A network traffic prediction method, device and storage medium based on spatiotemporal feature fusion

Abstract

The present invention discloses a network traffic prediction method, device and storage medium based on spatiotemporal feature fusion, which belongs to the field of smart grid technology. The method includes: inputting the historical network traffic and spatial topology of each network device into a pre-built GAN model, and the generator in the GAN model performs fitting training and outputs the generated traffic; inputting the generated traffic and the corresponding real value of the network traffic into the discriminator in the GAN model, performing identification training, and outputting the corresponding identification probability; updating the training parameters in the generator and the discriminator according to the identification probability; if the training parameters are stable and converged, the GAN model outputs the predicted traffic, otherwise, continue training until the training parameters are stable and converged. This method can improve the performance of the dynamic resource scheduling and orchestration algorithm of FlexE equipment involved in the energy Internet slice intelligent management and control platform through accurate network traffic prediction.

Description

Technical Field

[0001] The present invention relates to a network traffic prediction method, device and storage medium based on spatiotemporal feature fusion, and belongs to the technical field of smart grid. Background Art

[0002] Currently, most network traffic predictions are processed and modeled as time series data. However, in reality, network devices exist in a huge and interconnected network, so it is necessary to mine the spatiotemporal correlation of network traffic prediction.

[0003] At present, the technology of configuring bandwidth according to network traffic is relatively real-time or lagging. We need to advance the bandwidth configuration time and combine it with network traffic prediction to predict the traffic in future time periods. Then, we can use the predicted network traffic to configure bandwidth, which can save time and improve work efficiency in some tasks.

[0004] GAN has achieved great success in computer vision and has become a mainstream method in image generation and image denoising. The generator in GAN is used to generate realistic data or images, so this method can also be used to generate various data. GAN also shows certain potential in time series prediction, so GAN can be used in time and space to better predict data.

[0005] FlexE schedules and allocates the available time slot resources in the FlexE Group according to the bandwidth required by each FlexE Client service flow and the distribution of the 5G granularity time slots (Slots) of each PHY corresponding to the FlexE Shim, forming a mapping of the FlexE Client service flow to one or more time slot resources. Therefore, the core issue affecting FlexE service slicing is that the scheduling and management of time slot resources need to adapt to the ever-changing service characteristics of the FlexE Client side. Fast and accurate resource scheduling and orchestration algorithms are the key to solving this core problem, and are also an important part of building an energy Internet slice intelligent management and control platform based on Flex-E technology. If the network traffic of different FlexE Clients can be accurately predicted some time in advance before the resource scheduling and orchestration algorithms work, the future traffic information of different FlexE Clients can be provided to the resource scheduling and orchestration algorithms, further improving the performance of the resource scheduling and orchestration algorithms in the slice intelligent management and control platform. Summary of the invention

[0006] The purpose of the present invention is to provide a network traffic prediction method, device and storage medium based on the fusion of spatiotemporal features, which performs network traffic prediction by combining the time information and spatial information of network traffic through a GAN model. Through accurate network traffic prediction, the performance of the dynamic resource scheduling and orchestration algorithm of FlexE devices involved in the intelligent management and control platform of energy Internet slices is improved.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] In a first aspect, the present invention provides a network traffic prediction method based on spatiotemporal feature fusion, comprising:

[0009] Input the historical network traffic and spatial topology of each network device into a pre-built GAN model, and the generator in the GAN model performs fitting training and outputs the generated traffic;

[0010] Inputting the generated traffic and the corresponding real value of the network traffic into the discriminator in the GAN model, performing identification training, and outputting the corresponding identification probability;

[0011] According to the identification probability, updating the training parameters in the generator and the discriminator;

[0012] If the training parameters are stable and converged, the GAN model outputs the predicted traffic;

[0013] Otherwise, continue training until the training parameters are stable and converged;

[0014] The identification probability represents the matching degree between the generated traffic and the corresponding real value of the network traffic; and the predicted traffic is the generated traffic.

[0015] In combination with the first aspect, further, the generator includes: a time information learning module for outputting a time information sequence, a spatial information learning module for outputting a spatial information sequence, and a spatiotemporal fusion module for splicing the time information sequence and the spatial information sequence.

[0016] In combination with the first aspect, further, the time information learning module includes:

[0017] Using a segmented aggregation operation, the time series data of the historical network traffic is compressed and aggregated exponentially;

[0018] Use the corresponding angle and / or corresponding angle difference operation in the Gram angle field to perform the corresponding angle difference or sum on the compressed and aggregated time series data, and use the cosine operation to obtain the time series matrix;

[0019] Expressing the timing matrix in the form of an image to obtain a timing feature graph;

[0020] Using different convolution kernels in the convolution operation to learn the time correlation of different time intervals in the time series feature graph to output a time information sequence;

[0021] The time series data of the historical network traffic is compressed and aggregated exponentially, and the multiple used is set by segmentation.

[0022] In combination with the first aspect, further, the spatial information learning module includes:

[0023] Using an unweighted graph to describe the spatial topology of each of the network devices, and considering each network device as a node;

[0024] Using an adjacency matrix to describe the connection relationship between each of the network devices;

[0025] A filter is established in the Fourier domain using a graph convolutional neural network, and the filter is applied to each of the nodes, the spatial information between the nodes is captured through a first-order neighborhood, and a GCN model is constructed by stacking multiple convolutional layers;

[0026] The topological relationship between the central device and its surrounding devices is obtained through the GCN model, and the spatial dependency relationship of each of the network devices is obtained by encoding, so as to output a spatial information sequence;

[0027] The expression of the unweighted graph is shown in formula (1):

[0028] G=(V,E) (1)

[0029] In formula (1), G is the set of unweighted graphs, V is the set of network devices, V = {ν 1 ,ν 2 ,…,ν N}, N is the total number of network devices, E is the set of edges;

[0030] The expression of the adjacency matrix is ​​shown in formula (2):

[0031] A∈R N×N (2)

[0032] In formula (2), A is the adjacency matrix, R is a set of real numbers, and N is the total number of network devices;

[0033] The adjacency matrix includes a 0 element for indicating that there is no connection between any two network devices, and a non-0 element for indicating that there is a connection between any two network devices;

[0034] The expression of the GCN model is shown in formula (3):

[0035]

[0036] In formula (3), H (l+1) is the output of the l+1th layer of the GCN model, H (l0 is the output of the lth layer of the GCN model, θ (l) is the parameter of the lth layer of the GCN model, A is the adjacency matrix, I N is the identity matrix, Add the self-connected adjacency matrix to each node. is the connection relationship between node i and node j, is the degree matrix, and σ(·) is the nonlinear Sigmoid function.

[0037] In combination with the first aspect, further, the time-space fusion module includes: using the ID of each of the network devices as a keyword, performing one-to-one splicing processing on the time information sequence and the space information sequence.

[0038] In combination with the first aspect, further, the discriminator includes: an LSTM model for learning temporal information and a GCN model for learning spatial information, and the LSTM model is connected in parallel with the GCN model.

[0039] In combination with the first aspect, further, the function expression of the LSTM model is shown in formula (4):

[0040]

[0041] In formula (4), [h t-1 ,x t ] is the input information matrix at time t, h t-1 is the hidden information at time t-1, x t is the original eigenvector at time t, f t is the forget gate, W f is the forget gate parameter, b f is the forget gate offset, i t is the input gate, W i is the input gate parameter, b i is the input gate offset, is the information that needs to be remembered at time t, W c is the memory parameter, b c is the memory offset, C t For overall memory information, C t-1 is the overall memory information before time t, t is the output gate, W ο is the output gate parameter, b ο is the output gate offset, h t Hide information at time t.

[0042] In a second aspect, the present invention provides a network traffic prediction device based on spatiotemporal feature fusion, comprising:

[0043] Generation module: inputs the historical network traffic and spatial topology of each network device into the pre-built GAN model, and the generator in the GAN model performs fitting training and outputs the generated traffic;

[0044] Identification module: input the generated traffic and the corresponding real value of the network traffic into the discriminator in the GAN model, perform identification training, and output the corresponding identification probability;

[0045] Update module: update the training parameters in the generator and the discriminator according to the identification probability;

[0046] Discrimination module: if the training parameters are stable and converged, the GAN model outputs the predicted flow, otherwise, the training continues until the training parameters are stable and converged;

[0047] The identification probability represents the matching degree between the generated traffic and the corresponding real value of the network traffic; and the predicted traffic is the generated traffic.

[0048] In a third aspect, the present invention provides a network traffic prediction device based on spatiotemporal feature fusion, including a processor and a storage medium;

[0049] The storage medium is used to store instructions;

[0050] The processor is used to operate according to the instructions to execute the steps of the method according to any one of the first aspects.

[0051] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the methods described in the first aspect.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] The network traffic prediction method based on spatiotemporal feature fusion provided by the present invention can not only obtain the spatiotemporal correlation of data through the time information learning module, the space information learning module, and the spatiotemporal fusion module, but also can be trained through the GNA model to obtain more accurate prediction results; the time series is converted into an image for processing, and the time information is learned by using the Gram angular field, combined with the GCN model, through the linear layer, the image technology and the graph convolution technology can be integrated to play a better role. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1It is a flow chart of a network traffic prediction method based on spatiotemporal feature fusion provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The technical solution of this patent is further described in detail below in conjunction with specific implementation methods.

[0056] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limitations on the present invention. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0057] Embodiment 1:

[0058] Figure 1 This is a flow chart of a network traffic prediction method based on spatiotemporal feature fusion provided in the first embodiment of the present invention. This flow chart only shows the logical sequence of the method in this embodiment. In other possible embodiments of the present invention, different methods may be used without conflict. Figure 1 The steps shown or described are accomplished in the order shown.

[0059] The network traffic prediction method based on spatiotemporal feature fusion provided in this embodiment can be applied to a terminal and can be executed by a network traffic prediction device based on spatiotemporal feature fusion. The device can be implemented by software and / or hardware. The device can be integrated in a terminal, for example, any tablet computer or computer device with communication function. Figure 1 The method of this embodiment specifically includes the following steps:

[0060] Step 1: Input the historical network traffic and spatial topology of each network device into the pre-built GAN model, and the generator in the GAN model performs fitting training and outputs the generated traffic;

[0061] The generator includes: a time information learning module for outputting a time information sequence, a space information learning module for outputting a space information sequence, and a space-time fusion module for splicing the time information sequence and the space information sequence.

[0062] The time information learning module includes:

[0063] Step A: Use segmented aggregation operations to compress and aggregate the time series data of historical network traffic exponentially;

[0064] Step B: Use the corresponding angle and / or corresponding angle difference operation in the Gram angle field to perform corresponding angle difference or sum on the compressed and aggregated time series data, and use the cosine operation to obtain a time series matrix;

[0065] Step C: Express the time series matrix in the form of an image to obtain a time series feature graph;

[0066] Step D: Use different convolution kernels in the convolution operation to learn the temporal correlation of different time intervals in the temporal feature map to output a temporal information sequence;

[0067] Among them, the time series data of historical network traffic is compressed and aggregated exponentially, and the multiple used is set through segmentation.

[0068] Spatial information learning modules include:

[0069] Step I: Use an unweighted graph to describe the spatial topology of each network device, and regard each network device as a node;

[0070] Step II: Use the adjacency matrix to describe the connection relationship between each network device;

[0071] Step III: Use graph convolutional neural network to build filters in Fourier domain and apply the filters to each node, capture the spatial information between nodes through the first-order neighborhood, and build the GCN model by stacking multiple convolutional layers;

[0072] Step IV: Obtain the topological relationship between the central device and its surrounding devices through the GCN model, encode it, obtain the spatial dependency relationship of each network device, and output the spatial information sequence;

[0073] The expression of the unweighted graph is shown in formula (1):

[0074] G=(V,E) (1)

[0075] In formula (1), G is the set of unweighted graphs, V is the set of network devices, V = {ν 1 ,ν 2 ,…,ν N}, N is the total number of network devices, E is the set of edges;

[0076] The expression of the adjacency matrix is ​​shown in formula (2):

[0077] A∈R N×N (2)

[0078] In formula (2), A is the adjacency matrix, R is a set of real numbers, and N is the total number of network devices;

[0079] The adjacency matrix includes a 0 element for indicating that there is no connection between any two network devices, and a non-0 element for indicating that there is a connection between any two network devices;

[0080] The expression of the GCN model is shown in formula (3):

[0081]

[0082] In formula (3), H (l+1) is the output of the l+1th layer of the GCN model, H (l) is the output of the lth layer of the GCN model, θ (l) is the parameter of the lth layer of the GCN model, A is the adjacency matrix, I N is the identity matrix, Add the self-connected adjacency matrix to each node. is the connection relationship between node i and node j, is the degree matrix, and σ(·) is the nonlinear Sigmoid function.

[0083] The spatiotemporal fusion module includes: using the ID of each network device as the keyword, splicing the time information sequence and the space information sequence one by one. First, spatiotemporal splicing is performed based on the Gram angular field and the graph convolutional neural network. The spatiotemporal splicing method is in a parallel relationship in the overall generator model as the key point of splicing; secondly, the splicing details are spliced ​​based on the principle of "few parameters, fast training, and good results". The splicing is performed on the basis of separate time and space components, so that the traffic timing information and device space information of each network device are linearly spliced ​​one by one, and finally the prediction results are obtained through linear processing. Each spliced ​​network device contains an information pair consisting of time information and space information, in which the dimensions of time information and space information are consistent. After passing through the linear layer, the information pairs output by the spatiotemporal fusion module finally output the generated traffic.

[0084] Step 2: Input the generated traffic and the corresponding real value of the network traffic into the discriminator in the GAN model, perform identification training, and output the corresponding identification probability;

[0085] The discriminator includes: an LSTM model for learning temporal information and a GCN model for learning spatial information. The LSTM model is connected in parallel with the GCN model.

[0086] The function expression of the LSTM model is shown in formula (4):

[0087]

[0088] In formula (4), [h t-1 ,x t] is the input information matrix at time t, h t-1 is the hidden information at time t-1, x t is the original eigenvector at time t, f t is the forget gate, W f is the forget gate parameter, b f is the forget gate offset, i t is the input gate, W i is the input gate parameter, b i is the input gate offset, is the information that needs to be remembered at time t, W c is the memory parameter, b c is the memory offset, C t For overall memory information, C t-1 is the overall memory information before time t, t is the output gate, W ο is the output gate parameter, b ο is the output gate offset, h t Hide information at time t.

[0089] The design of the GCN model in the discriminator refers to the GCN model in the generator.

[0090] Step 3: Update the training parameters in the generator and discriminator according to the identification probability;

[0091] Step 4: If the training parameters are stable and converged, the GAN model outputs the predicted traffic. Otherwise, continue training until the training parameters are stable and converged.

[0092] By training a generative adversarial network (GAN), the trained GAN is used to predict the network traffic of different FlexE clients. Then, the predicted network traffic results are transmitted to the resource scheduling and orchestration algorithm in the intelligent management and control platform for reference by the algorithm in the process of FlexE resource scheduling and orchestration.

[0093] like Figure 1As shown in the figure, the historical network traffic and spatial topology of many network devices are first used as the overall input of the GAN model. The two first enter the generator to fit and train the generated traffic, and then the generated traffic and the true value of the network traffic enter the discriminator for gap identification, and the identification probability is used to reflect the identification gap between the two values; at the same time, the identification probability is used to update the training parameters in the discriminator and the generator respectively. The ultimate goal of the update is to reduce the gap between the two values ​​(generated traffic and the true value of network traffic) so as to achieve the goal of the generated traffic approaching the true value of the network traffic; during the training process, when the training parameters reach a stable and convergent state, the generator outputs the predicted traffic of the model, that is, the generated traffic, and the resource scheduling and orchestration algorithms use the predicted traffic to schedule and orchestrate resources.

[0094] Among them, FlexE technology is a flexible Ethernet technology based on a high-speed Ethernet interface. It decouples the Ethernet MAC layer from the PHY layer to achieve flexible control of the interface rate to adapt to different network transmission structures.

[0095] GAN is a generative adversarial network, a deep learning model, and one of the most promising methods for unsupervised learning on complex distributions in recent years. The GAN model produces fairly good output through the mutual game learning of (at least) two modules in the framework: the generative model and the discriminative model. In the original GAN ​​theory, it is not required that G and D are both neural networks, but only that they can fit the corresponding generation and discrimination functions.

[0096] The Gram matrix formula is shown in formula (5):

[0097]

[0098] Considering the time series data of network traffic as a set of vectors V, the Gram matrix is ​​the inner product matrix of each pair of vectors from V. Each element in the matrix <ν i ,ν j > is the vector ν i and ν j The vector product between .

[0099] GCN is a graph convolutional network, which generalizes the convolution operation from traditional data (image or grid) to graph data. The key is to learn a function f by aggregating its own features X i and the neighbor's feature X j To generate node υ i , where j∈N(υ i). Graph convolutional networks play a central role in building many other complex graph neural network models, including autoencoder-based models, generative models, and spatiotemporal networks.

[0100] LSTM is a long short-term memory network (Long Short-Term Memory), which is a time recurrent neural network specially designed to solve the long-term dependency problem of general recurrent neural networks (RNN). All RNNs have a chain form of repeated neural network modules. In the standard RNN, this repeated structural module has only a very simple structure, such as a tanh layer. Due to its unique design structure, LSTM is suitable for processing and predicting important events with very long intervals and delays in time series. There are three main gates in LSTM that play an important role: forget gate: determines the unit state C at the previous moment; t-1 How many are retained to the current moment C t ; Input gate: determines the input x of the network at the current moment t How much is saved to the cell state C t ; Output gate: controls the cell state C t How many outputs are there to the current output value h of the LSTM t .

[0101] The network traffic prediction method based on spatiotemporal feature fusion adopted in this embodiment uses segmented aggregation operation to compress and aggregate the time series data of historical network traffic exponentially, which can reduce the longitudinal dimension of the traffic data, greatly reduce the pressure of model processing without damaging the traffic time series, and improve the processing speed; using the corresponding angle (GASF) and / or corresponding angle difference (GADF) operation in the Gram angular field, the compressed and aggregated time series data is subjected to corresponding angle difference or sum, and a cosine operation is performed to form a time series matrix, which can be converted into an operation "similar to" inner product in a rectangular coordinate system; the competition between the generator and the discriminator makes them opponents, but their strength is also a problem. If the generator is too strong, the discriminator will not be able to distinguish and thus the generator will not be able to learn, and if the discriminator is too strong, it will also cause such a problem, so the discriminator is selected as a GCN+LSTM model, the discriminator can also learn the spatiotemporal correlation of the data and the complexity is lower than the generator model, which is conducive to better training of the generator model.

[0102] Embodiment 2:

[0103] This embodiment provides a network traffic prediction device based on spatiotemporal feature fusion, including:

[0104] Generation module: The historical network traffic and spatial topology of each network device are input into the pre-built GAN model, and the generator in the GAN model performs fitting training and outputs the generated traffic;

[0105] Identification module: inputs the generated traffic and the corresponding real value of the network traffic into the discriminator in the GAN model, performs identification training, and outputs the corresponding identification probability;

[0106] Update module: Update the training parameters in the generator and discriminator according to the identification probability;

[0107] Discrimination module: If the training parameters are stable and converged, the GAN model outputs the predicted traffic. Otherwise, the training continues until the training parameters are stable and converged.

[0108] Among them, the identification probability represents the matching degree between the generated traffic and the corresponding real value of the network traffic; the predicted traffic is the generated traffic.

[0109] Embodiment three:

[0110] This embodiment provides a network traffic prediction device based on spatiotemporal feature fusion, including a processor and a storage medium;

[0111] The storage medium is used to store instructions;

[0112] The processor is used to operate according to the instructions to execute the steps of the method in embodiment 1.

[0113] Embodiment 4:

[0114] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the method in the first embodiment are implemented.

[0115] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0116] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0117] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0118] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0119] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A network traffic prediction method based on spatiotemporal feature fusion, characterized in that: include: Input the historical network traffic and spatial topology of each network device into a pre-built GAN model, and the generator in the GAN model performs fitting training and outputs the generated traffic; Inputting the generated traffic and the corresponding real value of the network traffic into the discriminator in the GAN model, performing identification training, and outputting the corresponding identification probability; According to the identification probability, updating the training parameters in the generator and the discriminator; If the training parameters are stable and converged, the GAN model outputs the predicted flow, otherwise, the training continues until the training parameters are stable and converged; Wherein, the identification probability represents the matching degree between the generated traffic and the corresponding real value of the network traffic; the predicted traffic is the generated traffic; The generator includes: a time information learning module for outputting a time information sequence, a space information learning module for outputting a space information sequence, and a time-space fusion module for splicing the time information sequence and the space information sequence; The time information learning module includes: Using a segmented aggregation operation, the time series data of the historical network traffic is compressed and aggregated exponentially; Use the corresponding angle and / or corresponding angle difference operation in the Gram angle field to perform the corresponding angle difference or sum on the compressed and aggregated time series data, and use the cosine operation to obtain the time series matrix; Expressing the timing matrix in the form of an image to obtain a timing feature graph; Using different convolution kernels in the convolution operation to learn the time correlation of different time intervals in the time series feature graph to output a time information sequence; The time series data of the historical network traffic is compressed and aggregated exponentially, and the multiple used is set by segmentation.

2. The network traffic prediction method based on spatiotemporal feature fusion according to claim 1 is characterized in that: The spatial information learning module includes: Using an unweighted graph to describe the spatial topology of each of the network devices, and considering each network device as a node; Using an adjacency matrix to describe the connection relationship between each of the network devices; A filter is established in the Fourier domain using a graph convolutional neural network, and the filter is applied to each of the nodes, the spatial information between the nodes is captured through a first-order neighborhood, and a GCN model is constructed by stacking multiple convolutional layers; The topological relationship between the central device and its surrounding devices is obtained through the GCN model, and the spatial dependency relationship of each of the network devices is obtained by encoding, so as to output a spatial information sequence; The expression of the unweighted graph is shown in formula (1): （1）； In formula (1), is a set of unweighted graphs, A collection of network devices. , is the total number of network devices, is the set of edges; The expression of the adjacency matrix is ​​shown in formula (2): （2）； In formula (2), is the adjacency matrix, is the set of real numbers, is the total number of network devices; The adjacency matrix includes a 0 element for indicating that there is no connection between any two network devices, and a non-0 element for indicating that there is a connection between any two network devices; The expression of the GCN model is shown in formula (3): （3）； In formula (3), For the GCN model The output of the layer, For the GCN model The output of the layer, For the GCN model The parameters of the layer, is the adjacency matrix, is the identity matrix, Add the self-connected adjacency matrix to each node. For Node With node connection relationship, is the degree matrix, is the nonlinear Sigmoid function.

3. The network traffic prediction method based on spatiotemporal feature fusion according to claim 1 is characterized in that: The time-space fusion module includes: using the ID of each of the network devices as a keyword, performing one-to-one splicing processing on the time information sequence and the space information sequence.

4. The network traffic prediction method based on spatiotemporal feature fusion according to claim 1 is characterized in that: The discriminator includes: an LSTM model for learning temporal information and a GCN model for learning spatial information, and the LSTM model is connected in parallel with the GCN model.

5. The network traffic prediction method based on spatiotemporal feature fusion according to claim 4 is characterized in that: The functional expression of the LSTM model is shown in formula (4): （4）； In formula (4), for Input information matrix at all times, for Hide information at all times. for The original feature vector at time, For the Gate of Forgetfulness, is the forget gate parameter, is the forget gate offset, is the input gate, is the input gate parameter, is the input gate offset, for Information needs to be remembered at all times. is the memory parameter, is the memory offset, To remember the information as a whole, for The overall memory information before the moment, is the output gate, is the output gate parameter, is the output gate offset, for Always hide information.

6. A network traffic prediction device based on spatiotemporal feature fusion, characterized in that: include: Generation module: inputs the historical network traffic and spatial topology of each network device into the pre-built GAN model, and the generator in the GAN model performs fitting training and outputs the generated traffic; Identification module: input the generated traffic and the corresponding real value of the network traffic into the discriminator in the GAN model, perform identification training, and output the corresponding identification probability; Update module: update the training parameters in the generator and the discriminator according to the identification probability; Discrimination module: if the training parameters are stable and converged, the GAN model outputs the predicted flow, otherwise, the training continues until the training parameters are stable and converged; Wherein, the identification probability represents the matching degree between the generated traffic and the corresponding real value of the network traffic; the predicted traffic is the generated traffic; The generator includes: a time information learning module for outputting a time information sequence, a space information learning module for outputting a space information sequence, and a time-space fusion module for splicing the time information sequence and the space information sequence; The time information learning module includes: Using a segmented aggregation operation, the time series data of the historical network traffic is compressed and aggregated exponentially; Use the corresponding angle and / or corresponding angle difference operation in the Gram angle field to perform the corresponding angle difference or sum on the compressed and aggregated time series data, and use the cosine operation to obtain the time series matrix; Expressing the timing matrix in the form of an image to obtain a timing feature graph; Using different convolution kernels in the convolution operation to learn the time correlation of different time intervals in the time series feature graph to output a time information sequence; The time series data of the historical network traffic is compressed and aggregated exponentially, and the multiple used is set by segmentation.

7. A network traffic prediction device based on spatiotemporal feature fusion, characterized in that: including processor and storage medium; The storage medium is used to store instructions; The processor is used to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method described in any one of claims 1 to 5 are implemented.

Images