Intelligent cellular network flow prediction method and system based on recurrent neural network

Abstract

The invention relates to an intelligent cellular network flow prediction method and system based on a recurrent neural network. The method comprises the following steps: S1, establishing an intelligent cellular network model; s2, establishing a task model of cellular traffic prediction, and analyzing a task solution; s3, analyzing historical cellular traffic data and a cross-domain data set; s4, establishing an STCNet prediction algorithm model, and predicting the regional flow on the urban scale; s5, the STCNet is improved, and a transfer learning strategy is explored; and S6, evaluating the prediction error. According to the method, the problem of complexity and difficulty in predicting the regional cellular network flow is solved, the method is more reasonable and flexible, and the prediction accuracy can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of mobile communication technology, and in particular relates to an intelligent cellular network traffic prediction method and system based on a recurrent neural network. Background Art

[0002] With the development of 5G mobile communications and the advent of the intelligent era, achieving accurate cellular network traffic prediction has become a necessary step for communication systems to move towards intelligence. Cellular network traffic prediction is not only conducive to load balancing and reasonable allocation of network resources, but also can assist mobile terminals in edge computing and task offloading.

[0003] However, due to the complex external environment, the complexity of dynamic mobile behavior, and the randomness of network service requests, the task of cellular network traffic prediction is difficult. Especially when it comes to regional traffic prediction at the city scale, the complex spatiotemporal dependencies between cells make it more difficult to capture the traffic characteristics. Finding a method that can combine cross-domain data sets to accurately predict regional cellular network traffic is an important problem that needs to be solved urgently.

[0004] In addition, driven by artificial intelligence, cellular networks are gradually developing towards intelligence. More and more studies have shown that deep learning models have excellent performance in time series prediction, and the introduction of artificial intelligence into the prediction of cellular network traffic has become a research hotspot. Summary of the invention

[0005] In view of the above problems, the present invention proposes an intelligent cellular network traffic prediction method and system based on recurrent neural network, by introducing a recurrent neural network deep learning framework STCNet (Spatial-Temporal Cross-domainNetwork) based on ConvLSTM (Convolutional Long Short Term Memory) to capture its complex dependencies in the spatiotemporal domain, and adopting multiple input networks, while taking cellular network traffic data, cross-domain data and time feature data as input, so as to simultaneously extract the autocorrelation characteristics of the traffic data and its relationship with external influencing factors, so as to perform traffic prediction.

[0006] In order to achieve the above purpose, the present invention adopts the following technical solutions: The intelligent cellular network traffic prediction method based on recurrent neural network includes the following steps: S1, building an intelligent cellular network model; S2, using the intelligent cellular network model and combining it with the current urban wide area cellular network traffic data set, establish a task model for cellular traffic prediction, analyze the solution to the task, and describe the cellular traffic prediction problem; S3, according to the cellular traffic prediction problem, propose cellular traffic data and cross-domain data sets, and perform correlation analysis; S4, based on the problem analysis results of step S3, establish a STCNet prediction algorithm model to predict regional traffic at the city scale; S5, improve the STCNet prediction algorithm model, explore transfer learning strategies, and achieve the best prediction of traffic; S6, evaluating the prediction error of the best prediction.

[0007] Preferably, in step S1, the intelligent cellular network model includes a data center, a cloud server, an intelligent base station, an intelligent drone and a user; wherein the cloud server is used to provide cloud computing services; the data center is used to collect and store cellular traffic and cross-domain information data; the intelligent base station constitutes the basic unit for establishing a network link; and the intelligent drone is assigned to task offloading and edge computing tasks according to real-time conditions.

[0008] Preferably, step S2 specifically includes the following steps: S21, for a given T The mobile traffic data of the entire network is observed in the time interval, and the corresponding setting is expressed as A space-time sequence of D yes t At all times X*Y Mobile traffic data in a geographical area represented by a grid, namely: (1); in, Represents a coordinate ( X , Y )Data traffic in the grid; S22, the problem of using a simple spatiotemporal sequence algorithm to predict cellular network traffic is modeled as predicting the most likely given the previous S observations. k Step data point sequence, that is: (2); in, , , The predicted t, t+k, t-S+1 Data point at time; function For the spatiotemporal series prediction algorithm, the parameters learned by the algorithm is the function's parameter; S23, considering that the generation of cellular network traffic is not only affected by time and space factors, but also by cross-domain factors, we focus on the following three cross-domain factors: Base station distribution information BSs, city point of interest information POI and social activity level information.

[0009] Preferably, in step S3, the historical cellular traffic data adopts a city wide area cellular network traffic data set, which specifically includes three types: traffic data Internet, call service data Call and short message service data SMS; the historical cellular traffic data divides the selected city into a set number of grids, each grid covering the same urban area; the ID, time interval and detailed business records of each grid are recorded in each sample; In step S3, the cross-domain data set includes base station information, city point of interest information and social activity level information; the base station information covers a variety of base station-related information, including the specific location of the base station, the mobile country code and the coverage of each base station; the city point of interest information includes 13 types of points of interest, and the detailed description of each category is added together with the number of each category to form the final representation; the data obtained by the social activity level information includes information generated by users when using Twitter software, which is used to reflect the overall demand level of users for network services.

[0010] Preferably, in step S4, the STCNet prediction algorithm model includes the following two parts: The modeling and feature learning part based on multiple ConvLSTMs and the prediction part based on the densely connected convolutional neural network DensNet; Step S4 specifically includes the following steps: S41, the modeling and feature learning part based on multiple ConvLSTMs includes three network models, each of which receives cellular network traffic data, time feature data and cross-domain data respectively; when processing input tasks, the STCNet prediction algorithm model adopts a structure of a combination of three network models, and each network model performs feature learning on different input data respectively; the preliminary feature outputs of the three network models are spliced ​​according to the specified dimensions as the input of the densely connected convolutional neural network DensNet, and the output of DensNet is the final prediction value; S42, splicing feature learning: construct a spatiotemporal convolutional network STCNet to effectively extract the complex features of cellular network traffic data through dense convolution.

[0011] Preferably, step S41 includes the following steps: S411, constructing a spatiotemporal correlation model for receiving cellular network traffic data; using a two-layer ConvLSTM network for spatiotemporal modeling; the output of ConvLSTM is expressed as: (3); Among them, H represents the hidden state, F represents the forget gate, W represents the weight to be learned, and R represents the observed feature domain; S412, constructing a time feature information embedding model for receiving date and time interval information of cellular network traffic; using a two-layer linear neural network to build a model and process the corresponding feature vectors. The specific processing process is as follows: Extract four features of date and time, including day of the week, time of day, whether it is a weekday, and whether it is a weekend; encode the four data to form a feature vector m ; The feature vector m It is sent to a two-layer fully connected neural network for embedding learning to perform feature fusion; after passing through the fully connected neural network, m The specific calculation process is as follows: (4); in, and are the weights and biases that need to be learned in the neural network, l=1,2 , indicating the l Layer network; Indicates that the Sigmoid function is used as the activation function. is the output of the feature vector m; Finally, the reshape function is used again, and the output of feature embedding is finally displayed as: (5); S413, construct a cross-domain data model for receiving cross-domain data information; the cross-domain data model body includes a two-layer convolutional neural network; the base station data, point of interest data and social activity data are spliced ​​according to their respective specific dimensions to form a new data tensor ; Then the integrated data tensor Input into the convolutional neural network, through convolution operation and nonlinear transformation, extract the preliminary features of cross-domain data; the preliminary features will serve as an important input for the subsequent prediction model. The specific formula is as follows: (6); (7); in, Represents a splicing operation, are the parameters that need to be learned; It is a composite function consisting of regularization, activation function and convolution operation; is the traffic data of the base station, is the traffic data of the city's points of interest, It is the traffic data at the level of social activities.

[0012] Preferably, step S42 includes the following steps: S421, splicing the preliminary feature outputs of the spatiotemporal correlation model, the temporal feature information embedding model and the cross-domain data model in step S41 to obtain a unified representation after the preliminary feature fusion , the specific formula is as follows: (8); S422, input the spliced ​​features into into a dense neural convolutional neural network; the dense convolutional neural network consists of l Layers, each layer through a composite function Performing feature transformation; the composite function is composed of batch normalization, activation function and convolution operation; S423, considering that the input of each layer in the dense convolutional neural network contains the output of all previous layers, we set l Output of the layer It is expressed as: (9); Finally, a highly abstract feature representation is generated at layer L ; The feature representation is fed into an activation function to produce the final prediction result ; S424, define the optimization goal as minimizing the predicted value The Frobenius norm between the true value Y , used to train STCNet: (10); in, Represents the set of parameters that need to be optimized in STCNet.

[0013] Preferably, step S5 specifically includes the following process: S51, selecting a regional clustering-based transfer learning strategy to improve the prediction performance of the model; considering that there is a certain correlation between cellular network traffic data in different regions, the regions with similar cellular regional services are divided into the same group, and the models with different parameters are used to optimize the regions in different groups; S52, using a spectral clustering algorithm to cluster the regions, so as to maintain the spatial positions of the cells while clustering, as follows: Cellular Network Traffic Dataset With cross-domain datasets , , Combined, an overall feature representation D is formed as the data representation of all areas in the city; S53, the entire area is modeled as a graph, and each cell becomes a node in the graph. The edge information is determined by whether the cells are connected. According to the relationship between vertices and edges, the corresponding adjacency matrix is ​​calculated. A , and the adjacency matrix A The corresponding gradient is the overall feature representation D of the composed cellular network traffic dataset and the cross-domain dataset; then, the Laplace matrix L is constructed as follows: (11); in, P is a diagonal matrix, P In Element Representative A No. i The sum of the rows; then, we get the Laplace matrix L k The largest eigenvector , and after stacking into columns we get the matrix ; S54, calculating the matrix X obtained in step S53 to obtain a cluster label corresponding to each cell; S55, train STCNet in the first family and save the model parameters trained in the corresponding family; S56, using the model parameters output from step S55, initialize the network parameters of the training model on the second category, and continue training; the cycle is repeated until the training is completed. k Learning from class data.

[0014] Preferably, step S6 specifically includes the following process: S61, three classic criteria are selected to evaluate the prediction error, including the mean root error RMSE , mean absolute error MAE and the coefficient of determination R2; in the evaluation of the prediction results, RMSE and MAE The smaller the value of is, the better the performance of the STCNet prediction algorithm model is; and the closer the value of R2 is to 1, the better the fitting effect of the STCNet prediction algorithm model is and the better the performance is.

[0015] The present invention also discloses an intelligent cellular network traffic prediction system based on a recurrent neural network, comprising the following modules: An initial intelligent cellular network model module is used to construct an initial intelligent cellular network model based on data traffic distribution and network system component relationships; The task model analysis module of cellular traffic prediction is used to establish the task model of cellular traffic prediction and analyze the solution of the task; Traffic data analysis module, used to analyze historical cellular traffic data and cross-domain data sets; STCNet algorithm prediction module, used to establish the STCNet prediction algorithm model and predict regional traffic at the city scale; STCNet improvement and transfer learning module, used to improve STCNet and consider combining transfer learning strategies; The prediction error evaluation module is used to evaluate the errors generated by the prediction results.

[0016] Compared with the prior art, the present invention has the following technical effects: (1) The present invention uses the spatiotemporal cross-domain neural network STCNet to predict cross-domain cellular network traffic. It cleverly combines the characteristics of convolutional neural networks and long short-term neural networks to use ConvLSTM for spatiotemporal modeling, while using a linear layer to capture the temporal characteristics of cellular network traffic data and a feature splicing method to capture the features of cross-domain data for auxiliary prediction. (2) The present invention expands transfer learning between different businesses, avoids the cumbersome and complex data caused by coupling regional differentiation factors in different regions, and improves the accuracy of traffic prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A flow chart of the intelligent cellular network traffic prediction method based on recurrent neural network in the present invention; Figure 2 A network model diagram for intelligent cellular network traffic prediction based on recurrent neural network provided by an embodiment of the present invention; Figure 3 A network model diagram based on regional clustering transfer learning provided by an embodiment of the present invention; Figure 4 It is a principle block diagram of the intelligent cellular network traffic prediction system based on recurrent neural network in the present invention. DETAILED DESCRIPTION

[0018] The following describes the embodiments of the present invention through specific embodiments, 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 following embodiments and features in the embodiments can be combined with each other without conflict.

[0019] like Figure 1As shown, this embodiment proposes an intelligent cellular network traffic prediction method based on a recurrent neural network, which combines the recurrent neural network deep learning framework STCNet (Spatial-Temporal Cross-domainNetwork) and transfer learning. First, an intelligent cellular network model is constructed, and at the same time, a task model for cellular traffic prediction is established in combination with the current mainstream data sets, the solution to the task is analyzed, the cellular traffic prediction problem is described, and cellular traffic data and cross-domain data sets are proposed and analyzed based on the problem. Based on the results of the above problem analysis, an STCNet prediction algorithm model is established to predict regional traffic at the city scale. At the same time, transfer learning is added on the basis of the STCNet prediction algorithm model to achieve the best prediction of traffic.

[0020] This embodiment provides a network model diagram of intelligent cellular network traffic prediction using a recurrent neural network, using the STCNet learning framework. The model diagram refers to Figure 2 As shown in the figure, the model is used to capture the complex dependencies in the time and space domains, and multiple input networks are used. At the same time, cellular network traffic data, cross-domain data and time feature data are used as input to simultaneously extract the autocorrelation characteristics of traffic data and its relationship with external influencing factors to perform traffic prediction. The specific steps include: 1. Establish an intelligent cellular network model, including: The intelligent cellular network consists of data centers, cloud servers, intelligent base stations, intelligent drones, and users. The cloud servers provide cloud computing services, the data centers collect and store previous cellular traffic and cross-domain information, and intelligent base stations form the basic unit for establishing network links. Intelligent drones are assigned to tasks such as task offloading and edge computing according to real-time conditions.

[0021] 2. Establish a task model for cellular traffic prediction and analyze the solution to the task, including: 2-1. For the mobile traffic data observed in the entire network in a given time interval, it is expressed as A space-time sequence of D yes t At all times X*Y Mobile traffic data in a geographical area represented by a grid, namely: ; in, Represents a coordinate ( X , Y )Data traffic in the grid.

[0022] 2-2. The simple problem of using spatiotemporal sequence algorithms to predict cellular network traffic is modeled as predicting the most likely given the previous S observations.k Step data point sequence, that is: ; in, , , The predicted t, t+k, t-S+1 Data point at time; function For the spatiotemporal series prediction algorithm, the parameters learned by the algorithm is the function's parameter; 2-3. The generation of cellular network traffic is not only affected by time and space factors, but also by a variety of cross-domain factors. The present invention mainly focuses on the following three cross-domain factors: the distribution information of base stations (BSs), urban point of interest information (POI) and social activity level information: ; ; ; in , , They are t At all times H*W The distribution information of base stations (BSs), urban points of interest (POI) and social activity level information in the geographical area represented by the grid, , , Respectively represent the data flow in the coordinate grid.

[0023] 3. Analyze historical cellular traffic data and cross-domain datasets, including: 3-1. The dataset used is the urban wide-area cellular network traffic dataset released by a certain country's telecommunications company in 2015. This dataset covers the aggregated mobile service traffic sent or received in each cell of a certain city from November 1, 2013 to December 31, 2013. It includes three types: traffic data (Internet), call service data (Ca1l), and short message service data (SMS). This dataset divides a certain city into 100x100 grids, each covering an area of ​​235x235 square meters. In each sample, the ID, time interval, and detailed service records of each grid are recorded separately: a) The total number of SMS messages received by users in the area; b) SMS sent: the total number of SMS messages sent by users in the area; c) Received calls: the total number of calls received by users in the area; d) Outgoing calls: the total number of outgoing calls made by users in the area; e) Internet data: Internet traffic data of users in the area, recorded every 15 minutes.

[0024] 3-2. The cross-domain dataset contains base station information, point of interest distribution, and social activity level. Base station information covers a variety of information related to base stations, including their specific locations (given in the form of longitude and latitude), mobile country codes, and the approximate coverage of each base station. The city point of interest distribution information includes 13 types of points of interest, such as subway stations, shops, restaurants, etc. The detailed description of each category is added together with the number of each category to form the final representation. The data obtained from the community's social level information contains information generated by users when using the Twitter software, such as location and keywords, reflecting the overall demand for network services by users.

[0025] 4. Establish the STCNet prediction algorithm model to predict regional traffic at the city scale, including: The STCNet network mainly consists of two parts: the modeling and feature learning part based on multiple ConvLSTMs and the prediction part based on densely connected convolutional neural network (DensNet): 4-1. The first part of STCNet has three components, each of which receives cellular network traffic data, time feature data, and cross-domain data. When processing input tasks, the model adopts a structure that combines three network models, and each network model performs feature learning on different input data. The preliminary feature outputs of the three models are spliced ​​according to the specified dimensions as the input of DensNet. The output of DensNet is the final prediction value. The specific process is as follows: 4-1-1. Temporal and spatial correlation modeling. Receive cellular network traffic data, which is composed of multi-level ConvLSTM.

[0026] enter , represents the data flow matrix. It is analogous to a p Frames of video data, each of which can be regarded as single-channel image data. Convolutional neural networks have shown strong capabilities in modeling spatial characteristics, but are insufficient in capturing temporal characteristics; while long short-term memory networks are known for their excellent time series prediction capabilities, but are slightly inferior in modeling spatial characteristics. In order to complement the advantages of the two networks, STCNet uses a two-layer network of ConvLSTM for spatiotemporal modeling. The calculation method of ConvLSTM is: ; in: Represents input, represents the unit output, represents the hidden state, represents the input gate, represents the forget gate, Represents the output gate. and are the weights and biases that need to be learned. It should be noted that the ConvLSTM unit are all three-dimensional tensors. F Denoted as the number of feature maps, the final output of ConvLSTM is expressed as: ; Among them, H represents the hidden state, F represents the forget gate, W represents the weight to be learned, and R represents the observed feature domain; 4-1-2. Time feature information embedding. The date and time interval information of cellular network traffic is received and this feature vector is processed using a two-layer linear neural network.

[0027] Four features of the date and time will be extracted: day of the week, time of the day, whether it is a weekday, and whether it is a weekend. For example, 15:00 on May 16, 2024, will become: Thursday, 15:00, a weekday, and not a weekend after feature extraction. These four data are encoded to form a feature vector m . This feature vector m It is fed into a two-layer fully connected neural network for embedding learning to perform feature fusion.

[0028] After passing through the fully connected neural network, m The specific calculation process is as follows: ; in, and are the weights and biases that need to be learned in the neural network, l=1,2 , indicating the l Layer network; Indicates that the Sigmoid function is used as the activation function. is the output of the feature vector m; Finally, the reshape function is used again, and the output of feature embedding is finally displayed as: .

[0029] Considering the potential connection between regional cellular network traffic and external data, a two-layer convolutional neural network model was selected to fuse and process these two types of information. First, the cross-domain datasets were integrated. The specific steps include splicing base station data, point of interest data, and social activity data according to their respective specific dimensions to form a new data tensor. ; Then this integrated The data tensor is input into the convolutional neural network, and the preliminary features of the cross-domain data are extracted through convolution operations and nonlinear transformations. These features will serve as important inputs for subsequent prediction models to improve the accuracy and reliability of cellular network traffic prediction: ; ; in, Represents a splicing operation, are the parameters that need to be learned. It is a composite function consisting of regularization, activation function and convolution operation; is the traffic data of the base station, is the traffic data of the city's points of interest, It is the traffic data at the level of social activities.

[0030] 4-2. Splicing feature learning. Construct STCNet (Spatiotemporal Convolutional Network) to effectively extract the complex features of cellular network traffic data through dense convolution. The specific process is as follows: 4-2-1. First, concatenate the preliminary feature outputs of the three modules, ConvLSTM, display time feature encoding, and cross-domain data modeling, to obtain a unified representation after the preliminary feature fusion. . Specifically expressed as: ; 4-2-2. A spliced ​​feature Input into a dense neural convolutional neural network. This dense convolutional neural network consists of l Layers, each layer through a composite function Perform feature transformation. This function is composed of modules such as batch normalization, activation function, and convolution operation, ensuring that the model has strong generalization capabilities when processing cross-domain datasets.

[0031] 4-2-3. The input of each layer includes the output of all previous layers. l Output of the layer It can be expressed as: ; In this way, the model can gradually accumulate and fuse feature information at different levels, and finally L Layer produces a highly abstract feature representation ; This feature representation is fed into an activation function to produce the final prediction result : ; in is the activation function; 4-2-4. Define the optimization goal as minimizing the predicted value The Frobenius norm between the true value Y To train STCNet: ; in, Represents the set of parameters that need to be optimized in STCNet.

[0032] This embodiment also provides a network model diagram based on transfer learning. Figure 3 As shown in the figure, in this network, STCNet is improved and the transfer learning strategy is explored, which specifically includes the following processes: 5-1. Considering that there is a certain correlation between cellular network traffic in different regions and different services, a regional clustering transfer learning strategy is selected to improve the prediction performance of the model. Cellular network traffic data in different regions have a certain correlation. Regions with small differences are divided into the same group, and then models with different parameters are used to optimize regions in different groups.

[0033] 5-2. Use spectral clustering algorithm to cluster the regions, so as to maintain the spatial location of the cells while clustering. First, the cellular network traffic dataset With cross-domain datasets , , Combined, an overall feature representation D is formed as the data representation of all areas in the city.

[0034] 5-3. The entire area is modeled as a graph, and each cell becomes a node in the graph. The edge information is determined by whether the cells are connected. Specifically, two cells that are connected horizontally or vertically will form an edge between their landmark nodes. In this way, the graph topology of the entire area is obtained. According to the relationship between vertices and edges, the corresponding adjacency matrix is ​​calculated. A , and its gradient is the overall feature representation D of the composed cellular network traffic dataset and cross-domain dataset. Then, the Laplacian matrix can be constructed: ; in, P is a diagonal matrix, P In Element Representative A No. i The sum of the rows; then, we get the Laplace matrix L k The largest eigenvector , and after stacking into columns we get the matrix ; 5-4. Get the cluster label corresponding to each cell.

[0035] 5-5. Train STCNet in the first cluster and save the model parameters trained in this cluster.

[0036] 5-6. Use the parameters output by S55 to initialize the network parameters of the training model for the second category and continue training. Repeat this process until the training is completed. k Learning from class data.

[0037] 6. Evaluate the forecast error, including: Three classic criteria are selected to evaluate the prediction error, namely, root mean square error RMSE , mean absolute error MAE And the coefficient of determination R2 . It is defined as: ; ; ; In the evaluation of the prediction results, RMSE and MAE The smaller the value of , the better the performance of the model; and the closer the value of is to 1, the better the model fitting effect and the better the performance.

[0038] like Figure 4 As shown, the present invention also provides an intelligent cellular network traffic prediction system based on a recurrent neural network, which is used to execute the method of the above embodiment, and includes the following modules: An initial intelligent cellular network model module is used to construct an initial intelligent cellular network model based on data traffic distribution and network system component relationships; The task model analysis module of cellular traffic prediction is used to establish the task model of cellular traffic prediction and analyze the solution of the task; Traffic data analysis module, used to analyze historical cellular traffic data and cross-domain data sets; STCNet algorithm prediction module, used to establish the STCNet prediction algorithm model and predict regional traffic at the city scale; STCNet improvement and transfer learning module, used to improve STCNet and consider combining transfer learning strategies; The prediction error evaluation module is used to evaluate the errors generated by the prediction results.

[0039] The above description is only a detailed description of the preferred embodiments and principles of the present invention. For ordinary technicians in this field, according to the ideas provided by the present invention, there will be changes in the specific implementation methods, and these changes should also be regarded as the protection scope of the present invention.

Claims

1. An intelligent cellular network traffic prediction method based on recurrent neural network, characterized in that: The following steps are involved: S1, building an intelligent cellular network model; S2, using the intelligent cellular network model and combining it with the current urban wide area cellular network traffic data set, establish a task model for cellular traffic prediction, analyze the solution to the task, and describe the cellular traffic prediction problem; S3, according to the cellular traffic prediction problem, propose cellular traffic data and cross-domain data sets, and perform correlation analysis; S4, based on the problem analysis results of step S3, establish a STCNet prediction algorithm model to predict regional traffic at the city scale; S5, improve the STCNet prediction algorithm model, explore transfer learning strategies, and achieve the best prediction of traffic; S6, evaluating the prediction error of the best prediction.

2. The intelligent cellular network traffic prediction method based on recurrent neural network according to claim 1 is characterized in that: In step S1, the intelligent cellular network model includes a data center, a cloud server, an intelligent base station, an intelligent drone and a user; wherein the cloud server is used to provide cloud computing services; the data center is used to collect and store cellular traffic and cross-domain information data; the intelligent base station constitutes the basic unit for establishing a network link; the intelligent drone is assigned to task offloading and edge computing tasks according to real-time conditions.

3. The intelligent cellular network traffic prediction method based on recurrent neural network according to claim 1 is characterized in that: Step S2 specifically includes the following steps: S21, for a given T The mobile traffic data of the entire network is observed in the time interval, and the corresponding setting is expressed as A space-time sequence of D yes t Time X*Y Mobile traffic data in a geographical area represented by a grid, namely: （1）； in, Represents a coordinate ( X , Y )Data traffic in the grid; S22, the problem of using a simple spatiotemporal sequence algorithm to predict cellular network traffic is modeled as predicting the most likely given the previous S observations. k Step data point sequence, that is: （2）； in, , , The predicted t, t+k, t-S+1 Data point at time; function For the spatiotemporal series prediction algorithm, the parameters learned by the algorithm is the function's parameter; S23, considering that the generation of cellular network traffic is not only affected by time and space factors, but also by cross-domain factors, we focus on the following three cross-domain factors: Base station distribution information BSs, city point of interest information POI and social activity level information.

4. The intelligent cellular network traffic prediction method based on recurrent neural network according to claim 1 is characterized in that: In step S3, the historical cellular traffic data adopts the urban wide area cellular network traffic data set, which specifically includes three types: traffic data Internet, call service data Call and short message service data SMS; The historical cellular traffic data divides the selected city into a set number of grids, each grid covering the same urban area; In each sample, the ID, time interval and detailed business records of each grid are recorded; In step S3, the cross-domain data set includes base station information, city point of interest information and social activity level information; the base station information covers a variety of base station-related information, including the specific location of the base station, the mobile country code and the coverage of each base station; the city point of interest information includes 13 types of points of interest, and the detailed description of each category is added together with the number of each category to form the final representation; the data obtained by the social activity level information includes information generated by users when using Twitter software, which is used to reflect the overall demand level of users for network services.

5. The intelligent cellular network traffic prediction method based on recurrent neural network according to claim 1 is characterized in that: In step S4, the STCNet prediction algorithm model includes the following two parts: The modeling and feature learning part based on multiple ConvLSTMs and the prediction part based on the densely connected convolutional neural network DensNet; Step S4 specifically includes the following steps: S41, the modeling and feature learning part based on multiple ConvLSTMs includes three network models, each of which receives cellular network traffic data, time feature data and cross-domain data respectively; when processing input tasks, the STCNet prediction algorithm model adopts a structure of a combination of three network models, and each network model performs feature learning on different input data respectively; the preliminary feature outputs of the three network models are spliced ​​according to the specified dimensions as the input of the densely connected convolutional neural network DensNet, and the output of DensNet is the final prediction value; S42, splicing feature learning: construct a spatiotemporal convolutional network STCNet to effectively extract the complex features of cellular network traffic data through dense convolution.

6. The intelligent cellular network traffic prediction method based on recurrent neural network according to claim 5 is characterized in that: Step S41 includes the following steps: S411, constructing a spatiotemporal correlation model for receiving cellular network traffic data; using a two-layer ConvLSTM network for spatiotemporal modeling; the output of ConvLSTM is expressed as: （3）； Among them, H represents the hidden state, F represents the forget gate, W represents the weight to be learned, and R represents the observed feature domain; S412, constructing a time feature information embedding model for receiving date and time interval information of cellular network traffic; using a two-layer linear neural network to build a model and process the corresponding feature vectors. The specific processing process is as follows: Extract four features of date and time, including day of the week, time of day, whether it is a weekday, and whether it is a weekend; encode the four data to form a feature vector m ; The feature vector m It is sent to a two-layer fully connected neural network for embedding learning to perform feature fusion; after passing through the fully connected neural network, m The specific calculation process is as follows: （4）； in, and are the weights and biases that need to be learned in the neural network, l=1,2 , indicating the l Layer network; Indicates that the Sigmoid function is used as the activation function. is the output of the feature vector m; Finally, the reshape function is used again, and the output of feature embedding is finally displayed as: （5）； S413, construct a cross-domain data model for receiving cross-domain data information; the cross-domain data model body includes a two-layer convolutional neural network; the base station data, point of interest data and social activity data are spliced ​​according to their respective specific dimensions to form a new data tensor ; Then the integrated data tensor Input into the convolutional neural network, through convolution operation and nonlinear transformation, extract the preliminary features of cross-domain data; the preliminary features will serve as an important input for the subsequent prediction model. The specific formula is as follows: （6）； （7）； in, Represents a splicing operation, are the parameters that need to be learned; It is a composite function consisting of regularization, activation function and convolution operation; is the traffic data of the base station, is the traffic data of the city's points of interest, It is the traffic data at the level of social activities.

7. The intelligent cellular network traffic prediction method based on recurrent neural network according to claim 6 is characterized in that: Step S42 includes the following steps: S421, splicing the preliminary feature outputs of the spatiotemporal correlation model, the temporal feature information embedding model and the cross-domain data model in step S41 to obtain a unified representation after the preliminary feature fusion , the specific formula is as follows: （8）； S422, input the spliced ​​features into into a dense neural convolutional neural network; the dense convolutional neural network consists of l Layers, each layer through a composite function Performing feature transformation; the composite function is composed of batch normalization, activation function and convolution operation; S423, considering that the input of each layer in the dense convolutional neural network contains the output of all previous layers, we set l Output of the layer It is expressed as: （9）； Finally, a highly abstract feature representation is generated at layer L ; The feature representation is fed into an activation function to produce the final prediction result ; S424, define the optimization goal as minimizing the predicted value The Frobenius norm between the true value Y , used to train STCNet: （10）； in, Represents the set of parameters that need to be optimized in STCNet.

8. The intelligent cellular network traffic prediction method based on recurrent neural network according to claim 1 is characterized in that: Step S5 specifically includes the following process: S51, selecting a regional clustering-based transfer learning strategy to improve the prediction performance of the model; considering that there is a certain correlation between cellular network traffic data in different regions, the regions with similar cellular regional services are divided into the same group, and the models with different parameters are used to optimize the regions in different groups; S52, using a spectral clustering algorithm to cluster the regions, so as to maintain the spatial positions of the cells while clustering, as follows: Cellular Network Traffic Dataset With cross-domain datasets , , Combined, an overall feature representation D is formed as the data representation of all areas in the city; S53, the entire area is modeled as a graph, and each cell becomes a node in the graph. The edge information is determined by whether the cells are connected. According to the relationship between vertices and edges, the corresponding adjacency matrix is ​​calculated. A , and the adjacency matrix A The corresponding gradient is the overall feature representation D of the composed cellular network traffic dataset and the cross-domain dataset; Then, the Laplace matrix L is constructed as follows: （11）； in, P is a diagonal matrix, P In Element Representative A No. i The sum of the rows; then, we get the Laplace matrix L k The largest eigenvector , and after stacking into columns we get the matrix ; S54, calculating the matrix X obtained in step S53 to obtain a cluster label corresponding to each cell; S55, train STCNet in the first family and save the model parameters trained in the corresponding family; S56, using the model parameters output from step S55, initialize the network parameters of the training model on the second category, and continue training; the cycle is repeated until the training is completed. k Learning from class data.

9. The intelligent cellular network traffic prediction method based on recurrent neural network according to claim 1 is characterized in that: Step S6 specifically includes the following process: S61, three classic criteria are selected to evaluate the prediction error, including the mean root error RMSE , mean absolute error MAE And the coefficient of determination R2 ; In the evaluation of the prediction results, RMSE and MAE The smaller the value of, the better the performance of the STCNet prediction algorithm model; R2 The closer the value is to 1, the better the fitting effect of the STCNet prediction algorithm model is and the better the performance is.

10. An intelligent cellular network traffic prediction system based on a recurrent neural network, used to execute the method according to any one of claims 1 to 9, characterized in that: Includes the following modules: An initial intelligent cellular network model module is used to construct an initial intelligent cellular network model based on data traffic distribution and network system component relationships; The task model analysis module of cellular traffic prediction is used to establish the task model of cellular traffic prediction and analyze the solution of the task; Traffic data analysis module, used to analyze historical cellular traffic data and cross-domain data sets; STCNet algorithm prediction module, used to establish the STCNet prediction algorithm model and predict regional traffic at the city scale; STCNet improvement and transfer learning module, used to improve STCNet and consider combining transfer learning strategies; The prediction error evaluation module is used to evaluate the errors generated by the prediction results.

Images