Satellite time-varying network performance prediction method based on graph neural network

Abstract

The invention relates to the technical field of computer communication, and provides a satellite time-varying network performance prediction method based on a graph neural network. Specifically, the method comprises the following steps: obtaining network topology and state data of an LEO satellite network from a data plane through an SDN technology; constructing a graph network with heterogeneous nodes according to the network topology and the network flow; the efficient prediction of the LEO satellite network performance is realized through a network performance prediction model based on a message passing neural network (MPNN) and added with a node feature LSTM mechanism and a graph attention mechanism. The method can accurately predict the key performance indexes (KPI) of the end-to-end network flow in the satellite network, such as delay, jitter and packet loss rate, thereby optimizing the network quality of service (QoS).

Description

Technical Field

[0001] The present invention relates to the field of computer communication technology, and in particular to a method for predicting satellite time-varying network performance based on graph neural network. Background Art

[0002] Low Earth Orbit (LEO) satellite networks have become an indispensable part of modern communication architecture due to their wide coverage, ability to be unrestricted by geographical and natural conditions, advantages in long-distance communication, and significant communication capacity. With the continuous evolution of communication technology, the importance of LEO satellite networks has become increasingly prominent, providing new possibilities for global interconnection.

[0003] In order to achieve effective management of these dynamic networks, software-defined networking (SDN) technology has emerged. Through SDN, we can obtain and analyze the data plane characteristics of satellite networks to build an accurate network model. This model not only helps us deeply understand the network operation mechanism, but also predicts key performance indicators (KPIs) such as traffic distribution and transmission delay, making it possible to optimize the network for specific needs, thereby improving network efficiency and reducing the possibility of failure.

[0004] However, due to the high-speed motion of LEO satellites in their predetermined orbits, the network topology changes rapidly and inter-satellite links are frequently established and disconnected. This phenomenon poses a huge challenge to traditional static network modeling methods - the effectiveness of traditional methods in ground networks is inadequate when facing LEO satellite networks.

[0005] Early attempts to model SDN networks used mathematical models based on queuing theory, but such models usually assume that network traffic follows a Poisson distribution or that routing strategies are probabilistically selected, which is obviously inconsistent with actual operation and therefore cannot provide accurate performance predictions. In addition, although detailed simulations can be performed using network packet-level simulators, this method is extremely computationally expensive and difficult to apply on a large scale.

[0006] In recent years, the advancement of deep learning technology has opened up new paths for SDN network modeling. Advanced algorithms such as fully connected networks, convolutional neural networks (CNNs), graph neural networks (GNNs), recurrent neural networks (RNNs), and variational autoencoders (VAEs) have been introduced into the field of network modeling. Thanks to their excellent feature extraction capabilities, these deep learning-based methods have significantly improved the accuracy of performance prediction. However, it is worth noting that most of the above-mentioned technologies are applicable to relatively static ground SDN environments; when it comes to network topologies and link states that change over time, existing performance prediction algorithms face severe challenges because dynamic changes have complex effects on network flow performance on a time scale. Summary of the invention

[0007] In view of the time-varying network prediction problem caused by route changes and link changes brought about by the high-speed movement of the above-mentioned low-orbit (LEO) satellite network, the purpose of the present invention is to provide a satellite time-varying network performance prediction method, obtain network status information from the SDN status plane, and model the satellite network through a graph model with heterogeneous nodes, and use a network performance prediction model based on a message passing neural network (MPNN) and with the addition of node feature LSTM and graph attention mechanism to achieve efficient prediction of LEO satellite time-varying network performance.

[0008] The present invention achieves the above-mentioned purpose by the following technical scheme, a satellite time-varying network performance prediction method based on graph neural network, specifically comprising the following technical steps:

[0009] Step 1: Divide the low-orbit (LEO) satellite operation cycle into multiple time slices. In each time slice, treat the logical satellite network as a static network, obtain the satellite operation position information and inter-satellite link connection status, and calculate the routing information offline.

[0010] Step 2: for the satellite network of each time slice, based on the software defined network SDN architecture, obtain the network status data from the data plane of the LEO satellite network; for the satellite network of each time slice, convert the graph structure of the network topology mode into a graph structure with a queue-link-data flow heterogeneous node mode; after the data preprocessing, obtain the graph time series data of the satellite network;

[0011] Step 3: For the network status data, embed the status statistical data into a feature vector, construct a network performance prediction model based on a message passing neural network (MPNN) and adding a node feature LSTM mechanism and a graph attention mechanism, and use the graph time series data of the satellite network for model training;

[0012] Step 4: Based on the network prediction performance model, the final network performance prediction results are output from the actual network status of the LEO satellite, including network delay, jitter, and packet loss indicators.

[0013] Compared with the prior art, the present invention has the following significant advantages:

[0014] (1) The present invention uses node feature LSTM (v-LSTM) to model the temporal characteristics of the satellite network to deal with the impact of network topology changes and link quality changes caused by satellite movement on the prediction results;

[0015] (2) Using a link-queue-path heterogeneous graph structure instead of the original network topology as the model input data, it more accurately describes the network operation mode and can handle arbitrary topologies, routing schemes, and traffic intensities to generalize to satellite networks that have not been seen in training; BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 : Flowchart of satellite time-varying network performance prediction method.

[0017] Figure 2 : Schematic diagram of heterogeneous graph network model.

[0018] Figure 3 : Schematic diagram of the prediction model structure.

[0019] Figure 4 : Schematic diagram of message propagation mechanism. DETAILED DESCRIPTION

[0020] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0021] The dynamic network routing optimization method and system based on deep learning prediction of the present invention comprises the following steps:

[0022] Step 1: Divide the low-orbit (LEO) satellite operation cycle into multiple time slices. In each time slice, treat the logical satellite network as a static network, obtain the satellite operation position information and inter-satellite link connection status, and calculate the routing information offline.

[0023] Step 2: for the satellite network of each time slice, based on the software defined network SDN architecture, obtain the network status data from the data plane of the LEO satellite network; for the satellite network of each time slice, convert the graph structure of the network topology mode into a graph structure with a queue-link-data flow heterogeneous node mode; after the data preprocessing, obtain the graph time series data of the satellite network;

[0024] Step 3: For the network status data, embed the status statistical data into a feature vector, construct a network performance prediction model based on a message passing neural network (MPNN) and adding a node feature LSTM mechanism and a graph attention mechanism, and use the graph time series data of the satellite network for model training;

[0025] Step 4: Based on the network prediction performance model, the final network performance prediction results are output from the actual network status of the LEO satellite, including network delay, jitter, and packet loss indicators.

[0026] Furthermore, in step one, the entire operation cycle T of the satellite is divided into multiple time slices. In each time slice, the logical satellite network is treated as a static network, and the state at the start time of each time slice is taken as the satellite network operation state in the time slice. The satellite operation simulator STK is used to simulate the operation constellation and obtain the position data of the satellite in the constellation at that moment, and the connection state of the inter-satellite link (ISL) is obtained. The routing table information is calculated offline based on the routing algorithm to obtain the routing forwarding table of the satellite network.

[0027] Furthermore, in step 2, for each time slice of satellite network status data input, the satellite buffer queue in the network (denoted as q), the end-to-end data flow (denoted as f), and the link between satellite nodes (denoted as l) are used as nodes of the heterogeneous graph. Then, based on the routing table information, the link through which the data flow runs and the corresponding satellite node (denoted as f) are obtained. i ={(q i，1 , l i，1 ),...,(q i，M , l i，M )}). Finally, the edges between heterogeneous nodes can be formed according to the relationship between the links and corresponding nodes through which the data flow passes, and the various data flows, links and queues of the network topology can be organized into a heterogeneous graph network, such as Figure 2 shown.

[0028] According to the time slice start time in step 1, the traffic information of the data packets in the network topology in multi-hop routing forwarding and global end-to-end is collected to represent the data packet throughput and bit rate statistics of the end-to-end data flow at the time, thereby forming the traffic characteristics of the heterogeneous graph network. ;

[0029] The specific process of collecting traffic characteristics is as follows:

[0030] Step 2.1: Record the routing forwarding matrix (denoted as R) of each time slice network topology;

[0031] Given a fixed routing table for network data flow, record the forwarding matrix R, R i,j When the element is 1, it means that nodes i and j are directly reachable;

[0032] Step 2.2: Record the traffic matrix of each time slice network topology (denoted as P);

[0033] Considering the average bandwidth of the data flow between two nodes at the current moment, the total number of data packets, the distribution type and distribution parameters of the arrival time of data packets, and the distribution type and distribution parameters of the size of data packets, a dictionary structure is formed and saved as the matrix element P i,j ;

[0034] Step 2.3: Record the performance matrix of each time slice network topology (denoted as T);

[0035] The information such as packet loss and delay quantile between two nodes at the current moment is used as T i,j Elements (real packet loss or

[0036] The delay index is denoted as as the true label);

[0037] Step 2.4: Record the static features of each heterogeneous node and mark the edge types connecting the heterogeneous nodes;

[0038] Record the characteristics of each node in the heterogeneous graph: link node characteristics include the bandwidth of the two nodes, link load, and scheduling policy (for serving QoS queues); queue node characteristics include buffer queue size, queue priority, and scheduling policy; data stream node characteristics include scheduling policy type, specific parameters of the traffic model, and average traffic of data stream transmission, scheduling policy type, service type (ToS) of data stream, data packet arrival time distribution parameters, data packet size distribution parameters, average bandwidth, average number of data packets, delay, and packet loss;

[0039] Mark the edge types in the heterogeneous graph: the edge type between the queue node and the data flow node is marked as 0; the edge type between the data flow node and the link node is marked as 1; the edge type between the link node and the queue node is marked as 2.

[0040] Furthermore, in step three, a network performance prediction model based on graph neural network is constructed, such as Figure 3 As shown in the figure, a message passing neural network is used to describe the relationship model between nodes. First, the data stream sequence of queue features and link features is combined to obtain the time series representation of the data stream. Then, the data stream representation is used as the queue sequence to update the queue representation. Finally, the queue is used as the link sequence to update the link representation. The feature representation of the path node is obtained through multiple iterations. The node feature LSTM is used to capture the temporal change pattern of the dynamic graph. The graph attention mechanism is used to aggregate the node feature LSTM to update the data stream node and adjacent node features.

[0041] The specific steps for building a network performance prediction model are as follows:

[0042] Step 3.1: Construct a message propagation neural network and a node feature LSTM module to update the node feature representation;

[0043] like Figure 3 As shown in the figure, in the message passing process, the path sequence of queue features and link features is first used as a message update sequence to update the data stream feature representation, then the data stream feature is used as a message update sequence to update the queue feature representation, and finally the queue feature is used as a message update sequence to update the link feature representation, and the message passing process is completed through multiple iterations. In the node feature LSTM, the queue feature sequence on the time scale is used as the input sequence of the long short-term memory to update the queue feature of the next time step.

[0044] The specific node feature update process is as follows:

[0045] Step 3.1.1: Initialize feature representation;

[0046] The feature input of a given heterogeneous node is converted into the feature representation of three heterogeneous nodes, and the data flow node features are initialized as average bandwidth and average number of data packets (denoted as X f ), the link node characteristics are the number of links, link capacity and scheduling strategy type (denoted as X l ), initialize the queue node features as queue size, scheduling strategy and weight (denoted as X q ); The features are converted into feature vectors through the embedding layer, which can be expressed as follows:

[0047] h f =E f (X f );h q =E q (X q );h l =E l (X l )

[0048] Where E f 、E q 、E l The embedded functions represent the data flow characteristics, queue characteristics, and link characteristics respectively;

[0049] Step 3.1.2: By Figure 4 The message propagation mechanism shown updates the node feature representation;

[0050] The specific message propagation mechanism is as follows:

[0051] Step 3.1.2.1: Update data flow node features from link node and queue node features;

[0052] The edge types in the heterogeneous graph are selected as 0 and 1, which represent the link and queue node features that have a relationship with the data flow node. The recurrent neural network (RNN) is used to update the data flow node features from the adjacent node messages, which can be expressed as ml，q ,h f =RNN([h l ||h q ],h f ), where [h l ||h q ] means that the links and queues that the path passes through are spliced ​​according to the feature dimension and used as the input sequence; m q,l Indicates the sequence characteristics of the data flow path through the queue and link; the input h f Represents the initial hidden state in the RNN. In the RNN, a new round of path state representation is obtained through the update of the recurrent gated unit (GRU), which is used for the subsequent prediction of the readout function performance index. The GRU includes a message function and an update function, which can be expressed as follows:

[0053]

[0054] in represents the feature vector of node v updated in step t, represents the feature vector of node w adjacent to node v, represents the message vector updated in step t+1, M t represents the message function, N(v) represents the set of adjacent nodes of node v, and U t represents the update function;

[0055] Step 3.1.2.2: Update queue node features from data stream node features;

[0056] Select the heterogeneous graph edge type as 0, which means the link and queue node features that are related to the data flow node, and use a recurrent neural network (RNN) to update the data flow node features from the adjacent node messages. The GRU unit in the RNN is the same as step 3.1.2.1;

[0057] Step 3.1.2.3: Update link node features from queue node features;

[0058] Select heterogeneous graph edge type 2, which represents the queue node feature that has a relationship with the link node, and use a recurrent neural network (RNN) to update the link node feature from the adjacent node message. The GRU unit in the RNN is the same as step 3.1.2.1;

[0059] Step 3.1.3: Node feature LSTM updates node feature representation;

[0060] The node feature LTSM (v-LSTM) is used to model the feature changes of nodes on a time scale. The queue nodes are relatively stable in the satellite network, while the intersatellite links are dynamically connected and disconnected at each time step due to satellite movement. Therefore, only the queue node feature vector is used as the v-LSTM input for iterative update, which can be expressed as follows:

[0061]

[0062] H i Represents the matrix composed of queue node feature vectors at time step i, v-LSTM N LSTM by the number of graph nodes N Composition, each LSTM N represents an LSTM unit with N output nodes, and LSTM N In each LSTM input and output node, there are calculation gates such as forget gate, input gate, and output gate, and the transmission of timing information is completed between nodes by updating the cell state vector and hidden state vector.

[0063] Step 3.2 predict network performance indicators;

[0064] After the above modules iteratively update the feature vectors of each node, in the readout function, the graph attention module is used to perform the final representation of the data stream node, so that each data stream node aggregates information from the state of the adjacent nodes whose states have been updated. The graph attention module function can be expressed as follows:

[0065] a ij =Softmax(LeakyReLU(a T [Wh i ||Wh j ]))

[0066]

[0067] Where [Wh i ||Wh j ] indicates the splicing operation, h i and h j represents the feature vector of two adjacent nodes i and j, W is the learnable parameter matrix, a is the learnable parameter vector, a ij represents the attention parameter between nodes i and j, and N(i) represents the set of adjacent nodes of node i;

[0068] Finally, a linear layer MLP is used to output the final network performance prediction result from the feature vector of the data flow node.

[0069] Step 3.3: Construct the loss function of the training model, use the network flow path performance matrix obtained in step 2.1.3 as the true label for model training, and train the model until convergence;

[0070] Using MAPE as the regression loss function, it can be expressed as the following formula:

[0071]

[0072] in represents the true label, Represents the model prediction results;

[0073] Furthermore, in step 4, by periodically acquiring network status data from the data plane of the SDN, the link information includes the bandwidth of the two nodes, the scheduling strategy (for serving the QoS queue), etc.; the satellite node queue information includes the buffer queue size, scheduling strategy, etc.; the data flow information includes the scheduling strategy type, the service type (ToS) of the data flow, the arrival time distribution parameters of the data packets, the size distribution parameters of the data packets, the average bandwidth, the average number of data packets, etc. Each time point is recorded as the time series data of the target prediction network, which is used as the input of the trained model, and finally the network performance prediction results such as delay, jitter, and packet loss of the satellite network at the latest time point are output.

[0074] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A satellite time-varying network performance method based on graph neural network, characterized in that: The following steps are involved: Step 1: Divide the low-orbit (LEO) satellite operation cycle into multiple time slices. In each time slice, treat the logical satellite network as a static network, obtain the satellite operation position information and inter-satellite link connection status, and calculate the routing information offline. Step 2: for the satellite network of each time slice, based on the software defined network SDN architecture, obtain the network status data from the data plane of the LEO satellite network; for the satellite network of each time slice, convert the graph structure of the network topology mode into a graph structure with a queue-link-data flow heterogeneous node mode; after the data preprocessing, obtain the graph time series data of the satellite network; Step 3: For the network status data, embed the status statistical data into a feature vector, construct a network performance prediction model based on a message passing neural network (MPNN) and adding a node feature LSTM mechanism and a graph attention mechanism, and use the graph time series data of the satellite network for model training; Step 4: Based on the network prediction performance model, the final network performance prediction results are output from the actual network status of the LEO satellite, including network delay, jitter, and packet loss indicators.

2. The method for predicting satellite time-varying network performance based on graph neural network according to claim 1, characterized in that: In step one, the entire operation cycle T of the satellite is divided into multiple time slices. In each time slice, the logical satellite network is treated as a static network. The state at the start of each time slice is taken as the satellite network operation state in the time slice. The satellite operation simulator STK is used to simulate the operation constellation and obtain the position data of the satellite in the constellation at that moment, and the connection state of the inter-satellite link (ISL) is obtained. The routing information is calculated offline based on the routing algorithm to obtain the routing forwarding table of the satellite network.

3. The method for predicting satellite time-varying network performance based on graph neural network according to claim 1, characterized in that: In step 2, for each time slice of satellite network status data input, the satellite buffer queue (denoted as q j ∈Q, j∈(1,...,n q ), the end-to-end data flow (denoted as ), the link between satellite nodes (denoted as k∈(1,...,n l )) as nodes of the heterogeneous graph. Then, based on the routing table information, the links and corresponding satellite nodes (denoted as f) that the data flow running path passes through are obtained. i ={(q i，1 , l i，1 ),...,(q i，M , l i ，M )}); Finally, the edges between heterogeneous nodes can be formed according to the relationship between the links and corresponding nodes through which the data flows pass, and the various data flows, links and queues of the network topology can be organized into a heterogeneous graph network; according to the starting time of the time slice described in step one, the traffic information of the data packets in the network topology in multi-hop routing forwarding and global end-to-end is collected to represent the data packet throughput and bit rate statistics of the end-to-end data flow at the said time, thereby forming the traffic characteristics of the heterogeneous graph network.

4. The method for predicting satellite time-varying network performance based on graph neural network according to claim 3 is characterized in that: The data collection process in step 2 includes the following steps: Step 2.1: Record the routing forwarding matrix (denoted as R) of each time slice network topology; Given a fixed routing table for network data flow, record the forwarding matrix R, R i,j When the element is 1, it means that nodes i and j are directly reachable; Step 2.2: Record the traffic matrix of each time slice network topology (denoted as P); Get the average bandwidth of the data flow between two nodes at the current moment, the total number of data packets, the data packet arrival time distribution type and its distribution parameters, and the data packet size distribution type and its distribution parameters, form a dictionary structure, and save it as the matrix element P i,j ; Step 2.3: Record the performance matrix of each time slice network topology (denoted as T); The packet loss, delay quantile, and information between two satellite nodes at the current moment are taken as T i,j Elements (real packet loss or The delay index is denoted as as the true label); Step 2.4: Record the static features of each heterogeneous node and mark the edge types connecting the heterogeneous nodes; Record the characteristics of each node in the heterogeneous graph: link node characteristics include the bandwidth of the two nodes, link load, and scheduling policy (for serving QoS queues); queue node characteristics include buffer queue size, queue priority, and scheduling policy; data stream node characteristics include scheduling policy type, specific parameters of the traffic model, and average traffic of data stream transmission, scheduling policy type, service type (ToS) of data stream, data packet arrival time distribution parameters, data packet size distribution parameters, average bandwidth, average number of data packets, delay, and packet loss; Mark the edge type in the heterogeneous graph: the edge type between the queue node and the data flow point is marked as 0; The edge type between a data flow node and a link node is marked as 1; the edge type between a link node and a queue node is marked as 2.

5. The method for predicting satellite time-varying network performance based on graph neural network according to claim 1, characterized in that: In step three, a network performance prediction model based on graph neural network is constructed, in which a message passing neural network is used to describe the relationship pattern between nodes. First, the data flow sequence of queue features and link features is combined to obtain the time series representation of the data flow. Secondly, the data flow representation is used as the queue sequence to update the queue representation. Finally, the queue is used as the link sequence to update the link representation. Multiple iterations are performed to obtain the feature representation of the path node. The node feature LSTM is used to capture the temporal change pattern of the dynamic graph nodes. The graph attention mechanism is used to aggregate the node feature LSTM to update the data flow node and adjacent node features.

6. The method for predicting satellite time-varying network performance based on graph neural network according to claim 5, characterized in that: In the process of constructing a network performance prediction model based on a graph neural network in step 3, the following steps are included: Step 3.1: Construct a message propagation neural network and a node feature LSTM module to update the node feature representation; In the process of message passing, the path sequence of queue features and link features is first used as a message update sequence to update the data stream feature representation, then the data stream features are used as a message update sequence to update the queue feature representation, and finally the queue features are used as a message update sequence to update the link feature representation. After multiple rounds of iterations, the message passing process is completed; in the node feature LSTM, the queue feature sequence on the time scale is used as the input sequence of the long short-term memory to update the queue feature vector of the next time step; Step 3.2 predict network performance indicators; After the above modules iteratively update the feature vectors of each node, in the readout function, the graph attention module is used to perform the final representation of the data stream node, so that each data stream node aggregates information from the state of the adjacent nodes whose states have been updated. The graph attention module function can be expressed as the following formula: a ij =Softmax(LeakyReLU(a T [Wh i ||Wh j ])) Where [Wh i ||Wh j ] indicates the splicing operation, h i and h j represents the feature vector of two adjacent nodes i and j, W is the learnable parameter matrix, a is the learnable parameter vector, a ij represents the attention parameter between nodes i and j, and N(i) represents the set of adjacent nodes of node i; Finally, a linear layer MLP is used to output the final network performance prediction result from the feature vector of the data flow node; Step 3.3: Construct a loss function for the training model, use the network flow path performance matrix obtained in step 2.1.3 as the true label for model training, and train the model until the loss converges; Using MAPE as the regression loss function, it can be expressed as the following formula: in represents the true label, Represents the model prediction results.

7. The method for predicting satellite time-varying network performance based on graph neural network according to claim 6, characterized in that: In step 3.1, use: Step 3.1.1: Initialize feature representation; The feature input of a given heterogeneous node is converted into the feature representation of three heterogeneous nodes, and the initial data flow node features are the average bandwidth and the average number of packets (denoted as X f ), the link node characteristics are the number of links, link capacity and scheduling strategy type (denoted as X l ), initialize the queue node features as queue size, scheduling strategy and weight (denoted as X q ); The features are converted into feature vectors through the embedding layer, which can be expressed as follows: h f =E f (X f );h q =E q (X q );h l =E l (X l ) Where E f 、E q 、E l Respectively represent the data flow characteristics, Embedding functions of queue features and link features; Step 3.1.2: Update node feature representation through message propagation mechanism; Step 3.1.3: Node feature LSTM updates node feature representation; The node feature LTSM (v-LSTM) is used to model the feature changes of nodes on a time scale. The queue nodes are relatively stable in the satellite network, while the intersatellite links are dynamically connected and disconnected at each time step due to satellite movement. Therefore, only the queue node feature vector is used as the v-LSTM input for iterative update, which can be expressed as follows: H i Represents the matrix composed of queue node feature vectors at time step i, v-LSTM N LSTM by the number of graph nodes N Composition, each LSTM N represents an LSTM unit with N output nodes, and LSTM N In each LSTM input and output node, there are calculation gates such as forget gate, input gate, and output gate, and the transmission of timing information is completed between nodes by updating the cell state vector and hidden state vector.

8. The method for predicting satellite time-varying network performance based on graph neural network according to claim 7, characterized in that: In step 3.1.2, use: Step 3.1.2.1: Update data flow node features from link node and queue node features; The edge types in the heterogeneous graph are selected as 0 and 1, which represent the link and queue node features that have a relationship with the data flow node. The recurrent neural network (RNN) is used to update the data flow node features from the adjacent node messages, which can be expressed as m l，q ,h f =RNN([h l ||h q ],h f ), where [h l ||h q ] means that the links and queues that the path passes through are spliced ​​according to the feature dimension and used as the input sequence; m q,l Indicates the sequence characteristics of the data flow path through the queue and link; the input h f Represents the initial hidden state in the RNN. In the RNN, a new round of path state representation is obtained through the update of the recurrent gated unit (GRU), which is used for the subsequent prediction of the readout function performance index. The GRU includes a message function and an update function, which can be expressed as follows: in represents the feature vector of node v updated in step t, represents the feature vector of node w adjacent to node v, represents the message vector updated in step t+1, M t represents the message function, N(v) represents the set of adjacent nodes of node v, and U t represents the update function; Step 3.1.2.2: Update queue node features from data stream node features; Select the heterogeneous graph edge type as 0, which represents the link and queue node features that have a relationship with the data flow node, and use a recurrent neural network (RNN) to update the data flow node features from the adjacent node messages, where the GRU unit is the same as in step 3.1.2.1; Step 3.1.2.3: Update link node features from queue node features; Select heterogeneous graph edge type 2, which represents the queue node feature that has a relationship with the link node, and use a recurrent neural network (RNN) to update the link feature from the adjacent node message, where the GRU unit is the same as step 3.1.2.

1.

9. The method for predicting satellite time-varying network performance based on graph neural network according to claim 1, characterized in that: In step 4, the network status data is obtained from the data plane of the SDN periodically. The link information includes the bandwidth of the two nodes, the scheduling policy (for serving the QoS queue), etc. Satellite node queue information includes buffer queue size, scheduling strategy, etc. The data flow information includes the scheduling strategy type, the service type (ToS) of the data flow, the arrival time distribution parameters of the data packets, the size distribution parameters of the data packets, the average bandwidth, the average number of data packets, etc. Each time point is recorded as the time series data of the target prediction network, which is used as the input of the trained model, and finally the network performance prediction results such as delay, jitter, packet loss, etc. of the satellite network at the latest time point are output.

Images