Broadband link traffic prediction method and system based on spectral domain graph neural network

Through the method based on the spectral domain graph neural network, the problem of low accuracy and complex mode capture in broadband link traffic prediction is solved, and high-accuracy traffic prediction is achieved, which enhances the understanding and prediction ability of inter-link dependencies. .

CN118984304AInactive Publication Date: 2024-11-19CENT SOUTH UNIV

Patent Information

Application Number
CN202411264709.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-10
Publication Date
2024-11-19
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The prior art has low accuracy and short cycles in broadband link traffic prediction, making it difficult to capture the interdependence between complex traffic variation patterns and multiple broadband links.

Method used

Using a spectral domain graph neural network method, through data exception processing, clustering analysis and spectral domain graph neural network model construction, it automatically learns the potential relationships and dependencies between links, and captures the complex patterns and autocorrelations between traffic timings. feature.

Benefits of technology

It significantly improves the accuracy and reliability of traffic prediction, enhances the ability to capture complex traffic variation modes, and improves prediction accuracy and robustness.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118984304A_ABST
    Figure CN118984304A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of network traffic prediction, in particular to a broadband link traffic prediction method and system based on a spectral domain graph neural network, and the method comprises the following steps: S1, constructing a data processing and analysis module; s2, constructing a broadband link traffic prediction model based on a spectral domain graph neural network; s3, carrying out model training and prediction by using broadband link time sequence traffic data; the method has the following beneficial effects that the quality and efficiency of data processing are remarkably improved by comprehensively applying a periodic exception processing mechanism and a clustering analysis technology; certain similarity and relevance of flow changes between different broadband links are concerned, a self-attention structure is adopted, the capacity of a model in the aspect of capturing a complex flow change mode is remarkably enhanced, and prediction accuracy and reliability are further improved; according to the method, a spectral domain graph neural network modeling strategy is introduced, so that the robustness of the model to the traffic change uncertainty is remarkably enhanced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of network traffic prediction, and in particular to a method and system for predicting broadband link traffic based on a spectral temporal graph neural network (STGNN). Background Art

[0002] With the rapid development of Internet technology, the network has been deeply embedded in modern society and has become an indispensable key infrastructure to support the operation of the information society. The explosive growth in the number of network users and the widespread popularity of various innovative network applications have jointly driven the exponential growth of network traffic. This growth trend is not only reflected in the sharp increase in the total amount of network traffic, which challenges the carrying capacity of the existing network architecture, but also in the diversification and dynamic changes of traffic patterns, such as burst traffic during peak hours and specific traffic patterns brought by new applications, which greatly increase the difficulty and uncertainty of network management. Faced with such a complex and changeable traffic environment, network managers urgently need a scientific, efficient and long-term predictive network traffic method system, and based on this system, formulate forward-looking expansion plans, optimize resource allocation, ensure the stability and reliability of network services, and thus improve user experience and meet the growing needs of informatization.

[0003] The link between BRAS (Broadband Remote Access Server) and OLT (Optical Line Terminal) is a key part of the broadband network, responsible for connecting user devices (such as home routers or modems) to the Internet through the optical fiber network. The uplink traffic (users uploading data) and downlink traffic (servers sending data to users) on this link together constitute the main data flow of the network. Its traffic changes not only reflect the user activity patterns, but are also significantly affected by time periods (such as the difference between day and night, weekdays and weekends) and random events (such as network usage peaks caused by emergencies).

[0004] At present, the mainstream network traffic prediction technology covers statistical theory, traditional machine learning algorithms and cutting-edge deep learning methods. Although the existing traffic prediction technology has achieved certain results, there are still some limitations in practical applications. First, there are relatively few studies on network traffic, especially broadband link traffic prediction. Moreover, as a key component of network infrastructure, broadband links have complex and changeable traffic characteristics. Most studies have poor prediction results when dealing with non-stationary, nonlinear or mutational traffic data. Secondly, most studies focus on short-term or medium-term traffic prediction, lacking in-depth insights into long-term traffic change trends, and it is difficult to provide network managers with comprehensive information on future network load changes. At the same time, related studies often only focus on the changes in traffic on a single broadband link, ignoring the similarities and correlations between traffic changes between different broadband links caused by the commonality of user behavior patterns (such as time periodicity and daily activity patterns), the similarity of network topology structures, and the overlap of business types and service needs. Summary of the invention

[0005] The present invention relates to a broadband link traffic prediction method and system based on spectral domain graph neural network. The method aims to solve the technical problems that traditional broadband link traffic prediction methods have low accuracy, short cycle, and difficulty in effectively capturing complex traffic change patterns and the interdependence between multiple broadband link traffic.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A broadband link traffic prediction method based on spectral domain graph neural network is divided into the following steps:

[0008] S1 builds a data processing and analysis module, including a data anomaly processing module and a cluster analysis module. The input data of this module is the peak downlink flow data of broadband links between multiple BRAS devices and OLT devices, recorded at a time granularity of days, which is referred to as time series flow data downFlow; the details are as follows:

[0009] The data anomaly processing module described in S1.1 is used to perform spectrum characteristic analysis, outlier identification, and outlier replacement on each broadband link in the time series traffic data downFlow. The specific implementation steps are as follows:

[0010] S1.1.1 For each broadband link timing flow data, the Fast Fourier Transform (FFT) technology is used to analyze the spectrum characteristics and identify the main period in the timing flow data of each broadband link. This period will be used to set the sliding window of its corresponding link.

[0011] S1.1.2 Based on the identified main cycle, the detection threshold of abnormal values ​​in the broadband link time series traffic data is set according to the normal distribution assumption. Specifically, the detection threshold is set to three times the standard deviation of the data in the sliding window of the current link (i.e., the size of the main cycle), which is used as a quantitative standard to determine whether the data is abnormal. Once it is detected that the difference between the value of a data point in the current sliding window and the value of the corresponding data point in the previous and next adjacent cycles exceeds the preset detection threshold, it is determined to be abnormal data;

[0012] S1.1.3 For these abnormal data, linear interpolation method is used to repair them. The specific operation is to use the values ​​of corresponding data points in the front and back windows for linear interpolation to replace the abnormal data values, so as to ensure the continuity and accuracy of the data, and thus improve the accuracy of subsequent data analysis and prediction. The time series flow data after abnormal processing is obtained by the above method Where N is the number of broadband links, T is the number of timestamps, represents the traffic value of the nth broadband link at timestamp t. For the convenience of subsequent description, the traffic value at timestamp t is recorded as The timing flow of each link is recorded as

[0013] The cluster analysis module described in S1.2 is used for feature selection, cluster analysis, PCA dimension reduction and link traffic feature analysis. The specific implementation steps are as follows:

[0014] S1.2.1 Calculate the timing flow of each link The mean, standard deviation, rate of change, trend line slope, autocorrelation, peak average, etc. are used as the characteristic information of the link traffic;

[0015] S1.2.2 Use KMeans algorithm [MacQueen J. Some methods for classification and analysis of multivariate observations [C] / / Proceedings of the fifth Berkeley symposium on mathematical statistics and probability. 1967, 1 (14): 281-2970.] to perform cluster analysis on the characteristic information of the above link traffic;

[0016] S1.2.3 Use principal component analysis (PCA) [Hotelling H. Analysis of a complex of statistical variables into principal components [J]. Journal of educational psychology, 1933, 24 (6): 417.] to reduce the high-dimensional feature information data of link traffic to two dimensions;

[0017] S1.2.4 Draw a clustering result diagram and analyze the characteristics of link traffic in different clusters based on the clustering results.

[0018] S2 constructs a broadband link traffic prediction model based on a spectral domain graph neural network, which includes a forward normalization module, a spatial self-attention module, a graph frequency domain convolution module, a frequency domain time pattern learning module, a spatiotemporal restoration module, and a prediction module; the details are as follows:

[0019] The forward normalization module described in S2.1 is used for data standardization and affine normalization, and the input is the time series flow data after abnormal processing. The output is normalized time series traffic data The function of this module is to solve the distribution drift problem in time series traffic data (due to the statistical characteristics of time series traffic data, such as mean, variance, distribution shape, etc., which will change during the training and testing phases, this phenomenon is called the distribution drift problem). The specific implementation steps are as follows:

[0020] S2.1.1 Calculate input time series flow data The mean and standard deviation of each broadband link in the data are calculated, and data normalization is performed, that is, the input time series traffic data is subtracted from the mean of the corresponding link and then divided by the standard deviation.

[0021] S2.1.2 Multiply the standardized data by the affine weight and add the affine bias to further adjust the data distribution to achieve normalization. The initial values ​​of the affine weight and affine bias are vectors of all 1 and all 0 respectively, which are continuously learned and gradually optimized and updated during the entire model training. After the above processing, the normalized time series traffic data is obtained

[0022] The spatial self-attention module described in S2.2 includes a gated recurrent unit (GRU) layer and a self-attention structure, and the input is the normalized time series flow data The output is the adjacency matrix of the graph calculated by the self-attention mechanism where wnm Indicates the strength of association between link n and link m. The function of this module is to automatically learn and capture the potential associations and dependencies between different links. The specific implementation steps are as follows:

[0023] S2.2.1 Normalized time series traffic data It is sent to the GRU layer, which can process the input sequence timestamp by timestamp and finally obtain the last hidden state R of all input sequences.

[0024] S2.2.2 Multiply the last hidden state R of all input sequences obtained in the previous step by two different weight matrices, and the weight matrices are represented by W Q and W k Indicates that the weight matrix is ​​randomly initialized, continuously learned and updated during model training, and the multiplication results are the query vector (Query) and the key vector (Key), represented by Q and K respectively. Then, the dot product of the query vector Q and the key vector K is calculated and divided by the scaling factor The above process obtains the attention score matrix, which is then normalized by the Softmax function to obtain the normalized attention score matrix. The normalized attention score matrix is ​​regarded as the adjacency matrix of the graph.

[0025] S2.3 The graph frequency domain convolution module includes a Laplace transform module and a graph Fourier transform module, and the input is the normalized time series flow data and the adjacency matrix of the graph The output is the time series F of the relationship between traffic time series in the frequency domain after the above transformation. The function of this module is to capture the spatial dependencies in the graph structure and combine these spatial dependencies with the frequency information of the time series traffic data to learn the relationship between traffic time series. The specific implementation steps are as follows:

[0026] S2.3.1 Computing the adjacency matrix of a graph The Laplace matrix L of Among them I N is the N×N identity matrix, D is a diagonal matrix, and the diagonal elements D nn It represents the number of edges connected to node n (corresponding to link n) in the adjacency matrix W, that is, the degree of node n (corresponding to link n), and the calculation formula is All off-diagonal elements D nm The value of (n≠m) is 0. After the above calculation, the Laplace matrix L is obtained, which describes the overall structural characteristics of the graph.

[0027] S2.3.2 Perform a graph Fourier transform on the Laplace matrix L. First, perform eigendecomposition on the Laplace matrix L. The decomposition formula is: L = UΛU T , we get the eigenvector matrix U and the eigenvalue diagonal matrix Λ. The eigenvalue diagonal matrix Λ represents the frequency components of the graph, and the eigenvector matrix U is equivalent to the "basis" vector of the graph; next, the normalized time series traffic data Multiplying it with the transpose of the eigenvector matrix U, we get the time series F of the relationship between traffic time series in the frequency domain.

[0028] The frequency domain time pattern learning module described in S2.4 includes a discrete Fourier transform (DFT) module, a gated linear unit (GLU) layer, and an inverse discrete Fourier transform module (IDFT). The input is the time series F of the relationship between traffic time series in the frequency domain, and the output is the learned time pattern feature representation TF. The function of this module is to learn the repetitive pattern of each individual time series in the time series F of the relationship between traffic time series in the frequency domain and the autocorrelation features between different timestamps. The specific steps are as follows:

[0029] S2.4.1 First, discrete Fourier transform is applied to each individual time series in the time series F of the relationship between flow time series in the frequency domain to process the frequency information of each individual time series, and the frequency component of each individual time series is obtained.

[0030] S2.4.2 applies the GLU layer to learn the frequency components of each individual time series respectively, and obtains the feature representation learned in the frequency domain for each individual time series.

[0031] S2.4.3 applies the inverse discrete Fourier transform module to convert the feature representation learned in the frequency domain of each individual time series back to the time domain to obtain the feature representation of each individual time series, and then concatenates the feature representation of each individual time series to finally obtain the learned time pattern feature representation TF.

[0032] The spatiotemporal restoration module described in S2.5 is an inverse Fourier transform, which is used to restore the original state of the data and retain the learned feature representation. The input is the learned temporal pattern feature representation TF, and the output is the final result Z restored to the time domain. The learned temporal pattern feature representation TF is remapped back to the time series by applying the inverse Fourier transform to obtain the final result Z.

[0033] S2.6 The prediction module includes a fully connected layer (FC) and a reverse normalization module. The input is the final result Z obtained by the spatiotemporal restoration module, and the output is the link traffic prediction value. The function of this module is to obtain the predicted value of the traffic of each link.

[0034] The specific steps are as follows:

[0035] S2.6.1 The fully connected layer consists of two linear layers and a Leaky ReLU activation function. First, the final result Z is input into the first linear layer, which performs a linear transformation on the final result Z. Then, the features processed by the first linear layer are nonlinearly transformed using the Leaky ReLU activation function. Finally, the second linear layer is used to convert the features obtained after the nonlinear transformation into the required output dimension to obtain the predicted value.

[0036] S2.6.2 Since the forward normalization module is used to input the time series flow data The data is normalized, so the model actually learns the normalized data features. Therefore, the prediction values ​​obtained by the model are also calculated based on these normalized data features. In order to convert the prediction results back to the scale of the original data, the prediction values ​​obtained above need to be To restore, specifically, the predicted value Multiply by the affine bias used in forward normalization and divide by the affine weight to cancel the normalization effect, then multiply the processed data by the standard deviation used in forward normalization and add the mean to restore the scale and distribution of the original data, and finally get the link traffic prediction value

[0037] S3 uses broadband link time series traffic data for model training and prediction; the details are as follows:

[0038] S3.1 Before model training, first process the broadband link time series flow data according to the steps described in S1, and the output in S1.1 is used as the input of the model constructed by S2. Then all broadband link time series flow data are divided into training data sets and test data sets in proportion, and the proportion is adjusted according to the actual application requirements, for example, the ratio of training data sets to test data sets is 6:4. In the model training stage, the broadband time series flow data in the training data set is used as the input of the model constructed by S2, and then the model is trained to obtain a trained broadband link flow prediction model based on spectral domain graph neural network.

[0039] S3.4 In the model prediction stage, the broadband time series traffic data in the test data set is input into the trained broadband link traffic prediction model based on the spectral domain graph neural network to obtain the link traffic value of the prediction period.

[0040] The present invention also proposes a computer system based on the above method, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the steps of the above method are implemented when the processor executes the computer program.

[0041] The present invention has the following beneficial effects:

[0042] 1. The present invention significantly improves the quality and efficiency of data processing by comprehensively using the periodic exception handling mechanism and cluster analysis technology. The periodic exception handling mechanism can accurately identify and remove outliers in the time series traffic data, ensure the data purity of the model training and prediction process, and thus enhance the reliability of the prediction results. The application of cluster analysis technology further visualizes the distribution characteristics of different links in the time series traffic data, reveals the inherent characteristics of the time series traffic change patterns of different links, and provides a more intuitive data basis for subsequent model construction.

[0043] 2. The present invention pays attention to the fact that there are certain similarities and correlations in the traffic changes between different broadband links. In order to make full use of this feature, the present invention adopts a self-attention structure. The self-attention structure allows the model to automatically pay attention to the similarities and correlations of different links in time when processing time series traffic data without relying on external information or labeled data. This mechanism significantly enhances the model's ability to capture complex traffic change patterns and further improves the accuracy and reliability of predictions.

[0044] 3. The present invention introduces a spectral domain graph neural network (STGNN) modeling strategy, which effectively captures the complex spatial dependencies and temporal dynamics in broadband link timing traffic data by making full use of the advantages of graph neural networks in processing complex network structure data and combining the ability of time series feature learning in the spectral domain. This improves the prediction accuracy while significantly enhancing the robustness of the model to the uncertainty of traffic changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flow chart of a broadband link traffic prediction method based on spectral domain graph neural network of the present invention;

[0046] Figure 2 It is a framework diagram of a broadband link traffic prediction model based on a spectral domain graph neural network in an embodiment of the present invention;

[0047] Figure 3 This is a bandwidth 1G link clustering result diagram of the clustering analysis module in an embodiment of the present invention;

[0048] Figure 4 This is a graph showing the bandwidth 10G link clustering results of the clustering analysis module in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The embodiments of the present invention are described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0050] Embodiment 1:

[0051] This embodiment discloses a broadband link traffic prediction method based on a spectral domain graph neural network, comprising the following steps:

[0052] Construct a data processing and analysis module, including a data anomaly processing module and a cluster analysis module. The data anomaly processing module includes analyzing the spectral characteristics of time series traffic data, outlier identification and outlier replacement; the cluster analysis module includes feature selection, cluster analysis, PCA dimensionality reduction and link traffic feature analysis.

[0053] A broadband link traffic prediction model based on spectral domain graph neural network is constructed, and the training data in the training sample set is used to train the model, so that it can automatically learn and capture the potential associations and dependencies between different links, and learn the repetitive patterns and autocorrelation features in the time series in the frequency domain based on the formed graph structure, thereby realizing the prediction of broadband link traffic values.

[0054] In addition, in this embodiment, a computer system is also disclosed, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above method are implemented.

[0055] The present invention improves data quality by integrating outlier processing under spectral characteristic analysis, and then uses data normalization processing to solve the distribution drift problem. It then uses the self-attention mechanism to capture the potential correlation and dependency between broadband links, and integrates graph neural networks and temporal pattern learning in the spectral domain to accurately capture the spatiotemporal characteristics of time series traffic data, significantly improving the prediction accuracy.

[0056] Embodiment 2:

[0057] Embodiment 2 is a preferred embodiment of Embodiment 1, which focuses on solving the problem of inaccurate prediction of peak downlink traffic of bandwidth links between multiple BRAS (Broadband Remote Access Server) devices and Optical Line Terminal (OLT) devices at a provincial level in real scenarios. Based on the broadband link traffic prediction method based on spectral domain graph neural network, this example expands the specific implementation steps to better meet actual needs.

[0058] Specifically, this example processes and analyzes the time series traffic data, and then uses self-attention structure, graph neural network, spectrum transformation and other technical means to capture the correlation and association of traffic changes between broadband links in a more comprehensive way, and learn the time series pattern and autocorrelation characteristics of broadband link traffic. This method can effectively improve the prediction accuracy while realizing broadband link traffic prediction.

[0059] In this embodiment, the broadband link traffic prediction based on the spectral domain graph neural network is applied to the prediction of broadband link peak downstream traffic data recorded at a daily time granularity between multiple BRAS devices and OLT devices at a provincial level, specifically including the following steps:

[0060] 1. Data Preprocessing

[0061] The dataset used in this example is the peak downstream traffic data of broadband links between multiple BRAS devices and OLT devices at a provincial level, recorded at a daily time granularity. The time range of this dataset is from May 14, 2023 to May 10, 2024, about one year, involving 8,869 broadband links, 322 BRAS devices, and 3,307 OLT devices. Among them, broadband links are divided into 1G and 10G broadband links according to bandwidth size, with 775 and 8,094 links respectively.

[0062] The data preprocessing is as follows:

[0063] The daily peak downstream traffic data of the bandwidth link has some problems such as missing record fields due to equipment failure, network anomalies, etc., so it is necessary to perform preprocessing such as data cleaning and merging. The first column of the preprocessed daily peak downstream traffic data of the bandwidth link represents different link names, which are formed by the concatenation of the BRAS equipment IP and port number and the OLT equipment IP and port number at both ends of the link. The first row represents the timestamp with a daily time granularity from May 14, 2023 to May 10, 2024. According to the different link broadband capacities, the data set is subdivided into a bandwidth 1G link downstream traffic data subset and a bandwidth 10G link downstream traffic data subset. In view of the high consistency of the operations required to be performed by the two subsets in the subsequent operation stage, I will directly use the time series traffic data downFlow to refer to the above data set in the future.

[0064] In this example, all data sets are divided according to corresponding ratios, where the ratio of training set to test set is 6:4.

[0065] 2. Data Processing and Analysis

[0066] It is crucial to perform proper data processing and analysis before inputting the data into the prediction model to ensure the accuracy and reliability of the model. This example uses the data processing and analysis module to process and analyze the data. This module mainly includes multiple steps including the data anomaly processing module and the cluster analysis module, corresponding to Figure 2 The data processing and analysis module of the prediction model diagram.

[0067] In time series traffic data, sometimes some outliers may appear in the data due to various reasons such as equipment failure, network attacks, and measurement errors. If not processed, these outliers will seriously affect the accuracy and reliability of subsequent predictions. In order to prevent outliers in the data from affecting the accuracy of predictions, this example uses Fast Fourier Transform (FFT) technology to identify and analyze the spectral characteristics of the time series traffic data downFlow, determine its main period, and then set the detection threshold of abnormal data based on the identified main period and the normal distribution assumption. Specifically, let the current sliding window i be the size of the detected main period, and the set of time series traffic data downFlow under this window is expressed as A k represents the traffic record of link k in the current sliding window, represents the traffic value of link k at time t1. Next, we calculate the mean μ and standard deviation σ of the traffic data in this window and set the anomaly detection threshold T = 3σ. If a data point satisfy It is judged as an outlier. For the detected outlier Assume that the normal data points at the corresponding positions of the two adjacent periods before and after are and The corresponding time points are t before and t after , the time of the current abnormal data point is t abnormal , then the value of the data point after linear interpolation is:

[0068]

[0069] By processing the time series flow data downFlow according to the above method, the stability and consistency of the data can be ensured. The output of this module is the time series flow data. Where N is the number of links, T is the number of timestamps, represents the peak flow of the nth broadband link at timestamp t. For the convenience of subsequent description, we record the observed value at timestamp t as The timing flow of each link is recorded as

[0070] In order to better understand the distribution characteristics of time series traffic data X and identify the time change pattern of data, we perform cluster analysis on the time series traffic data X. The advantage of this is that links with similar characteristics can be classified into one category, which makes it easier to analyze the traffic change trends and characteristics of these links in the same category, providing strong support for the design of the model. First, we select the time series traffic of each link Then, the mean of the traffic flow of each link is calculated in turn. This feature reflects the average traffic level of each link; the standard deviation of the traffic flow of each link. This feature measures the discrete degree of the traffic distribution of each link; the rate of change of the traffic flow of each link. This feature calculates the rate of change between adjacent traffic values ​​by difference and then takes the average value, reflecting the degree of change of the traffic flow of each link; the slope of the trend line of the traffic flow of each link. This feature calculates the slope of the trend line of the traffic flow of each link by linear regression, reflecting the long-term change trend of the link traffic; the autocorrelation of the traffic flow of each link. This feature measures the similarity between the traffic flow of each link and its own past traffic value. Reflects the periodic changes of link traffic; Peak average value, this feature calculates the average value of the traffic exceeding the peak threshold in each link traffic. Here, the peak threshold traffic is set to 45% of the bandwidth size, reflecting the extreme fluctuation of link traffic. According to the above features, the KMeans algorithm is used to cluster the time series traffic data X. Since the data set is subdivided into bandwidth 1G link downlink traffic data subsets and bandwidth 10G link downlink traffic data subsets, they are clustered separately to obtain the bandwidth 1G link time series traffic data clustering results and bandwidth 10G link time series traffic data clustering results. The following table shows the clustering results:

[0071] Table 1 Clustering results of bandwidth 1G link timing traffic data

[0072]

[0073] Table 2 Bandwidth 10G link timing traffic data clustering results

[0074]

[0075] In order to show the clustering results more intuitively, we use principal component analysis (PCA) to reduce the high-dimensional feature data to two-dimensional space. The bandwidth 1G link clustering result graph and the bandwidth 10G link clustering result graph are shown as follows: Figure 3 , Figure 4 shown.

[0076] According to the clustering results in Table 1 and Table 2, we can know that there are two clusters of bandwidth 1G link timing flow data: Cluster 0 has a small and stable average flow, with an upward trend in the short term, but a slight downward trend in the long term, while Cluster 1 has a large average flow and large fluctuations, showing a slight growth trend, with significant peak values; for bandwidth 10G link timing flow data, there are two clusters: Cluster 0 has a low average flow, small fluctuations, a short-term downward trend, but a long-term upward trend; Cluster 1 has a high average flow, large fluctuations, a downward trend, strong periodicity, and a significant peak value. According to Figure 3 and Figure 4It can be clearly seen that there are significant differences between the two clusters of bandwidth 1G links and bandwidth 10G links in principal component 1 and principal component 2.

[0077] 3. Forward Normalization Module

[0078] Since the time series traffic data The statistical characteristics of the data, such as mean, variance, and distribution shape, often change between the training and testing phases. This phenomenon is called data distribution drift, which can have a significant impact on the accuracy of predictions. To eliminate this effect, this example will When inputting the model, the data is normalized to improve the prediction accuracy. This module corresponds to Figure 2 Forward normalization module in the prediction model.

[0079] Specifically, the mean and standard deviation of the timing traffic of each link are calculated and then standardized. The formula is as follows:

[0080]

[0081] where t is the number of timestamps for link k, μ k represents the mean flow of link k, σ k represents the standard deviation flow value of link k, represents the normalized flow value of link k at time t1.

[0082] Next, affine transformation normalization is performed, the formula is as follows:

[0083]

[0084] in is the data after affine transformation, ω is the affine weight, and b is the affine bias. The initial values ​​of the affine weight and affine bias are vectors of all 1 and all 0 respectively, which are continuously learned and gradually optimized and updated during the entire model training. The time series flow values ​​of all links are transformed according to the above method to obtain the normalized time series flow data

[0085] 4. Spatial Self-Attention Module

[0086] Given that multiple links in a broadband link network structure may share the same broadband remote access server (BRAS) or optical line terminal (OLT), the traffic change patterns between them are highly similar. In addition, specific links may also show similar dynamic change characteristics due to complex factors such as network congestion or user behavior patterns during specific periods of time. In order to accurately capture and model the implicit interdependencies and dynamic interaction patterns of these links in time series data, this example cleverly combines the gated recurrent unit (GRU) layer and the self-attention structure as a spatial self-attention module to construct a graph adjacency matrix that can reflect the strength of association between links. This module corresponds to Figure 2 Spatial self-attention module in the prediction model graph.

[0087] Specifically, firstly, the normalized time series traffic data The input is fed into the Gated Recurrent Unit (GRU) layer. The GRU layer processes the input sequence timestamp by timestamp through its unique gating mechanism, namely the update gate and the reset gate, and obtains the hidden state of its input time step. In this process, the GRU layer not only captures the temporal dependencies in the sequence, but also integrates historical information through its internal state update mechanism to form a deep understanding of the current time step. Finally, we obtain the global feature containing the time series, that is, the last hidden state R of all time series.

[0088] Subsequently, in order to further analyze and quantify the dynamic interaction between different links, we introduced a self-attention structure to construct the adjacency weight matrix of the graph. Under the self-attention structure, we first multiply the last hidden state R of all time series processed by the GRU layer with two different weight matrices. The weight matrices are represented by W Q and W k Indicates that the weight matrix is ​​randomly initialized, continuously learned and updated during model training, and the multiplication results are the query vector (Query) and the key vector (Key), represented by Q and K. Then we calculate the dot product between the query vector and the key vector, and divide it by a scaling factor (usually the square root of the key vector dimension, i.e. Where d is the dimension of the key vector) to alleviate the gradient vanishing problem that may be caused by the dot product operation. This process realizes the mapping from the original hidden state space to the space required for self-attention calculation, allowing the model to focus on the potential interactions between different links. After the above processing, the attention score matrix is ​​obtained, and then the Softmax function is applied to it for normalization, so as to obtain a normalized attention score matrix W = The calculation formula is:

[0089] Q=RW Q

[0090] K=RW K

[0091]

[0092] The final result is viewed as an adjacency matrix of a graph, where w nm represents the strength of association between link n and link m. This adjacency matrix not only captures the direct association between links, but also reveals the underlying and complex interaction patterns between links through the self-attention mechanism.

[0093] 5. Graph frequency domain convolution module

[0094] In order to capture the spatial dependencies in the graph structure and combine these spatial dependencies with the frequency information of the time series traffic data, this example uses the graph spectral convolution module to learn the relationship between traffic time series. This module mainly includes two parts: Laplace transform and graph Fourier transform, corresponding to Figure 2 Graph frequency domain convolution module for the prediction model graph in .

[0095] Based on the adjacency matrix of the above graph The corresponding Laplace matrix L is calculated according to the following formula:

[0096]

[0097] Among them I N is the N×N identity matrix, W is the adjacency matrix of the graph, D is a diagonal matrix, and the diagonal elements D nn It represents the number of edges connected to node n (corresponding to link n) in the adjacency matrix W, that is, the degree of node n (corresponding to link n), and the calculation formula is:

[0098]

[0099] After the above calculation steps, we get the Laplace matrix L, which describes the overall structural characteristics of the graph. Next, we perform graph Fourier transform on the Laplace matrix L. First, we perform eigendecomposition on it to get the eigenvector matrix U and the eigenvalue diagonal matrix Λ. The calculation formula is as follows:

[0100] L=UΛU T

[0101] The eigenvalue diagonal matrix Λ represents the frequency components of the graph, and the eigenvector matrix U is equivalent to the "basis" vector of the graph. Next, the normalized time series traffic data Multiplying by the transpose of the eigenvector matrix U, we get the time series F of the relationship between traffic time series in the frequency domain. The calculation formula is as follows:

[0102] F=GFT(X′)=UT X′

[0103] Here U is the time series traffic data Projected into the spectral domain, we obtain the time series F of the relationship between traffic time series in the frequency domain.

[0104] 6. Frequency Domain Time Pattern Learning Module

[0105] In order to learn the repetitive patterns in each individual time series in the time series F and the autocorrelation features between different timestamps in the frequency domain, this example uses the frequency domain time pattern learning module to learn the time pattern feature representation TF of the input time series F. This module mainly includes a discrete Fourier transform module (DFT), a gated linear unit (GLU) layer, and an inverse discrete Fourier transform module (IDFT), corresponding to Figure 2 A frequency-domain temporal pattern learning module for predictive model graphs in .

[0106] For the time series F, the frequency domain time pattern learning module first applies discrete Fourier transform to each individual time series to process the frequency information on each individual time series to obtain the frequency components of each individual time series, and then applies GLU layer to the frequency components of each individual time series to perform feature learning to obtain the feature representation learned in the frequency domain of each individual time series, and applies the inverse discrete Fourier transform module to convert the feature representation learned in the frequency domain of each individual time series back to the time domain, and splices the feature representations of multiple time series to finally obtain the time pattern feature representation TF. The calculation formula is as follows:

[0107] TF n (t) = IDFT(GLU(DFT(F i (t)))

[0108] TF=Concat(TF 1 (t),TF 2 (t),...,TF N (t))

[0109] where F n (t) represents each individual time series in the time series F, where n = 1, 2, ... N, t is the timestamp, TF n (t) represents each individual time series feature representation.

[0110] 7. Time and Space Restoration Module

[0111] In order to restore the original state of the data and retain the learned feature representation, this example restores the above temporal pattern feature representation TF to the time domain through the spatiotemporal restoration module to obtain the final result Z. This module is an inverse Fourier transform corresponding to Figure 2 The spatiotemporal restoration module of the prediction model graph in . The formula is as follows:

[0112] Z=GFT -1 (TF)=UTF

[0113] Here U represents the eigenvector matrix obtained by eigendecomposing the Laplace matrix L in the above-mentioned graph frequency domain convolution module.

[0114] 8. Prediction Module

[0115] In the prediction part of the model, this example outputs the predicted value through the fully connected layer Then, the inverse normalization module is used to restore the scale and distribution of the original data to obtain the link traffic prediction value. correspond Figure 2 The prediction module in the prediction model diagram.

[0116] First, we input the final result Z into Figure 2 In the first linear layer Linear in the prediction module shown in the figure, the function of this layer is to perform a linear transformation on the final result Z, then use the Leaky ReLU activation function to perform a nonlinear transformation on the features processed by the first linear layer, and finally use the second linear layer Linear to convert the features obtained after the nonlinear transformation into the required output dimension to obtain the predicted value The calculation formula is as follows:

[0117]

[0118] Since the forward normalization module mentioned above processes the input time series flow data The data is normalized, so the model actually learns the normalized data features. Therefore, after the model makes a prediction, the predicted value is also calculated based on these normalized features. In order to convert the prediction result back to the scale of the original data, the predicted value obtained above needs to be To restore, specifically, the predicted value Multiply by the affine bias used in forward normalization and divide by the radiation weight to cancel the normalization effect, then multiply the processed data by the standard deviation used in forward normalization and add the mean to restore the scale and distribution of the original data to obtain the link traffic prediction value The calculation formula is as follows:

[0119]

[0120] in Represents the predicted value The predicted value of link k at timestamp t1, ω and b represent the affine weight and affine bias calculated during forward normalization, ∈ represents a very small positive number to avoid division by zero error. represents the predicted value after inverse affine normalization, σ k and μ k They represent the mean flow and standard deviation flow of link k calculated during forward normalization, Represents the original prediction value after inverse normalization. The above processing can eliminate the trend changes caused by moment statistics (statistical characteristics of time points), while retaining the statistical information that can be used to restore the final prediction result, thereby significantly improving the performance of time series prediction and obtaining the link traffic prediction value

[0121] 9. Model Performance Evaluation Module

[0122] In order to verify the prediction performance of the broadband link traffic prediction model (STGNN) based on spectral domain graph neural network proposed in this example, five methods including historical mean method (HA), differential autoregressive moving average model (ARIMA), support vector regression (SVR), long short-term memory network (LSTM) and Fourier graph neural network (FourierGNN) were selected as the comparison baselines of the experiment.

[0123] The historical mean method is a simple model based on the average of past observations as a prediction, which is suitable for data with unclear trends or weak periodicity.

[0124] The difference autoregressive moving average model uses the difference, autoregressive and moving average components of the time series to build a model, which is suitable for stable non-seasonal data.

[0125] Support vector regression uses the idea of ​​support vector machine to perform nonlinear regression analysis by constructing a hyperplane. It is suitable for data with complex nonlinear relationships.

[0126] Long short-term memory networks are a variant of recurrent neural networks that are specifically designed to process sequential data and can capture long-term dependencies.

[0127] Fourier graph neural network captures the spatiotemporal dependencies of multivariate time series in Fourier space by constructing hypervariable graphs and introducing Fourier graph operators. It is suitable for complex scenarios that need to capture both spatial dynamics and temporal dependencies.

[0128] For model performance evaluation, we mainly compare from two aspects. From the perspective of quantitative evaluation, we use Mean Absolute Percentage Error (MAPE) as the main evaluation indicator. MAPE is a common method to measure the difference between the predicted value and the actual value, and is particularly suitable for the prediction effect evaluation of time series traffic data. Its calculation formula is as follows:

[0129]

[0130] Among them, y i represents the time series traffic value of the ith real broadband link, represents the predicted value of the time series traffic of the ith broadband link, and n represents the total number of predictions. The smaller the value of MAPE is, the closer the predicted value is to the actual value, and the higher the prediction accuracy is.

[0131] From the perspective of classification accuracy evaluation, the broadband link time series traffic values ​​are classified according to the preset bandwidth utilization standard (0% to 40% is normal, 40% to 50% is slightly congested, 50% to 60% is moderately congested, and more than 60% is severely congested). That is, for the bandwidth 1G link downlink traffic data subset, the traffic value range of 0G to 0.4G is set as Class 1, 0.4G to 0.5G is set as Class 2, 0.5G to 0.6G is set as Class 3, and values ​​above 0.6G are set as Class 4. For the bandwidth 10G link downlink traffic data subset, the traffic value range of 0G to 4G is set as Class 1, 4G to 5G is set as Class 2, 5G to 6G is set as Class 3, and values ​​above 6G are set as Class 4. The predicted value classification of the time series traffic data is compared with the actual value classification of the time series traffic data to evaluate the model's ability to capture the degree and trend of traffic changes. The classification accuracy (ACC) formula is calculated as follows:

[0132]

[0133] n is the total number of time series flow values. The closer the ACC value is to 1, the more accurate the predicted value of the time series flow data is in reflecting the trend and degree of flow change. The classification accuracy evaluation not only examines the numerical proximity of the predicted values ​​of the time series flow data, but also evaluates the model's ability to judge the flow change category from a more macro perspective.

[0134] The performance of the model proposed in this example will be verified and compared from the above two perspectives, including the following experimental results:

[0135] The STGNN model proposed in this example is experimented on the bandwidth 1G link downlink traffic data subset and the bandwidth 10G link downlink traffic data subset, and the performance comparison results of the model proposed in this example and other baseline predictions are compared, including the mean absolute percentage error (MAPE) and classification accuracy (ACC) two evaluation indicators. Table 3 and Table 4 show the prediction results of the bandwidth 1G link downlink traffic data subset and the bandwidth 10G link downlink traffic data subset using each model in terms of mean absolute percentage error and classification accuracy. It can be seen that the STGNN model proposed in this example shows excellent performance in both mean absolute percentage error and classification accuracy. Compared with other baseline models, its MAPE and ACC evaluation indicators are improved, which fully proves the effectiveness and superiority of our model in the prediction tasks of the bandwidth 1G link downlink traffic data subset and the bandwidth 10G link downlink traffic data subset.

[0136] Table 3 Prediction results of bandwidth 1G link downlink traffic data subset

[0137]

[0138] Table 4 Prediction results of bandwidth 10G link downlink traffic data subset

[0139]

Claims

1. A broadband link traffic prediction method based on spectral domain graph neural network, characterized in that: The method is divided into the following steps: S1 builds a data processing and analysis module, including a data anomaly processing module and a cluster analysis module. The input data of this module is the peak downlink flow data of broadband links between multiple BRAS devices and OLT devices, recorded at a time granularity of days, which is referred to as time series flow data downFlow. The details are as follows: The data anomaly processing module described in S1.1 is used to perform spectrum characteristic analysis, outlier identification, and outlier replacement on each broadband link in the time series traffic data downFlow. The specific implementation steps are as follows: S1.1.1 For each broadband link’s time series traffic data, the fast Fourier transform (FFT) technology is used to analyze the spectrum characteristics and identify the main period in the time series traffic data of each broadband link. This period will be used to set the sliding window of its corresponding link. S1.1.2 Based on the identified main cycle, the detection threshold of abnormal values ​​in the broadband link timing traffic data is set according to the normal distribution assumption. Once it is detected that the difference between the value of a data point in the current sliding window and the value of the corresponding position of the data point in the previous and next adjacent cycles exceeds the preset detection threshold, it is determined to be abnormal data; S1.1.3 For these abnormal data, linear interpolation method is used to repair them. The specific operation is to use the values ​​of corresponding data points in the front and back windows for linear interpolation to replace the abnormal data values, so as to ensure the continuity and accuracy of the data, and thus improve the accuracy of subsequent data analysis and prediction; the time series flow data after abnormal processing is obtained through the above method Where N is the number of broadband links, T is the number of timestamps, represents the traffic value of the nth broadband link at timestamp t; for the convenience of subsequent description, the traffic value at timestamp t is recorded as The timing flow of each link is recorded as The cluster analysis module described in S1.2 is used for feature selection, cluster analysis, PCA dimension reduction and link traffic feature analysis. The specific implementation steps are as follows: S1.2.1 Calculate the timing flow of each link The mean, standard deviation, rate of change, trend line slope, autocorrelation, and peak average are used as the characteristic information of the link traffic; S1.2.2 Use KMeans algorithm to perform cluster analysis on the characteristic information of the above link traffic; S1.2.3 Use principal component analysis PCA to reduce the high-dimensional feature information data of link traffic to two dimensions; S1.2.4 draw a clustering result diagram, and analyze the characteristics of link traffic in different clusters based on the clustering results; S2 constructs a broadband link traffic prediction model based on a spectral domain graph neural network, which includes a forward normalization module, a spatial self-attention module, a graph frequency domain convolution module, a frequency domain time pattern learning module, a spatiotemporal restoration module, and a prediction module; the details are as follows: The forward normalization module described in S2.1 is used for data standardization and affine normalization, and the input is the time series flow data after abnormal processing. The output is normalized time series traffic data The function of this module is to solve the distribution drift problem in time series traffic data; the specific implementation steps are as follows: S2.1.1 Calculate input time series flow data The mean and standard deviation of each broadband link in the data are used to perform data standardization, that is, the input time series traffic data is subtracted from the mean of the corresponding link and then divided by the standard deviation; S2.1.2 Multiply the standardized data by the affine weight and add the affine bias to further adjust the data distribution to achieve normalization; after the above processing, the normalized time series flow data is obtained S2.2 The spatial self-attention module includes a gated recurrent unit GRU layer and a self-attention structure, and the input is the normalized time series flow data The output is the adjacency matrix of the graph calculated by the self-attention mechanism where w nm Indicates the strength of association between link n and link m. The function of this module is to automatically learn and capture the potential associations and dependencies between different links. The specific implementation steps are as follows: S2.2.1 Normalized time series traffic data Send it to the GRU layer, which can process the input sequence timestamp by timestamp and finally get the last hidden state R of all input sequences; S2.2.2 Multiply the last hidden state R of all input sequences obtained in the previous step by two different weight matrices, and the weight matrices are represented by W Q and W k The multiplication results are the query vector and the key vector, denoted by Q and K respectively. Then the dot product of the query vector Q and the key vector K is calculated and divided by the scaling factor to scale, where d is the dimension of the key vector; after the above processing, the attention score matrix is ​​obtained, and then normalized by the Softmax function to obtain a normalized attention score matrix, which is regarded as the adjacency matrix of the graph S2.3 The graph frequency domain convolution module includes a Laplace transform module and a graph Fourier transform module, and the input is the normalized time series flow data and the adjacency matrix of the graph The output is the time series F of the relationship between traffic time series in the frequency domain after the above transformation processing; the function of this module is to capture the spatial dependencies in the graph structure and combine these spatial dependencies with the frequency information of the time series traffic data to learn the relationship between traffic time series. The specific implementation steps are as follows: S2.3.1 Computing the adjacency matrix of a graph The Laplace matrix L of Among them I N is the N×N identity matrix, D is a diagonal matrix, and the diagonal elements D nn It represents the number of edges connected to node n in the adjacency matrix W, that is, the degree of node n, and the calculation formula is All off-diagonal elements D nm The values ​​of are all 0, n≠m; after the above calculation, the Laplace matrix L is obtained, which describes the overall structural characteristics of the graph; S2.3.2 Perform a graph Fourier transform on the Laplace matrix L; first perform eigendecomposition on the Laplace matrix L, and the decomposition formula is: L = UΛU T , we get the eigenvector matrix U and the eigenvalue diagonal matrix Λ. The eigenvalue diagonal matrix Λ represents the frequency components of the graph, and the eigenvector matrix U is equivalent to the "basis" vector of the graph; next, the normalized time series traffic data Multiplying with the transpose of the eigenvector matrix U, we get the time series F of the relationship between the flow time series in the frequency domain; The frequency domain time pattern learning module described in S2.4 includes a discrete Fourier transform module, a gated linear unit GLU layer and an inverse discrete Fourier transform module. The input is the time series F of the relationship between traffic time series in the frequency domain, and the output is the learned time pattern feature representation TF. The function of this module is to learn the repetitive pattern of each individual time series in the time series F of the relationship between traffic time series in the frequency domain and the autocorrelation features between different timestamps. The specific steps are as follows: S2.4.1 First, discrete Fourier transform is applied to each individual time series in the time series F of the relationship between flow time series in the frequency domain to process the frequency information of each individual time series, and the frequency component of each individual time series is obtained; S2.4.2 applies the GLU layer to learn the frequency components of each individual time series, and obtains the feature representation learned in the frequency domain for each individual time series; S2.4.3 Apply the inverse discrete Fourier transform module to convert the feature representation learned in the frequency domain of each individual time series back to the time domain to obtain the feature representation of each individual time series, and then concatenate the feature representation of each individual time series to finally obtain the learned time pattern feature representation TF; S2.5 The spatiotemporal restoration module is an inverse Fourier transform, which is used to restore the original state of the data and retain the learned feature representation. The input is the learned temporal pattern feature representation TF, and the output is the final result Z restored to the time domain. The learned temporal pattern feature representation TF is remapped back to the time series by applying the inverse Fourier transform to obtain the final result Z. S2.6 The prediction module includes a fully connected layer and a reverse normalization module. The input is the final result Z obtained by the spatiotemporal restoration module, and the output is the link traffic prediction value. The function of this module is to obtain the predicted value of the traffic of each link; the specific steps are as follows: S2.6.1 The fully connected layer consists of two linear layers and a Leaky ReLU activation function. First, the final result Z is input into the first linear layer, which performs a linear transformation on the final result Z. Then, the features processed by the first linear layer are nonlinearly transformed using the Leaky ReLU activation function. Finally, the second linear layer is used to convert the features obtained after the nonlinear transformation into the required output dimension to obtain the predicted value. S2.6.2 Since the forward normalization module is used to input the time series flow data The data is normalized, so the model actually learns the normalized data features. Therefore, the prediction values ​​obtained by the model are also calculated based on these normalized data features. In order to convert the prediction results back to the scale of the original data, the prediction values ​​obtained above need to be To restore, specifically, the predicted value Multiply by the affine bias used in forward normalization and divide by the affine weight to cancel the normalization effect, then multiply the processed data by the standard deviation used in forward normalization and add the mean to restore the scale and distribution of the original data, and finally get the link traffic prediction value S3 uses broadband link time series traffic data for model training and prediction; the details are as follows: S3.1 Before model training, firstly process the broadband link time series flow data according to the steps described in S1, and the output in S1.1 is used as the input of the model constructed by S2; then all broadband link time series flow data are divided into training data set and test data set in proportion; in the model training stage, the broadband time series flow data in the training data set is used as the input of the model constructed by S2, and then the model is trained to obtain a trained broadband link flow prediction model based on spectral domain graph neural network; S3.4 In the model prediction stage, the broadband time series traffic data in the test data set is input into the trained broadband link traffic prediction model based on the spectral domain graph neural network to obtain the link traffic value of the prediction period.

2. A broadband link traffic prediction method based on spectral domain graph neural network according to claim 1, characterized in that: In S1.1.2, the detection threshold is set to three times the standard deviation of the data in the sliding window of the current link, which is used as a quantitative standard to determine whether the data is abnormal.

3. A broadband link traffic prediction method based on spectral domain graph neural network according to claim 1, characterized in that: In S2.1.2, the initial values ​​of the affine weights and affine biases are vectors of all 1s and all 0s, respectively, and are continuously learned and gradually optimized and updated during the entire model training.

4. A broadband link traffic prediction method based on spectral domain graph neural network according to claim 1, characterized in that: In S2.2.2, the weight matrix is ​​randomly initialized and is continuously learned and updated during model training.

5. A broadband link traffic prediction method based on spectral domain graph neural network according to claim 1, characterized in that: In S3.1, the ratio of dividing the training data set and the test data set is adjusted according to the actual application requirements.

6. A broadband link traffic prediction method based on spectral domain graph neural network according to claim 5, characterized in that: In S3.1, the ratio of the training data set to the test data set is 6:

4.

7. A computer system based on the method according to any one of claims 1 to 6, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above method are implemented.

Citation Information

Patent Citations

  • Internet of Things system

    CN116368355A

  • Network traffic prediction method based on space-time diagram multi-attention mechanism

    CN117880871A

  • Network traffic prediction method based on graph convolutional neural network

    CN118488005A

Cited By

  • Network acceleration strategy optimization system and method based on data flow

    CN119629100A

  • Time sequence prediction method and device, program product, storage medium and application of time sequence prediction method and device

    CN119939396A

  • Cellular traffic prediction method and system based on multilayer meta-learning model

    CN120201481A

  • User portrait construction method and system based on user habit analysis

    CN120317918A

  • Multivariable time sequence prediction method and device based on spatio-temporal feature fusion

    CN120910514A