Network traffic prediction method based on graph convolutional neural network

Abstract

The invention provides a network traffic prediction method based on a graph convolutional neural network, and the method comprises the steps: carrying out the prediction of the network traffic based on the periodic features of network traffic data; network traffic data of a geographic area having spatial correlation with a target area in the same time period of a previous week, in the same time period of a previous day and in two hours before the current day are adopted as target time sequence data to predict network traffic of the target area in two hours in the future, and the method specifically comprises the following steps: representing the network traffic data as a byte traffic diagram; performing multi-level feature extraction on the time sequence data by using an average pooling method and a maximum pooling method to obtain a coarse-grained change trend and a fine-grained change trend of the time sequence data, and performing splicing to obtain a sequence feature map; on the basis of a graph convolutional neural network, combining a long-short-term memory network and a gating circulation unit, performing graph convolution operation on the sequence feature graph to extract spatial-temporal features of the sequence feature graph; and after carrying out layer linear transformation, outputting a network flow prediction result. According to the method, the network traffic prediction precision can be improved.

Description

Technical Field

[0001] The present application belongs to the technical field of network traffic prediction, and in particular, relates to a network traffic prediction method based on graph convolutional neural network. Background Art

[0002] With the surge in network users and network traffic, network applications are constantly updated and iterated, the Internet environment has become more complex, and the rapid development of network information technology has affected all aspects of society. Network traffic prediction technology is of great significance both in academia and in real life. Early network traffic prediction methods mainly relied on naive methods and parameterized models based on time series, among which the historical average method and ARIMA model were more common. However, these traditional methods have insufficient performance when facing complex and changeable network traffic, especially it is difficult to capture the nonlinear and spatial characteristics of traffic. With the development of deep learning, models such as RNN, GRU, LSTM and CNN have been gradually applied to traffic prediction, and the prediction accuracy has been improved by combining time and space features. However, CNN has limitations in dealing with complex network topologies. The introduction of GCN and attention mechanism effectively improves the extraction of spatiotemporal features of traffic, but there are still problems such as insufficient feature extraction and insufficient capture of periodicity and fluctuation trends. Summary of the invention

[0003] In view of this, the purpose of this application is to provide a network traffic prediction method based on graph convolutional neural network, which fully captures the periodicity and fluctuation trend of network traffic data, and fully extracts the sequence characteristics, time characteristics, and spatial characteristics of network traffic data to improve the accuracy of network traffic prediction.

[0004] The present application provides a network traffic prediction method based on a graph convolutional neural network, the method comprising:

[0005] Based on the periodic characteristics of network traffic data, the network traffic data of the same time period of the previous week, the same time period of the previous day, and the first two hours of the current day in a geographical area with spatial correlation with the target area are used as target time series data to predict the network traffic of the target area in the next two hours. Specifically:

[0006] representing the network traffic data as a byte traffic graph;

[0007] Using average pooling and maximum pooling methods, multi-level feature extraction is performed on the byte flow graph to obtain coarse-grained change trends and fine-grained change trends of the time series data;

[0008] The coarse-grained change trend and the fine-grained change trend are spliced ​​to obtain a sequence feature graph of the time series data;

[0009] Based on a graph convolutional neural network, combined with a long short-term memory network and a gated recurrent unit, a graph convolution operation is performed on the sequence feature graph to extract the spatiotemporal features of the time series data;

[0010] After performing layer-by-layer linear transformation on the spatiotemporal features of the time series data, the network traffic prediction result of the target area is output.

[0011] Further, the network traffic data is represented as a byte traffic graph, including:

[0012] Using bytes in the network flow data as nodes of the byte flow graph;

[0013] Using the correlation between byte pairs in the network traffic data as the edges between nodes in the byte traffic graph, and using an adjacency matrix to represent the connection relationship between nodes;

[0014] The original values ​​of bytes in the network traffic data are used as features of nodes in the byte traffic graph.

[0015] Furthermore, the adjacency matrix of the byte flow graph is obtained in the following way:

[0016] For each byte pair, the distance similarity of the byte pair is calculated by counting the frequency of the byte pair appearing simultaneously in the sliding window;

[0017] Calculate the first-order difference of each byte in the sliding window to obtain the numerical similarity of the byte pairs;

[0018] Based on the distance similarity and the numerical similarity, a correlation value of the byte pair is calculated, and the correlation value is transformed into a range of [0, 1];

[0019] The association value of the transformed byte pair is compared with a predetermined threshold to retain or discard the byte pair, thereby obtaining the adjacency matrix.

[0020] Furthermore, the average pooling and maximum pooling methods are used to perform multi-level feature extraction on the time series data, including:

[0021] Using average pooling technology, by calculating the average value in a local window, the time series data is downsampled to obtain first feature data that can reflect global trend information;

[0022] Using the maximum pooling technique, by selecting the maximum value in the pooling window, the time series data is downsampled to obtain second feature data that can highlight local features;

[0023] Calculating the difference between the second feature data and the first feature data to obtain a coarse-grained change trend to characterize the global features of the time series data;

[0024] The difference between the second feature data and the time series data is calculated to obtain a fine-grained change trend to characterize the local features of each data point in the time series data.

[0025] Furthermore, the spatial correlation is jointly determined by geographic proximity, user mobility patterns, coverage of communication infrastructure and regional functional differences.

[0026] Furthermore, the graph convolution neural network is combined with a long short-term memory network and a gated recurrent unit to perform a graph convolution operation on the sequence feature graph to extract the spatiotemporal features of the time series data, including:

[0027] Based on a single graph convolution network and a long short-term memory network, an adjacency matrix of the time series data is used to perform a graph convolution operation on the sequence feature graph to obtain a first spatiotemporal feature;

[0028] Based on a multi-graph convolutional network and a gated recurrent unit, a graph convolution operation is performed on the sequence feature graph using an adjacency matrix that considers the fluctuation of the time series data to obtain a second spatiotemporal feature; wherein the fluctuation includes: a trough, a stable state, and a peak;

[0029] The first spatiotemporal feature and the second spatiotemporal feature are concatenated to obtain the spatiotemporal feature of the time series data.

[0030] Furthermore, the fluctuation of time series data is determined by:

[0031] Calculate the average value of network traffic data for each timestamp;

[0032] Calculate the standard deviation of the network traffic data within the sliding window to reflect the fluctuation of the time series data;

[0033] The time series data is divided into trough data, stable data and peak data according to the fluctuation by using quartiles.

[0034] Furthermore, the adjacency matrix of the time series data and the adjacency matrix considering the fluctuation of the time series data are obtained in the following ways:

[0035] Using a fast dynamic regularization distance algorithm, based on time series data, valley data, stable data and peak data, the similarity between each region is calculated respectively; wherein each region includes the target region and a geographical region having spatial correlation with the target region:

[0036] According to the similarity between the target area and the geographical area having spatial correlation with the target area, adjacency matrices of time series data, trough data, stable data and peak data are constructed respectively.

[0037] Furthermore, the calculation method of the fast dynamic time warping distance algorithm is as follows:

[0038] Simplify the two time series data, downsample layer by layer, and generate multi-level resolution representation;

[0039] At the lowest resolution level, the dynamic time warping distance algorithm is used to find the preliminary alignment path of the two time series data;

[0040] After mapping the preliminary alignment path to the previous level, the dynamic time warping distance algorithm is used to adjust the preliminary alignment path until the optimal alignment path is found to characterize the similarity of the two time series data.

[0041] The present application provides a network traffic prediction method based on a graph convolutional neural network. Based on the periodic characteristics of network traffic data, time series data of a specific time period is used as input, sequence features are extracted through average pooling and maximum pooling methods, and coarse-grained and fine-grained change trends are introduced to fully extract sequence features. Furthermore, a fast dynamic time warping distance is introduced to measure the difference between byte distributions, a graph convolutional neural network is used to extract position features in network traffic data, and a long short-term memory network and a gated recurrent unit are used to extract traffic time features respectively. The above method can fully capture the periodicity and fluctuation trends of network traffic data to improve the accuracy of network traffic prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A schematic diagram of selecting a specific time period based on periodic characteristics provided in an embodiment of the present application is shown;

[0043] Figure 2 A heat map of spatial correlation of network traffic data of 15 mutually adjacent geographical areas provided in an embodiment of the present application is shown;

[0044] Figure 3 A flowchart of a network traffic prediction method based on a graph convolutional neural network provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solution and advantages of the technical solution more clear, the technical solution is further described in detail below in conjunction with specific implementation methods. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of the technical solution.

[0046] Example 1: Periodic feature analysis of network traffic data:

[0047] The periodicity of network traffic data is manifested as similar traffic patterns at different time points, including the following types: 1) Daily cycle, usually a 24-hour cycle, reflecting the rhythm of people's daily activities. Generally speaking, due to human living habits, the usage of network services reaches a peak during the day and decreases significantly at night. 2) Weekly cycle, that is, different traffic patterns are shown in different time periods of the week. For example, on weekdays, network traffic usually peaks from morning to afternoon because users frequently use the Internet in work and study; on weekends, traffic peaks may occur during evening entertainment activities due to holidays. 3) Monthly cycle and seasonal cycle, some applications or services will show obvious periodic changes within a monthly or quarterly range. For example, traffic fluctuations on e-commerce platforms are usually associated with specific time nodes such as shopping festivals and holidays.

[0048] In view of the periodic characteristics of network traffic data, that is, the network traffic data in the same time period are similar, this paper uses the network traffic data in the same time period of the previous week, the same time period of the previous day, and the first two hours of the current day to predict the network traffic in the next two hours.

[0049] Here, please refer to Figure 1 The schematic diagram of selecting a specific time period based on periodic characteristics provided by the embodiment of the present application is shown. Figure 2 As shown in the figure, assuming that the network traffic from 9:00 to 11:00 on Tuesday, November 12 is predicted, the network traffic data of three time periods are selected as input, namely, 9:00 to 11:00 on Tuesday, November 5 (the same time period of the previous week), 9:00 to 11:00 on Monday, November 11 (the same time period of the previous day), and 7:00 to 9:00 on November 12 (the first two hours of the day). This method can effectively utilize historical periodic patterns and improve the accuracy of network traffic prediction.

[0050] Example 2: Spatial correlation analysis of network traffic data:

[0051] The spatial correlation of network traffic data refers to the mutual influence and dependence of network traffic activities between different geographical areas. This correlation may be caused by multiple factors. In order to explore the spatial correlation of network traffic data, the spatial correlation analysis of network traffic data was conducted for 15 adjacent geographical areas. The results are as follows: Figure 2 The heat map of spatial correlation of network traffic data of 15 mutually adjacent geographical regions provided in the embodiment of the present application is shown. In the figure, region represents a region, and region2022 is a region number.

[0052] like Figure 2As shown in the figure, most adjacent areas have high positive correlation, such as 2025 and 2026, 2024 and 2124, etc., which indicates that the traffic patterns of these areas are highly consistent. This consistency may be due to geographical proximity or similar functional zoning, resulting in convergence of user behavior and network demand. In addition, some adjacent areas have low correlation, such as 2123 and 2124, which indicates that there are significant differences in the traffic patterns of these areas. Although these areas are geographically adjacent, their traffic peak periods and usage patterns are different due to their different functions (for example, one is a commercial area and the other is a residential area), resulting in low traffic correlation. For non-adjacent areas, it can be seen from the heat map that most non-adjacent areas show strong negative correlation, such as 2122 and 2226, which indicates that the traffic change trends of these areas are opposite. This reflects the different functions of the areas, such as one is an office area and the other is a residential area, resulting in traffic peaks in different periods of time for the two. On the other hand, a few non-adjacent areas show strong positive correlation, such as 2022 and 2223, although they are not geographically adjacent, their traffic patterns are highly consistent. This phenomenon may be due to the fact that these areas have similar user behavior patterns or share similar network requirements. For example, they are both commercial areas or office areas. Although they are physically far apart, they still show synchronized traffic change trends.

[0053] Through the above analysis, we can draw the following conclusions: 1) Geographically adjacent areas usually have high similarity in network traffic, which is due to the similarity of user behavior and activity patterns between adjacent areas. For example, in a city, users frequently move between adjacent areas, resulting in strong correlation in the network traffic patterns of these areas. At the same time, the coverage of communication base stations often spans multiple adjacent grid areas, which makes the traffic fluctuations of adjacent grids show synchronization to a certain extent.

[0054] 2) The spatial correlation of network traffic can also be manifested between different functional areas. For example, the traffic patterns of commercial and residential areas often show interdependence characteristics within a specific time period. During the day, the traffic in commercial areas is usually high, due to a large number of communication needs caused by work and business activities, while the adjacent residential areas may have lower network traffic due to people working outside; on the contrary, at night, as people return to the residential areas, the network traffic in the residential areas increases, while the traffic in the commercial areas decreases significantly. This linkage of traffic patterns based on functional area division reflects the spatial correlation between different areas.

[0055] 3) The distribution of communication infrastructure is also a key factor affecting spatial correlation. The layout and coverage of base stations directly determine the network connection quality and traffic distribution in different areas. During network peak hours, some base stations may suffer from overload, resulting in reduced service quality, forcing users to connect to base stations in adjacent areas, which in turn causes traffic transfer and redistribution in space.

[0056] Based on this, the embodiments of the present application use geographical proximity, user mobility patterns, coverage of communication infrastructure, and differences in regional functions to jointly characterize the spatial correlation of network traffic data. The spatial correlation of network traffic data is an important feature in network traffic prediction, which not only reflects the traffic linkage effect between regions, but also provides a basis for optimizing network resource allocation and improving traffic prediction accuracy.

[0057] Example 3: Network traffic prediction method based on graph convolutional neural network:

[0058] Based on the periodic characteristic analysis results and spatial correlation analysis results of the above network traffic data, the network traffic data of the same time period of the previous week, the same time period of the previous day, and the first two hours of the current day in the geographical area with spatial correlation with the target area are used as the target time series data to predict the network traffic of the target area in the next two hours. Please refer to Figure 3 The flowchart of the network traffic prediction method based on graph convolutional neural network provided by the embodiment of the present application is shown, and the specific steps are as follows:

[0059] S101. Represent the network traffic data as a byte traffic graph.

[0060] In this step, the network data traffic is represented as a graph structure, namely, a byte traffic graph.

[0061] In specific implementation, the byte flow graph can be obtained by the following methods:

[0062] Step 1011: Use the bytes in the network traffic data as nodes of the byte traffic graph.

[0063] In this step, each node of the byte flow graph represents a byte in the network flow data, which are bytes in the header or payload of the network flow data packet.

[0064] Step 1012: Use the correlation between byte pairs in the network traffic data as the edges between nodes in the byte traffic graph, and use an adjacency matrix to represent the connection relationship between the nodes.

[0065] In this step, in order to better describe the relationship between bytes, an adjacency matrix is ​​introduced to represent the connection relationship between nodes. Specifically, the elements of the adjacency matrix represent whether there is an edge connection between the node pairs. If the element value is 1, it means that there is an edge between the node pairs, otherwise there is no edge connection.

[0066] Specifically, the adjacency matrix of the byte flow graph is obtained in the following way:

[0067] Step 201: For each byte pair, the distance similarity of the byte pair is calculated by counting the frequency of the byte pair appearing simultaneously in the sliding window.

[0068] Step 202: Calculate the first-order difference of each byte in the sliding window to obtain the numerical similarity of the byte pairs.

[0069] Step 203: Calculate the association value of the byte pair based on the distance similarity and the numerical similarity, and transform the association value into the range of [0, 1].

[0070] Step 204: Compare the association value of the transformed byte pair with a predetermined threshold value to retain or discard the byte pair, thereby obtaining the adjacency matrix.

[0071] Step 1013: Use the original value of the byte in the network traffic data as the feature of the node in the byte traffic graph.

[0072] S102, using average pooling and maximum pooling methods, performing multi-level feature extraction on the byte flow graph to obtain coarse-grained change trends and fine-grained change trends of the time series data.

[0073] In this step, by combining average pooling and maximum pooling techniques, multi-level feature extraction is performed on the byte traffic graph to improve the performance of the model in network traffic prediction.

[0074] In specific implementation, the coarse-grained change trend and fine-grained change trend of time series data can be obtained by the following methods:

[0075] Step 1021: using the average pooling technique, down-sample the time series data by calculating the average value in a local window to obtain first feature data that can reflect global trend information.

[0076] In this step, average pooling downsamples the sequence by calculating the average value in the local window, thereby effectively reducing the size of the sequence features, reducing the computational complexity, and smoothing the sequence features to a certain extent, weakening the impact of noise, and retaining global trend information. Average pooling weakens local fluctuations and retains global trend information, thereby enhancing the robustness of the model when processing complex data with more noise.

[0077] Step 1022: using the maximum pooling technique, down-sample the time series data by selecting the maximum value in the pooling window to obtain second feature data that can highlight local features.

[0078] In this step, max pooling is performed by downsampling by selecting the maximum value in the pooling window, which can not only reduce the length of the time series and reduce the amount of calculation, but also highlight the most significant features in the local area and strengthen the extreme value information in the time series. Max pooling is particularly suitable for extracting key features in the data and highlighting those values ​​in the local area that play a decisive role in the overall trend of the sequence. By highlighting local features, max pooling can effectively capture the extreme value information in the sequence, making the model more sensitive and accurate in dealing with extreme situations such as outliers or peaks.

[0079] Step 1023: Calculate the difference between the second feature data and the first feature data to obtain a coarse-grained change trend to characterize the global features of the time series data.

[0080] In this step, the coarse-grained extreme value trend analysis reveals the relationship between the extreme value and the average value of the time series at the macro level by calculating the difference between the maximum pooling value and the average pooling value, that is, the difference between the second feature data and the first feature data. In this way, the model can capture the deviation between the extreme value and the average level of the sequence at the global level, analyze the global trend, and identify the overall change pattern of the time series.

[0081] Step 1024: Calculate the difference between the second feature data and the time series data to obtain a fine-grained change trend to characterize the local features of each data point in the time series data.

[0082] In this step, the fine-grained extreme value trend reveals the degree of deviation of each data point from the local maximum value by calculating the difference between the maximum pooled value and the original sequence value, that is, the difference between the second feature data and the time series data. Fine-grained analysis can show the relative relationship between each data point and the extreme value in its window in more detail, reflecting the specific difference of each data point in the sequence relative to the extreme value, and is suitable for fine analysis of local features.

[0083] S103: splicing the coarse-grained change trend and the fine-grained change trend to obtain a sequence feature graph of the time series data.

[0084] S104. Based on a graph convolutional neural network, combined with a long short-term memory network and a gated recurrent unit, a graph convolution operation is performed on the sequence feature graph to extract the spatiotemporal features of the time series data.

[0085] In this step, a graph neural network model is constructed based on a graph convolutional neural network, combined with a long short-term memory network and a gated recurrent unit to predict network traffic.

[0086] In specific implementation, the spatiotemporal features of the time series data can be extracted in the following ways:

[0087] Step 1041: Based on a single graph convolutional network and a long short-term memory network, use the adjacency matrix of the time series data to perform a graph convolution operation on the sequence feature graph to obtain a first spatiotemporal feature.

[0088] In this step, the adjacency matrix of the time series data is first used to perform weighted aggregation on the sequence feature graph; then the GLU activation function is used to perform nonlinear transformation on the aggregated features to extract effective spatial features; the time series data is then processed using a long short-term memory network to further capture the temporal characteristics of network traffic. Here, the long short-term memory network can effectively handle the long-term dependency problems in the time series data, which helps to improve the accuracy of the prediction; finally, the spatial features and temporal features extracted separately are concatenated to obtain the first spatiotemporal features as the input of the subsequent model.

[0089] Step 1042: Based on the multi-graph convolutional network and the gated recurrent unit, a graph convolution operation is performed on the sequence feature graph using an adjacency matrix that takes into account the fluctuation of the time series data to obtain a second spatiotemporal feature.

[0090] The fluctuation conditions include: trough, stability and peak;

[0091] In this step, the gated recurrent unit is an activation function used in neural networks. The core idea is to introduce a gating mechanism to selectively control the input features through the Sigmoid function to adjust part of the input output; that is, use part of the input as the main information flow, and the other part of the input generates a gating signal through the Sigmoid function to selectively filter the information flow. This gating mechanism enables the model to capture fine-grained dependency features, and enhances the model's ability to handle complex dependencies by selectively retaining important information and suppressing noise.

[0092] The core of the multi-graph convolutional network is to select the adjacency matrix corresponding to the fluctuation of the input features and perform multi-graph convolution operations to further extract spatial features.

[0093] Specifically, an additional parameter is set in the input to select a specific fluctuation situation graph. During the training process, the model will automatically select the corresponding fluctuation situation association graph for convolution operation according to which fluctuation situation the time period of the traffic data to be predicted belongs to. The model extracts the spatial features of the traffic data under different fluctuation characteristics. The gated recurrent unit is used to extract the temporal features, and then the output results of the multi-graph convolution and the gated recurrent unit are spliced ​​to obtain the second spatiotemporal features.

[0094] In specific implementation, the fluctuation of time series data can be determined in the following ways:

[0095] Step 301: Calculate the average value of the network traffic data at each timestamp.

[0096] Step 302: Calculate the standard deviation of the network traffic data in the sliding window to reflect the fluctuation of the time series data.

[0097] In this step, the standard deviation reflects the variation of network traffic data. The larger the standard deviation, the greater the fluctuation of network traffic data and the greater the deviation of data value from the average value. The smaller the standard deviation, the relatively stable network traffic data and the closer the data value is to the average value.

[0098] It should be noted that the sliding window size is set to 12, and the standard deviation of each time point is calculated based on the network traffic data of the current time point and the 11 time points after it, for a total of 12 time points (i.e., 2 hours). That is, the fluctuation at the current time records the fluctuation in the next 2 hours. This way of recording future fluctuations helps predict future fluctuations in network traffic data.

[0099] Step 303: using quartiles, the time series data is divided into valley data, stable data and peak data according to the fluctuation.

[0100] In this step, the quartile is a statistic that describes the distribution of data and helps to understand the distribution characteristics of the data. Quartiles are usually divided into three key values: the first quartile (Q1), the second quartile (Q2), and the third quartile (Q3).

[0101] Specifically, Q1 is the value of the first quarter (smallest 25%) of the data points in the data set, Q2 is the median of the data set, and Q3 is the value of the last quarter of the data points in the data set. According to the degree of fluctuation, the quartiles are used to set the fluctuation threshold, that is, the boundary between the peak and trough periods, and the time series data is divided into three types of trough data, stable data, and peak data in a ratio of 1:2:1. The fluctuations in each area at the same time are the same. By using quartiles to reasonably divide the time period and using the distribution characteristics of the time series data itself to determine the classification criteria, rather than relying on fixed thresholds, the accuracy and generalization of network traffic prediction can be improved.

[0102] Furthermore, the adjacency matrix of the time series data and the adjacency matrix considering the fluctuation of the time series data can be obtained in the following ways:

[0103] Step 501: Calculate the similarity between regions based on time series data, valley data, stable data and peak data using a fast dynamic regularization distance algorithm.

[0104] Wherein, each of the regions includes the target region and a geographical region having spatial correlation with the target region:

[0105] In this step, the fast dynamic regularization distance algorithm reduces the computational complexity through hierarchical approximation and multi-resolution methods.

[0106] Specifically, the calculation method of the fast dynamic time warping distance algorithm is as follows:

[0107] Step 601: simplify the two time series data, downsample them layer by layer, and generate multi-level resolution representation.

[0108] As an example, every two data points can be merged into one data point to obtain a shorter sequence, which can form a series of multiple hierarchical sequences from the original sequence to the most simplified sequence.

[0109] Step 602: At the lowest resolution level, use the dynamic time warping distance algorithm to find a preliminary alignment path for the two time series data.

[0110] Step 603: After mapping the preliminary alignment path to the previous level, the preliminary alignment path is adjusted using a dynamic time warping distance algorithm until an optimal alignment path is found to characterize the similarity of the two time series data.

[0111] In this step, after multiple levels of adjustments, a final fine-tuning is performed at the original and highest resolution levels to obtain the approximate best matching path between the two time series data as the best alignment path.

[0112] Step 502: construct adjacency matrices of time series data, trough data, stable data and peak data respectively according to the similarity between the target area and a geographical area having spatial correlation with the target area.

[0113] Step 1043: Concatenate the first spatiotemporal feature and the second spatiotemporal feature to obtain the spatiotemporal feature of the time series data.

[0114] S105 , after performing layer-by-layer linear transformation on the spatial features of the time series data, output the network traffic prediction result of the target area.

[0115] The above contents are only preferred embodiments of the present invention. For ordinary technicians in this field, many changes can be made in the specific implementation methods and application scopes based on the ideas of the present technical content. As long as these changes do not deviate from the concept of the present invention, they all fall within the scope of protection of this patent.

Claims

1. A network traffic prediction method based on graph convolutional neural network, characterized in that: The method comprises: Based on the periodic characteristics of network traffic data, the network traffic data of the same time period of the previous week, the same time period of the previous day, and the first two hours of the current day in a geographical area with spatial correlation with the target area are used as target time series data to predict the network traffic of the target area in the next two hours. Specifically: representing the network traffic data as a byte traffic graph; Using average pooling and maximum pooling methods, multi-level feature extraction is performed on the byte flow graph to obtain coarse-grained change trends and fine-grained change trends of the time series data; The coarse-grained change trend and the fine-grained change trend are spliced ​​to obtain a sequence feature graph of the time series data; Based on a graph convolutional neural network, combined with a long short-term memory network and a gated recurrent unit, a graph convolution operation is performed on the sequence feature graph to extract the spatiotemporal features of the time series data; After performing layer-by-layer linear transformation on the spatiotemporal features of the time series data, the network traffic prediction result of the target area is output.

2. The method according to claim 1, characterized in that The step of representing the network traffic data as a byte traffic graph comprises: Using bytes in the network flow data as nodes of the byte flow graph; Using the correlation between byte pairs in the network traffic data as the edges between nodes in the byte traffic graph, and using an adjacency matrix to represent the connection relationship between nodes; The original values ​​of bytes in the network traffic data are used as features of nodes in the byte traffic graph.

3. The method according to claim 2, characterized in that The adjacency matrix of the byte flow graph is obtained in the following way: For each byte pair, the distance similarity of the byte pair is calculated by counting the frequency of the byte pair appearing simultaneously in the sliding window; Calculate the first-order difference of each byte in the sliding window to obtain the numerical similarity of the byte pairs; Based on the distance similarity and the numerical similarity, calculating the association value of the byte pair, and transforming the association value into the range of [0, 1]; The association value of the transformed byte pair is compared with a predetermined threshold to retain or discard the byte pair, thereby obtaining the adjacency matrix.

4. The method according to claim 1, characterized in that The method of using average pooling and maximum pooling to extract multi-level features from the time series data includes: Using average pooling technology, by calculating the average value in a local window, the time series data is downsampled to obtain first feature data that can reflect global trend information; Using the maximum pooling technique, by selecting the maximum value in the pooling window, the time series data is downsampled to obtain second feature data that can highlight local features; Calculating the difference between the second feature data and the first feature data to obtain a coarse-grained change trend to characterize the global features of the time series data; The difference between the second feature data and the time series data is calculated to obtain a fine-grained change trend to characterize the local features of each data point in the time series data.

5. The method according to claim 1, characterized in that The spatial correlation is determined by geographic proximity, user mobility patterns, coverage of communication infrastructure and regional functional differences.

6. The method according to claim 1, characterized in that The graph convolution neural network is based on a long short-term memory network and a gated recurrent unit, and a graph convolution operation is performed on the sequence feature graph to extract the spatiotemporal features of the time series data, including: Based on a single graph convolution network and a long short-term memory network, an adjacency matrix of the time series data is used to perform a graph convolution operation on the sequence feature graph to obtain a first spatiotemporal feature; Based on a multi-graph convolutional network and a gated recurrent unit, a graph convolution operation is performed on the sequence feature graph using an adjacency matrix that considers the fluctuation of the time series data to obtain a second spatiotemporal feature; wherein the fluctuation includes: a trough, a stable state, and a peak; The first spatiotemporal feature and the second spatiotemporal feature are concatenated to obtain the spatiotemporal feature of the time series data.

7. The method according to claim 6, characterized in that Determine the volatility of time series data by: Calculate the average value of network traffic data for each timestamp; Calculate the standard deviation of the network traffic data within the sliding window to reflect the fluctuation of the time series data; The time series data is divided into trough data, stable data and peak data according to the fluctuation by using quartiles.

8. The method according to claim 7, characterized in that The adjacency matrix of the time series data and the adjacency matrix considering the fluctuation of the time series data are obtained in the following ways: Using a fast dynamic regularization distance algorithm, based on time series data, valley data, stable data and peak data, the similarity between each region is calculated respectively; wherein each region includes the target region and a geographical region having spatial correlation with the target region: According to the similarity between the target area and the geographical area having spatial correlation with the target area, adjacency matrices of time series data, trough data, stable data and peak data are constructed respectively.

9. The method according to claim 8, characterized in that The calculation method of the fast dynamic time warping distance algorithm is as follows: Simplify the two time series data, downsample them layer by layer, and generate multi-level resolution representations; At the lowest resolution level, the dynamic time warping distance algorithm is used to find the preliminary alignment path of the two time series data; After mapping the preliminary alignment path to the previous level, the dynamic time warping distance algorithm is used to adjust the preliminary alignment path until the optimal alignment path is found to characterize the similarity of the two time series data.

Images