Wireless network flow prediction method based on multi-source data feature fusion

Abstract

The invention relates to a wireless network traffic prediction method based on multi-source data feature fusion, which belongs to the technical field of data prediction, and comprises the following steps: S1, obtaining historical traffic data of a network at a historical moment, and converting the network traffic into a thermodynamic diagram; s2, fusing the traffic features of the adjacent blocks and the historical traffic features under different time spans by adopting a residual convolution model to obtain traffic fusion features; s3, fusing the weather features into external features by using a convolutional network, and fusing the traffic fusion features and the external features to obtain multi-source data fusion features; and S4, dividing the multi-source data fusion features into a training set and a test set, and training and testing the wireless network traffic prediction model based on the graph convolutional network.

Description

Technical Field

[0001] The present invention belongs to the technical field of data prediction and relates to a wireless network traffic prediction method by fusing multi-source data features. Background Art

[0002] Traffic forecasting is a technology and method used to predict traffic conditions within a certain period of time in the future. Traffic can refer to data flows in different fields, such as network traffic, traffic flow, customer traffic, etc. The purpose of traffic forecasting is to predict future traffic trends by analyzing historical data and related factors, so as to better allocate resources, plan and manage.

[0003] Foreign researchers have conducted systematic research on methods for traffic flow prediction, which can be mainly divided into methods based on mathematical modeling and methods based on machine learning. Methods based on mathematical modeling, such as the ARIMA method and the Kalman filter method, generally use mathematical modeling to set corresponding parameters and use mathematical expressions to calculate the predicted value. The advantage of this type of method is that the model is simple, but it is difficult to explore the deep-level impact of dynamics; methods based on machine learning, especially models based on deep learning, such as convolutional neural network (CNN) models, recurrent neural network (RNN) models, long short-term memory networks (LSTM) models, etc., can learn the weights of each feature through training with a large amount of data to achieve traffic flow prediction.

[0004] However, most existing network traffic predictions are for a network node or a communication link, and they mostly focus on the spatiotemporal characteristics of network traffic, while ignoring the impact of weather factors on wireless communication channels. According to actual analysis, when encountering adverse weather conditions, people will spontaneously reduce their travel volume, reducing the actual network traffic in shopping malls, tourist attractions and other places. In addition, according to the principle of communication, adverse weather conditions such as rain, snow and fog will absorb, scatter and reflect wireless transmission signals, increase signal loss, reduce the signal-to-noise ratio, and then reduce the maximum information transmission rate, thereby affecting the actual network transmission rate. In wireless communication systems, the impact of various weather environmental factors on the actual network traffic affects the accuracy of the prediction results. Summary of the invention

[0005] In view of this, an object of the present invention is to provide a wireless network traffic prediction method by fusing multi-source data features.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A wireless network traffic prediction method based on multi-source data feature fusion includes the following steps:

[0008] S1: Obtain historical traffic data of the network at historical moments and convert the network traffic into a heat map;

[0009] S2: Use the residual convolution model to fuse the traffic features of adjacent blocks and the historical traffic features under different time spans to obtain the traffic fusion features;

[0010] S3: Use convolutional networks to fuse weather features into external features, and fuse traffic fusion features with external features to obtain multi-source data fusion features;

[0011] S4: Divide the multi-source data fusion features into training sets and test sets, and train and test the wireless network traffic prediction model based on graph convolutional networks.

[0012] Further, the step S1 specifically includes the following steps:

[0013] S11: Divide the city into I×J equally divided grids according to longitude and latitude, each grid represents a block of the city, the position (i, j) represents the block in the i-th row and j-th column, and there are N=I×J areas in total;

[0014] S12: Divide a day into 1440 / T different timestamps with T minutes as the time unit, calculate the average flow of each block (i, j) within the timestamp, which represents the flow size of the block during this period, and pre-process the data;

[0015] S13: Construct a three-dimensional coordinate system, where the x-axis and y-axis represent the longitude and latitude of the block, and the z-axis represents the flow size of the corresponding block. That is, in the three-dimensional coordinate system, the point Represents the size of the flow at block (a, b)

[0016] S14: When the network traffic data of a block (a, b) is missing at time t, the following formula is used to fill the traffic data to ensure data integrity:

[0017]

[0018] In the formula, represents the size of the flow at time t in block (a, b); T is the number of data samples at time t in the data set; Indicates the size of the flow at time t in block (a, b) in the kth data sample;

[0019] S15: Find the flow size of each block and fill it into a matrix V of size I×J to represent the flow size at time t at (a,b):

[0020]

[0021] In the formula, Indicates the size of the flow at time t at block (a, b); the total number of blocks is I×J, where I is the number of rows and J is the number of columns;

[0022] S16: Map the flow size of the matrix V(t) into colors, where whiter colors represent larger flow rates, to form a flow heat map C(t) of size I×J.

[0023] Furthermore, in step S2, the traffic heat map C(t) of the adjacent blocks is used as input, and the residual convolution model ResNet18 network structure is used to extract the network traffic feature X to obtain the spatiotemporal correlation of the traffic;

[0024] S21: The features of adjacent blocks are fused by weight addition method from the spatial scale to extract the traffic feature relationship of adjacent blocks:

[0025]

[0026] Where, X net Represents the flow characteristics after fusion of adjacent blocks; conv is the convolution function, where the convolution kernel size is 3×3; W stands for Hadamard product, that is, the elements at each position of the matrix are equal to the product of the elements at the same position of the two input matrices; up , W down , W left , W right Respectively represent the influence weights of the upper, lower, left and right blocks; X up , X down , X left , X right Respectively represent the flow characteristics of each adjacent block above, below, left and right;

[0027] S22: In the same block, historical traffic characteristics of different time spans are integrated from the time scale:

[0028] α+β+χ=1

[0029] Where, X tra represents the integration of traffic characteristics under different time spans; α, β, χ represent the traffic impact weights of one hour, one day, and one week respectively; Xhour net, Xday net, and Xweek net represent the Xhour net, one hour, one day, and one week respectively. net , that is, after fusing adjacent blocks X up , X down , X left , X right flow characteristics.

[0030] Furthermore, in step S3, the weather indicators are divided into I×J blocks according to the longitude and latitude, and the residual convolution model ResNet18 network structure is used to extract the rainfall R rain , Temperature T temp 、Humidity humi Features; Use the splicing method to integrate various meteorological features into weather features Z atmos :

[0031]

[0032] In the formula, Z atmos is the weather characteristics, R rain , T temp , H humi are the features of rainfall, temperature, and humidity respectively, concat is the concatenation operator; conv is the convolution function, where the convolution kernel size is 3×3;

[0033] The weather features and traffic features are added using the Hadamard product to obtain the multi-source data fusion features:

[0034]

[0035] Where, X c Represents the multi-source data fusion feature, W represents the Hadamard product; tra , W ext are feature weights; X tra Indicates the fusion of traffic characteristics under different time spans; Z atmos For weather characteristics.

[0036] Furthermore, the wireless network traffic prediction model based on graph convolutional network described in step S4 is as follows:

[0037] If two nodes are adjacent in space, connect them with an edge to construct an undirected graph G, where the adjacency relationship between blocks is represented by A; the undirected graph G is in a grid shape, and the degree matrix of the node is:

[0038] D=∑J j=1A ij

[0039] The Laplacian matrix of the undirected graph G is:

[0040] L=D -1 / 2 (A+E)D -1 / 2

[0041] Where D is the degree matrix of A, and E is the identity matrix;

[0042] In the GCN model, when l = 0, its matrix H (0) is the multi-source data fusion feature Xc ,Right now:

[0043] H (0) =X c

[0044] By H (0) From the Laplace matrix L, we can see that the first-order hidden layer expression is written as:

[0045] H (1) =σ(E(A+E)EH (0) W (0) )

[0046] Where σ represents the nonlinear activation function; E is the unit matrix; A is the adjacency matrix between blocks; W (0) is a first-order weight matrix;

[0047] At the second layer and above, the input of each layer is the Laplacian matrix L and the output H of the previous layer. (l) , through these two input parameters and the weight matrix W (l) Convolution is performed to obtain the result H (l+1) ;

[0048] Multiple convolutional layers are stacked to capture multi-order neighborhood node features. The complete formula of all layers is expressed as:

[0049] H (l+1) =σ(LH (l) W (l) )

[0050] In the formula, σ represents the nonlinear activation function; L is the Laplace matrix; H (l) is the input feature of order l; W (l) is the weight matrix of order l;

[0051] A Dropout module is added after the activation function ReLU to prevent the model from overfitting. After being processed by the Dropout module, the forward propagation formula of the two-layer GCN is:

[0052] H (2) =conv(softmax(L×Dropout(ReLU(Dropout(LX c W (0) )))×W (1) ))

[0053] In the formula, H (2) represents the output of the GCN model; L represents the Laplacian matrix; X c represents the fusion feature, that is, the traffic feature fused with external features; W (0) , W (1)Represent the weight parameters of the first-order and second-order hidden layers respectively; Dropout represents the Dropout module processing function; conv represents the convolution operation, where the convolution layer size is 3x3; ReLU is the activation function, which is called the linear rectifier unit;

[0054] Get the output H of the GCN model (2) Afterwards, the traffic prediction value at the next moment is obtained through feature mapping:

[0055] N net =conv(ReLU(H (2) ))

[0056] Where N net It represents the predicted value of the traffic at the next moment; conv represents the convolution operation, where the convolution layer size is 3x3; ReLU is the activation function, which is called the linear rectifier unit.

[0057] Furthermore, the wireless network traffic prediction model based on graph convolutional network uses Adam optimizer and mean square error MSE as the loss function to obtain the trained network traffic prediction model:

[0058]

[0059] In the formula, n represents the number of training data; y i Represents the true flow value of each training data; is the predicted value of the model;

[0060] The test set is processed into the multi-source data fusion features required by the model to ensure that the features of the test samples are consistent with the training data; the trained model is loaded and the features of the test samples are input into the model; after the model receives the input, it calculates and outputs the prediction results. These prediction values ​​are the network traffic predicted by the model at the next moment.

[0061] The beneficial effects of the present invention are as follows: the present invention uses multi-source spatiotemporal data for unified modeling, fully integrates multimodal spatiotemporal data, effectively captures the spatiotemporal characteristics of urban area traffic, establishes an effective and accurate prediction model, and improves the accuracy of traffic prediction results.

[0062] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0064] Figure 1 This is a flow chart of the wireless network traffic prediction method based on multi-source data feature fusion according to the present invention;

[0065] Figure 2 It is a characteristic processing flow chart of the present invention. DETAILED DESCRIPTION

[0066] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0067] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0068] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0069] like Figure 1-2 As shown, the present invention provides a wireless network traffic prediction method based on multi-source data feature fusion, the method comprising:

[0070] Step A: Obtain historical traffic data of the network at historical moments and convert the network traffic into a heat map;

[0071] Get the historical traffic data and trajectory data of the network at historical moments from the dataset.

[0072] The city is divided into I×J equally divided grids according to longitude and latitude. Each grid represents a block in the city. The position (i, j) represents the block in the i-th row and j-th column. There are N=I×J areas in total. A day is divided into 1440 / T different timestamps with T minutes as the time unit. The average flow of each block (i, j) within the timestamp is calculated to represent the flow size of the block during this period. The data is preprocessed. A three-dimensional coordinate system is constructed, where the x-axis and y-axis represent the longitude and latitude of the block, and the z-axis represents the flow size of the corresponding block. That is, in the three-dimensional coordinate system, the point Represents the size of the flow at block (a, b)

[0073] There is a situation where some data is missing in the data set. When the network flow data of a block (a, b) is missing at time t, the following formula is used to fill the flow data to ensure the integrity of the data:

[0074]

[0075] In the formula, It represents the size of the flow at time t in block (a, b); T is the number of data samples at time t in the data set; Represents the size of the flow at time t in block (a, b) in the kth data sample.

[0076] Finally, the flow size of each block is calculated. The total number of blocks is I×J, and the matrix V of size I×J can be filled to represent the flow size at (a,b) at time t:

[0077]

[0078] In the formula, Represents the size of the flow at time t in block (a, b); the total number of blocks is I×J, where I is the number of rows and J is the number of columns.

[0079] The flow size of the matrix V(t) is mapped into colors. The whiter the color, the larger the flow, forming a flow heat map C(t) of size I×J.

[0080] Step B: Use the residual convolution model to extract network traffic features, and fuse the traffic features of adjacent blocks and the historical traffic features under different time spans to obtain traffic fusion features;

[0081] When extracting the periodic features of network traffic data, the traffic heat map C(t) of adjacent blocks is used as input, and the residual convolution model ResNet18 network structure is used to extract the network traffic feature X to obtain the spatiotemporal correlation of the traffic. ResNet18 means that the basic architecture of the network is ResNet, and the depth of the network is 18 layers, which are 18 layers with weights, excluding BN layers and pooling layers. The basic residual unit used by ResNet18, each unit consists of two 3x3 convolution layers, with a BN layer and a ReLU activation function in the middle.

[0082] The traffic characteristics of adjacent blocks show similar trends, and the weighted addition method is used to fuse the characteristics of adjacent blocks from a spatial scale. The spatiotemporal characteristics of adjacent blocks have different weights on the current traffic, so the weighted addition method is used to fuse the characteristics of adjacent blocks to extract the traffic characteristic relationship of adjacent blocks:

[0083]

[0084] Where, X net Represents the flow characteristics after fusion of adjacent blocks; conv is the convolution function, where the convolution kernel size is 3×3; W stands for Hadamard product, that is, the elements at each position of the matrix are equal to the product of the elements at the same position of the two input matrices; up , W down , W left , W right Respectively represent the influence weights of the upper, lower, left and right blocks; X up , X down , X left , X right Respectively represent the traffic characteristics of adjacent blocks above, below, left and right.

[0085] In the same block, the change trend of current traffic is similar to that of historical traffic, but with different degrees of influence. By integrating the historical traffic characteristics of different time spans from the time scale, we have:

[0086] where α+β+χ=1

[0087] Where, X tra represents the integration of traffic characteristics under different time spans; α, β, χ represent the traffic impact weights of one hour, one day, and one week respectively; Xhour net, Xday net, and Xweek net represent the Xhour net, one hour, one day, and one week respectively. net , that is, after fusing adjacent blocks X up , X down , X left , X rightflow characteristics.

[0088] Step C: Use the convolutional network to fuse the weather features into external features, add and fuse the traffic fusion features and external features to obtain multi-source data fusion features;

[0089] The weather station center records the historical climate data of each region, including wind speed, relative humidity, rainfall and other indicators. Similar to the flow heat map formation method in step A, the weather indicators are also divided into I×J blocks according to longitude and latitude, and the residual convolution model ResNet18 network structure is used to extract the rainfall R rain , Temperature T temp 、Humidity humi Features. Use the splicing method to integrate various meteorological features into weather features Z atmos :

[0090]

[0091] In the formula, Z atmos is the weather characteristics, R rain 、T temp , H humi are the features of rainfall, temperature, and humidity respectively; concat is the concatenation operator; conv is the convolution function, where the convolution kernel size is 3×3.

[0092] The weather features and traffic features are added using the Hadamard product to obtain the multi-source data fusion features:

[0093]

[0094] Where, X c Represents the multi-source data fusion feature, W represents the Hadamard product; tra , W ext are feature weights; X tra Indicates the fusion of traffic characteristics under different time spans; Z atmos For weather characteristics.

[0095] Step D: 80% of the multi-source data fusion features are used as training samples of the graph convolutional network to train the prediction model;

[0096] Construct a spatial graph and obtain the spatial feature vector of each region through pre-training. If two nodes are adjacent in space, connect them with an edge to construct an undirected spatial graph G, where the adjacency relationship between blocks is represented by A. The undirected graph G is in a grid shape, and the degree matrix of the node can be obtained as D = ∑J j = 1A ij .

[0097] The Laplacian matrix of an undirected graph G is L = D -1 / 2 (A+E)D -1 / 2 , where D is the degree matrix of A and E is the identity matrix. In the GCN model, when l = 0, its matrix H (0) is the multi-source data fusion feature X c ,Right now:

[0098] H (0) =X c

[0099] By H (0) From the Laplace matrix L, we can see that the first-order hidden layer expression can be written as:

[0100] H (1) =σ(E(A+E)EH (0) W (0) )

[0101] Where σ represents the nonlinear activation function; E is the unit matrix; A is the adjacency matrix between blocks; W (0) is a first-order weight matrix.

[0102] At the second layer and above, the input of each layer becomes the Laplacian matrix L and the output H of the previous layer. (l) , through these two input parameters and the weight matrix W (l) Convolution is performed to obtain the result H (l+1) The stacking of multiple convolutional layers can capture multi-order neighborhood node features. The complete formula of all layers is expressed as:

[0103] H (l+1) =σ(LH (l) W (l) )

[0104] In the formula, σ represents the nonlinear activation function; L is the Laplace matrix; H (l) is the input feature of order l; W (l) is a weight matrix of order l.

[0105] Deeper convolution operations may not necessarily be able to extract richer features. Generally speaking, a two-layer GCN model can achieve relatively good results, so this patent adopts a two-layer GCN model.

[0106] Considering the current complex network topology and X c Fusion of FlowX tra , rainfall R rain , Temperature T tmep 、Humidity humiDue to various spatiotemporal features, the GCN algorithm is prone to overfitting problems, so the Dropout module is added after the activation function ReLU. The Dropout module is a regularization technique that is mainly used to prevent the model from overfitting. Overfitting means that the model performs well on the training data, but performs poorly on unseen test data. Dropout randomly discards (i.e. "drops") a part of the neurons so that the model does not rely too much on certain specific neurons during training, thereby improving the generalization ability of the model. After Dropout is processed by the Dropout layer, the formula for the forward propagation of the two-layer GCN is:

[0107] H (2) =conv(softmax(L×Dropout(ReLU(Dropout(LX c W (0) )))×W (1) ))

[0108] In the formula, H (2) represents the output of the GCN model; L represents the Laplacian matrix; X c represents the fusion feature, that is, the traffic feature fused with external features; W (0) , W (1) They represent the weight parameters of the first-order and second-order hidden layers respectively; Dropout represents the Dropout module processing function; conv represents the convolution operation, where the convolution layer size is 3x3; ReLU is the activation function, which is called the linear rectifier unit.

[0109] Get the output H of the GCN model (2) Afterwards, the traffic prediction value at the next moment can be obtained through feature mapping:

[0110] N net =conv(ReLU(H (2) ))

[0111] Where N net It represents the predicted value of the traffic at the next moment; conv represents the convolution operation, where the convolution layer size is 3x3; ReLU is the activation function, which is called the linear rectifier unit.

[0112] During training, the multiple groups of fused data of each sample node are divided into training set and test set samples in a ratio of 8:2. The Adam optimizer is used and the mean square error MSE is used as the loss function to obtain the trained network traffic prediction model:

[0113]

[0114] In the formula, n represents the number of training data; y iRepresents the true flow value of each training data; is the predicted value of the model.

[0115] Step E: Input the remaining 20% ​​of multi-source data fusion features into the model as test samples to predict the network traffic at the next moment;

[0116] The remaining 20% ​​of multi-source data is processed into the multi-source data fusion features required by the model to ensure that the features of the test sample are consistent with the training data. The trained model is loaded and the features of the test sample are input into the model. After receiving the input, the model calculates and outputs the prediction results. These prediction values ​​are the network traffic predicted by the model at the next moment.

[0117] A person skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing related hardware through a program. The program can be stored in a computer-readable storage medium. When the program is executed, the steps of the method can be implemented. The storage medium, such as ROM / RAM, a disk, an optical disk, etc.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A wireless network traffic prediction method based on multi-source data feature fusion, characterized by: The following steps are involved: S1: Obtain historical traffic data of the network at historical moments and convert the network traffic into a heat map; S2: Use the residual convolution model to fuse the traffic features of adjacent blocks and the historical traffic features under different time spans to obtain the traffic fusion features; S3: Use convolutional networks to fuse weather features into external features, and fuse traffic fusion features with external features to obtain multi-source data fusion features; S4: Divide the multi-source data fusion features into training sets and test sets, and train and test the wireless network traffic prediction model based on graph convolutional networks.

2. The wireless network traffic prediction method based on multi-source data feature fusion according to claim 1 is characterized by: The step S1 specifically includes the following steps: S11: Divide the city into I×J equally divided grids according to longitude and latitude, each grid represents a block of the city, the position (i, j) represents the block in the i-th row and j-th column, and there are N=I×J areas in total; S12: Divide a day into 1440 / T different timestamps with T minutes as the time unit, calculate the average flow of each block (i, j) within the timestamp, which represents the flow size of the block during this period, and pre-process the data; S13: Construct a three-dimensional coordinate system, where the x-axis and y-axis represent the longitude and latitude of the block, and the z-axis represents the flow size of the corresponding block. That is, in the three-dimensional coordinate system, the point Represents the size of the flow at block (a, b) S14: When the network traffic data of a block (a, b) is missing at time t, the following formula is used to fill the traffic data to ensure data integrity: In the formula, represents the size of the flow at time t in block (a, b); T is the number of data samples at time t in the data set; Indicates the size of the flow at time t in block (a, b) in the kth data sample; S15: Find the flow size of each block and fill it into a matrix V of size I×J to represent the flow size at time t at (a,b): In the formula, Indicates the size of the flow at time t at block (a, b); the total number of blocks is I×J, where I is the number of rows and J is the number of columns; S16: Map the flow size of the matrix V(t) into colors, where whiter colors represent larger flow rates, to form a flow heat map C(t) of size I×J.

3. The wireless network traffic prediction method based on multi-source data feature fusion according to claim 1 is characterized by: In step S2, the traffic heat map C(t) of the adjacent blocks is used as input, and the residual convolution model ResNet18 network structure is used to extract the network traffic feature X to obtain the spatiotemporal correlation of the traffic; S21: The features of adjacent blocks are fused by weight addition method from the spatial scale to extract the traffic feature relationship of adjacent blocks: In the formula, X net Represents the flow characteristics after fusion of adjacent blocks; conv is the convolution function, where the convolution kernel size is 3×3; W stands for Hadamard product, that is, the elements at each position of the matrix are equal to the product of the elements at the same position of the two input matrices; up , W down , W left , W right Respectively represent the influence weights of the upper, lower, left and right blocks; X up , X down , X left , X right Respectively represent the flow characteristics of each adjacent block above, below, left and right; S22: In the same block, historical traffic characteristics of different time spans are integrated from the time scale: Where, X tra represents the integration of traffic characteristics under different time spans; α, β, χ represent the traffic impact weights of one hour, one day, and one week respectively; Xhour net, Xday net, and Xweek net represent the Xhour net, one hour, one day, and one week respectively. net , that is, after fusing adjacent blocks X up , X down , X left , X right flow characteristics.

4. The wireless network traffic prediction method based on multi-source data feature fusion according to claim 1 is characterized by: In step S3, the weather indicators are divided into I×J blocks according to the longitude and latitude, and the residual convolution model ResNet18 network structure is used to extract the rainfall R rain , Temperature T temp 、Humidity humi feature; Use the splicing method to merge various meteorological features into weather feature Z atmos : In the formula, Z atmos is the weather characteristics, R rain , T temp , H humi are the features of rainfall, temperature, and humidity respectively, concat is the concatenation operator; conv is the convolution function, where the convolution kernel size is 3×3; The weather features and traffic features are added using the Hadamard product to obtain the multi-source data fusion features: Where, X c Represents the multi-source data fusion feature, W represents the Hadamard product; tra , W ext are feature weights; X tra Indicates the fusion of traffic characteristics under different time spans; Z atmos For weather characteristics.

5. The wireless network traffic prediction method based on multi-source data feature fusion according to claim 1 is characterized by: The wireless network traffic prediction model based on graph convolutional network described in step S4 is as follows: If two nodes are adjacent in space, connect them with an edge to construct an undirected graph G, where the adjacency relationship between blocks is represented by A; the undirected graph G is in a grid shape, and the degree matrix of the node is: D=∑J j=1A ij The Laplacian matrix of the undirected graph G is: L=D -1 / 2 (A+E)D -1 / 2 Where D is the degree matrix of A and E is the identity matrix; In the GCN model, when l = 0, its matrix H (0) is the multi-source data fusion feature X c ,Right now: H (0) =X c By H (0) From the Laplace matrix L, we can see that the first-order hidden layer expression is written as: H (1) =σ(E(A+E)EH (0) IN (0) ) In the formula, σ represents the nonlinear activation function; E is the unit matrix; A is the adjacency matrix between blocks; W (0) is a first-order weight matrix; At the second layer and above, the input of each layer is the Laplacian matrix L and the output H of the previous layer. (l) , through these two input parameters and the weight matrix W (l) Convolution is performed to obtain the result H (l+1) ; Multiple convolutional layers are stacked to capture multi-order neighborhood node features. The complete formula of all layers is expressed as: H (l+1) =σ(LH (l) W (l) ) In the formula, σ represents the nonlinear activation function; L is the Laplace matrix; H (l) is the input feature of order l; W (l) is the weight matrix of order l; A Dropout module is added after the activation function ReLU to prevent the model from overfitting. After being processed by the Dropout module, the forward propagation formula of the two-layer GCN is: H (2) =conv(softmax(L×Dropout(ReLU(Dropout(LX c W (0) )))×W (1) )) In the formula, H (2) represents the output of the GCN model; L represents the Laplacian matrix; X c Represents fused features, i.e., traffic features fused with external features; W (0) , W (1) Represent the weight parameters of the first-order and second-order hidden layers respectively; Dropout represents the Dropout module processing function; conv represents the convolution operation, where the convolution layer size is 3x3; ReLU is the activation function, which is called the linear rectifier unit; Get the output H of the GCN model (2) Afterwards, the traffic prediction value at the next moment is obtained through feature mapping: N net =conv(ReLU(H (2) )) Where N net It represents the predicted value of the traffic at the next moment; conv represents the convolution operation, where the convolution layer size is 3x3; ReLU is the activation function, which is called the linear rectifier unit.

6. The wireless network traffic prediction method based on multi-source data feature fusion according to claim 1 is characterized by: The wireless network traffic prediction model based on graph convolutional network uses Adam optimizer and mean square error MSE as the loss function to obtain the trained network traffic prediction model: In the formula, n represents the number of training data; y i Represents the true flow value of each training data; is the predicted value of the model; The test set is processed into the multi-source data fusion features required by the model to ensure that the features of the test samples are consistent with the training data; the trained model is loaded and the features of the test samples are input into the model; after the model receives the input, it calculates and outputs the prediction results. These prediction values ​​are the network traffic predicted by the model at the next moment.

Images