A network traffic prediction method and system based on spatiotemporal series

Abstract

The present invention discloses a network traffic prediction method and system based on spatiotemporal sequence, the method comprising the following steps: matrixing the traffic data of the target area in time and space; obtaining the spatial features in the matrixed data through a densely connected convolutional neural network; obtaining the temporal features in the matrixed data through a diffusion convolution gated cyclic unit; obtaining the influence features in the external influence factor data through a convolution method; combining the spatial features, temporal features and influence features into a matrix after convolution respectively to obtain a combined matrix; multiplying the combined matrix with the weight matrix, activating the multiplication result through a sigmoid function, and obtaining the network traffic prediction result. The present invention solves the technical problems of insufficient spatiotemporal factor analysis, insufficient nonlinear feature extraction, and low long-term prediction accuracy in the existing methods, and can accurately predict network traffic.

Description

Technical Field

[0001] The present invention relates to the field of network traffic prediction, and in particular to a network traffic prediction method and system based on spatiotemporal sequences. Background Art

[0002] With the development of technologies such as 5G / B5G, edge computing, and NFV, the number of Internet users and the time they spend online have continued to grow, which has led to a surge in network traffic in recent years. Refined, automated, and intelligent network operation and management will become a new challenge. To meet this challenge, it is necessary to accurately perceive application-level network traffic. Wireless network traffic prediction is one of the core technologies. Accurate and timely traffic prediction can help improve communication network management, optimize network resource allocation, and improve resource waste and network congestion, thereby improving communication network performance.

[0003] Traditional traffic prediction methods can be divided into statistical methods and machine learning methods. Statistical network traffic prediction methods regard traffic prediction as a time series prediction problem. They can learn the linear and nonlinear characteristics of network traffic, such as periodicity and seasonality, from a large amount of traffic data to make predictions. Such methods include classic autoregression (AR) and moving average (MA) as well as autoregressive integrated moving average (ARIMA) and α-stable models. However, the structure of wireless network traffic is actually very complex, including multiple influencing factors such as the spatial topology of network base stations, the number of base stations, user mobility, user behavior requirements, and arrival patterns. These factors make the spatiotemporal dependence between cellular traffic at different base stations complicated. Statistical models rely on stationary assumptions, cannot simultaneously model the impact of multiple factors, cannot overcome the interference of random events, and cannot handle multi-dimensional complex data.

[0004] In order to capture the complex and nonlinear dependencies hidden in wireless traffic data, machine learning models have been used to improve the accuracy of network traffic prediction. Prediction methods based on machine learning mainly include methods based on decision trees (DT), support vector machines (SVM), Bayesian networks (BN), hidden Markov models (HMM), etc. Prediction methods based on machine learning take into account the temporal and spatial characteristics of network traffic data at the same time, and have achieved good results in short-term sequence prediction. However, for network traffic data with multi-dimensional and dynamic spatiotemporal characteristics, machine learning methods are difficult to capture hidden complex features, and it is difficult to achieve network-wide and medium- and long-term traffic prediction. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a network traffic prediction method and system based on spatiotemporal series, which solves the problem that the existing network traffic prediction methods are difficult to take into account spatial characteristics, temporal characteristics and external influencing factors, resulting in poor prediction effect.

[0006] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0007] A network traffic prediction method based on spatiotemporal sequence is provided, which comprises the following steps:

[0008] S1, acquiring and matrixing the flow data of the target area in time and space to obtain matrixed data;

[0009] S2, obtain spatial features in the matrixed data through densely connected convolutional neural networks;

[0010] S3, using the spatial features in the matrixed data as the input of the diffusion convolution gated recurrent unit to obtain the temporal features in the matrixed data;

[0011] S4, modeling the external influencing factor data in the traffic data of the target area, and obtaining the influencing features in the external influencing factor data through a convolution method;

[0012] S5. Convolve the spatial features in the matrixed data, the temporal features in the matrixed data, and the influence features in the external influencing factor data respectively and combine them into a matrix to obtain a combined matrix; multiply the combined matrix with the weight matrix, activate the multiplication result through the sigmoid function, and obtain the network traffic prediction result.

[0013] Furthermore, the specific method of obtaining and matrixing the flow data of the target area in time and space in step S1 includes the following sub-steps:

[0014] S1-1. Divide the target area into N×N units to obtain a set of spatial geographic units.

[0015] S1-2, SMS service, phone call service and Internet access are considered as three different types of services;

[0016] S1-3, respectively obtain the traffic data of three different types of services in each spatial geographic unit;

[0017] S1-4. The cellular traffic of the kth service type in the T time interval is represented as a set of spatiotemporal matrices: Where V = 1, 2, 3, corresponding to SMS service, phone call service and Internet access respectively; Each element in represents the traffic data of the kth service type at the corresponding spatial geographic unit location at time t; represents the real number domain; v (N,N) Represents a collection of spatial geographic units The spatial geographic unit at the Nth row and Nth column in .

[0018] Furthermore, in step S2, the convolutional neural network includes L basic modules with dense connections, each of which includes a convolutional layer, a batch normalization layer, and an activation layer connected in sequence; the calculation expression of the lth basic module is:

[0019]

[0020] in Represents the output of the lth basic module in the convolutional neural network; Represents the input of the convolutional neural network; Represents the output of the l-1th basic module in the convolutional neural network; the output of the Lth basic module in the convolutional neural network That is, the spatial features in the matrixed data.

[0021] Furthermore, the diffusion convolution gated recurrent unit in step S3 includes M computing subunits, and the expression of the mth computing subunit is:

[0022]

[0023]

[0024]

[0025] H (m) =u (m) ⊙H (m-1) +(1-u (m) )⊙C (m)

[0026] r (m) is the output of the reset gate of the mth computing subunit; u (m) is the output of the update gate of the mth computing subunit; C (m) is the output of the candidate memory cell of the mth computing subunit, H (m) is the hidden state output by the mth computing subunit; and Both are diffusion convolution operations; X (m) is the input of the mth computing subunit; H (m-1) is the hidden state output by the m-1th computing subunit; tanh(.) represents the hyperbolic tangent function; b r , b u , b c are all bias terms; σ(.) represents the Sigmoid function; ⊙ represents the Hadamard product; the hidden state output by the Mth computing subunit is the temporal feature of the spatial feature in the matrixed data When m is 1, X (m) is the matrixed data, H (0) represents the initial hidden state, H (0) The value of is set by the all-zero initialization method.

[0027] Furthermore, the specific method of obtaining the influencing features in the external influencing factor data in step S4 includes the following sub-steps:

[0028] S4-1. Capture external influencing factor data in the target area through Google API; the external influencing factor data includes the number of base stations, distribution of points of interest, and level of social activities;

[0029] S4-2. Divide the data of the same external influencing factors into an N×N matrix and map it to a set of spatial geographic units The base station quantity matrix, interest point distribution matrix and social activity level matrix are obtained respectively;

[0030] S4-3, cascading the base station quantity matrix, the interest point distribution matrix and the social activity level matrix to obtain the external influence matrix;

[0031] S4-4, after convolving the external influence matrix with the tensor matrix, a nonlinear transformation is performed on the convolution operation result to obtain the influence feature O in the external influence factor data. cross .

[0032] Furthermore, the expressions of the base station number matrix, interest point distribution matrix and social activity level matrix are:

[0033]

[0034]

[0035]

[0036] Where D BS is the base station number matrix; D POI is the interest point distribution matrix; D Social is the social activity level matrix; represents the number of base stations in the spatial geographic unit at the Nth row and the Nth column; Represents the point of interest value in the spatial geographic unit at the Nth row and the Nth column; Represents the social activity level value in the spatial geographic unit at the Nth row and the Nth column.

[0037] Furthermore, the nonlinear transformation includes a composite function of three consecutive operations; wherein the composite functions of the three consecutive operations are batch normalization, rectified linearization, and convolution.

[0038] Furthermore, the calculation expression for multiplying the combination matrix and the weight matrix in step S5 is:

[0039]

[0040] Where Y is the network traffic prediction result, and W is the weight matrix; is the combination matrix; It is the result of convolution of spatial features in matrixed data; is the result of convolution of the time features in the matrixed data; cross ′ is the result of convolution of the influencing features in the external influencing factor data.

[0041] A network traffic prediction system based on spatiotemporal sequence is provided, which includes a target area data processing module, a space processing module, a time processing module, an external factor processing module and a parameter matrix fusion module;

[0042] The target area data processing module is used to obtain and matrix the flow data of the target area in time and space to obtain the matrixed data; and to model the external influencing factor data in the flow data of the target area;

[0043] The spatial processing module is used to obtain spatial features in the matrixed data through densely connected convolutional neural networks;

[0044] A time processing module, used to use the spatial features in the matrixed data as inputs of a diffusion convolution gated recurrent unit to obtain the time features in the matrixed data;

[0045] An external factor processing module is used to obtain the influencing features in the external influencing factor data through a convolution method;

[0046] The parameter matrix fusion module is used to convolve the spatial features in the matrixed data, the temporal features in the matrixed data, and the influence features in the external influencing factor data into a matrix to obtain a combined matrix; the combined matrix is ​​multiplied by the weight matrix, and the multiplication result is activated by the sigmoid function to obtain the network traffic prediction result.

[0047] The beneficial effects of the present invention are:

[0048] 1. The present invention solves the technical problems of insufficient spatiotemporal factor analysis, insufficient nonlinear feature extraction, and low long-term prediction accuracy in existing methods, and can accurately predict network traffic.

[0049] 2. The present invention comprehensively considers the temporal and spatial factors of traffic data and external influencing factors such as base station information, POI distribution and social activity level. Through this comprehensive consideration, the changing pattern of network traffic can be more comprehensively analyzed to improve the accuracy and reliability of prediction;

[0050] 3. The present invention adopts Huber error as the loss function, which not only considers the model accuracy within the small error range, but also reduces the adverse effects of outliers in a form similar to absolute error when facing large errors, thereby maintaining good performance and stability in many cases. Since network traffic data will have some outliers during holidays, using Huber error as the loss function improves the performance of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a schematic diagram of the process of this method;

[0052] Figure 2 This is the structural diagram of the system;

[0053] Figure 3 This is a schematic diagram of the prediction result of the network traffic corresponding to the SMS service of the present invention in the embodiment;

[0054] Figure 4 It is a schematic diagram of the prediction result of the network traffic corresponding to the telephone call service of the present invention in the embodiment;

[0055] Figure 5 It is a schematic diagram of the prediction result of the network traffic corresponding to the Internet access according to the present invention in the embodiment. DETAILED DESCRIPTION

[0056] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0057] like Figure 1 As shown, the network traffic prediction method based on spatiotemporal sequence includes the following steps:

[0058] S1, acquiring and matrixing the flow data of the target area in time and space to obtain matrixed data;

[0059] S2, obtain spatial features in the matrixed data through densely connected convolutional neural networks;

[0060] S3, using the spatial features in the matrixed data as the input of the diffusion convolution gated recurrent unit to obtain the temporal features in the matrixed data;

[0061] S4, modeling the external influencing factor data in the traffic data of the target area, and obtaining the influencing features in the external influencing factor data through a convolution method;

[0062] S5. Convolve the spatial features in the matrixed data, the temporal features in the matrixed data, and the influence features in the external influencing factor data respectively and combine them into a matrix to obtain a combined matrix; multiply the combined matrix with the weight matrix, activate the multiplication result through the sigmoid function, and obtain the network traffic prediction result.

[0063] like Figure 2 As shown, the network traffic prediction system based on spatiotemporal sequence includes a target area data processing module, a space processing module, a time processing module, an external factor processing module and a parameter matrix fusion module;

[0064] The target area data processing module is used to obtain and matrix the flow data of the target area in time and space to obtain the matrixed data; and to model the external influencing factor data in the flow data of the target area;

[0065] The spatial processing module is used to obtain spatial features in the matrixed data through densely connected convolutional neural networks;

[0066] A time processing module, used to use the spatial features in the matrixed data as inputs of a diffusion convolution gated recurrent unit to obtain the time features in the matrixed data;

[0067] An external factor processing module is used to obtain the influencing features in the external influencing factor data through a convolution method;

[0068] The parameter matrix fusion module is used to convolve the spatial features in the matrixed data, the temporal features in the matrixed data, and the influence features in the external influencing factor data into a matrix to obtain a combined matrix; the combined matrix is ​​multiplied by the weight matrix, and the multiplication result is activated by the sigmoid function to obtain the network traffic prediction result.

[0069] The specific method of obtaining and matrixing the flow data of the target area in time and space includes the following sub-steps:

[0070] S1-1. Divide the target area into N×N units to obtain a set of spatial geographic units.

[0071] S1-2, short message service (SMS), telephone call service (Call) and Internet access (Internet) are three different types of services;

[0072] S1-3, respectively obtain the traffic data of three different types of services in each spatial geographic unit;

[0073] S1-4. The cellular traffic of the kth service type in the T time interval is represented as a set of spatiotemporal matrices: Where k = 1, 2, 3, corresponding to SMS service, phone call service and Internet access respectively; Each element in represents the traffic data of the kth service type at the corresponding spatial geographic unit location at time t; represents the real number domain; v (N,N) Represents a collection of spatial geographic units The spatial geographic unit at the Nth row and Nth column in .

[0074] Convolutional neural network (CNN) is a deep learning model that can be applied to process data with obvious spatial structure. After a series of convolutions are performed at each layer, the local and global spatial dependencies of urban-wide traffic can be captured. Considering that the inbound and outbound network traffic of different units is spatially dependent on each other, CNN adopts a dense connection mode, the main feature of which is that there are dense connections between each layer, which means that each output of each layer is connected to all outputs of subsequent layers, not just to the output of the next layer. The design of dense connections can alleviate the gradient vanishing problem and enhance feature propagation, so that the network can learn the feature representation of data more deeply and densely, which helps to improve the performance and generalization ability of the model, and ultimately improve the prediction efficiency and accuracy. Based on the above considerations, the convolutional neural network in this embodiment includes L basic modules using dense connections, each of which includes a convolutional layer, a batch normalization layer and an activation layer connected in sequence; the calculation expression of the lth basic module is:

[0075]

[0076] in Represents the output of the lth basic module in the convolutional neural network; Represents the input of the convolutional neural network; Represents the output of the l-1th basic module in the convolutional neural network; the output of the Lth basic module in the convolutional neural network That is, the spatial features in the matrixed data.

[0077] The gated recurrent unit (GRU) controls the update and retention of hidden states by updating gates and resetting gates to capture temporal dependencies in sequence data. Diffused convolution is a convolution operation used to process graph-structured data. In the diffused convolution gated recurrent unit, diffused convolution is used to replace the matrix multiplication operation in the traditional GRU to capture spatial features in sequence data. Diffused convolution can effectively process the connection relationship between nodes, thereby better adapting to sequence data with complex spatiotemporal structures. In this embodiment, the diffused convolution gated recurrent unit includes M computing subunits, and the expression of the mth computing subunit is:

[0078]

[0079]

[0080]

[0081] H (m) =u (m) ⊙H (m-1) +(1-u (m) )⊙C (m)

[0082] r (m) is the output of the reset gate of the mth computing subunit, which is used to control the degree of retention of the hidden state of the previous computing subunit and the degree of influence of the current computing subunit input; u (m) is the output of the update gate of the mth computing subunit, which is used to control the update degree of the hidden state of the current computing subunit; C (m) is the output of the candidate memory cell of the mth computing subunit, indicating the candidate hidden state of the current computing subunit; H (m) is the hidden state output by the mth computing subunit, indicating the output or state of the current computing subunit; and Both are diffusion convolution operations; X (m) is the input of the mth computing subunit; H (m-1) is the hidden state output by the m-1th computing subunit; tanh(.) represents the hyperbolic tangent function; b r 、b u 、b c are all bias terms; σ(.) represents the Sigmoid function; ⊙ represents the Hadamard product; the hidden state output by the Mth computing subunit is the temporal feature of the spatial feature in the matrixed data When m is 1, X (m) is the matrixed data, H (0) represents the initial hidden state, H (0) The value of is set by the all-zero initialization method.

[0083] The specific method of obtaining the influencing features in the external influencing factor data includes the following sub-steps:

[0084] S4-1. Capture external influencing factor data in the target area through Google API; the external influencing factor data includes the number of base stations, distribution of points of interest, and level of social activities;

[0085] S4-2. Divide the data of the same external influencing factors into an N×N matrix and map it to a set of spatial geographic units The base station quantity matrix, interest point distribution matrix and social activity level matrix are obtained respectively, and the expressions are:

[0086]

[0087]

[0088]

[0089] Where D BS is the base station number matrix; D POI is the interest point distribution matrix; D Social is the social activity level matrix; represents the number of base stations in the spatial geographic unit at the Nth row and the Nth column; Represents the point of interest value in the spatial geographic unit at the Nth row and the Nth column; Represents the social activity level value in the spatial geographic unit at the Nth row and the Nth column.

[0090] S4-3, cascading the base station quantity matrix, the interest point distribution matrix and the social activity level matrix to obtain the external influence matrix;

[0091] S4-4, after convolving the external influence matrix with the tensor matrix, a nonlinear transformation is performed on the convolution operation result to obtain the influence feature O in the external influence factor data. cross .

[0092] The calculation expression for multiplying the combination matrix by the weight matrix is:

[0093]

[0094] Where Y is the network traffic prediction result, and W is the weight matrix; is the combination matrix; It is the result of convolution of spatial features in matrixed data; is the result of convolution of the time features in the matrixed data; cross ′ is the result of convolution of the influencing features in the external influencing factor data.

[0095] The expression for activating the multiplication result through the sigmoid function is:

[0096]

[0097] In the specific implementation process, taking the number of base stations (BS) as an example, the BS data set contains the geographical coordinates corresponding to the base station and the type and number of the base station. According to the geographical location, it can be corresponded to the spatial geographical unit in the traffic data, thereby obtaining the base station number matrix. Each element value in the base station number matrix represents the number of base stations contained in the corresponding area of ​​the city. In this embodiment, the points of interest include places or locations such as subway stations, shops, and restaurants.

[0098] The tensor matrix is ​​a tensor used to integrate the data of the three external influencing factors. The elements in the weight matrix W are determined by an optimization process to minimize the difference between the predicted results and the actual observed results. In this embodiment, the following steps are used to learn these weights:

[0099] ① Initialize weights: The elements of the W matrix can be initialized randomly or based on some heuristic method;

[0100] ② Loss function: Define a Haber error loss function, which measures the inconsistency between the fused prediction results and the actual results; the Huber error loss function combines the advantages of mean square error and absolute error to improve the robustness to outliers while maintaining the differentiability of the error. The Huber error loss function is expressed as:

[0101]

[0102] Among them, δ is a preset threshold to balance the conversion between mean square error and absolute error; represents the prediction result of the present invention, Y i Represents the true label of the sample.

[0103] ③Optimization system: Use gradient descent to iteratively update weights, with the goal of finding the minimum value of the loss function;

[0104] ④ Hyperparameter tuning: Use techniques such as cross-validation to find the best hyperparameter settings, which control the learning process (such as learning rate, regularization strength, etc.).

[0105] Through the above process, the value of each element in the weight matrix W can be determined, so that the present invention can achieve the best prediction output.

[0106] In one embodiment of the present invention, Figure 3 , Figure 4 and Figure 5 As shown, the traffic prediction values ​​of the three service types of the present invention are relatively close to the actual traffic values, the prediction accuracy can reach 95%, and the MAE can be limited to less than 30. Compared with basic methods such as LSTM, the performance of the present invention has been greatly improved.

[0107] In summary, the present invention solves the technical problems of insufficient spatiotemporal factor analysis, insufficient nonlinear feature extraction, and low long-term prediction accuracy in existing methods, and can accurately predict network traffic.

Claims

1. A network traffic prediction method based on spatiotemporal series, characterized in that: The following steps are involved: S1, acquiring and matrixing the flow data of the target area in time and space to obtain matrixed data; S2, obtain spatial features in the matrixed data through densely connected convolutional neural networks; S3, using the spatial features in the matrixed data as the input of the diffusion convolution gated recurrent unit to obtain the temporal features in the matrixed data; S4, modeling the external influencing factor data in the traffic data of the target area, and obtaining the influencing features in the external influencing factor data through a convolution method; S5, convolve the spatial features in the matrixed data, the time features in the matrixed data, and the influence features in the external influence factor data into a matrix to obtain a combined matrix; multiply the combined matrix with the weight matrix, activate the multiplication result through the sigmoid function, and obtain the network traffic prediction result; The specific method of obtaining and matrixing the flow data of the target area in time and space in step S1 includes the following sub-steps: S1-1. Divide the target area into N×N units to obtain a set of spatial geographic units. S1-2, SMS service, phone call service and Internet access are considered as three different types of services; S1-3, respectively obtain the traffic data of three different types of services in each spatial geographic unit; S1-4. The cellular traffic of the kth service type in the T time interval is represented as a set of spatiotemporal matrices: Where k = 1, 2, 3, corresponding to SMS service, phone call service and Internet access respectively; Each element in represents the traffic data of the kth service type at the corresponding spatial geographic unit location at time t; represents the real number domain; v (N，N) Represents a collection of spatial geographic units The spatial geographic unit at the Nth row and Nth column in ; In step S2, the convolutional neural network includes L basic modules with dense connections, each of which includes a convolution layer, a batch normalization layer, and an activation layer connected in sequence; the calculation expression of the lth basic module is: in Represents the output of the lth basic module in the convolutional neural network; Represents the input of the convolutional neural network; Represents the output of the l-1th basic module in the convolutional neural network; the output of the Lth basic module in the convolutional neural network That is, the spatial features in the matrixed data; The diffusion convolution gated recurrent unit in step S3 includes M computing subunits, and the expression of the mth computing subunit is: H (m) =u (m) ⊙H (m-1) +(1-u (m) )⊙C (m) r (m) is the output of the reset gate of the mth computing subunit; u (m) is the output of the update gate of the mth computing subunit; C (m) is the output of the candidate memory cell of the mth computing subunit, H (m) is the hidden state output by the mth computing subunit; and Both are diffusion convolution operations; X (m) is the input of the mth computing subunit; H (m -1) is the hidden state output by the m-1th computing subunit; tan h(.) represents the hyperbolic tangent function; b r 、b u 、b c are all bias terms; σ(.) represents the Sigmoid function; ⊙ represents the Hadamard product; the hidden state output by the Mth computing subunit is the temporal feature of the spatial feature in the matrixed data When m is 1, X (m) is the matrixed data, H (0) represents the initial hidden state, H (0) The value of is set by the all-zero initialization method; The specific method of obtaining the influencing features in the external influencing factor data in step S4 includes the following sub-steps: S4-1. Capture external influencing factor data in the target area through Google API; The external influencing factor data include the number of base stations, the distribution of points of interest, and the level of social activities; S4-2. Divide the data of the same external influencing factors into an N×N matrix and map it to a set of spatial geographic units The base station quantity matrix, interest point distribution matrix and social activity level matrix are obtained respectively; S4-3, cascading the base station quantity matrix, the interest point distribution matrix and the social activity level matrix to obtain the external influence matrix; S4-4, after convolving the external influence matrix with the tensor matrix, a nonlinear transformation is performed on the convolution operation result to obtain the influence feature O in the external influence factor data. cross .

2. The network traffic prediction method based on spatiotemporal series according to claim 1 is characterized in that: The expressions of base station quantity matrix, interest point distribution matrix and social activity level matrix are: Where D BS is the base station number matrix; D POI is the interest point distribution matrix; D Social is the social activity level matrix; represents the number of base stations in the spatial geographic unit at the Nth row and the Nth column; Represents the point of interest value in the spatial geographic unit at the Nth row and the Nth column; Represents the social activity level value in the spatial geographic unit at the Nth row and the Nth column.

3. The network traffic prediction method based on spatiotemporal series according to claim 1 is characterized in that: The nonlinear transformation includes a composite function of three consecutive operations; the composite functions of the three consecutive operations are batch normalization, rectified linearization, and convolution.

4. The network traffic prediction method based on spatiotemporal series according to claim 1 is characterized in that: The calculation expression for multiplying the combination matrix and the weight matrix in step S5 is: Where Y is the network traffic prediction result, and W is the weight matrix; is the combination matrix; It is the result of convolution of spatial features in matrixed data; is the result of convolution of the time features in the matrixed data; cross ′ is the result of convolution of the influencing features in the external influencing factor data.

5. A system based on the network traffic prediction method based on spatiotemporal series according to any one of claims 1 to 4, characterized in that: It includes target area data processing module, space processing module, time processing module, external factor processing module and parameter matrix fusion module; The target area data processing module is used to obtain and matrix the flow data of the target area in time and space to obtain matrixed data; Modeling the external influencing factor data in the traffic data of the target area; The spatial processing module is used to obtain spatial features in the matrixed data through densely connected convolutional neural networks; A time processing module, used to use the spatial features in the matrixed data as inputs of a diffusion convolution gated recurrent unit to obtain the time features in the matrixed data; An external factor processing module is used to obtain the influencing features in the external influencing factor data through a convolution method; The parameter matrix fusion module is used to convolve the spatial features in the matrixed data, the temporal features in the matrixed data, and the influence features in the external influencing factor data into a matrix to obtain a combined matrix; the combined matrix is ​​multiplied by the weight matrix, and the multiplication result is activated by the sigmoid function to obtain the network traffic prediction result.

Images