Dynamic deep space network traffic prediction method and system based on graph attention network

Abstract

The application discloses a dynamic deep space network traffic prediction method and system based on a graph attention network, and relates to the technical field of deep space network communication.The application comprises the following steps: receiving deep space network traffic data, dividing the network traffic data into a training set and a test set after preprocessing, taking an RNN recurrent neural network as a basic structure, constructing a deep space network traffic prediction model, and dividing the deep space network traffic prediction model into an encoder part and a decoder part.The application uses a spatial attention module and a sequence convolution module to analyze the spatial correlation and the time correlation of deep space network traffic information, fully analyzes the space-time correlation of network traffic, makes the prediction result more accurate, considers that there is also traffic correlation between unconnected nodes, introduces a graph generation module, and can obtain an adjacency matrix based on dynamic attributes according to current and historical traffic information, so that the correlation between nodes in the long term is reflected, and long-term prediction is facilitated.

Description

Technical Field

[0001] This invention relates to the field of deep space network communication technology, specifically to a dynamic deep space network traffic prediction method and system based on graph attention networks. Background Technology

[0002] Satellite networks play a vital role in various fields such as communications, aviation, maritime, and the military. In recent years, with the advancement of technology, human exploration of space has become increasingly in-depth, and artificial satellites and spacecraft are constantly venturing into deeper space, making deep space communication one of the focal points of scientific research. Compared with terrestrial communication and low-Earth orbit satellite communication, deep space communication faces significant technical challenges, including huge signal transmission delays and link discontinuities, which greatly affect the efficiency and reliability of communication.

[0003] In recent years, traffic prediction technologies have emerged in large numbers, which can be mainly divided into three categories: statistical models, traditional machine learning models, and neural network models. Statistical models rely too much on prior knowledge to obtain parameters and have a fixed model structure, making them unable to adapt to changes. This results in their inability to deeply explore the deep features of traffic information. Compared with the first two methods, neural network models have strong nonlinear fitting and adaptive learning capabilities, and can accurately predict the trend of traffic changes. However, traffic prediction algorithms based on neural networks lack comprehensive consideration of the dynamics of deep space network topology and the spatiotemporal correlation of traffic, resulting in poor prediction accuracy. Therefore, we propose a dynamic deep space network traffic prediction method and system based on graph attention networks. Summary of the Invention

[0004] The purpose of this invention is to provide a dynamic deep space network traffic prediction method based on graph attention networks to solve the problems mentioned in the background art.

[0005] According to a first aspect of the present invention, in order to achieve the above-mentioned objective, the present invention provides the following technical solution: a dynamic deep space network traffic prediction method based on graph attention networks, comprising the following steps:

[0006] Receive deep space network traffic data, and divide the network traffic data into training set and test set after preprocessing;

[0007] Based on the recurrent neural network (RNN) structure, a deep space network traffic prediction model is constructed. The deep space network traffic prediction model is divided into an encoder part and a decoder part. Both the encoder and the decoder include a graph generation module for generating a dynamic attribute adjacency matrix based on current and historical traffic information, a spatial attention module for capturing the spatial correlation of traffic data, and a sequence processing module for capturing the temporal correlation of traffic data.

[0008] The adjacency matrix and training set are input into the deep space network traffic prediction model, and an accelerated training algorithm is used to train it to obtain the trained deep space network traffic prediction model.

[0009] The test set is input into the trained deep space network traffic prediction model for testing, and the prediction results are obtained.

[0010] Furthermore, the deep space network traffic data is obtained in the following manner:

[0011] Based on the traffic monitoring process set for each network node, traffic data is periodically uploaded to the control terminal. The control terminal integrates the traffic data from each node as the dataset for the deep space network traffic prediction model, and divides the dataset into training and testing sets in a 7:3 ratio.

[0012] Furthermore, an adjacency matrix based on dynamic attributes is generated using the graph generation module, as follows:

[0013] (1) At each time step, the deep space network traffic X t The hidden state H of the previous time step t-1 They are stitched together and used as input for the graph generation module;

[0014] (2) Generate a dynamic filter DF through graph convolution operation. t Then embed the randomly initialized nodes into the DF. t The dynamic node embedding DE is obtained by multiplying point by point. The dynamic node embeddings DE1 and DE2 of the source node and the target node are obtained in this way respectively.

[0015] (3) Finally, the adjacency matrix DA based on dynamic attributes is calculated using node similarity. t .

[0016] Furthermore, the spatial attention module is implemented using a graph attention network (GAT), and the spatial attention module is used to capture the spatial correlation of deep space network traffic data, as detailed below:

[0017] (1) GAT uses the attention mechanism to replace the traditional graph convolution operation. Its input is a set of node features and its output is a new set of node features, where the node features are the traffic information of satellite nodes.

[0018] (2) In order to map the input features to a high-dimensional space, all nodes share the weight matrix W, and then use the shared attention mechanism β to calculate the attention coefficients between the nodes and their neighbors. Finally, the normalized attention coefficients m ij Numbers are fed into graph convolution;

[0019] (3) GAT contains a multi-head attention mechanism with multiple channels of convolutional neural networks. Each head is set with corresponding parameters independently. In the deep space network traffic prediction model, the K-head attention mechanism is adopted and the K-heads are aggregated by average aggregation.

[0020] Furthermore, the sequence processing module is implemented using a GRU, and it captures the temporal correlation of traffic information, as detailed below:

[0021] The GRU contains update gates and reset gates. The update gate determines how much information from the previous state needs to be retained in the current state. The larger the value of the update gate, the more historical information will be saved. The reset gate controls the extent to which information from the previous state is forgotten or ignored. The larger the value of the reset gate, the more previous state information is discarded, thus focusing on the new information of the current input.

[0022] Furthermore, the specific steps of the encoder part of the deep space network traffic prediction model are as follows:

[0023] (1) First, at each time step, the graph generation module obtains an adjacency matrix based on state attributes according to the current flow information;

[0024] (2) The adjacency matrix based on state attributes and the adjacency matrix based on satellite connection relationship are combined as a new adjacency matrix, which, together with the traffic information in the deep space network at the current moment, is used as the input of the space attention module to obtain the processed deep space network traffic information.

[0025] (3) The processed deep space network traffic information and the hidden information of historical moments are used as inputs to the time series processing module to obtain the processed feature sequence and the hidden output of the current moment.

[0026] Furthermore, the specific steps of the decoder part of the deep space network traffic prediction model are as follows:

[0027] (1) First, at each time step, the graph generation module obtains an adjacency matrix based on state attributes according to the hidden output of the previous time step;

[0028] (2) The adjacency matrix based on state attributes and the random topology are combined to form a new adjacency matrix, which, together with the flow information of the hidden output of the previous time step, is used as the input of the spatial attention module to obtain the processed sequence.

[0029] (3) The processed sequence and the hidden output of the previous time step are used together as input to the time series processing module to obtain the hidden output of the current time step and the deep space network traffic prediction result of the current time step.

[0030] Furthermore, an accelerated training algorithm is used to train the deep space network traffic prediction model, as follows:

[0031] (1) Initialize the loop count (cir) and the counter (count);

[0032] (2) When the number of loops cir is less than the set number of loops MT, execute (3)-(12);

[0033] (3) Initialize the hidden state H0 and randomly select a batch from the input to obtain...

[0034] (4) When the number of iterations is an integer multiple of the set step size and the count counter is less than the prediction time Q, the count counter is incremented by 1;

[0035] (5) Execute a loop to update H for p from 0 to P-1. p+1 ;

[0036] (6) Initialize the signals in the decoder

[0037] (7) Perform (8)-(10) on q from 0 to count-1;

[0038] (8) Update the predicted values

[0039] (9) Select a random number a;

[0040] (10) If a is less than the sampled value of the loop, then use the actual flow rate. renew Otherwise, use the predicted value. renew

[0041] (11) Update the loss value L, where the loss function is set to the masked mean absolute error;

[0042] (12) Update the model parameters and increment the counter by 1;

[0043] (13) End the loop.

[0044] According to a second aspect of the present invention, the present invention provides a dynamic deep space network traffic prediction system based on graph attention networks, for implementing the above-described dynamic deep space network traffic prediction method based on graph attention networks, comprising:

[0045] The preprocessing module receives deep space network traffic data, preprocesses the network traffic data, and divides it into training and testing sets.

[0046] The building module is used to construct a deep space network traffic prediction model based on the RNN recurrent neural network structure. The deep space network traffic prediction model is divided into an encoder part and a decoder part. Both the encoder and the decoder include a graph generation module for generating a dynamic attribute adjacency matrix based on current and historical traffic information, a spatial attention module for capturing the spatial correlation of traffic data, and a sequence processing module for capturing the temporal correlation of traffic data.

[0047] The training module is used to input the adjacency matrix and training set into the deep space network traffic prediction model, and to train it using an accelerated training algorithm to obtain the trained deep space network traffic prediction model.

[0048] The prediction output module is used to test the trained deep space network traffic prediction model by inputting the test set and obtain the prediction results.

[0049] According to a third aspect of the present invention, the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores the computer program capable of running on the processor, and when the processor loads and executes the computer program, it employs the above-described dynamic deep space network traffic prediction method based on graph attention networks.

[0050] The present invention has at least the following beneficial effects:

[0051] 1. This invention uses a spatial attention module and a sequence convolution module to analyze the spatial and temporal correlations of deep space network traffic information. The thorough analysis of the spatiotemporal correlations of network traffic makes the prediction results more accurate and is suitable for solving the problem of frequent link interruptions in deep space networks.

[0052] 2. This invention takes into account the traffic correlation between unconnected nodes and introduces a graph generation module. This module obtains an adjacency matrix based on dynamic attributes according to current and historical traffic information. This matrix reflects the correlation between nodes in the long term. Therefore, this invention can effectively deal with long-term prediction tasks.

[0053] 3. In the prediction model, the adjacency matrix is ​​set as a weighted sum of the adjacency matrix based on dynamic attributes and the adjacency matrix based on connection relationships. Therefore, the present invention can adaptively change the weighting value to adapt to different prediction tasks during the prediction process.

[0054] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating the method described in this invention;

[0056] Figure 2 This is a schematic diagram of the deep space network traffic prediction model in this invention;

[0057] Figure 3 This is a schematic diagram of the structure of the image generation module in this invention;

[0058] Figure 4 This is a schematic diagram of the spatial attention module in this invention;

[0059] Figure 5 This is a schematic diagram of the structure of a single sequence processing module in this invention;

[0060] Figure 6 This is a flowchart illustrating the accelerated training algorithm in this invention. Detailed Implementation

[0061] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0062] Significant transmission delays exist in real-world deep space networks, making real-time state-aware congestion control difficult and costly. Therefore, using a dynamic deep space network traffic prediction method based on graph attention networks to predict the traffic status of deep space networks in advance can effectively achieve congestion control, optimize traffic scheduling, and ensure efficient and stable communication.

[0063] Please see Figures 1-6 This invention provides a technical solution: a dynamic deep space network traffic prediction method based on graph attention networks, comprising the following steps:

[0064] S1. Receive deep space network traffic data, preprocess the network traffic data and divide it into training set and test set;

[0065] Deep space network traffic data is collected in the following ways:

[0066] Based on the traffic monitoring process set for each network node, traffic data is periodically uploaded to the control terminal. The control terminal integrates the traffic data from each node as the dataset for the deep space network traffic prediction model, and divides the dataset into a training set and a test set in a 7:3 ratio.

[0067] S2. Based on the recurrent neural network (RNN) structure, a deep space network traffic prediction model is constructed. This model consists of an encoder and a decoder. Both the encoder and decoder include a graph generation module for generating a dynamic attribute adjacency matrix based on current and historical traffic information, a spatial attention module for capturing the spatial correlation of traffic data, and a sequence processing module for capturing the temporal correlation of traffic data. Figure 2 As shown,;

[0068] (S21) Regarding the technical solution of this embodiment, such as Figure 3 As shown, the graph generation module generates an adjacency matrix based on dynamic attributes according to current and historical traffic information, as detailed below:

[0069] (S21.1) At each time step, the deep space network traffic X t The hidden state H of the previous time step t-1 They are stitched together and used as input for the graph generation module;

[0070] (S21.2) Generate a dynamic filter DF through graph convolution operation. t Then embed the randomly initialized nodes into the DF. t The dynamic node embedding DE is obtained by multiplying point by point. The dynamic node embeddings DE1 and DE2 of the source node and the target node are obtained in this way respectively.

[0071] (S21.3) Finally, the adjacency matrix DA based on dynamic attributes is calculated using node similarity. t ;

[0072] (S22) as Figure 4 As shown, the spatial attention module is implemented using the Graph Attention Network (GAT). The spatial attention module is used to capture the spatial correlation of deep space network traffic data, as detailed below:

[0073] (S22.1) GAT uses an attention mechanism to replace the traditional graph convolution operation. Its input is a set of node features and its output is a new set of node features, where the node features are the traffic information of satellite nodes.

[0074] (S22.2) To map the input features to a high-dimensional space, all nodes share the weight matrix W. Then, the attention coefficients between a node and its neighbors are calculated using the shared attention mechanism β. Finally, the normalized attention coefficients m ij Numbers are fed into graph convolution;

[0075] (S22.3)GAT includes a multi-head attention mechanism for convolutional neural networks with multiple channels. Each head is set with its own parameters. In the deep space network traffic prediction model, a K-head attention mechanism is used, and the K heads are aggregated by average aggregation.

[0076] (S24) The sequence processing module is implemented using GRU. The sequence processing module captures the temporal correlation of traffic information, as detailed below:

[0077] like Figure 5 As shown, an update gate r is set in the GRU. t and reset door z t These gates determine the state update of information. Specifically, the update gate determines how much information from the previous state needs to be retained in the current state. The larger the value of the update gate, the more historical information will be preserved. On the other hand, the reset gate controls the extent to which information from the previous state is forgotten or ignored. The larger the value of the reset gate, the more the model tends to discard more information from the previous state, thus focusing on the new information of the current input.

[0078] It should be noted that the encoder's operating steps are as follows:

[0079] (1) First, at each time step, the graph generation module obtains an adjacency matrix based on state attributes according to the current flow information;

[0080] (2) The adjacency matrix based on state attributes and the adjacency matrix based on satellite connection relationship are combined as a new adjacency matrix, which, together with the traffic information in the deep space network at the current moment, is used as the input of the space attention module to obtain the processed deep space network traffic information.

[0081] (3) The processed deep space network traffic information and the hidden information of historical moments are used together as input to the time series processing module to obtain the processed feature sequence and the hidden output of the current moment;

[0082] The working steps of the decoder are as follows:

[0083] (4) First, at each time step, the graph generation module obtains an adjacency matrix based on state attributes according to the hidden output of the previous time step.

[0084] (5) The adjacency matrix based on state attributes and the random topology are combined as a new adjacency matrix, which, together with the flow information of the hidden output of the previous time step, is used as the input of the spatial attention module to obtain the processed sequence.

[0085] (6) The processed sequence and the hidden output of the previous time step are used together as input to the time series processing module to obtain the hidden output of the current time step and the deep space network traffic prediction result of the current time step.

[0086] S3. Input the adjacency matrix and training set into the deep space network traffic prediction model, and train it using an accelerated training algorithm to obtain the trained deep space network traffic prediction model, such as... Figure 6 As shown, the details are as follows:

[0087] (1) Initialize the loop count (cir) and the counter (count);

[0088] (2) When the number of loops cir is less than the set number of loops MT, execute (3)-(12);

[0089] (3) Initialize the hidden state H0 and randomly select a batch from the input to obtain...

[0090] (4) When the number of iterations is an integer multiple of the set step size and the count counter is less than the prediction time Q, the count counter is incremented by 1. Figure 6 Steps 4-6;

[0091] (5) Execute a loop to update H for p from 0 to P-1. p+1 ,like Figure 6 Steps 7-9;

[0092] (6) Initialize the signals in the decoder

[0093] (7) Perform (8)-(10) on q from 0 to count-1;

[0094] (8) Figure 6 Step 12 Update Predicted Values

[0095] (9) such as Figure 6 In step 13, select a random number a;

[0096] (10) If a is less than the sampled value of the loop, then use the actual flow rate. renew Otherwise, use the predicted value. renew like Figure 6 Steps 14-18;

[0097] (11) Update the loss value L, where the loss function is set to the masked mean absolute error, such as Figure 6 Step 20;

[0098] (12) Update the model parameters and increment the counter by 1;

[0099] (13) End the loop.

[0100] To reduce the time and resource consumption during training, this invention proposes an accelerated training algorithm. First, during the forward propagation process before backpropagation, only the first few steps of the decoder are calculated, and these steps are recorded as counts. During training, the counts are continuously increased until the predicted sequence length is reached. This training method significantly reduces training time and saves a lot of memory resources in the early stages (lines 4-6). Furthermore, combining the sampling and training processes can further improve training performance (lines 14-18).

[0101] S4. Input the test set into the trained deep space network traffic prediction model to test and obtain the prediction results.

[0102] In summary, this embodiment proposes a dynamic deep space network traffic prediction method based on graph attention networks (TPDS-GAN). This method comprehensively considers the topological dynamics of deep space networks and the spatiotemporal correlation between network traffic. Considering that traffic correlation also exists between unconnected nodes, this embodiment introduces a graph generation module. This module obtains an adjacency matrix based on dynamic attributes based on current and historical traffic information. This matrix reflects the long-term correlation between nodes, thus this embodiment can effectively handle long-term prediction tasks. Furthermore, in the deep space network traffic prediction model of this embodiment, the adjacency matrix is ​​set as a weighted average of the adjacency matrix based on dynamic attributes and the adjacency matrix based on connection relationships. Therefore, the weighting value can be adaptively changed during the prediction process to adapt to different prediction tasks. Finally, this embodiment uses a spatial attention module and a sequence convolution module to analyze the spatial and temporal correlation of deep space network traffic information. A thorough analysis of the spatiotemporal correlation of network traffic makes the prediction results more accurate.

[0103] Example 2:

[0104] This embodiment provides a dynamic deep space network traffic prediction system based on graph attention networks, used to implement the aforementioned dynamic deep space network traffic prediction method based on graph attention networks, including:

[0105] The preprocessing module receives deep space network traffic data, preprocesses the network traffic data, and divides it into training and testing sets.

[0106] The building module is used to construct a deep space network traffic prediction model based on the RNN recurrent neural network structure. The deep space network traffic prediction model is divided into an encoder part and a decoder part. Both the encoder and the decoder include a graph generation module for generating a dynamic attribute adjacency matrix based on current and historical traffic information, a spatial attention module for capturing the spatial correlation of traffic data, and a sequence processing module for capturing the temporal correlation of traffic data.

[0107] The training module is used to input the adjacency matrix and training set into the deep space network traffic prediction model, and to train it using an accelerated training algorithm to obtain the trained deep space network traffic prediction model.

[0108] The prediction output module is used to test the trained deep space network traffic prediction model by inputting the test set and obtain the prediction results.

[0109] Specifically, the preprocessing module, construction module, training module, and prediction output module can be embedded into a computer processing system. The computer, based on the dynamic deep space network traffic prediction method based on graph attention networks provided above, calls each of the above modules to complete the task of predicting deep space network traffic. The preprocessing module, construction module, training module, and prediction output module can perform operations according to the specific steps given in the dynamic deep space network traffic prediction method based on graph attention networks.

[0110] It should be noted that the division of the various modules in the above system is merely a division of logical functions. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. These modules can be implemented entirely in software through processing element calls; they can be fully implemented in hardware; or some modules can be implemented by processing element calls to software, while others are implemented in hardware. For example, the preprocessing module can be a separate processing element or integrated into a chip in the aforementioned device. Alternatively, it can be stored as program code in the memory of the aforementioned device, and its functions can be called and executed by a processing element of the device. The implementation of other modules is similar. Furthermore, these modules can be fully or partially integrated together or implemented independently. The processing element mentioned here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each of the above modules can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.

[0111] For example, these modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more Digital Signal Processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). As another example, when a module is implemented using processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together to form a system-on-a-chip (SOC).

[0112] Example 3:

[0113] The present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores the computer program capable of running on the processor. When the processor loads and executes the computer program, it employs the above-mentioned dynamic deep space network traffic prediction method based on graph attention network.

[0114] It should be noted that the terminal device can be a computer device such as a desktop computer, a laptop computer, or a cloud server, and the terminal device includes, but is not limited to, a processor and a memory. For example, the terminal device may also include input / output devices, network access devices, and buses.

[0115] Furthermore, the processor can be a central processing unit (CPU). Of course, depending on the actual use, other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. can also be used. The general-purpose processor can be a microprocessor or any conventional processor, etc., and this application does not limit it in this regard.

[0116] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0117] For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances. When an element is referred to as being "assembled on," "mounted on," "fixed to," or "set on" another element, it may be directly on the other element or there may be an intermediate element present. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be an intermediate element present. The terms "vertical," "horizontal," "upper," "lower," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible embodiments.

[0118] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0119] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

Claims

1. A dynamic deep space network traffic prediction method based on graph attention networks, characterized in that, Includes the following steps: Receive deep space network traffic data, and divide the network traffic data into training set and test set after preprocessing; Based on the recurrent neural network (RNN) structure, a deep space network traffic prediction model is constructed. The deep space network traffic prediction model is divided into an encoder part and a decoder part. Both the encoder and the decoder include a graph generation module for generating a dynamic attribute adjacency matrix based on current and historical traffic information, a spatial attention module for capturing the spatial correlation of traffic data, and a sequence processing module for capturing the temporal correlation of traffic data. The adjacency matrix and training set are input into the deep space network traffic prediction model, and an accelerated training algorithm is used to train it to obtain the trained deep space network traffic prediction model. The test set is input into the trained deep space network traffic prediction model to test it and obtain the prediction results. The specific steps of the encoder part of the deep space network traffic prediction model are as follows: First, at each time step, the graph generation module obtains an adjacency matrix based on state attributes according to the current flow information; The adjacency matrix based on state attributes and the adjacency matrix based on satellite connectivity are combined to form a new adjacency matrix, which, together with the traffic information in the deep space network at the current moment, serves as the input to the space attention module to obtain the processed deep space network traffic information. The processed deep space network traffic information and the hidden information of historical moments are used together as input to the time series processing module to obtain the processed feature sequence and the hidden output of the current moment. The specific steps of the decoder part of the deep space network traffic prediction model are as follows: First, at each time step, the graph generation module obtains an adjacency matrix based on state attributes according to the hidden output of the previous time step; The adjacency matrix based on state attributes is combined with random topology to form a new adjacency matrix, which, together with the flow information of the hidden output of the previous time step, serves as the input of the spatial attention module to obtain the processed sequence. The processed sequence and the hidden output of the previous time step are used together as input to the time series processing module to obtain the hidden output of the current time step and the deep space network traffic prediction result of the current time step.

2. The dynamic deep space network traffic prediction method based on graph attention networks according to claim 1, characterized in that: The deep space network traffic data is collected in the following manner: Based on the traffic monitoring process set for each network node, traffic data is periodically uploaded to the control terminal. The control terminal integrates the traffic data from each node as the dataset for the deep space network traffic prediction model, and divides the dataset into a training set and a test set in a 7:3 ratio.

3. The dynamic deep space network traffic prediction method based on graph attention networks according to claim 2, characterized in that, The graph generation module is used to generate an adjacency matrix based on dynamic attributes, as follows: (1) At each time step, the deep space network traffic X t The hidden state H of the previous time step t-1 They are stitched together and used as input for the graph generation module; (2) Generate a dynamic filter DF through graph convolution operation. t Then embed the randomly initialized nodes into the DF. t The dynamic node embedding DE is obtained by multiplying point by point. The dynamic node embeddings DE1 and DE2 of the source node and the target node are obtained in this way respectively. (3) Finally, the adjacency matrix DA based on dynamic attributes is calculated using node similarity. t .

4. The dynamic deep space network traffic prediction method based on graph attention networks according to claim 3, characterized in that, The spatial attention module is implemented using a graph attention network (GAT). It captures the spatial correlation of deep space network traffic data, as detailed below: (1) GAT uses the attention mechanism to replace the traditional graph convolution operation. Its input is a set of node features and its output is a new set of node features, where the node features are the traffic information of satellite nodes. (2) In order to map the input features to a high-dimensional space, all nodes share the weight matrix W, and then use the shared attention mechanism β to calculate the attention coefficients between the nodes and their neighbors. Finally, the normalized attention coefficients m ij Numbers are fed into graph convolution; (3) GAT contains a multi-head attention mechanism with multiple channels of convolutional neural networks. Each head is set with corresponding parameters independently. In the deep space network traffic prediction model, the K-head attention mechanism is adopted and the K-heads are aggregated by average aggregation.

5. The dynamic deep space network traffic prediction method based on graph attention networks according to claim 4, characterized in that: The sequence processing module is implemented using a GRU and captures the temporal correlation of traffic information, as detailed below: The GRU contains update gates and reset gates. The update gate determines how much information from the previous state needs to be retained in the current state. The larger the value of the update gate, the more historical information will be saved. The reset gate controls the extent to which information from the previous state is forgotten or ignored. The larger the value of the reset gate, the more previous state information is discarded, thus focusing on the new information of the current input.

6. The dynamic deep space network traffic prediction method based on graph attention networks according to claim 5, characterized in that, An accelerated training algorithm is used to train the deep space network traffic prediction model. The specific method is as follows: (1) Initialize the loop count (cir) and the counter (count); (2) When the number of loops cir is less than the set number of loops MT, execute (3)-(12); (3) Initialize the hidden state H0 and randomly select a batch from the input to obtain... (4) When the number of iterations is an integer multiple of the set step size and the count counter is less than the prediction time Q, the count counter is incremented by 1; (5) Execute a loop to update H for p from 0 to P-1. p+1 ; (6) Initialize the signals in the decoder (7) Perform (8)-(10) on q from 0 to count-1; (8) Update the predicted values (9) Select a random number a; (10) If a is less than the sampled value of the loop, then use the actual flow rate. renew Otherwise, use the predicted value. renew (11) Update the loss value L, where the loss function is set to the masked mean absolute error; (12) Update the model parameters and increment the counter by 1; (13) End the loop.

7. A dynamic deep space network traffic prediction system based on graph attention networks, used to implement the dynamic deep space network traffic prediction method based on graph attention networks as described in any one of claims 1 to 6, characterized in that, include: The preprocessing module receives deep space network traffic data, preprocesses the network traffic data, and divides it into training and testing sets. The building module is used to construct a deep space network traffic prediction model based on the RNN recurrent neural network structure. The deep space network traffic prediction model is divided into an encoder part and a decoder part. Both the encoder and the decoder include a graph generation module for generating a dynamic attribute adjacency matrix based on current and historical traffic information, a spatial attention module for capturing the spatial correlation of traffic data, and a sequence processing module for capturing the temporal correlation of traffic data. The training module is used to input the adjacency matrix and training set into the deep space network traffic prediction model, and to train it using an accelerated training algorithm to obtain the trained deep space network traffic prediction model. The prediction output module is used to test the trained deep space network traffic prediction model by inputting the test set and obtain the prediction results.

8. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The memory stores a computer program that can run on a processor. When the processor loads and executes the computer program, it employs the dynamic deep space network traffic prediction method based on graph attention networks as described in any one of claims 1 to 6.

Images