Low earth orbit satellite network flow prediction method

Abstract

The invention provides a low earth orbit satellite network flow prediction method, and relates to the technical field of low earth orbit satellite communication. The method comprises the following steps: constructing an effective traffic mode in a low earth orbit satellite network scene; constructing a space matrix and a time sequence matrix for training; constructing a low earth orbit satellite network flow prediction model; training the low earth orbit satellite network traffic prediction model; and acquiring low-earth-orbit satellite network traffic in an actual scene, constructing a space matrix and a time sequence matrix, and inputting the space matrix and the time sequence matrix into the trained low-earth-orbit satellite network traffic prediction model to obtain a low-earth-orbit satellite network traffic prediction result. According to the method, the deep learning prediction model of the traffic time and space characteristics can be modeled at the same time, the time-space characteristics of the low earth orbit satellite network traffic are fully utilized, accurate prediction in a global range is realized, and key support is provided for intelligent management and control and optimization of the low earth orbit satellite network.

Description

Technical Field

[0001] The present invention relates to the technical field of low-orbit satellite communications, and in particular to a method for predicting low-orbit satellite network traffic. Background Art

[0002] With the rapid development and widespread application of low-orbit satellite networks, accurate prediction of satellite network traffic has become increasingly important. Accurate prediction of network traffic can not only help network operators with resource scheduling and capacity planning, but also provide key basis for tasks such as network optimization and anomaly detection.

[0003] However, the unique dynamic topology and complex traffic characteristics of low-orbit satellite networks pose huge challenges to network management and optimization. First, the topology of low-orbit satellite networks is highly dynamic, and the high-speed movement of satellite nodes causes network connections to constantly change. Traditional traffic prediction methods are difficult to capture this dynamic nature. Secondly, satellite network traffic exhibits complex correlations and non-stationarity in both time and space. Traffic data exhibits significant self-similarity, periodicity, and chaos. It is difficult to fully explore its inherent laws using only time series or statistical methods. In particular, the long correlation of traffic makes the correlation span in the time dimension larger, and the high-speed movement of satellite nodes causes the network topology and traffic distribution to constantly change, reflecting a strong spatial correlation. In addition, global traffic prediction needs to consider the traffic data of the entire network at the same time, which places higher requirements on the generalization ability and computational efficiency of the prediction algorithm.

[0004] Most existing traffic prediction methods are based on statistical models or shallow machine learning models, which make it difficult to fully exploit the complex spatiotemporal characteristics of traffic data. In low-orbit satellite networks, there is currently no traffic prediction solution that can simultaneously consider the spatial and temporal characteristics of satellite network traffic and predict global network traffic conditions. Specifically, existing technologies have the following limitations:

[0005] 1. Most existing traffic prediction algorithms only consider the temporal correlation of traffic and use a single time series model such as ARIMA and LSTM for modeling, ignoring the spatial characteristics of the satellite network topology. It is difficult to capture the correlation of traffic in the spatial dimension, resulting in limited prediction accuracy.

[0006] 2. Satellite network traffic data exhibits significant non-stationary and nonlinear characteristics. Traditional statistical models and shallow neural networks are difficult to describe its inherent complex time correlation. Existing deep learning methods mostly predict the traffic of a single node or link, lacking overall modeling of global traffic.

[0007] 3. The topology of low-orbit satellite networks is highly dynamic. Most existing prediction models are based on static topology or simplified dynamic models, which are difficult to adapt to the frequent topology changes in actual satellite networks, resulting in insufficient generalization ability of the prediction models.

[0008] Therefore, there is an urgent need for a deep learning model that can simultaneously capture the spatiotemporal characteristics of traffic, adapt to dynamic network topology, and achieve accurate global traffic prediction. Summary of the invention

[0009] In view of this, the present invention proposes a low-orbit satellite network traffic prediction method. This method can simultaneously model the deep learning prediction model of traffic temporal and spatial characteristics, make full use of the temporal and spatial characteristics of low-orbit satellite network traffic, achieve accurate prediction in a global scope, and provide key support for the intelligent management and optimization of low-orbit satellite networks.

[0010] In order to achieve the above object, the technical solution adopted by the present invention is:

[0011] A low-orbit satellite network traffic prediction method comprises the following steps:

[0012] Step 1: construct an effective traffic pattern in the low-orbit satellite network scenario and simulate the generation of low-orbit satellite network traffic;

[0013] Step 2: construct the simulated low-orbit satellite network traffic into a spatial matrix and a temporal matrix for training according to the network topology, and normalize all spatial matrices and all temporal matrices respectively;

[0014] Step 3: construct a low-orbit satellite network traffic prediction model, which includes a graph convolutional neural network, an LSTM network, and an Attention mechanism network. The graph convolutional neural network is used to mine the spatial topology information of network traffic from the spatial matrix, the LSTM network is used to capture the dynamic time series change characteristics of network traffic from the time series matrix, and the Attention mechanism network outputs the traffic prediction result according to the spatial topology information and the dynamic time series change characteristics.

[0015] Step 4, using the normalized spatial matrix and time series matrix in step 2 to train the low-orbit satellite network traffic prediction model constructed in step 3, to obtain a trained low-orbit satellite network traffic prediction model;

[0016] Step 5: Obtain the low-orbit satellite network traffic in the actual scenario, construct the spatial matrix and the time series matrix and perform normalization processing, and then input them into the trained low-orbit satellite network traffic prediction model to obtain the low-orbit satellite network traffic prediction results.

[0017] Furthermore, the specific method of step 1 is:

[0018] Using the open source software STK, a low-orbit polar satellite constellation is constructed based on the orbital parameters of the Iridium constellation. The constellation includes 10 polar orbits, each orbit contains 10 satellites, and a total of 100 low-orbit satellites. There is a reverse seam between orbit 0 and orbit 9, and there is no inter-satellite link. In addition, a delay is generated for each inter-satellite link according to the inter-satellite distance, and a 300M bandwidth is set for each inter-satellite link.

[0019] Import the constellation topology generated by STK simulation into the Mininet virtualized network simulation tool, select a total of 20 satellites on orbits 0 and 9 on both sides of the reverse seam as user terminals, specify the source node and the destination node to send and receive the specified pcap traffic packets to each other; construct the traffic pattern of each satellite terminal into a workday pattern, that is: 9:00-18:00 simulates the peak period, 18:00-22:00 simulates the gradually decreasing traffic, 22:00-6:00 simulates the low traffic, and 6:00-9:00 simulates the gradually increasing traffic; make the traffic transmitted in the satellite topology, so as to simulate the generation of low-orbit satellite network traffic and simulate the low-orbit satellite network load environment;

[0020] In the process of simulating the operation of the satellite network, POX is used in conjunction with Python scripts to monitor the inbound and outbound port traffic of each satellite. The simulated satellite traffic adopts the OpenFlow protocol, and all satellite nodes are sampled every 15 seconds.

[0021] Furthermore, the spatial matrix is ​​a 0, 1 matrix constructed according to the on-off relationship of the topology, and the temporal matrix is ​​composed of the flow size between nodes in the network.

[0022] Furthermore, in step 3, the graph convolutional neural network extracts spatial features through a convolution operation, and the convolution operation takes the form of spectral convolution or spatial convolution;

[0023] Given the feature matrix X and the adjacency matrix A, the graph convolutional neural network replaces the convolution operation in the convolutional neural network by performing spectral convolution operations. When performing spectral convolution, the graph convolutional neural network considers the graph nodes and their first-order neighbors to capture the spatial characteristics of the graph; in addition, the graph convolutional neural network uses a hierarchical propagation rule to superimpose two network layers. The two-layer graph convolutional neural network is expressed as:

[0024]

[0025] in, is a preprocessing step where is the adjacency matrix of the graph G with a self-connected structure, W 0 ∈R P×H is the weight matrix from the input layer to the hidden unit layer, where P is the time length, H is the number of hidden units, and W 1 ∈RH×T is the weight matrix from the hidden layer to the output layer, f(X,A)∈R N×T Represents the output of the predicted length T, and ReLU is a nonlinear activation function;

[0026] The graph convolutional neural network encodes the topological structure and node attributes of the satellite network by determining the topological relationship between the central node and the surrounding nodes.

[0027] Furthermore, in step 3, the LSTM network uses the hidden state of the previous moment and the traffic information of the current moment as input to determine the traffic state of the current moment; the LSTM network retains the change trend of the historical traffic information while capturing the traffic information of the current moment through the gating mechanism, thereby capturing the dynamic time series change characteristics from the traffic data; the specific method is:

[0028] (311) An LSTM network is used to capture temporal dependencies, and its calculation process is as follows:

[0029] u t =s(W u *[X t ,h t-1 ]+b u )

[0030] r t =s(W r *[X t ,h t-1 ]+b r )

[0031] c t =tanh(W c [X t ,(r t *h t-1 )]+b c )

[0032] h t =u t *h t-1 +(1-u t )*c t

[0033] Among them, h t-1 represents the hidden state at time t-1, X t Indicates the flow rate at the current moment; r t is the reset gate, which is used to adjust the influence of the state at the previous moment to filter out information useful for prediction; t is the update gate, which is used to determine the amount of information transferred from the previous state to the current state; c t Indicates the memory content stored at the current moment, h tis the output state representation at the current moment, W and b represent the weight and bias parameters respectively;

[0034] When the reset gate t When the output of is close to 0, the flow condition of the previous moment is eliminated; when the reset gate r t When the output of tends to 1, the flow data at the previous moment will completely affect the next state;

[0035] (312) When processing time series prediction tasks, the maximum and minimum scaling method is used to preprocess the network traffic data. The specific method is as follows:

[0036]

[0037] Among them, x t Represents X t The raw network traffic data points in x max is the maximum value of the flow data, x min is the minimum value of the flow data, the normalized new value x' t Will ensure that all data is mapped to the range [0,1].

[0038] Furthermore, in step 3, the Attention mechanism network evaluates the importance of traffic information at different time points, adaptively assigns importance weights to different time steps, and constructs a context vector that reflects the overall evolution trend of the traffic state. The context vector is used to obtain the final output result, and the predicted future traffic matrix is ​​obtained.

[0039] Suppose we introduce a time series x i , i = 1, 2, ..., n, where n represents the length of the time series, then the Attention mechanism network works as follows:

[0040] (321) Use convolutional neural network to calculate the hidden state h at different times i , i = 1, 2, ..., n, and represent it as H = {h 1 ,h 2 ,...,h n}; When calculating the weight of each hidden state, the feature h at each moment i As input, it passes through two hidden layers and gets the corresponding output;

[0041] (322) Construct a scoring function that uses a multilayer perceptron to evaluate the relative weight of each hidden state; the weight α of each feature i is calculated by the Softmax normalized exponential function, where w (1) and b (1) are the weight and bias of the first layer, w(2) and b (2) are the weights and biases of the second layer respectively:

[0042] e i =w (2) (w (1) H+b (1) )+b (2)

[0043]

[0044] (333) The attention mechanism is implemented through the attention function to generate a context vector C that can reflect the global traffic dynamics t , the calculation process is as follows:

[0045]

[0046] The beneficial effects achieved by adopting the above technical solution are:

[0047] 1. The present invention realizes the low-orbit satellite network traffic prediction based on graph convolutional neural network (GCN), long short-term memory network (LSTM) and attention mechanism. It makes full use of graph convolutional neural network to extract spatial features in the satellite network topology structure, and uses long short-term memory network to capture the evolution law of traffic in the time dimension. The attention mechanism is used to adaptively allocate the importance weights of different times, which can realize accurate prediction of the global traffic of the entire satellite network.

[0048] 2. Compared with the existing technology, the present invention comprehensively considers the spatiotemporal correlation of low-orbit satellite network traffic, overcomes the shortcomings of existing algorithms that ignore spatial characteristics and have difficulty in making global predictions of large-scale networks, and can more accurately predict the changing trend of network traffic, providing an important basis for satellite network management, control, optimization and other tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 Generate model simulation graphs and traffic pattern graphs for low-orbit satellite network datasets;

[0050] Figure 2 This is a traffic prediction performance comparison table of a low-orbit satellite network traffic prediction method and several benchmark solutions;

[0051] Figure 3 This is a traffic prediction RMSE performance comparison chart of a low-orbit satellite network traffic prediction method and several benchmark solutions;

[0052] Figure 4The following is a performance comparison chart of traffic prediction accuracy between a low-orbit satellite network traffic prediction method and several benchmark schemes at different constellation scales, as well as a performance comparison chart of traffic prediction accuracy at different prediction durations. DETAILED DESCRIPTION

[0053] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] A low-orbit satellite network traffic prediction method is proposed. This method combines graph convolutional networks, long short-term memory neural networks, and attention mechanisms, and fully considers the complex characteristics of low-orbit satellite network traffic in time and space dimensions.

[0055] This method first normalizes the network traffic matrix sequence and constructs an adjacency matrix containing topological information. Then, the GCN layer is used to extract the spatial features of the traffic data to form an embedding matrix containing node features and topological structure. The embedding matrix is ​​then sent to the LSTM layer to capture the long-term temporal dependencies of the traffic data. Finally, the Attention mechanism is introduced to weight the sequence data so that the model pays more attention to the key time steps that have an important impact on the prediction target.

[0056] The specific method is as follows:

[0057] Step 1: Build an effective traffic model in the low-orbit satellite network scenario and generate an input data set for the traffic prediction solution. The specific method is:

[0058] like Figure 1 As shown in the figure, the open source software STK is used to build a low-orbit polar satellite constellation. Based on the orbital parameters of the Iridium constellation, the self-built "X" constellation has a scale of 10*10=100 low-orbit satellites, that is, the constellation contains 10 polar orbits, and each orbit contains 10 satellites. There is a reverse seam between orbit 0 and orbit 9, and there is no inter-satellite link. At the same time, a delay is generated for each inter-satellite link according to the inter-satellite distance, and a bandwidth of 300M is set for each inter-satellite link. In order to simulate the generation of low-orbit satellite network traffic, the "X" constellation topology generated by STK simulation is imported into the Mininet virtualized network simulation tool.

[0059] In order to distribute the simulated traffic on all satellite nodes as much as possible, this method selects a total of 20 satellites on orbits 0 and 9 on both sides of the reverse seam as user terminals, and specifies source and destination nodes to send and receive specified pcap traffic packets to each other. The traffic is transmitted in the satellite topology to simulate the generation of low-orbit satellite network traffic and the load environment of the low-orbit satellite network. Among them, in order to simulate the real network usage scenario of satellite terminal users, this method constructs the traffic pattern of each satellite terminal into a weekday pattern, that is: the peak period is simulated during the day (9:00-18:00), the gradually decreasing traffic is simulated at night (18:00-22:00), the low traffic is simulated late at night (22:00-6:00), and the gradually increasing traffic is simulated in the morning (6:00-9:00);

[0060] POX is used with Python scripts to monitor the inbound and outbound traffic of each satellite. The simulated satellite traffic uses the OpenFlow protocol, and all satellite nodes are sampled every 15 seconds.

[0061] Step 2: construct the simulated low-orbit satellite network traffic into a spatial matrix and a time matrix for training, and normalize all spatial matrices and all time matrices respectively. The spatial matrix is ​​a 0, 1 matrix constructed according to the topological on-off relationship, and the time matrix is ​​composed of the traffic size between nodes in the network.

[0062] Step 3: Build a low-orbit satellite network traffic prediction model, which includes a graph convolutional neural network, an LSTM network, and an Attention mechanism network. The graph convolutional neural network is used to fully mine the spatial topology information of the spatial matrix and deeply analyze the spatial distribution law of network traffic. The specific method is:

[0063] Graph convolutional neural network (GCN) extracts spatial features through convolution operation. The convolution operation can be in the form of spectral convolution or spatial convolution. This method uses spectral convolution.

[0064] Given a feature matrix X and an adjacency matrix A, a graph convolutional network can replace the convolution operation in a traditional convolutional neural network (CNN) by performing a spectral convolution operation. When performing spectral convolution, GCN considers graph nodes and their first-order neighbors to capture the spatial characteristics of the graph. In addition, GCN uses a hierarchical propagation rule to superimpose two network layers. The two-layer GCN model can be expressed as:

[0065]

[0066] in, is a preprocessing step where is the adjacency matrix of the graph G with a self-connected structure, where I N is the identity matrix, A is the adjacency matrix, is the degree matrix. W 0 ∈R P×H is the weight matrix from the input layer to the hidden unit layer, where P is the time length and H is the number of hidden units. 1 ∈R H×T is the weight matrix from the hidden layer to the output layer. f(X,A)∈R N×T ReLU is a common nonlinear activation function. GCN can simultaneously encode the topological structure and node attributes of the satellite network by determining the topological relationship between the central node and the surrounding nodes. On this basis, spatial dependencies can be captured.

[0067] The LSTM network is used to capture dynamic time series change characteristics from the time series matrix and model the change trend of traffic data. The specific method is:

[0068] (311) captures the timing dependency, which is calculated as follows:

[0069] u t =s(W u *[X t ,h t-1 ]+b u )

[0070] r t =s(W r *[X t ,h t-1 ]+b r )

[0071] c t =tanh(W c [X t ,(r t *h t-1 )]+b c )

[0072] h t =u t *h t-1 +(1-u t )*c t

[0073] Among them, h t-1 represents the hidden state at time t-1, X t represents the flow rate at the current moment, and r tIt is a reset gate, whose function is to adjust the influence of the state at the previous moment to filter out the information that is crucial to the prediction. By resetting the gate, irrelevant previous state data can be effectively excluded. When the output of the reset gate is close to 0, the flow condition at the previous moment is almost excluded; if the output is close to 1, the flow data at the previous moment will completely affect the next state. t is the update gate, which determines the amount of information transferred from the previous state to the current state. In addition, c t represents the memory content stored at the current moment, and h t is the output state representation at the current moment. W and b represent the weight and bias parameters respectively.

[0074] (312) When dealing with time series prediction tasks, in order to alleviate the gradient vanishing problem and accelerate the convergence of the neural network, this method uses the min-max normalization technology to preprocess the network traffic data. The normalization steps are as follows:

[0075]

[0076] Among them, x t Represents X t The raw network traffic data points in x max is the maximum value of the flow data, x min is the minimum value of the flow data. The normalized new value x' t Will ensure that all data is mapped to the range [0,1].

[0077] LSTM uses the hidden state of the previous moment and the traffic information of the current moment as input to determine the current traffic state. Due to the gating mechanism, LSTM can retain the changing trend of historical traffic information when capturing the current traffic information. Therefore, the model can capture dynamic time series change characteristics from traffic data.

[0078] The Attention mechanism network is used to enhance the model's prediction capability. It combines graph convolutional networks and long short-term memory neural networks, and uses the attention mechanism to perform global feature learning to fully capture the complex spatial and temporal features of network routing nodes and network traffic.

[0079] The input of the attention mechanism network includes: the network topology connection output by the graph convolutional network, and the historical traffic matrix time series output by the long short-term memory neural network. By weighting the sequence data, the attention mechanism network can adaptively assign the importance weights of different time steps, so that the model pays more attention to the key time steps that have an important impact on the prediction target, and finally outputs the traffic matrix prediction results.

[0080] Specifically, the Attention mechanism network evaluates the importance of traffic information at different time points and constructs a context vector that can reflect the overall evolution trend of the traffic state to support subsequent traffic prediction tasks.

[0081] Suppose we introduce a time series x i (i=1,2,...,n), where n represents the length of the time series. The following is the working process of the Attention mechanism network:

[0082] (321) First, a convolutional neural network is used to calculate the hidden state h at different times. i (i=1,2,...,n), and represent it as H={h 1 ,h 2 ,...,h n Specifically, when calculating the weight of each hidden state, the feature h at each moment is i As input, through two hidden layers, the corresponding output can be obtained.

[0083] (322) Then, a scoring function is constructed using a multilayer perceptron to evaluate the relative weight of each hidden state. The weight α of each feature i is calculated by the Softmax normalized exponential function, where w (1) and b (1) are the weights and biases of the first layer, and w (2) and b (2) are the weights and biases of the second layer respectively.

[0084] e i =w (2) (w (1) H+b (1) )+b (2)

[0085]

[0086] (323) Design the attention function. By implementing the attention mechanism, a context vector C is generated that can reflect the global traffic dynamics. t The calculation process is as follows:

[0087]

[0088] By weighting the sequence data, the Attention mechanism can adaptively assign importance weights to different time steps, so that the model pays more attention to the key time steps that have an important impact on the prediction target. Finally, the context vector is used to obtain the final output result, and the model outputs the predicted future traffic matrix. Figure 2, 3 As shown in Figure 4, the low-orbit satellite network traffic prediction model realizes the effective prediction of low-orbit satellite network traffic by fully extracting the spatial topological characteristics and temporal dependence characteristics of network traffic, and improves the generalization ability and robustness of the global hybrid model.

[0089] In summary, the present invention achieves effective prediction of low-orbit satellite network traffic and improves the generalization and robustness of the global hybrid model. The present invention can simultaneously model the deep learning prediction model of traffic temporal and spatial characteristics, make full use of the temporal and spatial characteristics of low-orbit satellite network traffic, achieve accurate prediction in a global range, and provide key support for the intelligent management and optimization of low-orbit satellite networks.

Claims

1. A low-orbit satellite network traffic prediction method, characterized in that: The following steps are involved: Step 1: construct an effective traffic pattern in the low-orbit satellite network scenario and simulate the generation of low-orbit satellite network traffic; Step 2: construct the simulated low-orbit satellite network traffic into a spatial matrix and a temporal matrix for training according to the network topology, and normalize all spatial matrices and all temporal matrices respectively; Step 3: construct a low-orbit satellite network traffic prediction model, which includes a graph convolutional neural network, an LSTM network, and an Attention mechanism network. The graph convolutional neural network is used to mine the spatial topology information of network traffic from the spatial matrix, the LSTM network is used to capture the dynamic time series change characteristics of network traffic from the time series matrix, and the Attention mechanism network outputs the traffic prediction result according to the spatial topology information and the dynamic time series change characteristics. Step 4, using the normalized spatial matrix and time series matrix in step 2 to train the low-orbit satellite network traffic prediction model constructed in step 3, to obtain a trained low-orbit satellite network traffic prediction model; Step 5: Obtain the low-orbit satellite network traffic in the actual scenario, construct the spatial matrix and the time series matrix and perform normalization processing, and then input them into the trained low-orbit satellite network traffic prediction model to obtain the low-orbit satellite network traffic prediction results.

2. A low-orbit satellite network traffic prediction method according to claim 1, characterized in that: The specific method of step 1 is: Using the open source software STK, a low-orbit polar satellite constellation is constructed based on the orbital parameters of the Iridium constellation. The constellation includes 10 polar orbits, each orbit contains 10 satellites, and a total of 100 low-orbit satellites. There is a reverse seam between orbit 0 and orbit 9, and there is no inter-satellite link. In addition, a delay is generated for each inter-satellite link according to the inter-satellite distance, and a 300M bandwidth is set for each inter-satellite link. Import the constellation topology generated by STK simulation into the Mininet virtualized network simulation tool, select a total of 20 satellites on orbits 0 and 9 on both sides of the reverse seam as user terminals, specify the source node and the destination node to send and receive the specified pcap traffic packets to each other; construct the traffic pattern of each satellite terminal into a workday pattern, that is: 9:00-18:00 simulates the peak period, 18:00-22:00 simulates the gradually decreasing traffic, 22:00-6:00 simulates the low traffic, and 6:00-9:00 simulates the gradually increasing traffic; make the traffic transmitted in the satellite topology, so as to simulate the generation of low-orbit satellite network traffic and simulate the low-orbit satellite network load environment; In the process of simulating the operation of the satellite network, POX is used in conjunction with Python scripts to monitor the inbound and outbound port traffic of each satellite. The simulated satellite traffic adopts the OpenFlow protocol, and all satellite nodes are sampled every 15 seconds.

3. A low-orbit satellite network traffic prediction method according to claim 1, characterized in that: The spatial matrix is ​​a 0, 1 matrix constructed according to the on-off relationship of the topology, and the temporal matrix is ​​composed of the flow size between nodes in the network.

4. A low-orbit satellite network traffic prediction method according to claim 1, characterized in that: In step 3, the graph convolutional neural network extracts spatial features through convolution operations, and the convolution operations are in the form of spectral convolution or spatial convolution; Given the feature matrix X and the adjacency matrix A, the graph convolutional neural network replaces the convolution operation in the convolutional neural network by performing spectral convolution operations. When performing spectral convolution, the graph convolutional neural network considers the graph nodes and their first-order neighbors to capture the spatial characteristics of the graph; in addition, the graph convolutional neural network uses a hierarchical propagation rule to superimpose two network layers. The two-layer graph convolutional neural network is expressed as: in, is a preprocessing step where is the adjacency matrix of the graph G with a self-connected structure, W0∈R P×H is the weight matrix from the input layer to the hidden unit layer, where P is the time length, H is the number of hidden units, and W1∈R H×T is the weight matrix from the hidden layer to the output layer, f(X,A)∈R N×T Represents the output of the predicted length T, and ReLU is a nonlinear activation function; The graph convolutional neural network encodes the topological structure and node attributes of the satellite network by determining the topological relationship between the central node and the surrounding nodes.

5. A low-orbit satellite network traffic prediction method according to claim 1, characterized in that: In step 3, the LSTM network uses the hidden state of the previous moment and the traffic information of the current moment as input to determine the traffic state of the current moment. The LSTM network uses a gating mechanism to retain the change trend of historical traffic information while capturing the traffic information of the current moment, thereby capturing dynamic time series change characteristics from the traffic data. The specific method is: (311) An LSTM network is used to capture temporal dependencies, and its calculation process is as follows: u t =s(W u *[X t ,h t-1 ]+b u ) r t =s(W r *[X t ,h t-1 ]+b r ) c t =tanh(W c [X t ,(r t *h t-1 )]+b c ) h t =u t *h t-1 +(1-u t )*c t Among them, h t-1 represents the hidden state at time t-1, X t Indicates the flow rate at the current moment; r t is the reset gate, which is used to adjust the influence of the state at the previous moment to filter out information useful for prediction; t is the update gate, which is used to determine the amount of information transferred from the previous state to the current state; c t Indicates the memory content stored at the current moment, h t is the output state representation at the current moment, W and b represent the weight and bias parameters respectively; When the reset gate t When the output of is close to 0, the flow condition of the previous moment is eliminated; when the reset gate r t When the output of tends to 1, the flow data at the previous moment will completely affect the next state; (312) When processing time series prediction tasks, the maximum and minimum scaling method is used to preprocess the network traffic data. The specific method is as follows: Among them, x t Represents X t The raw network traffic data points in x max is the maximum value of the flow data, x min is the minimum value of the flow data, the normalized new value x' t Will ensure that all data is mapped to the range [0,1].

6. A low-orbit satellite network traffic prediction method according to claim 1, characterized in that: In step 3, the Attention mechanism network evaluates the importance of traffic information at different time points, adaptively assigns importance weights to different time steps, and constructs a context vector that reflects the overall evolution trend of the traffic state. The final output result is obtained using the context vector, and the predicted future traffic matrix is ​​obtained. Suppose we introduce a time series x i , i = 1, 2, ..., n, where n represents the length of the time series, then the Attention mechanism network works as follows: (321) Use convolutional neural network to calculate the hidden state h at different times i , i = 1, 2, ..., n, and represent it as H = {h1, h2, ..., h n }; When calculating the weight of each hidden state, the feature h at each moment i As input, it passes through two hidden layers and gets the corresponding output; (322) Construct a scoring function using a multilayer perceptron to evaluate the relative weight of each hidden state; The weight α of each feature i is calculated by the Softmax normalized exponential function, where w (1) and b (1) are the weight and bias of the first layer, w (2) and b (2) are the weights and biases of the second layer respectively: e i =w (2) (w (1) H+b (1) )+b (2) (333) The attention mechanism is implemented through the attention function to generate a context vector C that can reflect the global traffic dynamics t , the calculation process is as follows:

Images