A(t) Where φ(t) represents the network parameters of the Critic network at time slot t; simultaneously, a synchronized target Critic network is maintained. To improve the stability of Critic network training; during initialization, the network parameters of the target Critic network are... Subsequently, its network parameters are continuously updated through soft updates to improve the stability of Critic network training; Step S3-3: Train the parameters in the content caching action selection model based on multi-agent graph attention reinforcement learning. The specific steps are as follows: Step S3-3-1: Initialize the Actor network for all LEO satellites and centralized Critic network Q φ(t) and target network and Step S3-3-2: Each satellite inputs its local status. Spatial features h are obtained in the Multi-head GAT layer i (t); then, input its spatial features h. i (t) is used to obtain the spatiotemporal features h' in the GRU layer. i (t); Finally, input its local spatiotemporal features h' i (t) Obtain the local action A from the Actor network. i (t); The local actions of all satellites constitute the joint action A(t); Step S3-3-3: Execute the joint action A(t) to obtain the joint reward. The global state transitions to the state of the next time slot. Storage Quadruple To the replay buffer P; Step S3-3-4: All Actor network parameters θ i Both (t) and the Critic network parameter φ(t) are obtained by sampling mini-batch empirical tuples randomly from the replay buffer P. Update the learning process to reduce its relevance to experience and stabilize it. Step S3-3-5: Randomly sample a small batch of experience from the replay buffer P to train the Critic network parameters φ(t); the target value is calculated using the following formula: Where γ is the discount factor; A'(t+1) represents the next target action from all target Actor networks; the Critic network parameter φ(t) is optimized by minimizing the mean squared error (MSE) loss. Step S3-3-6: Randomly sample a mini-batch of experience from the replay buffer P to train the Actor network parameters θ. i (t); Update according to the policy gradient: in, The dominance function is used to represent action A. i (t) Relative gain compared to the expected action outcome; benchmark value Defined as: Step S3-3-7: After updating the Actor network and Critic network, the target network undergoes a soft update to improve training stability: Where τ∈(0,1) is a smoothing factor used to control the update rate of the target network; Step S3-3-6: Repeat S3-3-2 to S3-3-7 until the reward value converges, completing the training.

5. The content caching method for satellite edge computing considering dynamic network topology according to claim 4, characterized in that: In step S4, the observed system state is input. The action selection model outputs the action A(t) of the satellite using the trained multi-agent graph attention reinforcement learning content cache model, and each LEO satellite caches the corresponding content based on this action.

Images