A satellite time-varying network performance prediction method based on a graph neural network

Abstract

The application 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, network topology and state data of a LEO satellite network are acquired from a data plane through SDN technology; a graph network with heterogeneous nodes is constructed according to the network topology and network flow; and 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 is used to realize efficient prediction of the network performance of the LEO satellite network. The method can accurately predict key performance indicators (KPI) of end-to-end network flow in the satellite network, such as delay, jitter and packet loss rate, so as to optimize the quality of network service (QoS).

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer communication, and particularly relates to a satellite time-varying network performance prediction method based on a graph neural network. BACKGROUND

[0002] Low Earth Orbit (LEO) satellite networks have become an integral part of modern communication architectures due to their wide coverage, ability to operate without geographical and natural condition limitations, advantages in long-distance communication, and significant communication capacity. With the continuous evolution of communication technology, the importance of LEO satellite networks is increasingly highlighted, providing new possibilities for global interconnection.

[0003] To achieve effective management of these dynamic networks, Software Defined Network (SDN) technology has emerged. Through SDN, we can obtain and analyze the data plane characteristics of satellite networks, thereby constructing accurate network models. Such models not only help us understand the network operation mechanism in depth, but also predict 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 likelihood of failures.

[0004] However, due to the high-speed movement characteristics of LEO satellites in their predetermined orbits, the network topology changes rapidly and the inter-satellite links are frequently established and disconnected. This phenomenon poses a great challenge to traditional static network modeling methods - the effectiveness of traditional methods in ground networks is not sufficient when faced with LEO satellite networks.

[0005] Early attempts at SDN network modeling used queuing theory-based mathematical models, but such models usually assume that network traffic follows a Poisson distribution or that routing strategies are probabilistically selected, which obviously does not match the actual operating conditions, and thus cannot provide accurate performance predictions. In addition, although network packet-level simulators can be used for detailed simulation, this method is extremely computationally expensive and difficult to apply on a large scale.

[0006] In recent years, the progress of deep learning technology has opened up new paths for SDN network modeling. Advanced algorithms such as fully connected networks, Convolutional Neural Networks (CNN), Graph Neural Networks (GNN), Recurrent Neural Networks (RNN), and Variational Autoencoders (VAE) have been introduced into the field of network modeling. Benefiting from their excellent feature extraction capabilities, these deep learning-based methods have significantly improved the accuracy of performance prediction. However, it is worth noting that the above-mentioned technologies are mostly suitable for relatively static ground SDN environments; when it comes to time-varying network topology and link state, existing performance prediction algorithms face serious challenges because dynamic changes have complex effects on network flow performance on the time scale. SUMMARY

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

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

[0009] Step one, divide the low earth orbit (LEO) satellite operation period into multiple time slices, in each time slice, treat the logical satellite network as a static network, obtain satellite operation position information and inter-satellite link connection state, and calculate route information offline;

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

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

[0012] Step four, based on the network prediction performance model, output the final network performance prediction result from the real network state of the LEO satellite, including the network delay, jitter and packet loss indicators.

[0013] Compared with the existing technology, the present application has the following obvious advantages:

[0014] (1) The present application uses node feature LSTM (v-LSTM) to model the time sequence features of the satellite network, to process the influence of the network topology change and link quality change caused by the satellite movement on the prediction result;

[0015] (2) Using a link-queue-path heterogeneous graph structure instead of the original network topology as the model input data, more accurately describing the network operation mode, and being able to process any topology, routing scheme and traffic intensity, to generalize to satellite networks not seen in training; BRIEF DESCRIPTION OF DRAWINGS

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

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

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

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

[0020] The application will be described in detail below with specific embodiments. The following examples will help those skilled in the art to further understand the application, but do not limit the application in any form. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of changes and improvements can be made. These are within the scope of the present application.

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

[0022] Step one, divide the low earth orbit (LEO) satellite operation period into multiple time slices, in each time slice, treat the logical satellite network as a static network, obtain satellite operation position information and inter-satellite link connection state, and offline calculate routing information;

[0023] Step two, for the satellite network of each time slice, based on the software defined network (SDN) architecture, obtain the state data of the network 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 queue-link-data flow heterogeneous node mode; after the data preprocessing, obtain the graph time series data of the satellite network;

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

[0025] Step four, based on the network performance prediction model, output the final network performance prediction results from the real network state of LEO satellite, including network delay, jitter, packet loss indicators.

[0026] Further, 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, and the satellite operation simulator STK is used to simulate the operation of the constellation and obtain the position data of the satellites in the constellation at that time, to obtain the connection state of the inter-satellite link (ISL), and based on the routing algorithm, the routing table information is calculated offline to obtain the routing and forwarding table of the satellite network.

[0027] Further, in step two, for the input satellite network state data of each time slice, 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 taken as the nodes of the heterogeneous graph. Then based on the routing table information, the links and corresponding satellite nodes (denoted as f i = {(q i，1 , l i，1 ),..., (q i，M , l i，M )} are obtained. Finally, according to the relationship between the links and corresponding nodes through which the data flow passes, the edges between the heterogeneous nodes are formed, and the network topology of each data flow, link and queue is organized into a heterogeneous graph network, as shown in Figure 2

[0028] According to the time slice start time in step one, the packet traffic information in the network topology in multi-hop routing forwarding and global end-to-end is collected to represent the packet throughput and bit rate statistical information 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 and forwarding matrix (denoted as R) of the network topology of each time slice;

[0031] Given the fixed routing and forwarding table of the network data flow, record the forwarding matrix R, R i,j element is 1, indicating that node i and node j are directly reachable;

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

[0033] ​Consider the average bandwidth between two nodes at the current time, the total number of data packets, the arrival time distribution type and its distribution parameters of data packets, and the size distribution type and its distribution parameters of data packets, to form a dictionary structure, and save it as a matrix element P i,j ;

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

[0035] The packet loss and delay quantile information between the current time and two nodes are taken as the T i,j element (the real packet loss or

[0036] delay index, denoted as as the real label);

[0037] Step 2.4: Record the static characteristics of each heterogeneous node and mark the edge type between the heterogeneous nodes;

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

[0039] Mark the edge type 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; and the edge type between the link node and the queue node is marked as 2.

[0040] Further, in step three, a network performance prediction model based on graph neural network is constructed, as shown in Figure 3 . Among them, the message passing neural network is used to describe the relationship model between nodes, first the data flow sequence is combined with the queue characteristics and the link characteristics to obtain the time sequence 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, and the feature representation of the path node is obtained after multiple iterations; the node feature LSTM is used to capture the change pattern of the dynamic graph in time; the graph attention mechanism is used to aggregate the data flow node and adjacent node features after the node feature LSTM update.

[0041] The specific network performance prediction model construction steps are as follows:

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

[0043] like Figure 3 As shown, during message passing, the path sequence of queue features and link features is first used as the message update sequence to update the data flow feature representation. Then, the data flow features are used as the message update sequence to update the queue feature representation. Finally, the queue features are used as the message update sequence to update the link feature representation. This process is repeated multiple times to complete the message passing process. In the node feature LSTM, the queue feature sequence at the time scale is used as the input sequence for the Long Short-Term Memory (LSTM) to update the queue features at the next time step.

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

[0045] Step 3.1.1: Initialize the feature representation;

[0046] The feature input of a given heterogeneous node is transformed into three feature representations of heterogeneous nodes. The data flow node features are initialized with average bandwidth and average number of data packets, etc. (denoted as X). f The characteristics of a link node include the number of links, link capacity, and scheduling strategy type (denoted as X). l The initialization of queue node characteristics includes queue size, scheduling policy, and weight (denoted as X). q The embedding layer transforms features into feature vectors, which can be represented 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 These represent the embedding functions for data flow features, queue features, and link features, respectively.

[0049] Step 3.1.2: By means of... 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 the data stream node characteristics from the link node and queue node characteristics;

[0052] Choosing edge types 0 and 1 in the heterogeneous graph represents the characteristics of link and queue nodes that are related to data flow nodes. Updating these data flow node characteristics using a recurrent neural network (RNN) from adjacent node messages can be represented as m.l，q , h f l ||h q ], h f ), wherein [h l ||h q ] represents concatenating the link and queue features of the path in the feature dimension and as an input sequence; m q,l represents the sequence feature representation of the data flow path through the queue and link; the input h f represents the initial hidden state in the RNN, and a new round of path state representation is obtained through the update of the loop gating unit (GRU) in the RNN, which is used for subsequent readout function performance index prediction; the GRU includes a message function and an update function, which can be represented as follows:

[0053]

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

[0055] Step 3.1.2.2: updating the queue node feature from the data flow node feature;

[0056] The heterogeneous graph edge type is selected as 0, that is, the link and queue node features related to the data flow node, and the recurrent neural network (RNN) is used to update the data flow node feature from the adjacent node message, and the GRU unit in the RNN is the same as that in step 3.1.2.1;

[0057] Step 3.1.2.3: updating the link node feature from the queue node feature;

[0058] The heterogeneous graph edge type is selected as 2, that is, the queue node feature related to the link node, and the recurrent neural network (RNN) is used to update the link node feature from the adjacent node message, and the GRU unit in the RNN is the same as that in step 3.1.2.1;

[0059] Step 3.1.3: updating the node feature representation by the node feature LSTM;

[0060] ​The node feature LTSM (v-LSTM) is used to model the feature change of the node in the time scale. The queue node is relatively stable in the satellite network, while the inter-satellite link is dynamically connected and disconnected between each time step due to the movement of the satellite. Therefore, only the queue node feature vector is used as the input of the v-LSTM for iterative update, which can be represented as follows:

[0061]

[0062] H i represents the matrix composed of the queue node feature vector at time step i, and the v-LSTM N is composed of the LSTM N of the number of graph nodes, and each LSTM N represents an LSTM unit with N output nodes, and the LSTM N shares parameters between them. In each LSTM input and output node, the calculation gate including the forgetting gate, the input gate, the output gate, and the like is included, and the transmission of the timing information is completed between the nodes by updating the cell state vector and the hidden state vector.

[0063] Step 3.2 predicts the network performance index;

[0064] After the feature vector of each node is iteratively updated by the above-mentioned module, in the readout function, the final representation of the data flow node is performed by the graph attention module, so that each data flow node aggregates information from the state updated adjacent node state. The function of the graph attention module can be represented as follows:

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

[0066]

[0067] where [Wh i || Wh j ] represents the splicing operation, h i and h j represent the feature vectors of adjacent two nodes i, j, W is a learnable parameter matrix, a is a learnable parameter vector, and a ij represents the attention parameter between nodes i, j, and N(i) represents the adjacent node set 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 of model training, train the model until convergence;

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

[0071]

[0072] Wherein represents the true label, represents the model prediction result;

[0073] Further, in step four, by regularly obtaining network state data from the data plane of the SDN, link information includes the bandwidth of two nodes, scheduling policy (for service QoS queue), etc.; satellite node queue information includes buffer queue size, scheduling policy, etc.; data flow information includes scheduling policy type, data flow service type (ToS), data packet arrival time distribution parameter, data packet size distribution parameter, average bandwidth, average data packet number, etc. Record each time point as the time series data of the target prediction network, as the input of the trained model, and finally output the network performance prediction results of the satellite network at the latest time point, such as delay, jitter and packet loss.

[0074] In addition, it should be understood that although the present specification is described in terms of embodiments, not every embodiment contains only one independent technical solution, and the description of the specification is only for the sake of clarity, and those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can be combined appropriately to form other embodiments that those skilled in the art can understand.

Claims

1. A graph neural network based satellite time-varying network performance method, characterized in that, The method comprises the following steps: Step one, divide the low earth 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 state, and offline calculate the routing information; Step two, for the satellite network of each time slice, obtain the network state data from the data plane of the LEO satellite network based on the software defined network (SDN) architecture; for the satellite network of each time slice, convert the graph structure of the network topology mode into the graph structure with the queue-link-data flow heterogeneous node mode; through the data preprocessing, obtain the graph time series data of the satellite network; Step three, for the network state data, embed the state statistical data into the feature vector, build a network performance prediction model based on the message passing neural network (MPNN) and add the node feature LSTM mechanism and the graph attention mechanism, and use the graph time series data of the satellite network for model training; Step four, based on the network performance prediction model, output the final network performance prediction result from the real network state of the LEO satellite, including the network delay, jitter and packet loss indicators.

2. The satellite time-varying network performance prediction method based on graph neural network according to claim 1, wherein: 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 starting time of each time slice is taken as the satellite network operation state in the time slice, the satellite constellation is simulated using the satellite operation simulator STK to obtain the position data of the satellites in the constellation at the time, the connection state of the inter-satellite link (ISL) is obtained, and the routing information is calculated offline based on the routing algorithm to obtain the routing forwarding table of the satellite network.

3. The satellite time-varying network performance prediction method based on graph neural network according to claim 1, wherein: In step two, for each time slice input satellite network state data, the satellite buffer queue in the network is denoted as , the end-to-end data flow is denoted as , and the link between satellite nodes is denoted as , as a node of a heterogeneous graph. Subsequently, based on the routing table information, the link and corresponding satellite node through which the data flow runs are obtained, denoted as ; finally, the edges between the heterogeneous nodes are formed according to the relationship between the link and corresponding node through which the data flow passes, and the various data flows, links and queues of the network topology are organized into a heterogeneous graph network; according to the time slice starting time in step one, the traffic information of the data packets in the network topology in the multi-hop routing forwarding and the global end-to-end is collected to represent the data packet throughput and bit rate statistical information of the end-to-end data flow at the time, thereby forming the traffic characteristics of the heterogeneous graph network.

4. The satellite time-varying network performance prediction method based on graph neural network according to claim 1, wherein: In step three, the network performance prediction model based on graph neural network is constructed, wherein the message passing neural network is used to describe the relationship mode between nodes, first, the data flow sequence of the queue feature and the link feature is combined to obtain the time sequence representation of the data flow, second, the data flow representation is used as the queue sequence to update the queue representation, and finally the queue is used as the link sequence to update the link representation, and the feature representation of the path node is obtained through multiple iterations; the node feature LSTM is used to capture the change mode of the dynamic graph node in time; the graph attention mechanism is used to aggregate the data flow node and the adjacent node features after the node feature LSTM is updated.

5. The satellite time-varying network performance prediction method based on a graph neural network according to claim 4, characterized in that, In the process of constructing the network performance prediction model based on graph neural network in step three, the following steps are included: Step 3.1: construct the message passing neural network and the node feature LSTM module to update the node feature representation; In the message passing process, firstly, the path sequence of the queue feature and the link feature is taken as the message update sequence to update the data flow feature representation, secondly, the data flow feature is taken as the message update sequence to update the queue feature representation, and finally, the queue feature is taken as the 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 taken as the input sequence of the long short-term memory to update the queue feature vector at the next time step; Step 3.2 predicts the network performance index; After the feature vectors of the nodes are iteratively updated by the above modules, in the readout function, the final representation of the data flow node is obtained through the graph attention module, so that each data flow node aggregates information from the state updated adjacent node, and the function of the graph attention module can be expressed as the following formula: wherein , wherein denotes a concatenation operation, and denotes the feature vector of the two adjacent nodes i, j, is a learnable parameter matrix, is a learnable parameter vector, denotes the attention parameter between nodes i, j, denotes the set of neighboring 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: Constructing the loss function of the training model, using the performance matrix of each time slice network topology record As the true label of model training, the model is trained until the loss converges; The mean absolute percentage error is used as the regression loss function, abbreviated as MAPE, which can be expressed as the following formula: ， wherein denotes the true label, denotes the model prediction.

6. The satellite time-varying network performance prediction method based on a graph neural network according to claim 5, characterized in that, In step 3.1, the following is adopted: Step 3.1.1: initialization of feature representation; The feature input of a given heterogeneous node is converted into three kinds of feature representations of heterogeneous nodes, the data flow node feature is initialized as the average bandwidth and the average number of data packets, denoted as , the link node feature is the number of links, link capacity and scheduling strategy type, denoted as , and the queue node feature is the queue size, scheduling strategy and weight, denoted as ; the features are converted into feature vectors through an embedding layer, which can be expressed as follows: ； wherein , , respectively represent data stream features, Queue feature, link feature embedding function; Step 3.1.2: update the node feature representation through the message propagation mechanism; Step 3.1.3: update the node feature representation by node feature LSTM; The node feature LTSM (v-LSTM) is used to model the change of the node feature on the time scale, and the queue node is relatively stable in the satellite network, while the inter-satellite link is dynamically connected and disconnected at each time step due to the movement of the satellite. Therefore, only the queue node feature vector is used as the input of the v-LSTM for iterative update, which can be expressed as follows: ， denotes a matrix composed of queue node feature vectors at time step i, by the number of graph nodes , each of which denotes an LSTM unit with N output nodes, and share parameters between them; in each LSTM input / output node, there are calculation gates such as a forget gate, an input gate, and an output gate, and the transmission of time series information is completed between nodes by updating the cell state vector and the hidden state vector.

7. The satellite time-varying network performance prediction method based on a graph neural network according to claim 6, characterized in that, In step 3.1.2, the following is adopted: Step 3.1.2.1: update the data flow node feature from the link node and queue node feature; The edge types in the heterogeneous graph are selected as 0 and 1, that is, the link and queue node features representing the relationship with the data flow node, and the data flow node features are updated from the adjacent node messages using a recurrent neural network (RNN), which can be represented as wherein represents that the links and queues through which the path passes are spliced according to the feature dimension and used as an input sequence; represents the sequence feature representation of the data flow path passing through the queue and the link; the input represents the initial hidden state in the RNN, and a new round of path state representation is obtained through the update of a recurrent gate unit (GRU) in the RNN, which is used for subsequent readout function performance index prediction; the GRU includes a message function and an update function, which can be represented as follows: ， wherein denotes the eigenvector of the node v updated in the t-th step, denotes the eigenvector of the node w adjacent to the node v, denotes the message vector updated in the t+1-th step, denotes the message function, denotes the set of adjacent nodes of the node v, denotes the update function; Step 3.1.2.2: update the queue node feature from the data flow node feature; The heterogeneous graph edge type 0 is selected, which represents the link and queue node features related to the data flow node, and a recurrent neural network (RNN) is used to update the data flow node feature from the adjacent node message, wherein the GRU unit is the same as that in step 3.1.2.1; Step 3.1.2.3: update the link node feature from the queue node feature; The heterogeneous graph edge type 2 is selected, which represents the queue node feature related to the link node, and a recurrent neural network (RNN) is used to update the link feature from the adjacent node message, wherein the GRU unit is the same as that in step 3.1.2.

1.

8. The satellite time-varying network performance prediction method based on graph neural network according to claim 1, wherein: In step four, the network state data is obtained from the data plane of the SDN at regular intervals, the link information includes the bandwidth of the two nodes, the scheduling strategy for the service QoS queue, etc.; and the satellite node queue information includes the buffer queue size, the scheduling strategy, etc. In step 3.1, the following is adopted: The data flow information includes a scheduling strategy type, a service type (ToS) of the data flow, a data packet arrival time distribution parameter, a data packet size distribution parameter, an average bandwidth, an average data packet number, and the like; each time point is recorded as time sequence data of the target prediction network, as an input of the completed model, and finally outputs a network performance prediction result of the satellite network at the latest time point, such as delay, jitter, and packet loss.

Images