A topology prediction model training method and prediction method applied to a satellite network

Abstract

The application provides a topology prediction model training method applied to a satellite network, the satellite network comprising a plurality of satellites, a plurality of terminals and a plurality of ground stations, the topology prediction model comprising a space-time learning layer and a full connection layer, and the method comprises the following steps: S1, acquiring satellite network historical topology graph data in a target area, the satellite network historical topology graph data comprising a plurality of continuous satellite network topology graphs at different time points, and each satellite network topology graph at a time point comprising a plurality of nodes and a plurality of edges, wherein the nodes represent terminals, satellites or ground stations, and edges between any two nodes represent that a link is arranged between the two nodes; and S2, training the topology prediction model based on the satellite network historical topology graph data acquired in step S1 until convergence. The scheme provided by the application replaces huge storage consumption with weight memory parameters, greatly reduces storage occupation and guarantees the integrity of network topology information.

Description

Technical Field

[0001] This invention relates to the field of satellite communication technology, specifically to satellite network topology research in the field of satellite communication technology, and more specifically, to a training method and prediction method for a topology prediction model applied to satellite networks. Background Technology

[0002] Satellite communication networks are considered an important component of communication technology, boasting advantages such as large scale, wide coverage, and low latency, with the goal of providing ubiquitous internet access to users worldwide. However, due to the high-speed movement of satellite nodes within these networks, the satellite-to-ground links and inter-satellite links between satellites are frequently interrupted. Furthermore, the highly complex and dynamic nature of the satellite network topology makes it difficult to construct end-to-end service transmission paths, thus impacting the management of the satellite communication network.

[0003] To better manage satellite communication networks, researchers have proposed predicting satellite network topology based on its time-varying characteristics. This prediction, which forecasts the complete link between two terminals communicating via the satellite network, enables more efficient end-to-end link prediction. With this effective prediction, better end-to-end service construction can be achieved, leading to improved management and application of satellite communication networks.

[0004] Researchers have proposed several schemes for satellite network topology prediction, among which the following are representative: 1. Using snapshots to store periodically changing satellite network topology data to predict the satellite network topology at a certain moment. However, when the network scale is large, the number of subgraphs generated by the snapshot method is proportional to the time length, and the storage space is proportional to the network size, resulting in a large storage space consumption problem. 2. Using traversal methods to predict the satellite network topology at a certain time. However, this type of scheme has high topology calculation complexity, and it is necessary to traverse the satellite network topology of all time periods when obtaining the network topology at a certain time, resulting in a large time overhead. 3. The model can be trained based on the obtained inter-satellite link and satellite-to-ground link data at different times to build a link prediction model for link prediction. However, this type of scheme does not consider the changes in user links, which leads to incomplete satellite network topology information and affects the accuracy of link prediction.

[0005] Although existing technologies can predict satellite network topology, they still suffer from problems such as large storage space requirements, high computational complexity of topology, and incomplete satellite network topology information due to the failure to consider changes in user links, which are not conducive to satellite network topology prediction. Summary of the Invention

[0006] Therefore, the purpose of this invention is to overcome the shortcomings of the prior art and provide a topology prediction model training method for satellite networks and a topology prediction method for satellite networks.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] According to a first aspect of the present invention, a method for training a topology prediction model for a satellite network is provided. The satellite network includes multiple satellites, multiple terminals, and multiple ground stations. The topology prediction model includes a spatiotemporal learning layer and a fully connected layer. The method includes the following steps: S1, acquiring historical topology map data of the satellite network within a target area. The historical topology map data includes satellite network topology maps at multiple consecutive time points. Each time point's satellite network topology map includes multiple nodes and multiple edges, wherein a node represents a terminal, satellite, or ground station, and an edge between any two nodes indicates that a link is established between the two nodes; S2, training the topology prediction model based on the historical topology map data of the satellite network acquired in step S1 until convergence.

[0009] In some embodiments of the present invention, the spatiotemporal learning layer includes one or more spatiotemporal learning modules, which are used to acquire the spatial and spatiotemporal features of satellite network topology data.

[0010] In some embodiments of the present invention, the spatiotemporal learning module includes a graph convolution module and a temporal feature extraction module connected in series. The graph convolution module is used to extract spatial features of the satellite network topology map data, and the temporal feature extraction module is used to obtain the spatiotemporal features of the satellite network topology map data based on the output of the graph convolution module and the satellite network topology map data.

[0011] Preferably, the time feature extraction module employs a long short-term memory network, a recurrent neural network, or a gated recurrent unit.

[0012] In some embodiments of the present invention, step S1 includes: S11, dividing the globe into different regions; S12, taking any one of the regions obtained in step S11 as the target region, and obtaining historical satellite network topology data within the target region.

[0013] In some embodiments of the present invention, step S2 includes: S21, obtaining the corresponding adjacency matrix set based on the satellite network historical map data obtained in step S1; S22, processing each adjacency matrix in the adjacency matrix set obtained in step S21 according to preset data processing rules, and constructing a dataset based on the processing results of each adjacency matrix; S23, training the topology prediction model based on the dataset obtained in step S22 until convergence.

[0014] In some embodiments of the present invention, in step S22, the preset data processing rule is as follows: after adding the original adjacency matrix to the identity matrix, a symmetric normalization process is performed to obtain a new adjacency matrix; based on the new adjacency matrix, a degree matrix corresponding to it is constructed; and the new adjacency matrix and its corresponding degree matrix are used as the processing result of the original adjacency matrix.

[0015] In some embodiments of the present invention, step S23 includes: S231, performing multiple data cuts from the dataset obtained in step S22 to obtain multiple samples to form a training set, wherein each data cut is performed by selecting data corresponding to the satellite network topology map within a time window of a preset width that is continuous in time from the dataset; S232, training the topology prediction model using self-supervised learning based on the training set obtained in step S231 until the model converges.

[0016] Preferably, the preset width is 6.

[0017] According to a second aspect of the present invention, a topology prediction method for satellite networks is provided, the method comprising the following steps: T1, acquiring network topology data of the satellite network at the previous moment; T2, using a topology prediction model trained by the method described in the first aspect of the present invention to predict the network topology data of the satellite network at the next moment based on the network topology data of the satellite network acquired in step T1.

[0018] Compared with the prior art, the advantages of the present invention are as follows:

[0019] (1) Divide the world into different regions and fully consider the link relationships between terminals and satellites, ground stations and satellites, and satellites within the region to obtain historical topology data of satellite networks within the region. This accurately describes the interrelationships between satellites, ground stations and terminals within the region, which is beneficial for satellite resource allocation.

[0020] (2) It fully considers inter-satellite, satellite-to-ground and user link data, and uses a dataset composed of satellite-to-ground, inter-satellite and user link data to train the neural network model to learn the spatiotemporal dynamic characteristics of historical satellite network topology data. The trained neural network model predicts the satellite network topology of the next moment based on the satellite network topology data of the previous moment, so that it is not necessary to store satellite network topology data for all time periods. Only the weight memoization parameters of the trained model need to be stored to realize satellite network topology prediction, which greatly reduces storage occupation and ensures the integrity of network topology information. Attached Figure Description

[0021] The embodiments of the present invention will be further described below with reference to the accompanying drawings, wherein:

[0022] Figure 1 This is a schematic diagram illustrating an example of a satellite communication network according to the present invention;

[0023] Figure 2 This is a schematic diagram illustrating an example of an inter-satellite link according to the present invention;

[0024] Figure 3 This is a schematic diagram of the topology prediction model structure according to an embodiment of the present invention;

[0025] Figure 4 This is a schematic diagram of the spatiotemporal learning module structure according to an embodiment of the present invention;

[0026] Figure 5 This is a schematic diagram illustrating an example of the spatiotemporal learning module structure according to an embodiment of the present invention;

[0027] Figure 6 This is a schematic diagram of the training process of the topology prediction model according to an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention is further described in detail below through specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0029] As mentioned in the background section, existing solutions suffer from problems such as large storage space requirements, high topology computation complexity, and incomplete satellite network topology information due to the failure to consider changes in user links (links between satellites and terminals), which are detrimental to satellite network topology prediction.

[0030] To address the above shortcomings, the inventors discovered that the large storage space requirement stems from the need to store excessive historical satellite network topology data in existing technologies. The high computational complexity arises from the use of traversal methods to iterate through satellite network topology data across all time periods. Furthermore, the incomplete satellite network topology information is due to the lack of consideration for changes in user links. To resolve these deficiencies, this invention proposes a scheme that comprehensively utilizes satellite-to-ground, inter-satellite, and user link data to train a neural network model. By training the neural network model with complete satellite network topology data, it eliminates the need to store satellite network topology data for all time periods, significantly reducing storage requirements while ensuring the integrity of network topology information. In summary, this invention trains a neural network model to convergence using complete satellite network topology data including satellite-to-ground, inter-satellite, and user link data. The trained model then predicts the satellite network topology for the next time period based on the previous time period's topology data. This eliminates the need to store satellite network topology data for each time period; only the trained model parameters are required for satellite network topology prediction, thereby reducing storage requirements and computational complexity. Furthermore, because the training process uses complete satellite network topology data including satellite-to-ground, inter-satellite, and user links, the satellite network topology data predicted by the model can ensure the integrity of the satellite network topology.

[0031] To better understand the present invention, the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0032] To facilitate understanding, this invention will be described from the aspects of satellite communication networks, the link establishment constraints of satellite communication networks, the structure of topology prediction models, and the training process.

[0033] I. Satellite Communication Network

[0034] According to one example of the invention, such as Figure 1 As shown, it illustrates a satellite communication network comprising multiple satellites, multiple ground stations, and multiple terminals (vehicle-mounted terminals, airborne terminals, handheld terminals, and shipborne terminals, etc.). Inter-satellite links can be established between satellites, user links can be established between satellites and terminals, and satellite-to-ground links (feed links) can be established between satellites and ground stations. It should be noted that... Figure 1 The satellite communication network shown cannot be directly used as data for model training; it needs to be converted into a satellite network topology diagram for further processing. Therefore, the satellite communication network needs to be converted into its corresponding satellite network topology diagram. Based on the basic principles of graph theory (graph theory is existing technology and will not be discussed in detail here), the satellite network topology diagram corresponding to the satellite communication network is represented as follows: , ,in, Represents a set of nodes. Represents a node It represents a terminal, satellite, or ground station in a satellite communication network; express The set of edges at any given time; if there is an edge between two nodes, it means that a link exists between those two nodes. express Time Node (Terminal, ground station, or satellite) and nodes Links between (terminals, ground stations, or satellites). Although the basic principles of graph theory can convert a satellite communication network into a satellite network topology graph, the link connections within the satellite communication network are uncertain at different times. Therefore, before constructing the satellite network topology graph corresponding to a specific moment, it is necessary to determine the link connections within the satellite communication network, i.e., to determine all inter-satellite links, all satellite-to-ground links, and all user links included in the network. Since establishing links between satellites and terminals, or between satellites or ground stations, requires satisfying link establishment constraints, determining the link connections within a satellite communication network only requires judging whether there are inter-satellite links between two satellites, satellite-to-ground links between a ground station and a satellite, and user links between a terminal and a satellite based on these constraints.

[0035] II. Constraints on Establishing Satellite Communication Networks

[0036] 2.1 Constraints on Establishing Inter-Satellite Links

[0037] To establish an inter-satellite link between two satellites, both line-of-sight constraints and relative motion constraints must be met simultaneously.

[0038] The so-called communication line-of-sight constraint refers to the situation where the line of sight between two satellites cannot pass through the Earth, in which case the communication line-of-sight between the two satellites is unobstructed. According to an example of the present invention, such as... Figure 2 As shown, Figure 2 (a) shows the connection between two satellites at the same orbital altitude. Figure 2 (b) shows the connection between two satellites at different orbital altitudes, in order to Figure 2 The scenario shown illustrates the communication line-of-sight constraint. It should be noted that, although... Figure 2 (a) and Figure 2 (b) The situations shown are different, but the same line-of-sight constraint can be used to determine whether the line connecting the two satellites passes through the Earth. The line-of-sight constraint is:

[0039] when At that time, the line of sight between the two satellites does not pass through the Earth, and the line of sight between the two satellites is not obstructed;

[0040] when hour:

[0041]

[0042] when ,and At that time, the line of sight between the two satellites does not pass through the Earth, and the line of sight for communication between the two satellites is unobstructed; among them, This indicates the altitude of the center line connecting the two satellites. This indicates that the center of the Earth points towards the satellite with the lower position among the two satellites. vector, This indicates that the center of the Earth points towards the satellite with the higher position among the two satellites. vector, This represents the angle between the vectors of the two satellites. It represents the vertical distance from the line connecting the two satellites to the Earth's surface. Indicates the satellite with the lower position among the two satellites. The angle of elevation.

[0043] Relative motion constraints mean that the relative velocity between two satellites must be less than or equal to the maximum angular velocity of the antenna rotation of the satellite with the lower position. The relative velocity between the two satellites is calculated as follows:

[0044]

[0045]

[0046]

[0047] in, This represents the difference in the velocity vectors of the two satellites. Indicates the satellite with the higher position among the two satellites. The velocity vector of motion, , and They represent exist axis, shaft and Components on the axis, Indicates the satellite with the lower position among the two satellites. The velocity vector of motion, , and They represent exist axis, shaft and Components on the axis, This indicates the relative velocity between the two satellites. Indicates the satellite with the lower position among the two satellites. The maximum angular velocity of the antenna rotation.

[0048] 2.2 Constraints for Establishing a Satellite-to-Ground Link Between Ground Station and Satellite

[0049] For a ground station to establish a satellite-to-ground link with a satellite, the elevation angle between the ground station and the satellite needs to be greater than a preset elevation angle value. According to one example of the present invention, the preset elevation angle value ranges from 25° to 40°. According to one embodiment of the present invention, the elevation angle between the ground station and the satellite is calculated as follows:

[0050] in, Indicates the elevation angle between the ground station and the satellite. Represents the ground station position vector. This represents the satellite's position vector. It should be noted that a ground station can establish satellite-to-ground links with multiple satellites; according to one example of the present invention, a ground station can simultaneously access up to eight satellites.

[0051] 2.3 Constraints on Establishing User Links Between Terminals and Satellites

[0052] A terminal can only establish a user link with one satellite, but a terminal can access more than one satellite. Therefore, to establish a user link between a terminal and a satellite, the elevation angle between the terminal and the satellite must first be greater than a preset elevation angle value. Then, the satellite with the highest elevation angle and the shortest distance is selected from all satellites that meet this condition. It should be noted that the preset elevation angle range and the elevation angle calculation method are the same as those described in the section on satellite-to-ground links, and will not be repeated here.

[0053] III. Structure of Topology Prediction Model

[0054] According to one embodiment of the present invention, such as Figure 3 As shown, the topology prediction model includes one or more spatiotemporal learning modules, wherein, as Figure 4 As shown, each spatiotemporal learning module includes a cascaded graph convolution module and a temporal feature extraction module. Preferably, the temporal feature extraction module is a Long Short-Term Memory (LSTM) network, a Recurrent Neural Network (RNN), or a Gated Recurrent Unit (GRU). For simplicity, as... Figure 5 As shown, the spatiotemporal learning module constructed using GCN (two stacked graph convolutional layers) and GRU is used as an example in the embodiments of the present invention for detailed explanation. It should be noted that multiple graph convolutional layers can be stacked in GCN, and the present invention does not impose a specific limitation on the number of graph convolutional layers.

[0055] To better understand this, we will briefly introduce the data processing principle of the topology prediction model during training. Since the topology prediction model includes one or more spatiotemporal learning modules, and each spatiotemporal learning module includes GCN and GRU, this invention will only explain the data processing principles of GCN and GRU respectively.

[0056] 3.1 GCN Data Processing Principles

[0057] First, adopt Extract spatial features from the input data; among which, express Spatial characteristics of the input data at time t. express Input data at any given time (in this invention, the input data is satellite network topology data). express The adjacency matrix, Indicates to Normalize, Representing the adjacency matrix With the identity matrix The sum of, for The degree matrix, This represents the weight matrix of the first convolutional layer. This represents the weight matrix of the second convolutional layer; This indicates that it has been processed by GCN;

[0058] Then, the spatial features are concatenated with the original input data and fed into the GRU to obtain the spatiotemporal features:

[0059]

[0060] in, express The spatiotemporal characteristics of the input data at time t, where tanh is the hyperbolic tangent activation function. This represents the spatial features extracted by GCN. The original input data, This indicates that the data has been processed by GRU.

[0061] 3.2 Data Processing Principles of GRU

[0062] As is well known, the GRU model has two gates: the reset gate and the update gate.

[0063] The reset gate determines how new input information is combined with previous memory (output), which can be expressed by the formula:

[0064]

[0065] in, This indicates that the door is being reset. This represents the reset gate weight matrix. express The spatiotemporal characteristics of a moment express Input data at any given time, "For the accumulation of Hadama," It is the sigmoid activation function.

[0066] The update gate controls the extent to which information from the previous state is incorporated into the current state. In other words, the update gate helps the model decide how much past information to pass to the future; simply put, it's used to update memory. This can be expressed by the formula:

[0067]

[0068] in, Indicates an update to the door. This indicates updating the gate weight matrix. express The spatiotemporal characteristics of a moment express Input data at any time, For Hadama accumulation, It is the sigmoid activation function.

[0069] Calculate the undetermined spatiotemporal features at the current time step based on the reset gate, the spatiotemporal features of the previous time step, and the input data at the current time step:

[0070]

[0071] in, express The undetermined time-space characteristics at time step 1, where tanh is the hyperbolic tangent activation function. A weight matrix representing the degree of influence of the previous time step on the current time step. This indicates that the door is being reset. express The spatiotemporal characteristics of a moment express Input data at any given time.

[0072] Calculate the spatiotemporal characteristics of the current moment based on the update gate, the pending spatiotemporal characteristics, and the spatiotemporal characteristics of the previous moment:

[0073]

[0074] in, Indicated The spatiotemporal characteristics of a moment Indicates an update to the door. express The spatiotemporal characteristics of a moment express The undetermined time-space characteristics at a given moment.

[0075] IV. Topology Prediction Model Training Process

[0076] According to one embodiment of the present invention, such as Figure 6 As shown, the training process of the topology prediction model includes steps S1-S2, which are explained below.

[0077] In step S1, historical satellite network topology data within the target area is obtained. This historical topology data includes satellite network topology data from multiple consecutive time points. Each time point's satellite network topology data includes multiple nodes and multiple edges. Nodes represent terminals, satellites, or ground stations, and an edge between any two nodes indicates a link between them. According to an embodiment of the present invention, step S1 includes the following steps: S11, dividing the globe into different regions; S12, using any one of the regions obtained in step S11 as the target region, obtaining historical satellite network topology data within that target region.

[0078] In step S2, the topology prediction model is trained based on the historical satellite network topology data obtained in step S1 until convergence. According to an embodiment of the present invention, step S2 includes steps S21-S23, which are described below.

[0079] In step S21, the corresponding adjacency matrix set is obtained based on the satellite network historical map data obtained in step S1.

[0080] In step S22, each adjacency matrix in the adjacency matrix set obtained in step S21 is processed according to a preset data processing rule, and a dataset is constructed using the processing results of each adjacency matrix. According to an embodiment of the present invention, the preset data processing rule is as follows: the original adjacency matrix is ​​added to the identity matrix and then symmetrically normalized to obtain a new adjacency matrix; a degree matrix is ​​constructed based on the new adjacency matrix; the new adjacency matrix and its corresponding degree matrix are used as the processing result of the original adjacency matrix. It should be noted that after processing the adjacency matrix according to the preset data processing rule to obtain the dataset, the dataset is divided into a training set, a validation set, and a test set in a 7:1:2 ratio, and the topology prediction model is trained using the divided training set.

[0081] In step S23, the topology prediction model is trained based on the dataset obtained in step S22 until the model converges.

[0082] According to an embodiment of the present invention, step S23 includes: S231, performing multiple data cuts from the dataset obtained in step S22 to obtain multiple samples forming a training set, wherein each data cut selects data corresponding to the satellite network topology map within a time window of a preset width that is continuous in time from the dataset; according to an embodiment of the present invention, during each cut, data with a time window of T and data at time point T+1 are randomly selected from the dataset, wherein T is a preset time window width, such as 6, 7, 8, etc., and the data at time point T+1 is used as the true control value in the model prediction. S232, training the topology prediction model using self-supervised learning based on the training set obtained in step S231 until the model converges; according to an embodiment of the present invention, MSE (mean squared error) is used as the loss function during the training process, and the Adam algorithm is used to... The weights are updated based on the learning rate. Specifically, during training, data with a time window of T is input into the model each time, and the predicted value at time point T+1 is obtained from the model output; then, the error between the predicted value at time point T+1 and the true value at time point T+1 is calculated using the loss function; finally, the Adam algorithm is applied based on the error. Update the weights based on the learning rate. The Adam algorithm process is as follows:

[0083] First initialize the gradient First-order exponential moving weighted average and second-order exponential moving weighted average , yes A biased estimate of the first-order origin distance yes The biased estimate of the second-order origin distance is obtained; then gradient descent is continuously performed until the model parameters are obtained. Convergence. Wherein:

[0084] The following method is used to calculate the first... gradient of round iteration :

[0085]

[0086] in, For loss function, Represents the gradient operator, express Wheel model parameters.

[0087] Update using the following method First-order exponential moving weighted average and second-order exponential moving weighted average :

[0088]

[0089]

[0090] in, and These are the first-order and second-order exponentially moving weighted decay rates, respectively.

[0091] The following methods are used to perform First-order exponential moving weighted average and second-order exponential moving weighted average Perform deviation correction:

[0092]

[0093]

[0094] Update the model parameters as follows:

[0095]

[0096] in, For the learning rate, select ; To avoid the denominator being zero, a relatively small value is generally used. .

[0097] The beneficial effects of this invention are as follows:

[0098] (1) Divide the world into different regions and fully consider the link relationships between terminals and satellites, ground stations and satellites, and satellites within the region to obtain historical topology data of satellite networks within the region. This accurately describes the interrelationships between satellites, ground stations and terminals within the region, which is beneficial for satellite resource allocation.

[0099] (2) It fully considers inter-satellite, satellite-to-ground and user link data, and uses a dataset composed of satellite-to-ground, inter-satellite and user link data to train the neural network model to learn the spatiotemporal dynamic characteristics of historical satellite network topology data. The trained neural network model predicts the satellite network topology of the next moment based on the satellite network topology data of the previous moment, so that it is not necessary to store satellite network topology data for all time periods. Only the weight memoization parameters of the trained model need to be stored to realize satellite network topology prediction, which greatly reduces storage occupation and ensures the integrity of network topology information.

[0100] It should be noted that although the steps are described in a specific order above, it does not mean that the steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required function can be achieved.

[0101] This invention can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0102] Computer-readable storage media can be tangible devices that hold and store instructions for use by an instruction execution device. Computer-readable storage media can include, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof.

[0103] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for training a topology prediction model for a satellite network, the satellite network comprising multiple satellites, multiple terminals, and multiple ground stations, characterized in that, The topology prediction model includes a spatiotemporal learning layer and a fully connected layer, and the method includes the following steps: S1. Obtain historical satellite network topology data within the target area. This historical topology data includes satellite network topology data from multiple consecutive time points. Each time point's satellite network topology includes multiple nodes and multiple edges. Nodes represent terminals, satellites, or ground stations. An edge between any two nodes indicates a link between them. These links include inter-satellite links, satellite-to-ground links, and user links. Step S1 includes: S11. Divide the world into different regions; S12. Taking any region from all the regions obtained in step S11 as the target region, obtain the historical satellite network topology data within that target region, wherein: The existence of an inter-satellite link between any two satellites is determined based on the communication line-of-sight constraint and the relative motion constraint. The communication line-of-sight constraint means that the line connecting the two satellites does not pass through the Earth, and the relative motion constraint means that the relative motion speed between the two satellites is less than or equal to the maximum angular velocity of the antenna of the satellite with the lower position. The existence of a satellite-to-ground link between any ground station and any satellite is determined based on the elevation angle constraint, where the elevation angle constraint means that the elevation angle between the ground station or terminal and the satellite is greater than a preset elevation angle value. The existence of a user link between any terminal and any satellite is determined based on elevation and distance constraints. The distance constraint means selecting the satellite with the shortest distance to the terminal from all satellites that satisfy the elevation constraint to establish a user link. S2. Train the topology prediction model based on the historical satellite network topology data obtained in step S1 until convergence.

2. The method according to claim 1, characterized in that, The spatiotemporal learning layer includes one or more spatiotemporal learning modules, which are used to acquire the spatial and spatiotemporal features of satellite network topology data.

3. The method according to claim 2, characterized in that, The spatiotemporal learning module includes a graph convolution module and a temporal feature extraction module connected in series. The graph convolution module is used to extract the spatial features of the satellite network topology map data, and the temporal feature extraction module is used to obtain the spatiotemporal features of the satellite network topology map data based on the output of the graph convolution module and the satellite network topology map data.

4. The method according to claim 3, characterized in that, The time feature extraction module employs a long short-term memory network, a recurrent neural network, or a gated recurrent unit.

5. The method according to claim 4, characterized in that, Step S2 includes: S21. Based on the historical satellite network map data obtained in step S1, obtain the corresponding set of adjacency matrices. S22. Process each adjacency matrix in the adjacency matrix set obtained in step S21 according to the preset data processing rules, and construct a dataset based on the processing results of each adjacency matrix. S23. Train the topology prediction model based on the dataset obtained in step S22 until it converges.

6. The method according to claim 5, characterized in that, In step S22, the preset data processing rule is as follows: The new adjacency matrix is ​​obtained by adding the original adjacency matrix to the identity matrix and then performing symmetric normalization. Construct its corresponding degree matrix based on the new adjacency matrix; The new adjacency matrix and its corresponding degree matrix are used as the result of processing the original adjacency matrix.

7. The method according to claim 6, characterized in that, Step S23 includes: S231. Multiple data cuts are performed on the dataset obtained in step S22 to obtain multiple samples to form a training set. Each data cut is performed by selecting data corresponding to the satellite network topology map within a time window of a preset width that is continuous in time from the dataset. S232. Based on the training set obtained in step S231, the topology prediction model is trained using self-supervised learning until the model converges.

8. The method according to claim 7, characterized in that, The preset width is 6.

9. A topology prediction method applied to satellite networks, characterized in that, The method includes the following steps: T1. Obtain the network topology data of the satellite network at the previous moment; T2. The topology prediction model trained using any one of the methods described in claims 1-8 predicts the network topology data of the satellite network at the next moment based on the network topology data of the satellite network obtained in step T1 at the previous moment.

10. A computer-readable storage medium, characterized in that, It contains a computer program that can be executed by a processor to implement the steps of the method according to any one of claims 1-9.

11. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the electronic device to perform the steps of the method as described in any one of claims 1-9.

Images