Network traffic prediction method and apparatus, and electronic device

By combining static and dynamic attention models, considering the static and dynamic characteristics of network traffic, the dynamic problem of failure to effectively predict network traffic in the prior art is solved, and the accuracy and real-time prediction are improved.

CN118972259APending Publication Date: 2024-11-15CHINA TELECOM CORP LTD TECHNOLOGY INNOVATION CENTER +1
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411022443.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-29
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The prior art fails to effectively consider dynamic characteristics in cellular network traffic prediction, resulting in low accuracy of prediction results.

Method used

The static attention model and dynamic attention model are combined, and through interactive learning of static features and dynamic features, a comprehensive feature vector is constructed for network traffic prediction, considering the two dimensions of time and space.

Benefits of technology

It improves the accuracy of network traffic prediction, can accurately analyze and predict network traffic from two dimensions of time and space, and enhances the real-time data adaptability and response speed of the prediction system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118972259A_ABST
    Figure CN118972259A_ABST
Patent Text Reader

Abstract

The invention relates to a network traffic prediction method, a network traffic prediction device and electronic equipment, which are used for solving the dynamic problem of network traffic prediction. The method comprises the following steps: constructing a static attention model and a dynamic attention model; extracting a static feature vector and a dynamic feature vector of the network traffic based on the static attention model and the dynamic attention model respectively; splicing the static feature vector and the dynamic feature vector to obtain a comprehensive feature vector; performing interactive learning on the comprehensive feature vector to obtain a target feature vector; and performing prediction processing on the target feature vector to obtain a prediction result of the network traffic. Based on the method, the accuracy of network traffic prediction can be effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a network traffic prediction method, device and electronic device. Background Art

[0002] With the continuous emergence of new communication technologies and network facilities, especially the rapid development of cellular networks for 5G and its successor technologies, the requirements for smarter, more reliable and more efficient cellular network services in the Industry 5.0 blueprint have also increased. In order to meet these increasingly stringent requirements, it has become crucial to proactively and effectively allocate cellular resources. As an integral part of the cellular network resource management system, cellular traffic forecasting faces the challenge of strict accuracy and reliability requirements. One of the most critical issues is how to improve the forecasting performance by exploring the spatiotemporal information in cellular traffic data.

[0003] The existing technology combines the characteristics of wireless channels with the graph attention mechanism network to realize network traffic prediction, but this method only considers the static characteristics (long-term stable, time-invariant data characteristics) in network traffic prediction, and does not consider the dynamic characteristics (time-varying data characteristics) in network traffic prediction. Therefore, the accuracy of network traffic prediction results is not high. Summary of the invention

[0004] The purpose of this application is to provide a network traffic prediction method, device and electronic device to achieve accurate network traffic analysis and prediction from both time and space dimensions, thereby improving the accuracy of network traffic prediction.

[0005] In a first aspect, the present application provides a network traffic prediction method, the method comprising:

[0006] Construct static attention model and dynamic attention model;

[0007] Extracting a static feature vector and a dynamic feature vector of network traffic based on the static attention model and the dynamic attention model respectively;

[0008] Concatenating the static feature vector and the dynamic feature vector to obtain a comprehensive feature vector;

[0009] Interactively learning the comprehensive feature vector to obtain a target feature vector;

[0010] The target feature vector is predicted to obtain a prediction result of the network traffic.

[0011] Static attention mechanism and dynamic attention mechanism are used to obtain static characteristics and dynamic characteristics in network traffic data respectively. By splicing and interactively learning static features and dynamic features, network traffic prediction can be accurately analyzed and predicted from both time and space dimensions, thereby improving the accuracy of the prediction.

[0012] In one possible design, the construction of the static attention model and the dynamic attention model includes: constructing the static attention model according to each node, the characteristics of each node, and the characteristics of the edges between different nodes, wherein the node is a communication device and the edge represents the communication between two nodes; constructing the dynamic attention model according to the hidden state of each node and the specific event impact function.

[0013] By constructing static attention models and dynamic attention models, we can obtain static features of network traffic that do not change over time and dynamic features that change over time, which facilitates the subsequent accurate prediction of network traffic based on static and dynamic features, thereby improving the accuracy of the prediction.

[0014] In a possible design, the static attention model is constructed according to each node, the characteristics of each node, and the characteristics of the edges between different nodes, including: constructing a feature processing function of the edge according to the characteristics of the edges between different nodes, wherein the characteristics of the edge include: physical distance between different nodes, grid population density, radiation power, transmitting antenna gain, system loss factor, wavelength, obstruction environment parameters and multipath effect; constructing a static attention scoring function using the edge feature processing function and a feature conversion matrix, wherein the feature conversion matrix is ​​used to map the characteristics of different nodes to the same space; constructing a static attention mechanism weight function based on the static attention scoring function and the characteristics of each node; and optimizing the learnable parameters in the static attention mechanism weight function to obtain the static attention model.

[0015] A static attention model is constructed, which comprehensively considers the influence of grid population density, base station physical distance, radiation power, transmitting antenna gain, system loss factor, electromagnetic wave data (wavelength), obstruction environmental parameters and duration effect on the static attention weight. Through the static attention membrane model, static features with greater impact on network traffic prediction can be obtained.

[0016] In one possible design, the dynamic attention model is constructed according to the hidden state of each node and the specific event impact function, including: constructing a hidden state update function according to the hidden state and node characteristics of each node at the current time point; constructing a dynamic attention scoring function according to the specific event impact function and the hidden state of different nodes; generating a dynamic attention mechanism weight function based on the dynamic attention scoring function and the hidden state update function; and training the dynamic attention mechanism weight function using an incremental learning mechanism to obtain the dynamic attention model.

[0017] A dynamic attention model is constructed, which comprehensively considers time series dependency, real-time data adaptability, and the model's responsiveness to specific events. Through this dynamic attention model, dynamic features that have a greater impact on network traffic prediction can be obtained.

[0018] In one possible design, the static attention mechanism weight function is specifically:

[0019] w ij =softmax(f att (X i ,X j ,E ij ))

[0020] Among them, w ij represents the attention weight of node i to node j, X i and X j represent the features of node i and node j respectively, E ij represents the feature of the edge between node i and node j. Softmax is a normalized indicator function used to ensure that the sum of the attention weights from node i to all nodes is 1. att Scoring function for static attention.

[0021] In a possible design, the static attention scoring function is specifically:

[0022]

[0023] Among them, a is a learnable weight vector, W is the feature transformation matrix used to map different node features to the same space, σ is a nonlinear activation function, is the feature processing function of the edge between node i and node j.

[0024] In a possible design, the feature processing function of the edge is specifically:

[0025]

[0026] in, is the feature processing function of the edge, d ij is the physical distance between node i and node j, is the grid population density, P rad is the radiated power, G ant is the transmit antenna gain, L sys is the system loss factor, λ is the wavelength, O env is the obstruction environment parameter, M path is a quantitative representation of the multipath effect, a, β, γ, δ, n, θ, l, k are learnable parameters used to adjust the influence of each feature, and V is a learnable weight vector used to transform the representation of the feature.

[0027] In a possible design, the hidden state update function is specifically:

[0028] H t+1 =convLSTM(H t ,X t )

[0029] Among them, H t+1 According to the hidden state H t Update the hidden state at the next time point, X t is the node feature at time point t.

[0030] In a possible design, the dynamic attention scoring function is specifically:

[0031]

[0032] in, is the dynamic attention scoring function, E(t) is the specific event influence function, which represents the influence of a specific event at time point t, a and b are learnable weight vectors, and σ is the activation function.

[0033] In a possible design, the dynamic attention scoring function is specifically:

[0034]

[0035] in, is the dynamic attention scoring function, E(t) is the specific event influence function, which represents the influence of a specific event at time point t, a and b are learnable weight vectors, and σ is the activation function.

[0036] In one possible design, the dynamic attention mechanism weight function is specifically:

[0037]

[0038] in, represents the attention weight of node i to node j at time point t, is the dynamic attention scoring function, and They are the hidden states of node i and node j at time point t, respectively, generated by the convLSTM model that can capture the characteristics of time series, and △t is the preset time interval.

[0039] In a possible design, the incremental learning mechanism is specifically:

[0040]

[0041] Among them, Θ represents the parameters of the convLSTM model, η is the learning rate, L is the loss function, Yt is the true value at time point t, and Y^t is the predicted value at time point t. is the gradient of the loss function with respect to Θ.

[0042] In a possible design, the formula for interactive learning is:

[0043]

[0044] Among them, W C is the weight matrix, b C is the bias term, is the comprehensive feature vector, is the target feature vector.

[0045] After interactive learning, the target feature vector is obtained. The target feature vector is a feature vector that is more conducive to the network traffic prediction task after feature reorganization. Using the target feature vector for subsequent network traffic prediction analysis can make more accurate judgments and predictions.

[0046] In a possible design, the formula for the prediction process is:

[0047]

[0048] Among them, W Y is the weight of the output layer, b Y is the bias term of the output layer, is the prediction result, expressed as the flow of node i at the next time point, and ReLU is a linear rectification function.

[0049] In a second aspect, the present application provides a network traffic prediction device, the device comprising:

[0050] Model building unit, building static attention model and dynamic attention model;

[0051] A feature extraction unit, extracting a static feature vector and a dynamic feature vector of network traffic based on the static attention model and the dynamic attention model respectively;

[0052] A feature concatenation unit concatenates the static feature vector and the dynamic feature vector to obtain a comprehensive feature vector;

[0053] A feature learning unit interactively learns the comprehensive feature vector to obtain a target feature vector;

[0054] The prediction processing unit performs prediction processing on the target feature vector to obtain a prediction result of the network traffic.

[0055] In one possible design, the model building unit is specifically used to: build the static attention model according to each node, the characteristics of each node, and the characteristics of the edges between different nodes, wherein the node is a communication device and the edge represents the communication between two nodes; and build the dynamic attention model according to the hidden state of each node and the specific event impact function.

[0056] In a possible design, the device is also used to: construct a feature processing function of the edges between different nodes according to the features of the edges, wherein the features of the edges include: physical distance between different nodes, grid population density, radiation power, transmitting antenna gain, system loss factor, wavelength, obstruction environment parameters and multipath effect; construct a static attention scoring function using the feature processing function of the edges and a feature conversion matrix, wherein the feature conversion matrix is ​​used to map the features of different nodes to the same space; construct a static attention mechanism weight function based on the static attention scoring function and the features of each node; optimize the learnable parameters in the static attention mechanism weight function to obtain the static attention model.

[0057] In a possible design, the device is also used to: construct a hidden state update function based on the hidden state and node characteristics of each node at the current time point; construct a dynamic attention scoring function based on the specific event impact function and the hidden state of different nodes; generate a dynamic attention mechanism weight function based on the dynamic attention scoring function and the hidden state update function; and use an incremental learning mechanism to train the dynamic attention mechanism weight function to obtain the dynamic attention model.

[0058] In one possible design, the static attention mechanism weight function is specifically:

[0059] w ij =softmax(f att (X i ,X j,E ij ))

[0060] Among them, w ij represents the attention weight of node i to node j, X i and X j represent the features of node i and node j respectively, E ij represents the feature of the edge between node i and node j. Softmax is a normalized indicator function used to ensure that the sum of the attention weights from node i to all nodes is 1. att Scoring function for static attention.

[0061] In a possible design, the static attention scoring function is specifically:

[0062]

[0063] Among them, a is a learnable weight vector, W is the feature transformation matrix used to map different node features to the same space, σ is a nonlinear activation function, is the feature processing function of the edge between node i and node j.

[0064] In a possible design, the feature processing function of the edge is specifically:

[0065]

[0066] in, is the feature processing function of the edge, d ij is the physical distance between node i and node j, is the grid population density, P rad is the radiated power, G ant is the transmit antenna gain, L sys is the system loss factor, λ is the wavelength, O env is the obstruction environment parameter, M path is a quantitative representation of the multipath effect, a, β, γ, δ, n, θ, l, k are learnable parameters used to adjust the influence of each feature, and V is a learnable weight vector used to transform the representation of the feature.

[0067] In a possible design, the hidden state update function is specifically:

[0068] H t+1 =convLSTM(H t ,X t )

[0069] Among them, H t+1 According to the hidden state H t Update the hidden state at the next time point, Xt is the node feature at time point t.

[0070] In a possible design, the dynamic attention scoring function is specifically:

[0071]

[0072] in, is the dynamic attention scoring function, E(t) is the specific event influence function, which represents the influence of a specific event at time point t, a and b are learnable weight vectors, and σ is the activation function.

[0073] In a possible design, the dynamic attention scoring function is specifically:

[0074]

[0075] in, is the dynamic attention scoring function, E(t) is the specific event influence function, which represents the influence of a specific event at time point t, a and b are learnable weight vectors, and σ is the activation function.

[0076] In one possible design, the dynamic attention mechanism weight function is specifically:

[0077]

[0078] in, represents the attention weight of node i to node j at time point t, is the dynamic attention scoring function, and They are the hidden states of node i and node j at time point t, respectively, generated by the convLSTM model that can capture the characteristics of time series, and △t is the preset time interval.

[0079] In a possible design, the incremental learning mechanism is specifically:

[0080]

[0081] Among them, Θ represents the parameters of the convLSTM model, η is the learning rate, L is the loss function, Yt is the true value at time point t, and Y^t is the predicted value at time point t. is the gradient of the loss function with respect to Θ.

[0082] In a possible design, the formula for interactive learning is:

[0083]

[0084] Among them, W Cis the weight matrix, b C is the bias term, is the comprehensive feature vector, is the target feature vector.

[0085] In a possible design, the formula for the prediction process is:

[0086]

[0087] Among them, W Y is the weight of the output layer, b Y is the bias term of the output layer, is the prediction result, expressed as the flow of node i at the next time point, and ReLU is a linear rectification function.

[0088] In a third aspect, the present application provides an electronic device, the electronic device comprising:

[0089] Memory, used to store computer programs;

[0090] The processor is used to implement the above-mentioned network traffic prediction method steps when executing the computer program stored in the memory.

[0091] In a fourth aspect, the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned network traffic prediction method are implemented.

[0092] For each aspect from the second to the fourth aspect and the technical effects that may be achieved by each aspect, please refer to the above description of the technical effects that can be achieved by the first aspect or various possible schemes in the first aspect, and no further details will be given here. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 A schematic diagram of a possible wireless communication network provided for this application;

[0094] Figure 2 A flowchart of a network traffic prediction method provided by this application;

[0095] Figure 3 A flowchart for constructing a static attention model provided for this application;

[0096] Figure 4 A schematic diagram of a network traffic prediction device provided by the present application;

[0097] Figure 5 A schematic diagram of the structure of an electronic device provided in this application. DETAILED DESCRIPTION

[0098] In order to make the purpose, technical solution and advantages of the present application more clear, the present application will be further described in detail below in conjunction with the accompanying drawings. The specific operation method in the method embodiment can also be applied to the device embodiment or the system embodiment.

[0099] In the description of this application, "multiple" is understood to be "at least two". "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent three situations: A exists alone, A and B exist at the same time, and B exists alone. A is connected to B, which can represent two situations: A is directly connected to B and A is connected to B through C. In addition, in the description of this application, words such as "first" and "second" are only used for the purpose of distinguishing descriptions, and cannot be understood as indicating or implying relative importance, nor can they be understood as indicating or implying order.

[0100] The following is a brief introduction to the design concept of the embodiments of the present application.

[0101] At present, researchers have begun to explore the application of graph neural network (GNN) methods to network traffic prediction. Compared with traditional network traffic prediction methods, GNN provides a more intuitive way to process and analyze spatial data. Therefore, GNN is regarded as the key to improving the accuracy of network traffic prediction in the future. In the development of graph neural networks, the improved graph attention mechanism (GAT) network based on the graph convolutional network (GCN) is a mainstream model, which can adjust the connection weights of different edges according to different scenarios. The existing technology combines the characteristics of wireless channels with the graph attention mechanism network to realize network traffic prediction. However, this method only considers the static characteristics of network traffic prediction, and does not consider the more complex dynamic characteristics of network traffic prediction.

[0102] To this end, this application proposes a network traffic prediction method. Based on the graph convolutional neural network, this method proposes a hierarchical attention mechanism, which consists of a static attention mechanism and a dynamic attention mechanism. The static attention mechanism is responsible for capturing the spatial relationship between base stations, while the dynamic attention mechanism is responsible for capturing the dynamic changes of the time series on this basis. It is used to improve the graph attention mechanism network, solve the dynamic problem of network traffic prediction, and improve the accuracy of network traffic prediction.

[0103] The following briefly introduces the application scenarios in which the technical solutions of the embodiments of the present application can be used. It should be noted that the application scenarios introduced below are only used to illustrate the embodiments of the present application and are not limited. In the specific implementation process, the technical solutions provided by the embodiments of the present application can be flexibly applied according to actual needs.

[0104] The solution provided in the embodiment of the present application can be applied to most wireless communication networks, such as Figure 1 As shown, a wireless communication network provided by an embodiment of the present application includes base stations and communication between base stations. If the base stations are regarded as nodes and the communication between base stations is regarded as edges, a graph can be constructed based on this, which is the graph structure data required when using graph neural networks to process wireless communication problems.

[0105] See also Figure 2 As shown, it is a flow chart of a network traffic prediction method provided by an embodiment of the present application based on a graph convolutional neural network. The specific implementation process of the method is as follows:

[0106] Step 201: construct a static attention model and a dynamic attention model;

[0107] In an embodiment of the present application, a static attention model is constructed based on each node, the characteristics of each node, and the characteristics of the edges between different nodes. Among them, the node is a communication device, which can be but not limited to a base station, and the edges between different nodes represent the communication between different communication devices. The static attention model aims to comprehensively consider the long-term stable data characteristics that do not change over time, such as network population density, base station physical distance, radiation power, transmitting antenna gain, system loss factor, electromagnetic wave data (wavelength), obstruction environment parameters, and the impact of multipath effects on static attention weights.

[0108] For details, see Figure 3 As shown, it is a flowchart of constructing a static attention model provided by an embodiment of the present application, and the specific implementation process is as follows:

[0109] Step 301: Construct a weight function of a static attention mechanism;

[0110] In the embodiment of the present application, firstly, a static attention mechanism weight function is constructed according to the static attention scoring function and the characteristics of each node, and the weight function w of the static attention mechanism is specifically defined. ij The basic formula is as follows:

[0111] w ij =softmax(f att (X i ,X j ,E ij ))

[0112] Among them, w ij represents the attention weight of node i to node j, node represents the communication device, X i and X j represent the features of node i and node j respectively, E ij represents the characteristics of the edge between node i and node j (which may include but is not limited to physical distance, radiation power, etc.), the edge represents the communication between two nodes, softmax is a normalized indicator function, which is used to ensure that the sum of the attention weights from node i to all nodes is 1, f att Scoring function for static attention.

[0113] Step 302: Construct a static attention scoring function:

[0114] The edge feature processing function and feature conversion matrix are used to construct a static attention scoring function. The formula of the static attention scoring function is defined as follows:

[0115]

[0116] Among them, a is a learnable weight vector, W is a feature transformation matrix used to map different node features to the same space, σ is a nonlinear activation function, is the feature processing function of the edge between node i and node j.

[0117] Step 303: construct edge feature processing function;

[0118] Specifically, according to the characteristics of the edges between different nodes, the feature processing function of the edge is constructed. The characteristics of the edge can be the characteristics that need to be paid attention to in network traffic prediction, such as grid population density, radiation power, transmitting antenna gain, system loss factor, wavelength, obstruction environment parameters and multipath effect. The formula for defining the feature processing function of the edge is as follows:

[0119]

[0120] Among them, d ij is the physical distance between node i and node j, is the grid population density, P rad is the radiated power, G ant is the transmit antenna gain, L sys is the system loss factor, λ is the wavelength, O env is the obstruction environment parameter, M path is a quantitative representation of the multipath effect. a, β, γ, δ, n, θ, l, k are learnable parameters used to adjust the influence of each feature. V is a learnable weight vector used to convert the above features into the representation required by the static attention mechanism.

[0121] Step 304: Optimize the learnable parameters in the weight function to obtain a static attention model.

[0122] After constructing the weight function, static attention scoring function, and edge feature processing function of the static attention mechanism, the model is trained using sample data. During the model training process, the values ​​of features such as population density, radiation power, transmitting antenna gain, system loss factor, and wavelength are fixed, and the learnable parameters a, W, V, α, β, γ, δ, η, θ, l, and k are automatically adjusted through back propagation.

[0123] Specifically, the prediction error is calculated using a loss function, which may be, but is not limited to, a mean square error loss function, and the prediction error is minimized by continuously optimizing the learnable parameters a, W, V, α, β, γ, δ, η, θ, l, k. When the prediction error reaches the set minimum value, the values ​​of each learnable parameter are determined to obtain a static attention model.

[0124] For the dynamic attention model, in an embodiment of the present application, a dynamic attention model is constructed based on three aspects: time series dependency, real-time data adaptability, and event-driven dynamic adjustment.

[0125] Specifically, first, construct the weight function of the dynamic attention mechanism Represents the attention weight of node i to node j at time point t. The dynamic attention weight function defined The basic formula is as follows:

[0126]

[0127] in, is the dynamic attention scoring function, and are the hidden states of node i and node j at time point t, respectively. The hidden states can be generated by models that can capture the characteristics of time series, such as recurrent neural networks (RNNs) or long short-term memory networks (LSTMs). △t is a preset time interval used to capture the dependency of time series. The dependency of time series means that the model can identify and utilize the impact of historical data points on future predictions. When processing time series data, it is crucial to capture this dependency.

[0128] In this application, the ConvLSTM (Convolutional Long Short-Term Memory) model is used to generate hidden states. ConvLSTM is a special LSTM network. By adding convolution operations to the input gate, forget gate, and output gate of LSTM, the network can capture dependencies in the time dimension while also extracting features in the spatial dimension. This is especially important for spatiotemporal datasets because in network traffic prediction, the traffic of base stations is not only affected by time factors, but also by spatial location and surrounding environment. Therefore, this application uses ConvLSTM to capture the dependencies of time series, and constructs a hidden state update function based on the hidden state and node features of each node at the current time point. The specific formula of the hidden state update function is as follows:

[0129] H t+1 =convLSTM(H t ,X t )

[0130] Among them, H t represents the hidden state of the node at time point t, H t+1 According to the hidden state H t Update the hidden state at the next time point, X t is the node feature at time point t.

[0131] In an embodiment of the present application, the dynamic attention model also takes into account real-time data adaptability, which focuses on the model's ability to adapt to new input data, and uses an incremental learning mechanism to improve the real-time data adaptability of the dynamic attention model. The incremental learning mechanism is a strategy that allows the model to update its knowledge base when it receives new data, rather than retraining from scratch. This mechanism is particularly important for processing real-time streaming data because it can significantly reduce the consumption of computing resources and allow the model to dynamically adapt to data changes.

[0132] In incremental learning, the update of model parameters is usually based on a loss function. By calculating the loss between the true value and the predicted value and applying back propagation, the model parameters are gradually adjusted to better reflect the characteristics and patterns of the new data. This approach ensures that the model can keep the memory of previously learned knowledge while continuously learning new information. The formula of the incremental learning mechanism constructed in this application is as follows:

[0133]

[0134] Among them, Θ represents the parameters of the convLSTM model, η is the learning rate, L is the loss function, Yt is the true value at time point t, and Y^t is the predicted value at time point t. is the gradient of the loss function with respect to Θ.

[0135] The dynamic attention model also takes into account event-driven dynamic adjustment, which focuses on the model's responsiveness to specific times. In applications such as network traffic prediction, specific events (such as holidays, large events, etc.) may cause significant changes in traffic patterns. In order to capture such non-periodic changes, the model needs to be able to identify these events and adjust its behavior accordingly. In an embodiment of the present application, an event recognition component is introduced into the dynamic attention model, which can identify the occurrence of specific events based on input data. The model can then respond to these events by adjusting the dynamic attention weights, such as increasing attention to specific time points or areas. This adjustment can be based on predefined rules or a data-driven method that learns the relationship between events and traffic patterns.

[0136] For event-driven dynamic adjustment, the embodiment of the present application constructs a specific event impact function E(t), and constructs a dynamic attention scoring function based on the specific event impact function and the hidden states of different nodes. The specific formula of the dynamic attention scoring function is as follows:

[0137]

[0138] Among them, E(t) is the specific event influence function, which represents the influence of a specific event at time point t. It can be a model learned from historical data or a rule-based indicator function. a and b are learnable weight vectors, and σ is the activation function.

[0139] By constructing the static attention model and the dynamic attention model in the above way, the static attention model can capture the static characteristics of network traffic data, which reflect long-term stable and time-independent attributes, such as geographic location, infrastructure layout, etc. The dynamic attention model can capture the dynamic characteristics of network traffic data, which reflect information that changes over time, such as traffic fluctuations, event impacts, etc. Both the static attention model and the dynamic attention model capture the key influencing factors in network traffic prediction.

[0140] Step 202: extracting a static feature vector and a dynamic feature vector of network traffic based on a static attention model and a dynamic attention model respectively;

[0141] Specifically, the network traffic is input into the constructed static attention model to obtain the static feature vector of the network traffic, where Si is used to represent the static feature vector. The network traffic is input into the constructed dynamic attention model to obtain the dynamic feature vector of the network traffic, where Si is used to represent the dynamic feature vector. represents a dynamic feature vector.

[0142] Step 203: concatenate the static feature vector and the dynamic feature vector to obtain a comprehensive feature vector;

[0143] Specifically, the feature splicing method is used to extract the static feature vector Si and the dynamic feature vector After splicing, we get the comprehensive feature vector, which is expressed as The specific formula is as follows:

[0144]

[0145] Among them, || represents the concatenation operation of vectors.

[0146] The static attention mechanism and the dynamic attention mechanism are combined through the feature splicing method to obtain a comprehensive feature vector. The comprehensive feature vector contains both the static properties of space and the dynamic changes of time, which provides the ability to analyze and predict network traffic in time and space dimensions, thereby improving the accuracy of network traffic prediction.

[0147] Step 204: interactively learning the comprehensive feature vector to obtain a target feature vector;

[0148] Interactive learning further optimizes the interaction between different features through learning based on feature splicing. The core of interactive learning is to identify and utilize the intrinsic connection between static and dynamic features, and their joint impact on the target task. In the embodiment of this application, the target task is network traffic prediction.

[0149] Specifically, by introducing one or more fully connected layers (or other types of network structures) to adjust the relationship between features in the comprehensive feature vector, the feature combination method can be dynamically adjusted according to the needs of specific tasks. The formula for interactive learning is as follows:

[0150]

[0151] Among them, W C is the weight matrix, b C is the bias term, is the comprehensive feature vector, σ is the nonlinear activation function, It is the feature representation after interactive learning, that is, the target feature vector.

[0152] After interactive learning, the target feature vector is obtained. The target feature vector is a feature vector that is more conducive to the network traffic prediction task after feature reorganization. Using the target feature vector for subsequent network traffic prediction analysis can make more accurate judgments and predictions.

[0153] Step 205: Perform prediction processing on the target feature vector to obtain a prediction result of the network traffic.

[0154] Specifically, the target feature vector is predicted and processed. The goal is to predict the traffic of the node at the next time point based on the static information and dynamic feature information contained in the target feature vector. The specific formula for the prediction process is as follows:

[0155]

[0156] Among them, W Y is the weight of the output layer, b Y is the bias term of the output layer, is the prediction result, i.e. the flow of node i at the next time point. ReLU is a linear rectification function, which is used to ensure that the output of the task is regressive.

[0157] The network traffic prediction method provided in the embodiment of the present application uses a static attention mechanism and a dynamic attention mechanism to respectively obtain static characteristics and dynamic characteristics in network traffic data, and by splicing and interactively learning static features and dynamic features, it is possible to accurately analyze and predict network traffic from two dimensions, time and space, thereby improving the accuracy of the prediction. In addition, the dynamic attention mechanism constructed in the present application can respond to new data and burst time in a timely manner, ensuring that the prediction system can be updated in real time, and improving the response speed and practicality of the prediction system.

[0158] Based on the same inventive concept, the present application also provides a network traffic prediction device for accurately analyzing and predicting network traffic in both time and space dimensions, thereby improving the accuracy of network traffic prediction. Figure 4 , the device comprises:

[0159] A model building unit 401 builds a static attention model and a dynamic attention model;

[0160] A feature extraction unit 402 extracts a static feature vector and a dynamic feature vector of network traffic based on the static attention model and the dynamic attention model respectively;

[0161] A feature concatenation unit 403 concatenates the static feature vector and the dynamic feature vector to obtain a comprehensive feature vector;

[0162] A feature learning unit 404 performs interactive learning on the comprehensive feature vector to obtain a target feature vector;

[0163] The prediction processing unit 405 performs prediction processing on the target feature vector to obtain a prediction result of the network traffic.

[0164] In one possible design, the model building unit 401 is specifically used to: build the static attention model according to each node, the characteristics of each node, and the characteristics of the edges between different nodes, wherein the node is a communication device and the edge represents the communication between two nodes; and build the dynamic attention model according to the hidden state of each node and the specific event influence function.

[0165] In a possible design, the device is also used to: construct a feature processing function of the edges between different nodes according to the features of the edges, wherein the features of the edges include: physical distance between different nodes, grid population density, radiation power, transmitting antenna gain, system loss factor, wavelength, obstruction environment parameters and multipath effect; construct a static attention scoring function using the feature processing function of the edges and a feature conversion matrix, wherein the feature conversion matrix is ​​used to map the features of different nodes to the same space; construct a static attention mechanism weight function based on the static attention scoring function and the features of each node; optimize the learnable parameters in the static attention mechanism weight function to obtain the static attention model.

[0166] In a possible design, the device is also used to: construct a hidden state update function based on the hidden state and node characteristics of each node at the current time point; construct a dynamic attention scoring function based on the specific event impact function and the hidden state of different nodes; generate a dynamic attention mechanism weight function based on the dynamic attention scoring function and the hidden state update function; and use an incremental learning mechanism to train the dynamic attention mechanism weight function to obtain the dynamic attention model.

[0167] In one possible design, the static attention mechanism weight function is specifically:

[0168] w ij =softmax(f att (X i ,X j ,E ij ))

[0169] Among them, w ij represents the attention weight of node i to node j, X i and X j represent the features of node i and node j respectively, E ij represents the feature of the edge between node i and node j. Softmax is a normalized indicator function used to ensure that the sum of the attention weights from node i to all nodes is 1. att Scoring function for static attention.

[0170] In a possible design, the static attention scoring function is specifically:

[0171]

[0172] Among them, a is a learnable weight vector, W is the feature transformation matrix used to map different node features to the same space, σ is a nonlinear activation function, is the feature processing function of the edge between node i and node j.

[0173] In a possible design, the feature processing function of the edge is specifically:

[0174]

[0175] in, is the feature processing function of the edge, d ij is the physical distance between node i and node j, is the grid population density, P rad is the radiated power, G ant is the transmit antenna gain, L sys is the system loss factor, λ is the wavelength, O env is the obstruction environment parameter, M path is a quantitative representation of the multipath effect, a, β, γ, δ, n, θ, l, k are learnable parameters used to adjust the influence of each feature, and V is a learnable weight vector used to transform the representation of the feature.

[0176] In a possible design, the hidden state update function is specifically:

[0177] H t+1 =convLSTM(H t ,X t )

[0178] Among them, H t+1 According to the hidden state H t Update the hidden state at the next time point, X t is the node feature at time point t.

[0179] In a possible design, the dynamic attention scoring function is specifically:

[0180]

[0181] in, is the dynamic attention scoring function, E(t) is the specific event influence function, which represents the influence of a specific event at time point t, a and b are learnable weight vectors, and σ is the activation function.

[0182] In a possible design, the dynamic attention scoring function is specifically:

[0183]

[0184] in, is the dynamic attention scoring function, E(t) is the specific event influence function, which represents the influence of a specific event at time point t, a and b are learnable weight vectors, and σ is the activation function.

[0185] In one possible design, the dynamic attention mechanism weight function is specifically:

[0186]

[0187] in, represents the attention weight of node i to node j at time point t, is the dynamic attention scoring function, and They are the hidden states of node i and node j at time point t, respectively, generated by the convLSTM model that can capture the characteristics of time series, and △t is the preset time interval.

[0188] In a possible design, the incremental learning mechanism is specifically:

[0189]

[0190] Among them, Θ represents the parameters of the convLSTM model, η is the learning rate, L is the loss function, Yt is the true value at time point t, and Y^t is the predicted value at time point t. is the gradient of the loss function with respect to Θ.

[0191] In a possible design, the formula for interactive learning is:

[0192]

[0193] Among them, W C is the weight matrix, b C is the bias term, is the comprehensive feature vector, is the target feature vector.

[0194] In a possible design, the formula for the prediction process is:

[0195]

[0196] Among them, W Y is the weight of the output layer, b Y is the bias term of the output layer, is the prediction result, expressed as the flow of node i at the next time point, and ReLU is a linear rectification function.

[0197] Based on the same inventive concept, an electronic device is also provided in an embodiment of the present application, and the electronic device can realize the functions of the aforementioned network traffic prediction device, referring to Figure 5 , the electronic device comprises:

[0198] At least one processor 501, and a memory 502 connected to the at least one processor 501. The specific connection medium between the processor 501 and the memory 502 is not limited in the embodiment of the present application. Figure 5 In the example, the processor 501 and the memory 502 are connected via a bus 500. The bus 500 is Figure 5 The connections between other components are shown in bold lines, and are not intended to be limiting. The bus 500 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 5 Only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus. Alternatively, the processor 501 can also be called a controller, and there is no limitation on the name.

[0199] In the embodiment of the present application, the memory 502 stores instructions that can be executed by at least one processor 501. The at least one processor 501 can execute the network traffic prediction method discussed above by executing the instructions stored in the memory 502. The processor 501 can implement Figure 4 The functions of the various units in the device shown.

[0200] Among them, the processor 501 is the control center of the device, and can use various interfaces and lines to connect the various parts of the entire control device. By running or executing instructions stored in the memory 502 and calling data stored in the memory 502, the various functions of the device and processing data, the device can be monitored as a whole.

[0201] In one possible design, the processor 501 may include one or more processing units, and the processor 501 may integrate an application processor and a modem processor, wherein the application processor mainly processes an operating system, a user interface, and application programs, and the modem processor mainly processes wireless communications. It is understandable that the modem processor may not be integrated into the processor 501. In some embodiments, the processor 501 and the memory 502 may be implemented on the same chip, and in some embodiments, they may also be implemented separately on separate chips.

[0202] The processor 501 may be a general-purpose processor, such as a central processing unit (CPU), a digital signal processor, an application-specific integrated circuit, a field programmable gate array or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component, and may implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of the present application. A general-purpose processor may be a microprocessor or any conventional processor, etc. The steps of the network traffic prediction method disclosed in the embodiments of the present application may be directly embodied as being executed by a hardware processor, or may be executed by a combination of hardware and software modules in the processor.

[0203] The memory 502 is a non-volatile computer-readable storage medium that can be used to store non-volatile software programs, non-volatile computer executable programs and modules. The memory 502 may include at least one type of storage medium, such as a flash memory, a hard disk, a multimedia card, a card-type memory, a random access memory (Random Access Memory, RAM), a static random access memory (Static Random Access Memory, SRAM), a programmable read-only memory (Programmable Read Only Memory, PROM), a read-only memory (Read Only Memory, ROM), an electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, EEPROM), a magnetic memory, a disk, an optical disk, etc. The memory 502 is any other medium that can be used to carry or store a desired program code in the form of an instruction or data structure and can be accessed by a computer, but is not limited thereto. The memory 502 in the embodiment of the present application can also be a circuit or any other device that can realize a storage function, for storing program instructions and / or data.

[0204] By designing and programming the processor 501, the code corresponding to the network traffic prediction method described in the above embodiment can be fixed into the chip, so that the chip can execute the code when running. Figure 2 The steps of the network traffic prediction method of the embodiment shown are as follows: How to design and program the processor 501 is a technology well known to those skilled in the art and will not be described in detail here.

[0205] Based on the same inventive concept, an embodiment of the present application further provides a storage medium, which stores computer instructions. When the computer instructions are executed on a computer, the computer executes the network traffic prediction method discussed above.

[0206] In some possible implementations, various aspects of the network traffic prediction method provided by the present application can also be implemented in the form of a program product, which includes program code. When the program product is run on an apparatus, the program code is used to enable the control device to execute the steps of the network traffic prediction method according to various exemplary embodiments of the present application described above in this specification.

[0207] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, devices, 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 on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0208] 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.

[0209] 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.

[0210] 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 The steps for the functions specified in one or more boxes.

[0211] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A network traffic prediction method, characterized in that: The method comprises: Construct static attention model and dynamic attention model; Extracting a static feature vector and a dynamic feature vector of network traffic based on the static attention model and the dynamic attention model respectively; Concatenating the static feature vector and the dynamic feature vector to obtain a comprehensive feature vector; Interactively learning the comprehensive feature vector to obtain a target feature vector; The target feature vector is predicted to obtain a prediction result of the network traffic.

2. The method according to claim 1, characterized in that The construction of the static attention model and the dynamic attention model includes: Constructing the static attention model according to each node, the characteristics of each node, and the characteristics of the edges between different nodes, wherein the node is a communication device and the edge represents the communication between two nodes; The dynamic attention model is constructed according to the hidden state of each node and the specific event impact function.

3. The method according to claim 2, characterized in that The static attention model is constructed according to each node, the features of each node, and the features of the edges between different nodes, including: According to the characteristics of the edges between different nodes, a characteristic processing function of the edges is constructed, wherein the characteristics of the edges include: the physical distance between different nodes, the grid population density, the radiation power, the transmitting antenna gain, the system loss factor, the wavelength, the obstruction environment parameters and the multipath effect; A static attention scoring function is constructed using the feature processing function of the edge and the feature conversion matrix, wherein the feature conversion matrix is ​​used to map features of different nodes to the same space; Based on the static attention scoring function and the characteristics of each node, a static attention mechanism weight function is constructed; The learnable parameters in the static attention mechanism weight function are optimized to obtain the static attention model.

4. The method according to claim 2, characterized in that The dynamic attention model is constructed according to the hidden state of each node and the specific event impact function, including: Construct a hidden state update function based on the hidden state and node features of each node at the current time point; Constructing a dynamic attention scoring function according to the specific event impact function and the hidden states of different nodes; Generate a dynamic attention mechanism weight function based on the dynamic attention scoring function and the hidden state update function; The dynamic attention mechanism weight function is trained using an incremental learning mechanism to obtain the dynamic attention model.

5. The method according to claim 3, characterized in that The static attention mechanism weight function is specifically: w ij =softmax(f att (X i ,X j ,E ij )) Among them, w ij represents the attention weight of node i to node j, X i and X j represent the features of node i and node j respectively, E ij represents the feature of the edge between node i and node j. Softmax is a normalized indicator function used to ensure that the sum of the attention weights from node i to all nodes is 1. att Scoring function for static attention.

6. The method according to claim 3, characterized in that The static attention scoring function is specifically: Among them, a is a learnable weight vector, W is the feature transformation matrix used to map different node features to the same space, σ is a nonlinear activation function, is the feature processing function of the edge between node i and node j.

7. The method according to claim 3, characterized in that The feature processing function of the edge is specifically: in, is the feature processing function of the edge, d ij is the physical distance between node i and node j, is the grid population density, P rad is the radiated power, G ant is the transmit antenna gain, L sys is the system loss factor, λ is the wavelength, O env is the obstruction environment parameter, M path is a quantitative representation of the multipath effect, a, β, γ, δ, n, θ, l, k are learnable parameters used to adjust the influence of each feature, and V is a learnable weight vector used to transform the representation of the feature.

8. The method according to claim 4, characterized in that The hidden state update function is specifically: H t+1 =convLSTM(H t ,X t ) Among them, H t+1 According to the hidden state H t Update the hidden state at the next time point, X t is the node feature at time point t.

9. The method according to claim 4, characterized in that The dynamic attention scoring function is specifically: in, is the dynamic attention scoring function, E(t) is the specific event influence function, which represents the influence of a specific event at time point t, a and b are learnable weight vectors, and σ is the activation function.

10. The method according to claim 4, characterized in that The dynamic attention mechanism weight function is specifically: in, represents the attention weight of node i to node j at time point t, is the dynamic attention scoring function, and are the hidden states of node i and node j at time point t, respectively, generated by the convLSTM model that can capture the characteristics of time series, and Δt is the preset time interval.

11. The method according to claim 4, characterized in that The incremental learning mechanism is specifically: Among them, Θ represents the parameters of the convLSTM model, η is the learning rate, L is the loss function, Yt is the true value at time point t, and Y^t is the predicted value at time point t. is the gradient of the loss function with respect to Θ.

12. The method according to claim 1, characterized in that The formula for the prediction process is: Among them, W Y is the weight of the output layer, b Y is the bias term of the output layer, Y i t+1 is the prediction result, expressed as the flow of node i at the next time point, and ReLU is a linear rectification function.

13. A network traffic prediction device, characterized in that: The device comprises: Model building unit, building static attention model and dynamic attention model; A feature extraction unit, extracting a static feature vector and a dynamic feature vector of network traffic based on the static attention model and the dynamic attention model respectively; A feature concatenation unit concatenates the static feature vector and the dynamic feature vector to obtain a comprehensive feature vector; A feature learning unit interactively learns the comprehensive feature vector to obtain a target feature vector; The prediction processing unit performs prediction processing on the target feature vector to obtain a prediction result of the network traffic.

14. An electronic device, characterized in that: include: Memory, used to store computer programs; A processor, configured to implement the method steps of any one of claims 1 to 12 when executing a computer program stored in the memory.

15. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method steps described in any one of claims 1 to 12 are implemented.

Citation Information

Cited By

  • Reinforcement learning quadruped robot path planning method and system based on multi-head attention mechanism

    CN120538534A