Traffic prediction method and device, equipment and storage medium

Abstract

The invention provides a traffic prediction method and device, equipment and a storage medium. The method comprises the following steps: acquiring flow information, wherein the flow information comprises base station information and flow data corresponding to the base station information; according to the traffic information and a preset deep neural network model, predicted traffic information is obtained, the deep neural network model is based on space-time convolution block stacking and a time sequence decomposition network, space-time convolution blocks obtain first time sequence information according to the traffic information, first predicted traffic information is obtained, and output is obtained through residual connection; the time sequence decomposition network obtains second time sequence information and second prediction flow information according to the flow information, the first time sequence information comprises time feature information and spatial feature information in the flow information, and the second time sequence information comprises trend information and season information in the flow information; the predicted traffic information comprises first predicted traffic information and second predicted traffic information. And through combination of space-time convolution and time sequence decomposition, the flow prediction accuracy is improved.

Description

Technical Field

[0001] The present application relates to the field of communication technology, and in particular to a traffic prediction method, device, equipment and storage medium. Background Art

[0002] With the popularity of smartphones and the development of 5G communications, mobile traffic has exploded, posing a huge challenge to the resource allocation and maintenance of base stations. To solve the overload and tidal phenomena of base stations, the traditional method is for operation and maintenance personnel to pay attention to the traffic changes of each base station in real time and manually configure the carrier frequency switch, which is inefficient and ineffective. Therefore, base station traffic prediction is needed to achieve reasonable deployment of operation and maintenance personnel. Base station traffic prediction can essentially be regarded as a periodic non-stationary time series prediction. At the same time, it will be affected by factors such as weekends, holidays, and social events, and there is non-stationarity and uncertainty.

[0003] At present, the existing time series prediction methods mainly include: multi-time step prediction method based on lag feature derivative: convert multi-time step prediction into multiple single time step prediction problems, construct multiple lag features, sliding window features, etc. for each sample, try to model the relationship between the target value and the historical features, and make predictions based on this, and use the prediction results of the time step as input each time to predict the output of the next time step, multi-time step prediction method based on recurrent neural network (RNN), multi-time step prediction method based on temporal convolutional network (TCN): by stacking multiple layers to achieve continuous expansion of the receptive field, and using void convolution to effectively reduce the number of parameters.

[0004] However, existing traffic prediction results are inaccurate and require a lot of manual participation. Summary of the invention

[0005] The present application provides a flow prediction method, device, equipment and storage medium to solve the problem that flow prediction results are inaccurate and require a lot of manual participation.

[0006] In a first aspect, the present application provides a flow prediction method, comprising:

[0007] Acquire flow information, where the flow information includes base station information and flow data corresponding to the base station information;

[0008] The predicted traffic information is obtained according to the traffic information and a preset deep neural network model, wherein the deep neural network model is constructed based on a spatiotemporal convolutional network and a time series decomposition network, the spatiotemporal convolutional network includes a plurality of spatiotemporal convolutional blocks, the spatiotemporal convolutional blocks are used to obtain first time series information according to the traffic information, and obtain first predicted traffic information according to the first time series information, the output of the spatiotemporal convolutional network is obtained by integrating the output of each spatiotemporal convolutional block by residual connection, the time series decomposition network is used to obtain second time series information according to the traffic information, and obtain second predicted traffic information according to the second time series information, the first time series information includes time feature information and spatial feature information in the traffic information, the second time series information includes trend information and seasonal information in the traffic information, and the predicted traffic information includes first predicted traffic information and second predicted traffic information.

[0009] Among them, in this application, obtaining traffic information includes:

[0010] Acquire initial flow information, the initial flow information including base station information and initial flow data corresponding to the base station information;

[0011] Perform dimensional processing on the initial flow data to obtain the target flow data;

[0012] Obtain flow data according to the preset ratio and target flow data;

[0013] The flow information is obtained according to the base station information and flow data.

[0014] Among them, in this application, obtaining initial flow information includes:

[0015] Obtain target historical traffic information, where the target historical traffic information is the original traffic data corresponding to the base station information and within a preset time range;

[0016] If there are missing data in the original traffic data, clean the missing data to obtain the initial traffic data;

[0017] Initial flow information is obtained according to the initial flow data and base station information.

[0018] Among them, in the present application, before obtaining predicted traffic information according to the traffic information and the preset deep neural network model, the method also includes:

[0019] Acquire a training traffic sample, where the training traffic sample includes training base station information and training traffic sample data corresponding to the training base station information;

[0020] Perform two-dimensional average pooling on the training flow samples to obtain trend information and seasonal information;

[0021] According to the trend information and the seasonal information, the parameters of the initial time series decomposition network are adjusted to obtain the time series decomposition network;

[0022] Construct a spatiotemporal convolutional network based on training traffic samples;

[0023] Based on the temporal decomposition network and the spatiotemporal convolutional network, a deep neural network model is obtained.

[0024] Among them, in this application, a spatiotemporal convolutional network is constructed based on training traffic samples, including:

[0025] Extracting temporal and spatial feature information from training traffic samples;

[0026] Obtain target feature data according to the training flow sample data, time feature information and space feature information;

[0027] Obtaining an output result of the graph convolution unit according to the target feature data, a preset adaptive adjacency matrix, an adaptive adjacency matrix transpose corresponding to the adaptive adjacency matrix, and the graph convolution unit;

[0028] The output result of the graph convolution unit and the target feature data are processed by residual connection to obtain the output result of the initial spatiotemporal convolution block;

[0029] According to the output result of the initial spatiotemporal convolution block, the weight of the adaptive adjacency matrix is ​​adjusted to obtain the spatiotemporal convolution block;

[0030] According to the spatiotemporal convolutional blocks, a spatiotemporal convolutional network is constructed.

[0031] Among them, in the present application, according to the target feature data, the preset adaptive adjacency matrix, the adaptive adjacency matrix transpose corresponding to the adaptive adjacency matrix, and the graph convolution unit, the output result of the graph convolution unit is obtained, including:

[0032] Get the preset adaptive adjacency matrix;

[0033] Transpose the adaptive adjacency matrix to obtain the adaptive adjacency matrix transpose;

[0034] Obtaining a first output result of a first multi-hop graph convolutional layer according to the target feature data and the adaptive adjacency matrix;

[0035] Obtaining a second output result of a second multi-hop graph convolutional layer according to the target feature data and the adaptive adjacency matrix transposition;

[0036] An output result of the graph convolution unit is obtained according to a first output result of the first multi-hop graph convolution layer and a second output result of the second multi-hop graph convolution layer.

[0037] Among them, in the present application, after obtaining the deep neural network model according to the temporal decomposition network and the spatiotemporal convolutional network, the method further includes:

[0038] Get the loss function of the deep neural network model;

[0039] If the loss function converges, the trained deep neural network model is obtained;

[0040] If the loss function has not converged, other training traffic samples are obtained and used as training traffic samples, and the steps of performing two-dimensional average pooling on the training traffic samples to obtain trend information and seasonal information are re-executed until the loss function converges to obtain a trained deep neural network model.

[0041] Among them, in the present application, after obtaining predicted traffic information according to the traffic information and the preset deep neural network model, the method also includes:

[0042] Acquire verification traffic information and target predicted traffic information, where the target predicted traffic information is information in the predicted traffic information corresponding to the time of the verification traffic information;

[0043] According to the verification flow information and the target predicted flow information, the mean absolute error information, the mean absolute percentage error information, and the root mean square error information are obtained;

[0044] According to the mean absolute error information, the mean absolute percentage error information, and the root mean square error information, an evaluation result of the predicted flow information is obtained;

[0045] If the evaluation result of the predicted traffic information indicates that the predicted traffic information is accurate, the deep neural network model is determined to be the target deep neural network model.

[0046] In a second aspect, the present application provides a flow prediction device, comprising:

[0047] An acquisition module, for acquiring flow information, wherein the flow information includes base station information and flow data corresponding to the base station information;

[0048] A module is obtained, and predicted traffic information is obtained according to traffic information and a preset deep neural network model, wherein the deep neural network model is constructed based on a space-time convolutional network and a time series decomposition network, and the space-time convolutional network includes multiple space-time convolutional blocks, and the space-time convolutional block is used to obtain first time series information according to the traffic information, and obtain first predicted traffic information according to the first time series information, and the output of the space-time convolutional network is obtained by integrating the output of each space-time convolutional block through residual connection, and the time series decomposition network is used to obtain second time series information according to the traffic information, and obtain second predicted traffic information according to the second time series information, the first time series information includes time feature information and spatial feature information in the traffic information, the second time series information includes trend information and seasonal information in the traffic information, and the predicted traffic information includes first predicted traffic information and second predicted traffic information.

[0049] In a third aspect, the present application provides an electronic device, comprising: a processor, and a memory communicatively connected to the processor;

[0050] Memory stores computer-executable instructions;

[0051] The processor executes the computer-executable instructions stored in the memory to implement the method in the present application.

[0052] In a fourth aspect, the present application provides a computer-readable storage medium, in which computer-executable instructions are stored. When the computer-executable instructions are executed by a processor, they are used to implement the method of the present application.

[0053] The present application provides a traffic prediction method, apparatus, device and storage medium, which obtains target historical traffic, cleans missing data and obtains initial traffic information; performs dimension processing on the initial traffic information, and then obtains traffic information according to a preset ratio, wherein the traffic information includes base station information and traffic information corresponding to the base station; obtains predicted traffic information according to the traffic information and a preset deep neural network model, wherein the deep neural network model is constructed based on a spatiotemporal convolutional network and a time series decomposition network, the spatiotemporal convolutional network includes a plurality of spatiotemporal convolutional blocks, the spatiotemporal convolutional blocks are used to obtain first time series information according to the traffic information, and obtain first predicted traffic information according to the first time series information, the output of the spatiotemporal convolutional network is obtained by integrating the output of each spatiotemporal convolutional block by residual connection, the time series decomposition network is used to obtain second time series information according to the traffic information, and obtain first predicted traffic information according to the second time series information The first time series information is used to obtain the second predicted traffic information. The first time series information includes the time feature information and the space feature information in the traffic information. The second time series information includes the trend information and the season information in the traffic information. The predicted traffic information includes the first predicted traffic information and the second predicted traffic information. By constructing a time series decomposition block to mine the trend information and the season information and a time-space convolution block to mine the time information and the space information, when the time series information itself can be used, other effective information such as weather, holidays, and seasons can be introduced for traffic prediction to improve the prediction accuracy. At the same time, by constructing an adaptive matrix based on the attention mechanism, the correlation between different base stations is obtained in a graphical manner, thereby improving the accuracy of traffic prediction. At the same time, the time series decomposition and the adaptive adjacency matrix are used to construct a deep learning model, the historical indicators of the base stations are analyzed and a data set is constructed, and multiple rounds of historical feature parameters are optimized to avoid the problem of requiring a large amount of manual participation in constructing a large number of historical features. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0055] Figure 1 A flow chart of a flow prediction method provided in an embodiment of the present application;

[0056] Figure 2 A flow chart of another flow prediction method provided in an embodiment of the present application;

[0057] Figure 3 A schematic diagram of a flow prediction structure provided in an embodiment of the present application;

[0058] Figure 4 A schematic diagram of the electronic structure provided in an embodiment of the present application.

[0059] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0060] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0061] In the prior art, the multi-time-step prediction method derived from the lag feature converts the multi-time-step prediction into multiple single-time-step prediction problems or uses the prediction result of the time step as input each time to predict the output of the next time step. This type of method will cause error accumulation and inaccurate prediction. The multi-time-step prediction method based on the recurrent neural network fully considers the sequence order and continuously updates the hidden layer state during training. However, the disadvantage is that the number of parameters is large and the convergence is slow. When the input sequence is too long, it is easy to have gradient explosion or gradient diffusion problems. The time domain convolutional network uses the convolution method to process sequence data, and realizes the continuous expansion of the receptive field by superimposing multiple layers, and uses the dilated convolution to effectively reduce the number of parameters. However, the disadvantage is that this type of dilated convolution cannot fully utilize the sequence information, resulting in inaccurate prediction results. The prior art requires a large number of manual participation to construct a large number of historical features, has large parameters and slow convergence, does not consider the spatial relationship of time series prediction, and cannot introduce other information except the time series itself, resulting in inaccurate prediction results.

[0062] In order to solve the above problems, the present application provides a traffic prediction method, device, equipment and storage medium, which obtains the target historical traffic, cleans the missing data, and obtains the initial traffic information; performs dimensional processing on the initial traffic information, and then obtains the traffic information according to a preset ratio, wherein the traffic information includes base station information and the traffic information corresponding to the base station; obtains the predicted traffic information according to the traffic information and a preset deep neural network model, wherein the deep neural network model is constructed based on a spatiotemporal convolutional network and a time series decomposition network, the spatiotemporal convolutional network includes multiple spatiotemporal convolutional blocks, the spatiotemporal convolutional blocks are used to obtain first time series information according to the traffic information, and obtain first predicted traffic information according to the first time series information, the output of the spatiotemporal convolutional network is obtained by integrating the output of each spatiotemporal convolutional block by residual connection, the time series decomposition network is used to obtain second time series information according to the traffic information, and obtain first predicted traffic information based on the first time series information. According to the second time series information, the second predicted traffic information is obtained, the first time series information includes the time feature information and the space feature information in the traffic information, the second time series information includes the trend information and the season information in the traffic information, and the predicted traffic information includes the first predicted traffic information and the second predicted traffic information; by constructing a time series decomposition block to mine the trend information and the season information and a time-space convolution block to mine the time information and the space information, when the time series information itself can be used, other effective information such as weather, holidays, and seasons can be introduced for traffic prediction to improve the prediction accuracy, and at the same time, by constructing an adaptive matrix based on the attention mechanism, the correlation between different base stations is obtained in a graphical manner, thereby improving the accuracy of traffic prediction; at the same time, the time series decomposition and the adaptive adjacency matrix construct a deep learning model, analyze the historical indicators of the base stations and construct a data set, and optimize the historical feature parameters through multiple rounds to avoid the problem of requiring a large amount of manual participation in constructing a large number of historical features.

[0063] The data transmission method provided in this application is intended to solve the above technical problems in the prior art.

[0064] The technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0065] The execution subject of the determination method provided in the embodiment of the present application may be a server. Among them, the server may be a mobile phone, a tablet, a computer and other devices. This embodiment does not impose any special restrictions on the implementation method of the execution subject, as long as the execution subject can obtain traffic information, the traffic information includes base station information, and traffic data corresponding to the base station information; according to the traffic information and the preset deep neural network model, the predicted traffic information is obtained, wherein the deep neural network model is constructed based on the spatiotemporal convolution network and the time series decomposition network, the spatiotemporal convolution network includes a plurality of spatiotemporal convolution blocks, the spatiotemporal convolution blocks are used to obtain the first time series information according to the traffic information, and obtain the first predicted traffic information according to the first time series information, the output of the spatiotemporal convolution network is obtained by integrating the output of each spatiotemporal convolution block through a residual connection, the time series decomposition network is used to obtain the second time series information according to the traffic information, and obtain the second predicted traffic information according to the second time series information, the first time series information includes the time feature information and the spatial feature information in the traffic information, the second time series information includes the trend information and the seasonal information in the traffic information, and the predicted traffic information includes the first predicted traffic information and the second predicted traffic information.

[0066] First, the terms involved in this application are explained:

[0067] Time series decomposition: Time series decomposition is a method to explore the law of time series changes, mainly exploring periodicity and trend. Based on the results of time series decomposition, we can perform subsequent time prediction and anomaly detection.

[0068] Residual connection: Residual connection is one of the cores of deep residual learning, which improves the performance of neural networks. Many problems in deep learning, such as gradient disappearance and network degradation, can be solved by residual connection. Residual connection, commonly known as shortcut, ensures that the network has the ability to "skip layers" during training by directly adding the input signal to the output signal. Specifically, when constructing a neural network, the residual connection directly uses the output of the previous layer as one of the inputs of the next layer, and the other layers between two convolutional layers in the network are called residual blocks.

[0069] Adjacency matrix: An adjacency matrix is ​​a matrix that represents the connectivity of a graph or network, where each element indicates whether there is a connection between two nodes. An adjacency matrix is ​​a two-dimensional array, where the value of each element indicates whether there is a connection between the corresponding nodes, usually represented by 0 and 1. In graphics and network science, the adjacency matrix is ​​a commonly used data structure.

[0070] Adaptive Adjacency Matrix: Adaptive Adjacency Matrix is ​​an adjacency matrix used in Graph Neural Network (GNN). Adaptive Adjacency Matrix introduces an adaptive mechanism to dynamically organize the neighbor nodes of each node into an adjacency matrix of variable size. Specifically, each row of the adaptive adjacency matrix represents the neighbor nodes of a node, and the length of each row is the degree of this node.

[0071] Graph Convolution Neural Networks: Graph Convolution Neural Networks (GCNNs) methods are divided into two categories, spectral domain-based methods and spatial domain-based methods. Spectral domain-based methods define graph convolution by introducing filters from the perspective of graph signal processing, where the graph convolution operation is interpreted as removing noise from the graph signal. Spatial domain-based methods represent graph convolution as aggregating feature information from neighbors.

[0072] Convolutional layer: Each convolutional layer in a convolutional neural network consists of several convolutional units, and the parameters of each convolutional unit are optimized through the back-propagation algorithm. The purpose of the convolution operation is to extract different features of the input. The first convolutional layer may only extract some low-level features such as edges, lines, and corners. More layers of the network can iteratively extract more complex features from low-level features.

[0073] Time series: A time series is an ordered sequence of variables at different time points. For example, the average daily temperature in Beijing in April 2013 constitutes a time series. For convenience, we generally assume that adjacent elements in the sequence have the same time interval.

[0074] Time Series Prediction: Time series can also be called time series, historical complex or dynamic series. Time Series Prediction (TimeSeries Prediction) is the full name of time series prediction, which is actually the abbreviation of a method for predicting time series. The general definition of time series prediction is to compile and analyze time series, and then combine the development process, direction and trend reflected in the time series to make analogies or extensions, so as to predict the level that may be achieved in the next period of time or in the next few years. Time series prediction is to predict the change and predicted value of the variable in the future period of time or at a certain moment by analyzing the change of a certain quantity in a limited period of time. In addition to time series prediction, research directions on time series problems also include time series classification research, time series anomaly detection research, etc.

[0075] Figure 1 A flow chart of a flow prediction method provided in an embodiment of the present application. Figure 1As shown, the execution subject of the method can be a server or other servers, and this embodiment is not particularly limited here. Figure 1 As shown, the method includes:

[0076] S101. Obtain flow information, where the flow information includes base station information and flow data corresponding to the base station information.

[0077] Among them, traffic information can refer to uplink rate, downlink rate and prb (Physical Resource Block) utilization. Among them, the uplink rate refers to the data transmission rate when the mobile terminal sends information to the base station, such as the data transmission rate of wireless terminals such as mobile phones and laptops to the base station. The downlink rate refers to the transmission rate when the base station sends information to the mobile terminal, such as the rate at which wireless terminals such as mobile phones or laptops download data from the base station or the network. The prb utilization is used to evaluate the utilization of base station resources.

[0078] The base station information may refer to a coverage area centered on the base station as a cell, retaining the OID (Object Identifier) ​​of the cell as the unique identifier of the base station, and the traffic data corresponding to the base station traffic information may refer to specific data corresponding to the traffic information of the base station per hour.

[0079] In this embodiment of the present application, obtaining flow information includes:

[0080] Acquire initial flow information, the initial flow information including base station information and initial flow data corresponding to the base station information;

[0081] Perform dimensional processing on the initial flow data to obtain the target flow data;

[0082] Obtain flow data according to the preset ratio and target flow data;

[0083] The flow information is obtained according to the base station information and flow data.

[0084] Among them, the initial traffic information may refer to the uplink rate, downlink rate and prb (Physical Resource Block) utilization initially obtained from the base station, and the initial traffic data corresponding to the base station information refers to the specific data corresponding to the initial traffic information per hour.

[0085] Dimensional processing refers to standardizing the initial flow data, converting the initial flow data into dimensionless, non-magnitude standardized values, eliminating the impact of different magnitudes between different data, so as to make the results more comparable. In the embodiment of the present application, the dimensional processing can be:

[0086]

[0087] The preset ratio may refer to the proportion of the target traffic data in the training set and the validation set respectively. In some embodiments, the preset ratio may be any pre-set ratio. For example, the preset ratio may be 60% of the target traffic data as the training set and 40% of the target traffic data as the validation set; or 70% of the target traffic data as the training set and 30% of the target traffic data as the validation set.

[0088] By presetting the ratio, the flow data can be selected from the target flow data. Thus, the base station information and the flow data are combined to obtain the flow information.

[0089] In this embodiment of the present application, obtaining initial flow information includes:

[0090] Obtain target historical traffic information, where the target historical traffic information is the original traffic data corresponding to the base station information and within a preset time range;

[0091] If there are missing data in the original traffic data, clean the missing data to obtain the initial traffic data;

[0092] Initial flow information is obtained according to the initial flow data and base station information.

[0093] Among them, the target historical traffic is obtained by presetting the time range and specifying the area, and then obtaining the traffic information obtained at the specified time and the specified area. You can specify the hourly historical traffic information of all base stations in area A in the past month, or you can specify to obtain the hourly historical traffic information of all base stations in area B in the past week, or you can specify to obtain the historical traffic information of all base stations in area C in the past four days.

[0094] The raw traffic data refers to the unprocessed base station traffic data obtained from the base station. The traffic data includes uplink rate data, downlink rate data and prb (Physical Resource Block) utilization data.

[0095] The preset time range can be the past month, the past week, or the past two weeks, and the historical traffic information is obtained once an hour.

[0096] Data cleaning means using the mean of historical data or data from the same period as missing data. As long as the hours of historical data are the same, there is no restriction on the date. For example, if the data at 3 pm on Friday last week is missing, the data at 3 pm on Friday two weeks ago can be used as the data at 3 pm on Friday last week, or the data at 3 pm on Thursday of a week last month can be used as the data at 3 pm on Friday last week.

[0097] S102. Obtain predicted traffic information based on traffic information and a preset deep neural network model, wherein the deep neural network model is constructed based on a spatiotemporal convolutional network and a time series decomposition network, the spatiotemporal convolutional network includes multiple spatiotemporal convolutional blocks, the spatiotemporal convolutional blocks are used to obtain first time series information based on traffic information, and obtain first predicted traffic information based on the first time series information, the output of the spatiotemporal convolutional network is obtained by integrating the output of each spatiotemporal convolutional block through residual connection, the time series decomposition network is used to obtain second time series information based on traffic information, and obtain second predicted traffic information based on the second time series information, the first time series information includes time feature information and spatial feature information in the traffic information, the second time series information includes trend information and season information in the traffic information, and the predicted traffic information includes first predicted traffic information and second predicted traffic information.

[0098] In the embodiment of the present application, before obtaining predicted traffic information according to the traffic information and the preset deep neural network model, the method further includes:

[0099] Acquire a training traffic sample, where the training traffic sample includes training base station information and training traffic sample data corresponding to the training base station information;

[0100] Perform two-dimensional average pooling on the training flow samples to obtain trend information and seasonal information;

[0101] According to the trend information and the seasonal information, the parameters of the initial time series decomposition network are adjusted to obtain the time series decomposition network;

[0102] Construct a spatiotemporal convolutional network based on training traffic samples;

[0103] Based on the temporal decomposition network and the spatiotemporal convolutional network, a deep neural network model is obtained.

[0104] The training base station information refers to the base station unique identifier using the cell OID, and the training traffic sample data includes uplink rate data, downlink rate data, and PRB utilization data.

[0105] Two-dimensional average pooling, that is, taking the average value in the pooling area as the feature value output.

[0106] Time series can refer to training traffic samples, which characterize the changing trend. The time series is also affected by seasons and holidays. Here, the training traffic samples are obtained by performing two-dimensional average pooling by specifying the sliding window size for each base station node to obtain trend information and seasonal information. Then some external parameters are introduced. The external parameters include some simple functions, such as matrix transformation functions. The trend information and seasonal information are learned in blocks through the functions, and finally the trend information and seasonal information are output from two channels respectively, and the outputs are integrated. Among them, multiple convolution kernels of different sizes can be used as sliding windows, which can be four convolution kernels of different sizes.

[0107] The time series decomposition network continuously learns the trend information and seasonal information in the time series, and the parameters that are continuously adjusted are the trend information and seasonal information.

[0108] Among them, in the embodiment of the present application, a spatiotemporal convolutional network is constructed according to the training traffic samples, including:

[0109] Extracting temporal and spatial feature information from training traffic samples;

[0110] Obtain target feature data according to the training flow sample data, time feature information and space feature information;

[0111] Obtaining an output result of the graph convolution unit according to the target feature data, a preset adaptive adjacency matrix, an adaptive adjacency matrix transpose corresponding to the adaptive adjacency matrix, and the graph convolution unit;

[0112] The output result of the graph convolution unit and the target feature data are processed by residual connection to obtain the output result of the initial spatiotemporal convolution block;

[0113] According to the output result of the initial spatiotemporal convolution block, the weight of the adaptive adjacency matrix is ​​adjusted to obtain the spatiotemporal convolution block;

[0114] According to the spatiotemporal convolutional blocks, a spatiotemporal convolutional network is constructed.

[0115] The target feature data is the traffic data with time feature information and space feature information in the training traffic samples.

[0116] Among them, in the embodiment of the present application, according to the target feature data, the preset adaptive adjacency matrix, the adaptive adjacency matrix transpose corresponding to the adaptive adjacency matrix, and the graph convolution unit, the output result of the graph convolution unit is obtained, including:

[0117] Get the preset adaptive adjacency matrix;

[0118] Transpose the adaptive adjacency matrix to obtain the adaptive adjacency matrix transpose;

[0119] Obtaining a first output result of a first multi-hop graph convolutional layer according to the target feature data and the adaptive adjacency matrix;

[0120] Obtaining a second output result of a second multi-hop graph convolutional layer according to the target feature data and the adaptive adjacency matrix transposition;

[0121] An output result of the graph convolution unit is obtained according to a first output result of the first multi-hop graph convolution layer and a second output result of the second multi-hop graph convolution layer.

[0122] The target characteristic data may refer to traffic data with spatial characteristic information.

[0123] The adjacency matrix is ​​obtained by introducing the embedding layer and initializing the self-attention mechanism. The adjacency matrix and the target feature data are added to the graph convolution unit for learning. The adjacency matrix achieves an adaptive effect, that is, an adaptive adjacency matrix is ​​obtained.

[0124] If the adaptive adjacency matrix is ​​obtained by training the traffic data from a certain base station A to a certain base station B, then the adaptive adjacency matrix transpose corresponding to the adaptive adjacency matrix is ​​obtained by training the traffic from base station B to base station A.

[0125] Among them, in the embodiment of the present application, after obtaining the deep neural network model according to the temporal decomposition network and the spatiotemporal convolutional network, the method further includes:

[0126] Get the loss function of the deep neural network model;

[0127] If the loss function converges, the trained deep neural network model is obtained;

[0128] If the loss function has not converged, other training traffic samples are obtained and used as training traffic samples, and the steps of performing two-dimensional average pooling on the training traffic samples to obtain trend information and seasonal information are re-executed until the loss function converges to obtain a trained deep neural network model.

[0129] Among them, the convergence of the loss function indicates that the training parameters of the deep neural network model are accurate and the training is completed. By obtaining the loss of the current round of model training function and the loss of the previous round of model training function, the absolute value of the difference between the two model training function losses is obtained, and it is judged whether the absolute value is within the preset value. If it is within the preset value, it means that the training parameters of the deep neural network model are accurate and the training is completed. If it is not within the preset value, other data is needed to continue model training until all training traffic data are added to the model training. Among them, the preset value can be 0.1 or 0.01.

[0130] In the embodiment of the present application, after obtaining predicted traffic information according to the traffic information and the preset deep neural network model, the method further includes:

[0131] Acquire verification traffic information and target predicted traffic information, where the target predicted traffic information is information in the predicted traffic information corresponding to the time of the verification traffic information;

[0132] According to the verification flow information and the target predicted flow information, the mean absolute error information, the mean absolute percentage error information, and the root mean square error information are obtained;

[0133] According to the mean absolute error information, the mean absolute percentage error information, and the root mean square error information, an evaluation result of the predicted flow information is obtained;

[0134] If the evaluation result of the predicted traffic information indicates that the predicted traffic information is accurate, the deep neural network model is determined to be the target deep neural network model.

[0135] Among them, the mean absolute error information is MAE (mean absolute error), the mean absolute percentage error information is MAPE (Mean Absolute Percentage Error), and the root mean square error information is PMSE (Root, Mean Square Deviation).

[0136] By presetting MAE values, MAPE values, and MPSE values, and then comparing the actual MAE values, MAPE values, and MPSE values ​​with the preset MAE values, MAPE values, and MPSE values ​​for evaluation, if the evaluation results indicate that the predicted traffic information is accurate, it can be determined that the deep neural network model can continue to be used for traffic prediction. If it is inaccurate, the deep neural network model needs to be retrained.

[0137] Figure 2 A flow chart of another flow prediction method provided in an embodiment of the present application. Figure 2 As shown, the execution subject of the method can be a server or other servers, and this embodiment is not particularly limited here. Figure 2 As shown, the method includes:

[0138] S201. Obtain the base station traffic data of all base stations in a specified area for the past month, use the cell OID as the unique identifier, obtain the base station traffic data of the base station every hour, and if the traffic data is missing at a certain time, replace it with the average value of the same period. The base station traffic data includes: uplink traffic, downlink traffic, and prb utilization rate.

[0139] S202, split the base station traffic data in a ratio of 7:3 to obtain a training set and a validation set, and standardize the base station traffic data using a z-score method.

[0140] S203. Obtain a model, which mainly consists of two parts: a spatiotemporal convolution block stack and a temporal decomposition block.

[0141] Among them, the model method is as follows: the backbone architecture adopts the stacking of spatiotemporal convolution blocks, and integrates the output of each spatiotemporal convolution block into the output module through residual connection. The temporal decomposition is only done once in the shallow layer, and the result is also integrated into the output module.

[0142] The method for constructing the spatiotemporal convolution block is as follows: first, the input data is normalized in the time domain and the spatial domain to mine the temporal and spatial information; then a part of the data enters the dilated convolution layer to expand the receptive field without increasing the number of parameters, and then enters the gating mechanism to obtain data including valid information. The valid data passing through the dilated convolution layer is input into the graph convolution module through jump convolution. In the graph convolution module, the Embedding layer is introduced to initialize the node weights, and the adjacency matrix is ​​obtained through self-attention. The adjacency matrix and the data including valid information are added to the graph learning module, and the adjacency matrix achieves adaptive learning effect. The adaptive adjacency matrix is ​​transposed to obtain the corresponding adaptive adjacency matrix transposition. In the multi-hop graph convolution layer, the adaptive adjacency matrix and the data including valid information are learned separately, and finally the results are output through residual connection. The data that does not enter the dilated convolution layer is directly integrated with the output of the graph convolution module in the form of residual connection.

[0143] The method for constructing a time series decomposition block is as follows: four convolution kernels of different sizes are set as specified sliding windows, and then two-dimensional average pooling is performed through the specified sliding windows to obtain the trend information and seasonal information of the two-dimensional data, which are output from two paths respectively, and then parameters are introduced to learn and output the trend information and seasonal information respectively, and then all outputs are integrated; the introduced parameters can be some simple functions (such as matrix transformation functions).

[0144] S204. After obtaining the training set and validation set and building the model, use the training set data to train the model on NVIDIA GTX 3080TI.

[0145] Among them, the model training method using training set data on NVIDIA GTX 3080TI is: the model input format is batch_size x n_features x n_nodes x input_seq_len, and the model output format is batch_size x 1x n_nodes x output_seq_len, where n_features represents the number of input fields, which must include current indicators (uplink traffic, downlink traffic, prb utilization indicators), and other optional ones include holiday information, weather, etc., n_nodes represents the number of base stations, input_seq_len and output_seq_len represent the input time steps (i.e., historical base station traffic data used for prediction) and the expected output time steps (i.e., base station traffic data in the future that needs to be predicted), respectively.

[0146] S205. Obtain the model loss of the previous round and the model loss of this round. If the difference between the model loss of the previous round and the model loss of this round is less than the preset difference, the model training is completed. If the difference between the model loss of the previous round and the model loss of this round is not less than the preset difference, obtain other data in the training set for model training again until all the data in the training set are used up.

[0147] S206. Obtain the validation set and target predicted traffic data. The target predicted traffic information is the data corresponding to the validation set time. According to the validation set and target predicted traffic data, obtain the three evaluation indicators of MAE, MAPE, and RMSE.

[0148] Another traffic testing method provided by the embodiment of the present application is to build a deep learning model based on time series decomposition and adaptive adjacency matrix, analyze the historical indicators of the base station and build a data set, optimize the model parameters through multiple rounds of training, and finally predict the indicators of the base station in the future time period. The method supports base stations and additional feature inputs (such as holidays, seasonal weather, etc.), and meets the input and output lengths of different sizes; build a time series decomposition module to mine the trend information and cycle information of the time series in the node dimension; build an adaptive adjacency matrix based on the attention mechanism, and obtain the correlation between different base stations in a graph learning manner. And it meets the fact that the prediction effect of this method is significant in real scenarios, which can help operation and maintenance personnel improve deployment efficiency and save labor costs.

[0149] Figure 3 A schematic diagram of a flow prediction structure provided in an embodiment of the present application. Figure 3 As shown, the traffic prediction device 30 includes: an acquisition module 301 and an obtaining module 302. Among them:

[0150] The acquisition module 301 acquires flow information, where the flow information includes base station information and flow data corresponding to the base station information.

[0151] Obtain module 302, obtain predicted traffic information according to the traffic information and a preset deep neural network model, wherein the deep neural network model is constructed based on a spatiotemporal convolutional network and a time series decomposition network, the spatiotemporal convolutional network includes multiple spatiotemporal convolutional blocks, the spatiotemporal convolutional block is used to obtain first time series information according to the traffic information, and obtain first predicted traffic information according to the first time series information, the output of the spatiotemporal convolutional network is obtained by integrating the output of each spatiotemporal convolutional block by residual connection, the time series decomposition network is used to obtain second time series information according to the traffic information, and obtain second predicted traffic information according to the second time series information, the first time series information includes time feature information and spatial feature information in the traffic information, the second time series information includes trend information and seasonal information in the traffic information, and the predicted traffic information includes first predicted traffic information and second predicted traffic information.

[0152] In the embodiment of the present application, the acquisition module 301 is further specifically used for:

[0153] Acquire initial flow information, the initial flow information including base station information and initial flow data corresponding to the base station information;

[0154] Perform dimensional processing on the initial flow data to obtain the target flow data;

[0155] Obtain flow data according to the preset ratio and target flow data;

[0156] The flow information is obtained according to the base station information and flow data.

[0157] In the embodiment of the present application, the acquisition module 301 is further specifically used for:

[0158] Obtain target historical traffic information, where the target historical traffic information is the original traffic data corresponding to the base station information and within a preset time range;

[0159] If there are missing data in the original traffic data, clean the missing data to obtain the initial traffic data;

[0160] Initial flow information is obtained according to the initial flow data and base station information.

[0161] In the embodiment of the present application, the obtaining module 302 is further specifically used for:

[0162] Acquire a training traffic sample, where the training traffic sample includes training base station information and training traffic sample data corresponding to the training base station information;

[0163] Perform two-dimensional average pooling on the training flow samples to obtain trend information and seasonal information;

[0164] According to the trend information and the seasonal information, the parameters of the initial time series decomposition network are adjusted to obtain the time series decomposition network;

[0165] Construct a spatiotemporal convolutional network based on training traffic samples;

[0166] Based on the temporal decomposition network and the spatiotemporal convolutional network, a deep neural network model is obtained.

[0167] In the embodiment of the present application, the obtaining module 302 is further specifically used for:

[0168] Extracting temporal and spatial feature information from training traffic samples;

[0169] Obtain target feature data according to the training flow sample data, time feature information and space feature information;

[0170] Obtaining an output result of the graph convolution unit according to the target feature data, a preset adaptive adjacency matrix, an adaptive adjacency matrix transpose corresponding to the adaptive adjacency matrix, and the graph convolution unit;

[0171] The output result of the graph convolution unit and the target feature data are processed by residual connection to obtain the output result of the initial spatiotemporal convolution block;

[0172] According to the output result of the initial spatiotemporal convolution block, the weight of the adaptive adjacency matrix is ​​adjusted to obtain the spatiotemporal convolution block;

[0173] According to the spatiotemporal convolutional blocks, a spatiotemporal convolutional network is constructed.

[0174] In the embodiment of the present application, the obtaining module 302 is further specifically used for:

[0175] Get the preset adaptive adjacency matrix;

[0176] Transpose the adaptive adjacency matrix to obtain the adaptive adjacency matrix transpose;

[0177] Obtaining a first output result of a first multi-hop graph convolutional layer according to the target feature data and the adaptive adjacency matrix;

[0178] Obtaining a second output result of a second multi-hop graph convolutional layer according to the target feature data and the adaptive adjacency matrix transposition;

[0179] An output result of the graph convolution unit is obtained according to a first output result of the first multi-hop graph convolution layer and a second output result of the second multi-hop graph convolution layer.

[0180] In the embodiment of the present application, the obtaining module 302 is further specifically used for:

[0181] Get the loss function of the deep neural network model;

[0182] If the loss function converges, the trained deep neural network model is obtained;

[0183] If the loss function has not converged, other training traffic samples are obtained and used as training traffic samples, and the steps of performing two-dimensional average pooling on the training traffic samples to obtain trend information and seasonal information are re-executed until the loss function converges to obtain a trained deep neural network model.

[0184] In the embodiment of the present application, the obtaining module 302 is further specifically used for:

[0185] Acquire verification traffic information and target predicted traffic information, where the target predicted traffic information is information in the predicted traffic information corresponding to the time of the verification traffic information;

[0186] According to the verification flow information and the target predicted flow information, the mean absolute error information, the mean absolute percentage error information, and the root mean square error information are obtained;

[0187] According to the mean absolute error information, the mean absolute percentage error information, and the root mean square error information, an evaluation result of the predicted flow information is obtained;

[0188] If the evaluation result of the predicted traffic information indicates that the predicted traffic information is accurate, the deep neural network model is determined to be the target deep neural network model.

[0189] Figure 4 This is a schematic diagram of the electronic structure provided in the embodiment of the present application. Figure 4 As shown, the electronic device 40 includes:

[0190] The electronic device 40 may include a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, a communication component 403 and other components. The processor 401 , the memory 402 and the communication component 403 are connected via a bus 407 .

[0191] In a specific implementation process, at least one processor 401 executes computer-executable instructions stored in the memory 402, so that at least one processor 401 executes the above traffic prediction method.

[0192] The specific implementation process of the processor 401 can be found in the above method embodiment, and its implementation principle and technical effect are similar, so this embodiment will not be repeated here.

[0193] In the above Figure 4In the illustrated embodiment, it should be understood that the processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. A general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the invention may be directly implemented as being executed by a hardware processor, or may be executed by a combination of hardware and software modules in the processor.

[0194] The memory may include a high-speed memory (Random Access Memory, RAM), and may also include a non-volatile memory (Non-volatile Memory, NVM), such as at least one disk memory.

[0195] The bus may be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus in the drawings of the present application is not limited to only one bus or one type of bus.

[0196] In some embodiments, a computer program product is also proposed, including a computer program or instructions, which implement the steps of any of the above-mentioned traffic prediction methods when executed by a processor.

[0197] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. The present application is intended to cover any modification, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary techniques in the art that are not disclosed in the present application. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present application are indicated by the following claims.

[0198] It should be understood that the present application is not limited to the precise structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A flow prediction method, characterized in that: The method comprises: Acquire flow information, the flow information including base station information and flow data corresponding to the base station information; According to the traffic information and a preset deep neural network model, predicted traffic information is obtained, wherein the deep neural network model is constructed based on a spatiotemporal convolutional network and a time series decomposition network, the spatiotemporal convolutional network includes multiple spatiotemporal convolutional blocks, the spatiotemporal convolutional blocks are used to obtain first time series information according to the traffic information, and obtain first predicted traffic information according to the first time series information, the output of the spatiotemporal convolutional network is obtained by integrating the output of each of the spatiotemporal convolutional blocks through residual connection, the time series decomposition network is used to obtain second time series information according to the traffic information, and obtain second predicted traffic information according to the second time series information, the first time series information includes time feature information and spatial feature information in the traffic information, the second time series information includes trend information and season information in the traffic information, and the predicted traffic information includes the first predicted traffic information and the second predicted traffic information.

2. The method according to claim 1, characterized in that The obtaining of flow information includes: Acquire initial flow information, the initial flow information including the base station information and initial flow data corresponding to the base station information; Performing dimensional processing on the initial flow data to obtain target flow data; Obtaining the flow data according to a preset ratio and the target flow data; The flow information is obtained according to the base station information and the flow data.

3. The method according to claim 2, characterized in that The obtaining of initial flow information includes: Acquire target historical traffic information, where the target historical traffic information is original traffic data corresponding to the base station information and within a preset time range; If there are missing data in the original traffic data, clean the missing data to obtain initial traffic data; The initial flow information is obtained according to the initial flow data and the base station information.

4. The method according to claim 1, characterized in that: Before obtaining predicted traffic information according to the traffic information and a preset deep neural network model, the method further includes: Acquire a training traffic sample, wherein the training traffic sample includes training base station information and training traffic sample data corresponding to the training base station information; Performing two-dimensional average pooling processing on the training flow samples to obtain the trend information and the season information; According to the trend information and the season information, adjusting parameters of the initial time series decomposition network to obtain the time series decomposition network; Constructing the spatiotemporal convolutional network according to the training traffic samples; The deep neural network model is obtained according to the temporal decomposition network and the spatiotemporal convolutional network.

5. The method according to claim 4, characterized in that The step of constructing the spatiotemporal convolutional network according to the training traffic sample includes: Extracting the time characteristic information and the space characteristic information from the training traffic sample; Obtaining target feature data according to the training flow sample data, the time feature information and the space feature information; Obtaining an output result of the graph convolution unit according to the target feature data, a preset adaptive adjacency matrix, an adaptive adjacency matrix transpose corresponding to the adaptive adjacency matrix, and a graph convolution unit; Performing residual connection processing on the output result of the graph convolution unit and the target feature data to obtain the output result of the initial spatiotemporal convolution block; According to the output result of the initial spatiotemporal convolution block, the weight of the adaptive adjacency matrix is ​​adjusted to obtain the spatiotemporal convolution block; According to the spatiotemporal convolutional block, the spatiotemporal convolutional network is constructed.

6. The method according to claim 5, characterized in that The step of obtaining an output result of the graph convolution unit according to the target feature data, a preset adaptive adjacency matrix, an adaptive adjacency matrix transpose corresponding to the adaptive adjacency matrix, and a graph convolution unit includes: Get the preset adaptive adjacency matrix; Transposing the adaptive adjacency matrix to obtain an adaptive adjacency matrix transpose; Obtaining a first output result of a first multi-hop graph convolutional layer according to the target feature data and the adaptive adjacency matrix; Obtaining a second output result of a second multi-hop graph convolutional layer according to the target feature data and the adaptive adjacency matrix transposition; An output result of the graph convolution unit is obtained according to a first output result of the first multi-hop graph convolution layer and a second output result of the second multi-hop graph convolution layer.

7. The method according to claim 4, characterized in that After obtaining the deep neural network model according to the temporal decomposition network and the spatiotemporal convolutional network, the method further includes: Obtaining a loss function of the deep neural network model; If the loss function converges, the trained deep neural network model is obtained; If the loss function has not converged, other training traffic samples are obtained and used as the training traffic samples, and the step of performing two-dimensional average pooling on the training traffic samples to obtain the trend information and the seasonal information is re-executed until the loss function converges to obtain the trained deep neural network model.

8. The method according to claim 1, characterized in that After obtaining predicted traffic information according to the traffic information and a preset deep neural network model, the method further includes: Acquire verification flow information and target predicted flow information, wherein the target predicted flow information is information in the predicted flow information corresponding to the time of the verification flow information; Obtaining mean absolute error information, mean absolute percentage error information, and root mean square error information according to the verification flow information and the target predicted flow information; Obtaining an evaluation result of the predicted flow information according to the mean absolute error information, the mean absolute percentage error information, and the root mean square error information; If the evaluation result of the predicted traffic information indicates that the predicted traffic information is accurate, the deep neural network model is determined to be the target deep neural network model.

9. A flow prediction device, characterized in that: include: An acquisition module, which acquires flow information, wherein the flow information includes base station information and flow data corresponding to the base station information; A module is obtained, which obtains predicted traffic information according to the traffic information and a preset deep neural network model, wherein the deep neural network model is constructed based on a spatiotemporal convolutional network and a time series decomposition network, and the spatiotemporal convolutional network includes multiple spatiotemporal convolutional blocks, and the spatiotemporal convolutional blocks are used to obtain first time series information according to the traffic information, and obtain first predicted traffic information according to the first time series information. The output of the spatiotemporal convolutional network is obtained by integrating the output of each of the spatiotemporal convolutional blocks through residual connection. The time series decomposition network is used to obtain second time series information according to the traffic information, and obtain second predicted traffic information according to the second time series information. The first time series information includes time feature information and spatial feature information in the traffic information, and the second time series information includes trend information and seasonal information in the traffic information. The predicted traffic information includes the first predicted traffic information and the second predicted traffic information.

10. An electronic device, characterized in that: include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory to implement the method according to any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which are used to implement the method according to any one of claims 1 to 8 when executed by a processor.

Images