Prediction method and device of communication network traffic, electronic equipment and storage medium

Abstract

The invention provides a communication network traffic prediction method and device, electronic equipment and a computer readable storage medium, and relates to the technical field of communication networks. The prediction method comprises the steps that multivariate cross-domain data of all cells in a target area are acquired, and the multivariate cross-domain data comprise auxiliary feature data and spatial feature data; determining the data validity of each cell in the target area according to the auxiliary feature data; determining a low data validity sub-region set in the target region according to spatial feature data and the data validity; and constructing a communication network flow prediction model of each low-data-validity sub-region in the low-data-validity sub-region set, and predicting the communication network flow of each low-data-validity sub-region in the low-data-validity sub-region set based on the communication network flow prediction model. The problem of low accuracy of communication network data traffic prediction in the prior art is at least solved. The method is suitable for traffic prediction scenes.

Description

Technical Field

[0001] The present invention relates to the field of communication network technology, and in particular to a method, device, electronic device and computer-readable storage medium for predicting communication network traffic. Background Art

[0002] With the rapid development of communication networks, the data traffic of communication networks has increased dramatically. At present, 5G (5th Generation Mobile Communication Technology, the fifth generation of mobile communication technology) networks are being built on a large scale, 5.5G (5G-Advanced, the evolution and enhancement of 5G, a mobile communication technology in the transition stage between 5G and 6G) ​​networks are beginning to emerge, and 6G (6th Generation Mobile Communication Technology, the sixth generation of mobile communication technology) networks are ready to go. The construction of communication networks is essentially to meet the massive data service needs of users anytime and anywhere. However, as the number of base stations increases, the cost of communication network construction and maintenance will rise sharply. Communication operators cannot build base stations without restrictions or keep base stations open 24 hours a day just to meet user needs. Therefore, it is necessary to accurately predict the user's communication network traffic in order to select the appropriate base station address and equipment type in the subsequent planning stage and design an intelligent base station sleep strategy.

[0003] However, the existing technology has not fully considered the differences that arise in the process of predicting communication network traffic, including but not limited to: differences in multivariate cross-domain data caused by changes in time or space, and differences in the distribution of cells in the target area, which affect the division of communication network areas and thus affect the accuracy of communication network data traffic prediction, reduce the accuracy of selecting appropriate base station addresses and equipment types in the subsequent planning stage and designing intelligent base station sleep strategies, and increase the cost of communication network construction and maintenance. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a communication network traffic prediction method, device, electronic device and computer-readable storage medium in view of the above-mentioned deficiencies in the prior art. The method can realize accurate prediction of communication network traffic, improve the accuracy of selecting base station addresses and equipment types in the subsequent planning stage and designing intelligent base station sleep strategies, and reduce the construction and maintenance costs of communication networks.

[0005] In a first aspect, the present invention provides a method for predicting communication network traffic, comprising: obtaining multivariate cross-domain data of all cells in a target area, wherein the multivariate cross-domain data includes auxiliary feature data and spatial feature data; determining the data validity of each cell in the target area based on the auxiliary feature data, wherein the data validity is used to characterize the data quality of the multivariate cross-domain data of the cell that changes over time; determining a set of low data validity sub-regions in the target area based on the spatial feature data and the data validity, wherein the low data validity sub-regions are used to characterize the sub-regions under the cell with poor communication network traffic data quality; constructing a communication network traffic prediction model for each low data validity sub-region in the low data validity sub-region set, and predicting the communication network traffic of each low data validity sub-region in the low data validity sub-region set based on the communication network traffic prediction model.

[0006] Preferably, determining the data validity of each cell in the target area based on the auxiliary feature data specifically includes: obtaining the weight and sample size of the auxiliary feature data of each cell in the target area; evaluating the data completeness of the auxiliary feature data of each cell in the target area, wherein the auxiliary feature data includes one or more combinations of the following: natural attribute feature data, social attribute feature data, and the data completeness is used to characterize the proportion of non-empty feature data of the auxiliary feature data; calculating the data validity of each cell in the target area based on the weight, sample size and data completeness of the auxiliary feature data of each cell in the target area.

[0007] Preferably, the data validity of each cell in the target area is calculated according to the weight, sample size and data completeness of the auxiliary feature data of each cell in the target area, specifically including: calculating the data validity of each cell in the target area according to formula (1):

[0008]

[0009] Where T represents the current time, represents the sample capacity of the auxiliary feature data of the i-th cell in the target area at time t on the time scale S, represents the weight of the auxiliary feature data of the i-th cell in the target area at time t on the time scale S, Indicates the data completeness of the auxiliary feature data of the i-th cell in the target area at time t on the time scale S, It represents the data validity of the i-th cell in the target area at time T on the time scale S.

[0010] Preferably, determining a set of sub-regions with low data validity in the target area based on the spatial feature data and the data validity specifically includes: dividing the target area based on the Voronoi diagram method, the spatial feature data and the data validity to obtain target sub-regions under each cell, wherein the spatial feature data is used to characterize the physical location of the cell; sorting the data validity of the target sub-regions, and determining a set of sub-regions with low data validity in the target area according to the interquartile range method.

[0011] Preferably, the target area is divided based on the Voronoi diagram method, the spatial feature data and the data validity to obtain the target sub-area under each cell, specifically including: obtaining the neighbor relationship information of all cells in the target area, and determining the neighboring area of ​​the i-th cell in the target area according to the neighbor relationship information; calculating the Voronoi boundary points between the i-th cell and the neighboring area of ​​the i-th cell according to the spatial feature data and data validity of the i-th cell and the spatial feature data and data validity of the neighboring area of ​​the i-th cell; connecting the Voronoi boundary points to obtain the sub-area under the i-th cell; judging whether the i-th cell is within the sub-area under the i-th cell; in response to the i-th cell not being within the sub-area under the i-cell, connecting the Voronoi boundary points and the i-th cell to obtain the target sub-area under the i-cell; in response to the i-cell being within the sub-area under the i-cell, determining that the sub-area under the i-cell is the target sub-area under the i-cell.

[0012] Preferably, the calculating of the Voronoi boundary points between the i-th cell and the neighboring area of ​​the i-th cell according to the spatial feature data and data validity of the i-th cell and the spatial feature data and data validity of the neighboring area of ​​the i-th cell specifically includes: calculating the Voronoi boundary points between the i-th cell and the neighboring area of ​​the i-th cell according to formula (2):

[0013]

[0014] Among them, P i represents the Voronoi boundary point between the ith cell and the neighboring cell of the ith cell, (x i ,y i )、(x j ,y j ) represent the spatial feature data of the i-th cell and the j-th cell respectively, They respectively represent the data validity of the i-th cell and the j-th cell at time t in the S time scale, and the j-th cell refers to the neighboring cell of the i-th cell.

[0015] Preferably, the multivariate cross-domain data also includes communication network traffic data. After determining the set of low data validity sub-regions in the target area based on the spatial feature data and the data validity, and before constructing a communication network traffic prediction model for each low data validity sub-region in the set of low data validity sub-regions, the communication network traffic prediction method also includes: clustering the set of low data validity sub-regions based on the communication network traffic data and a clustering algorithm.

[0016] Preferably, clustering the set of low data validity sub-regions based on communication network traffic data and a clustering algorithm specifically includes: extracting the communication network traffic data of each low data validity sub-region in the set of low data validity sub-regions; based on a time series analysis algorithm, performing time series decomposition and fitting on the communication network traffic data of each low data validity sub-region, to obtain a communication network traffic time series corresponding to each low data validity sub-region; based on a K-means clustering algorithm, calculating the distance between the communication network traffic time series, and clustering the set of low data validity sub-regions based on the distance.

[0017] Preferably, the constructing of the communication network traffic prediction model for each low data validity sub-region in the low data validity sub-region set specifically includes: constructing the communication network traffic prediction model for each low data validity sub-region in the low data validity sub-region set based on the width learning algorithm; The mapping nodes, enhancement nodes and model parameters of each cell under the cluster are obtained to obtain the first The initial prediction model of each cell under the cluster, where N represents the number of clusters in the low data validity sub-region set; based on the reinforcement learning algorithm, optimize the The initial prediction model of each cell under the cluster is obtained in the low data validity sub-region set. Communication network traffic prediction model for each low data validity sub-area under the cluster.

[0018] Preferably, the step of optimizing the The initial prediction model of each cell under the cluster is obtained in the low data validity sub-region set. The communication network traffic prediction model of each low data validity sub-region under the cluster specifically includes: constructing the first cluster reinforcement learning environment; based on the reinforcement learning environment and deep Q learning algorithm, optimize the The initial prediction model of each cell under the cluster is obtained The optimized prediction model of each cell in the cluster is selected, and the long-term prediction error of the optimized prediction model of each cell is calculated; the optimized prediction model with the smallest long-term prediction error is taken as the first optimized prediction model in the low data validity sub-region set. Communication network traffic prediction model for each low data validity sub-area under the cluster.

[0019] Preferably, the reinforcement learning environment includes an agent, a state space, an action space, a reward function and a state transition probability.

[0020] The agent is the A virtual server of a cluster, wherein the The cluster's virtual servers act as The central server corresponding to all edge servers of the cluster,

[0021] State Space

[0022] Action Space

[0023] Reward Function

[0024] State transition probability

[0025] in, Indicates Cluster, T represents the current time, C i Indicates cluster The i-th cell under represents the Cartesian product, Represents C i The state at time T, Represents C i The data validity level at time T, Represents C i The root mean square logarithmic error level at time T, Represents C i The root mean square logarithmic error at time T, D i (T) means C i The action to be chosen at time T, D i (t)∈{TR,+E,-E,-M,+M,FL},D i (t) represents C i The action chosen at time t is i When (t) = TR, it means self-optimization C i The model parameters, when D i When (t) = +E, it means increasing Ci Enhanced nodes and optimized C i The model parameters, when D i When (t) = -E, it means reducing C i Enhanced nodes and optimized C i The model parameters, when D i When (t) = -M, it means reducing C i Mapping nodes and optimize C i The model parameters, when D i When (t) = +M, it means increasing C i Mapping nodes and optimize C i The model parameters, when D i When (t) = FL, it means that C is optimized based on the federated learning algorithm. i The model parameters.

[0026] Preferably, the step of optimizing the first The initial prediction model of each cell under the cluster is obtained The optimization prediction model of each cell under the cluster includes: obtaining C i Status at time T According to C i Status at time T Judgment C i Action D i (T); in response to D i (T) = FL, determine C i The initial prediction model is the optimized federated learning model parameter fusion participant at time T, and based on the federated learning algorithm, determines C i Optimization model parameters based on C i Mapping nodes, enhancement nodes and optimized model parameters to generate C i Optimized prediction model.

[0027] Preferably, the method is based on a federated learning algorithm to determine C i The optimization model parameters specifically include: summarizing the first cluster All optimized federated learning model parameter fusion participants at time T are obtained to obtain the set of federated learning model parameter fusion participants Calculate the model parameter acceptance rate, where the model parameter acceptance rate refers to the sum of the data validity of all optimized federated learning model parameter fusion participants at time T and the cluster The ratio of the sum of the data validity of all cells at the next time T; determine whether the model parameter acceptance rate is greater than or equal to the model parameter acceptance rate threshold; in response to the model parameter acceptance rate being greater than or equal to the model parameter acceptance rate threshold, based on the preset weights, fuse the federated learning model parameter fusion participant set The model parameters of all optimized federated learning model parameter fusion participants are fused, and the fusion results are used as the optimized model parameters of all optimized federated learning model parameter fusion participants at time T.

[0028] In a second aspect, the present invention also provides a communication network traffic prediction device, comprising: an acquisition module, a first determination module, a second determination module and a construction module, the acquisition module being used to acquire multivariate cross-domain data of all cells in a target area, wherein the multivariate cross-domain data comprises auxiliary feature data and spatial feature data, the first determination module being connected to the acquisition module, and being used to determine the data validity of each cell in the target area according to the auxiliary feature data, wherein the data validity is used to characterize the data quality of the multivariate cross-domain data of the cell that changes over time, the second determination module being connected to the first determination module, and being used to determine a set of low data validity sub-areas in the target area according to the spatial feature data and the data validity, wherein the low data validity sub-areas are used to characterize the sub-areas under the cell with poor communication network traffic data quality, the construction module being connected to the second determination module, and being used to construct a communication network traffic prediction model for each low data validity sub-area in the low data validity sub-area set, and predicting the communication network traffic of each low data validity sub-area in the low data validity sub-area set based on the communication network traffic prediction model.

[0029] In a third aspect, the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to implement the communication network traffic prediction method provided in the first aspect above.

[0030] In a fourth aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the method for predicting communication network traffic provided in the first aspect above is implemented.

[0031] The present invention provides a method, device, electronic device and computer-readable storage medium for predicting communication network traffic, which obtains auxiliary feature data of all cells in a target area, calculates the data quality of the auxiliary feature data of each cell that changes over time as the data validity of each cell in the target area, determines a set of low data validity sub-areas in the target area, and constructs a communication network traffic prediction model for each low data validity sub-area in the low data validity sub-area set, so as to predict the communication network traffic of each low data validity sub-area in the low data validity sub-area set based on the communication network traffic prediction model. Therefore, the present invention can realize accurate prediction of communication network traffic, improve the accuracy of selecting base station addresses and equipment types in the subsequent planning stage and designing intelligent base station sleep strategies, and reduce the cost of communication network construction and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a flow chart of a method for predicting communication network traffic according to Embodiment 1 of the present invention;

[0033] Figure 2 This is an example diagram of a neighboring area relationship Voronoi diagram in Embodiment 1 of the present invention;

[0034] Figure 3 This is an example diagram of a Voronoi boundary point in Embodiment 1 of the present invention;

[0035] Figure 4 This is an example diagram of a target sub-region in Embodiment 1 of the present invention;

[0036] Figure 5 This is an example diagram of calculating the similarity between communication network traffic time series in Embodiment 1 of the present invention;

[0037] Figure 6 This is an example diagram of mapping nodes, enhancement nodes and model parameters of a cell in Embodiment 1 of the present invention;

[0038] Figure 7 This is an example diagram of an action space in Embodiment 1 of the present invention;

[0039] Figure 8 This is a flow chart of a method for predicting communication network traffic according to Embodiment 2 of the present invention;

[0040] Fig. 9 This is a schematic diagram of the structure of a communication network traffic prediction device according to Embodiment 3 of the present invention. DETAILED DESCRIPTION

[0041] In order to enable those skilled in the art to better understand the technical solution of the present invention, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0042] It should be understood that the specific embodiments and drawings described herein are only used to explain the present invention rather than to limit the present invention.

[0043] It can be understood that, in the absence of conflict, the various embodiments of the present invention and the various features in the embodiments can be combined with each other.

[0044] It can be understood that, for the convenience of description, the drawings of the present invention only show the parts related to the present invention, while the parts irrelevant to the present invention are not shown in the drawings.

[0045] It can be understood that each unit and module involved in the embodiments of the present invention may correspond to only one physical structure, or may be composed of multiple physical structures, or multiple units and modules may be integrated into one physical structure.

[0046] It can be understood that, without conflict, the functions and steps marked in the flowcharts and block diagrams of the present invention may occur in an order different from that marked in the drawings.

[0047] It is understood that the flowcharts and block diagrams of the present invention illustrate the possible architectures, functions, and operations of the systems, devices, equipment, and methods according to the various embodiments of the present invention. Each box in the flowchart or block diagram may represent a unit, module, program segment, or code, which contains executable instructions for implementing the specified functions. Moreover, each box or combination of boxes in the block diagram and flowchart may be implemented by a hardware-based system that implements the specified functions, or may be implemented by a combination of hardware and computer instructions.

[0048] It can be understood that the units and modules involved in the embodiments of the present invention can be implemented by software or hardware. For example, the units and modules can be located in a processor.

[0049] Embodiment 1:

[0050] like Figure 1 As shown, this embodiment provides a method for predicting communication network traffic.

[0051] In this embodiment, the communication network refers to various networks that support the transmission of data, sound, and video information, including wired networks and wireless networks. A wireless network refers to a network that transmits data via wireless signals, including but not limited to mobile networks, Wi-Fi, and Bluetooth. In this embodiment, the communication network is taken as an example of a wireless network.

[0052] The communication network traffic prediction method includes:

[0053] S101, obtaining multivariate cross-domain data of all cells in a target area, wherein the multivariate cross-domain data includes auxiliary feature data and spatial feature data.

[0054] In this embodiment, the spatial feature data may be the location of the cell in two-dimensional space or three-dimensional space, such as: in two-dimensional space, the spatial feature data refers to the longitude and latitude of the location of the cell; in three-dimensional space, the spatial feature data refers to the longitude, latitude, and altitude of the location of the cell. The spatial feature data may be obtained through the engineering parameter system of the communication operator. This embodiment takes the location of the cell in two-dimensional space as an example of spatial feature data.

[0055] The auxiliary characteristic data includes one or more combinations of the following: natural attribute characteristic data and weather attribute characteristic data. The natural attribute characteristic data includes weather factor characteristic data and air quality characteristic data. Among them, the weather factor characteristic data includes: air pressure, wind direction, wind speed, temperature, relative humidity, and precipitation in the time and space of the community. The air quality characteristic data refers to the air quality index (AQI, Air Quality Index) in the time and space of the community, which can be calculated by the concentration of six major pollutants (i.e. fine particulate matter, inhalable particulate matter, sulfur dioxide, nitrogen dioxide, ozone, and carbon monoxide). The weather factor characteristic data can be obtained through the website of the National Meteorological Science Data Center; the AQI of the air quality characteristic data can be obtained through the ecological environment website. The social attribute characteristic data includes: day type, scene type, and scenario type. Among them, the day type includes: working day or holiday. The scene type refers to the scene classification of the time and space of the community, such as: residential area, hospital, school, office area, scenic spot. The scenario type refers to the scenario classification of the time and space of the community, such as: daily, cultural and sports activities, and major meetings. Day types can be obtained through government and enterprise websites; scene types can be obtained through the communication operator's engineering parameter system; scenario types can be obtained through the communication operator's major event support system.

[0056] Optionally, the multivariate cross-domain data also includes communication network traffic data.

[0057] In this embodiment, taking the communication network as a wireless network as an example, correspondingly, the communication network traffic data is the wireless network traffic data.

[0058] In addition to auxiliary feature data and spatial feature data, multivariate cross-domain data also includes time feature data and wireless network traffic data. Among them, time feature data includes time scale feature data and moment feature data. The time scale features include: year (YR), quarter (QT), month (MO), week (WK), day (DY), hour (HR), minute (MN), second (SE). Time feature data can be obtained through the network management system of communication operators.

[0059] This embodiment takes the wireless network area to be predicted for wireless network data traffic as the target area. The time scale feature data set is denoted as S = {YR, QT, MO, WK, DY, HR, MN, SE}, and the time is t. The set of all cells in the target area is denoted as C. For example, if the spatial feature data is in two-dimensional space, the longitude and latitude of the two-dimensional space points of all cells in the target area are recorded as (x, y), and C is recorded as i The longitude and latitude of a two-dimensional space point are (x i ,y i ); if the spatial feature data is in three-dimensional space, the longitude, latitude, and altitude of the three-dimensional space points of all cells in the target area are recorded as (x, y, z), and C is recorded as i The longitude, latitude, and altitude of the three-dimensional space point are (x i ,y i ,z i ). i The auxiliary feature set at time t in time scale s is Remember C i Auxiliary features The auxiliary feature data under auxiliary feature m is For example, auxiliary feature m is the day type. That is, the daily type data of the cell at time t under the time scale s. i The sample set of multivariate cross-domain data is in, C i Auxiliary feature data at time t under time scale s, that is, auxiliary feature set Auxiliary feature data under all auxiliary features in, such as: auxiliary feature set For wireless network traffic data, is the concatenation operator.

[0060] S102, determining data validity of each cell in the target area according to the auxiliary feature data, wherein the data validity is used to characterize the data quality of multivariate cross-domain data of the cell that changes over time.

[0061] In this embodiment, the prior art usually evaluates the data quality of the multivariate cross-domain data of each cell in the target area based only on the preset weights. However, the time-varying nature of the data quality is only reflected by the collection cycle of the multivariate cross-domain data. For example, the multivariate cross-domain data may be updated in real time or intermittently, resulting in differences in data quality at different time points. In addition, improper preset weights may lead to erroneous quality assessment results. In practical applications, it is crucial for the prior art to ensure the rationality of the preset weights.

[0062] Specifically, S102: determining the data validity of each cell in the target area according to the auxiliary feature data, including steps S1021 to S1023:

[0063] S1021, obtaining the weight and sample capacity of the auxiliary feature data of each cell in the target area.

[0064] In this embodiment, C i The weight of the auxiliary feature data at time t in time scale s is Among them, for have And there is a formula Constraint relationship. i The sample size at time t in time scale s is Among them, for

[0065] S1022, evaluating the data completeness of the auxiliary feature data of each cell in the target area, wherein the auxiliary feature data includes one or more combinations of the following: natural attribute feature data, social attribute feature data, and the data completeness is used to characterize the proportion of non-empty feature data in the auxiliary feature data.

[0066] In this embodiment, C i The data integrity of the auxiliary feature data at time t under the time scale s is for According to the formula Evaluate the data integrity of each auxiliary feature data of each cell in the target area, where the numerator represents C at time t under time scale s i The number of auxiliary features m that are not empty, 1{·} is the decision function, which returns 1 when {·} is true, otherwise it returns 0.

[0067] S1023, calculating the data validity of each cell in the target area according to the weight, sample size and data completeness of the auxiliary feature data of each cell in the target area.

[0068] Specifically, S1023: calculating the data validity of each cell in the target area according to the weight, sample size and data completeness of the auxiliary feature data of each cell in the target area, including: calculating the data validity of each cell in the target area according to formula (1):

[0069]

[0070] Where T represents the current time, represents the sample capacity of the auxiliary feature data of the i-th cell in the target area at time t on the time scale S, represents the weight of the auxiliary feature data of the i-th cell in the target area at time t on the time scale S, Indicates the data completeness of the auxiliary feature data of the i-th cell in the target area at time t on the time scale S, It represents the data validity of the i-th cell in the target area at time T on the time scale S.

[0071] In this embodiment, at the current time t=T, according to C i The weight of the auxiliary feature data at time t under the time scale s Sample size Data integrity Calculate C i Data validity Compared with the prior art, the data quality of the cell, i.e., the data validity, is calculated based on the fixed weights and static data missing rate of the multivariate cross-domain data. This does not take into account the differences in the weights of different data features in real scenarios and the time-varying nature of the data missing rate as the multivariate cross-domain data acquisition conditions and the number of users change, resulting in changes in data quality, which in turn affects the accuracy of subsequent communication network predictions. This embodiment dynamically calculates the data quality of the auxiliary feature data of the cell that changes with time based on the weights, sample capacity, and data completeness of the auxiliary feature data of the cell at different times, and then uses the data validity of the auxiliary feature data of the cell that changes with time as the data validity of the cell. It comprehensively considers the time-varying nature reflected by the changes in the collection cycle, acquisition conditions, and number of users of the multivariate cross-domain data, improves the reliability and effectiveness of the data quality, monitors the changes in data quality in real time, and helps to promptly discover and resolve potential problems, ensuring the continued effectiveness of the data.

[0072] S103, determining a set of low data validity sub-regions in the target area according to the spatial feature data and the data validity, wherein the low data validity sub-regions are used to characterize sub-regions under cells with poor communication network traffic data quality.

[0073] Specifically, S103: determining a set of low data validity sub-regions in the target region according to the spatial feature data and the data validity, including steps S1031 and S1032:

[0074] S1031, based on the Voronoi diagram method, spatial feature data and data validity, divide the target area to obtain target sub-areas under each cell, wherein the spatial feature data is used to characterize the physical location of the cell.

[0075] In this embodiment, the prior art mainly divides the wireless network area into multiple sub-areas through spatial feature data, and the commonly used division methods include the grid method and the Voronoi diagram method. Among them, the grid method is to divide the entire wireless network area into square sub-areas of equal size; the Voronoi diagram method is to divide the wireless network area into Thiessen polygon sub-areas of unequal size based on the physical distance of the neighboring area relationship. However, the prior art uses the grid method to divide the wireless network area, facing the problems of fixed grids and uneven space utilization, resulting in the inability to predict the wireless network data traffic of the sub-area due to the absence of cells; the prior art uses the Voronoi diagram method to divide the wireless network area, but does not take into account the situation that cells with similar physical distances in the wireless network area may not have neighboring area relationships. In the case that cells with similar physical distances do not have neighboring area relationships, if the wireless network area is divided only according to the physical distance of the cells with neighboring area relationships, the sub-area range will be lost, and the wireless network area will be divided according to all cells with similar physical distances. The sub-area range will be expanded, and both will produce wireless network area division errors and reduce the accuracy of wireless network area division.

[0076] Specifically, S1031: based on the Voronoi diagram method, spatial feature data and data validity, the target area is divided to obtain the target sub-area under each cell, including: obtaining the neighbor relationship information of all cells in the target area, and determining the neighboring area of ​​the i-th cell in the target area according to the neighbor relationship information; calculating the Voronoi boundary points between the i-th cell and the neighboring area of ​​the i-th cell according to the spatial feature data and data validity of the i-th cell and the spatial feature data and data validity of the neighboring area of ​​the i-th cell; connecting the Voronoi boundary points to obtain the sub-area under the i-th cell; judging whether the i-th cell is within the sub-area under the i-th cell; in response to the i-th cell not being within the sub-area under the i-th cell, connecting the Voronoi boundary points and the i-th cell to obtain the target sub-area under the i-th cell; in response to the i-th cell being within the sub-area under the i-th cell, determining that the sub-area under the i-th cell is the target sub-area under the i-th cell.

[0077] Specifically, according to the spatial feature data and data validity of the ith cell and the spatial feature data and data validity of the neighboring area of ​​the ith cell, the Voronoi boundary point between the ith cell and the neighboring area of ​​the ith cell is calculated, including: according to formula (2), the Voronoi boundary point between the ith cell and the neighboring area of ​​the ith cell is calculated:

[0078]

[0079] Among them, P i represents the Voronoi boundary point between the ith cell and its neighboring cells, (x i ,y i )、(xj ,y j ) represent the spatial feature data of the i-th cell and the j-th cell respectively, They respectively represent the data validity of the i-th cell and the j-th cell at time t on the time scale S, and the j-th cell refers to the neighboring cell of the i-th cell.

[0080] In this embodiment, it is assumed that all cells in the target area have completed the neighboring cell configuration, and C i The set of cells with neighboring cell relationships is C j For C i ,C j For example, let the edge between the two be <C i ,C j >, remember C i With C j The set of edges between Note The space composed of all undirected edges in is C i The wireless network space represented by the multi-domain cross-domain data is C i The target sub-area under the cell is composed of the point set P i The enclosed space.

[0081] Based on the Voronoi diagram method, spatial feature data and data validity, the target area is divided to obtain the target sub-areas under each cell, specifically:

[0082] (1) For Then we have:

[0083] (1-1) Point set P i The space enclosed by the formula Constructed.

[0084] (1-2) Let C←C-{C i}.

[0085] (1-3) If End, otherwise go to (1-1).

[0086] (2) For Then we have:

[0087] (2-1) Place all cells in C into a Delauany triangle (if it is a three-dimensional space, it is a Delauany pyramid).

[0088] (2-2) Point set P i The space enclosed by the formula Constructed.

[0089] (2-3) Let C←C-{C i},like Then the process ends; otherwise, go to (2-1).

[0090] This embodiment uses Take this as an example, the target sub-area under the i-th cell is determined as follows:

[0091] (I)If Figure 2 As shown, C j All the cells in the ,area are placed into the Delauany triangle to obtain the neighboring area relationship Voronoi diagram, where the Delauany triangle is Figure 2 The triangle formed by the thick solid and dotted lines, C j The coordinate positions of all cells in Figure 2 The black solid point in C i The coordinate position of Figure 2 The hollow point in the .

[0092] (II) Figure 3 As shown, according to formula (2), P is constructed i , get the Voronoi boundary points, where the Voronoi boundary points are Figure 3 The points with lines between the solid black points and the hollow black points.

[0093] (III) Figure 4 As shown, connect P i Form the Thiessen polygon and obtain the target sub-region, where the target sub-region is Figure 4 The polygons formed by thin solid lines in the figure. This embodiment calculates the Voronoi boundary points between the neighboring areas of the cell based on the spatial feature data and data validity of the cell and the spatial feature data and data validity of the neighboring areas of the cell to determine the target sub-area under the cell, adjusts the Voronoi boundary between the neighboring areas of the cell based on the data validity of the cell, and fully considers the situations where the cells in the target area have neighboring area relationships or do not have neighboring area relationships, thereby reducing the division error of the wireless network area and improving the accuracy of the wireless network area division.

[0094] S1032, sorting the data validity of the target sub-regions, and determining a set of sub-regions with low data validity in the target region according to the interquartile range method.

[0095] In this embodiment, the data validity of the cells in C is sorted from large to small, that is, an ordered set is constructed. for

[0096] The outlier detection method is used to treat the target sub-region with low data validity as an outlier, and a set of low data validity sub-regions in the target region is obtained. The outlier detection method includes but is not limited to the Z-score method, box plot, distribution model, isolation forest (Isolation Forest), local outlier factor (Local Outlier Factor, LOF), principal component analysis (Principal Component Analysis, PCA), clustering-based method, autoregressive model, deep learning method, interquartile range method. This embodiment takes the interquartile range method as an example of the outlier detection method. Record the low data validity sub-region set Based on statistics, this embodiment further divides the target sub-area set after the wireless network area is divided into a low data validity sub-area set and a non-low data validity sub-area, thereby achieving a clear division of low data validity and non-low data validity sub-areas, which is helpful for the subsequent construction of traffic prediction models for the low data validity and non-low data validity sub-areas respectively, and for conducting in-depth traffic prediction analysis, thereby improving the modeling efficiency and prediction accuracy of the traffic prediction model.

[0097] S104, constructing a communication network traffic prediction model for each low data validity sub-region in the low data validity sub-region set, and predicting the communication network traffic of each low data validity sub-region in the low data validity sub-region set based on the communication network traffic prediction model.

[0098] In this embodiment, since the high data quality sub-area usually has a stable signal strength, low interference, and a small number of users, the data collection of network traffic is more accurate and complete, and thus easier to model and predict. However, the low data quality sub-area may have problems such as unstable signals, diversified user behaviors, and large interference, which will lead to increased data noise, affect the accuracy of the model, and increase the error of the prediction result. Due to the inaccurate flow prediction accuracy of the low data quality sub-area, the flow prediction accuracy of the high data quality sub-area and the low data quality sub-area is greatly different, which in turn affects the overall flow prediction accuracy of the wireless network area. Usually, flow prediction models are constructed for low data validity and non-low data validity sub-areas to perform flow prediction. Therefore, in addition to constructing a wireless network flow prediction model for each low data validity sub-area in the low data validity sub-area set, this embodiment also constructs a wireless network flow prediction model for each non-low data validity sub-area. The wireless network flow prediction model for each non-low data validity sub-area can be constructed using the same algorithm or formula of the wireless network flow prediction model of the high data quality sub-area constructed by the prior art, including but not limited to: time series analysis algorithm and machine learning algorithm.

[0099] Optionally, after S103: determining a set of low data validity sub-regions in the target area according to the spatial feature data and the data validity, and before constructing a communication network traffic prediction model for each low data validity sub-region in the set of low data validity sub-regions, the communication network traffic prediction method further includes:

[0100] S105, clustering a set of low data validity sub-regions according to the communication network traffic data and a clustering algorithm.

[0101] In this embodiment, the prior art clustering of the target sub-areas after the wireless network area is divided usually only considers the distribution similarity of the wireless network flow data, that is, clustering the wireless network flow data with similar distance into one category.

[0102] Specifically, S105: clustering a set of low data validity sub-regions according to the communication network traffic data and the clustering algorithm, including steps S1051 to S1053:

[0103] S1051, extracting communication network traffic data of each low data validity sub-region in the low data validity sub-region set Δ.

[0104] In this embodiment, C i ∈C s (T). C i Wireless network traffic data at time t in time scale s.

[0105] S1052, based on the time series analysis algorithm, the communication network traffic data of each low data validity sub-region is respectively decomposed and fitted to obtain the communication network traffic time series corresponding to each low data validity sub-region.

[0106] In this embodiment, the time series analysis algorithm includes a time series decomposition algorithm and a time series fitting algorithm; based on the time series decomposition algorithm, the wireless network traffic data of the low data validity sub-area is decomposed into n time series; based on the time series fitting algorithm, the n time series are fitted in the form of B characteristic sub-item functions. The time series decomposition algorithm includes but is not limited to: empirical mode decomposition, seasonal decomposition, wavelet transform, Fourier transform, and trend decomposition. The time series fitting algorithm includes but is not limited to: Prophet model, ARIMA (autoregressive integrated moving average, autoregressive integrated moving average) model, SARIMA (Seasonal Autoregressive Integrated Moving Average, seasonal autoregressive integrated moving average) model, ETS (Exponential Smoothing, exponential smoothing method), LSTM (Long Short-Term Memory, long short-term memory) network. In this embodiment, the time series analysis algorithm is the empirical mode decomposition algorithm, and the time series fitting algorithm is the Prophet model as an example.

[0107] remember is the αth time series component after the wireless network traffic data is decomposed, is the residual term of the αth time series component. is the trend characteristic sub-item function used to fit the non-periodic time series, is the periodic characteristic sub-term function used to fit the periodic time series, is a fixed characteristic sub-term function used to fit a specific time series, is the noise characteristic sub-item function.

[0108] Based on the empirical mode decomposition algorithm, the low data validity sub-region C i The wireless network traffic data is decomposed into time series as follows: (I) Let α←1. (II) Let (III) Calculation (IV) Connect all the maxima and minima by an interpolation algorithm, wherein the interpolation algorithm includes but is not limited to the cubic spline method, and the constructed upper envelope is denoted as The lower envelope is (V) Calculate the mean of the upper and lower envelopes As the formula (VI) Calculate the intermediate time series As the formula As shown. (VII) Let σ be the standard deviation, let σ θ is the standard deviation threshold, if Does not satisfy the formula Condition, then (VII-i) order (VII-ii) Let α←α+1. (VII-iii) Go to step (III). (VIII) The time series after empirical mode decomposition is as follows: shown.

[0109] Based on the Prophet model, the wireless network traffic data of each low data validity sub-area is fitted with a time series, specifically: (I) Note C i The wireless network traffic carrying capacity is G i , the growth rate is k i , the offset parameter is m i Since wireless network traffic is almost saturated, The formula The fitting is performed in the manner shown. and (II) Record C i The wireless network traffic cycle is T i , the low-pass filter parameters are therefore, The formula The Fourier series shown is used for fitting. If s = WK, then Take at this time

[0110] (III) The number of holidays is The time interval of holiday pair i is The impact factor is κ i .therefore, The formula The fitting is performed in the manner shown. Where v∝the degree of influence of holidays on the model. (IV) Generate a random value that follows a normal distribution for fitting (V) The time series and residual vector after Prophet model fitting are as follows: α=1,2,…,n and α=1,2,…,n.

[0111] After time series fitting, according to the formula Calculate C i The characteristic sub-item functions of the wireless network traffic time series at time t under the time scale s. Normalized to get Ci The wireless network traffic time series vector at time t under time scale s, namely C i Wireless network traffic time series corresponding to the low data validity sub-area

[0112] S1053, based on the K-means clustering algorithm, calculate the distance between the communication network traffic time series, and cluster the low data validity sub-region set based on the distance.

[0113] In this embodiment, clustering algorithms include but are not limited to K-means clustering, hierarchical clustering, DBSCAN (Density-Based Spatial Clustering of Applications with Noise), and OPTICS (Ordering Points to identify the clustering structure). In this embodiment, the clustering algorithm is taken as an example of the K-means clustering algorithm.

[0114] Based on the K-means clustering algorithm, the low data validity sub-region set is clustered into N clusters, among which Cluster Specifically:

[0115] (1) Let N←1.

[0116] (2) The Davies-Bouldin index is make

[0117] (3) Randomly select N cells as cluster centers, and record the cluster centers as The cluster center set is denoted as C * (T).

[0118] (4) For According to the formula calculate and The similarity of Figure 5 As shown. Among them, is a cross-correlation sequence of length 2T-1,

[0119]

[0120] (5) For According to the formula calculate and The distance measure of .

[0121] (6) For According to the formula calculate and (7) According to the formula Will and C x The shortest distance to the cluster center Divide into a cluster.

[0122] (7) For According to the formula Calculate the average distance within a cluster.

[0123] (8) According to the formula Calculate the Davies-Bouldin index of the current clustering result. Otherwise, there is

[0124] (8-1)N←N+1.

[0125] (8-2) Order

[0126] (8-3) Go to (3). This embodiment performs time series decomposition and fitting on low data validity sub-regions based on a time series analysis algorithm, better captures the scale and translation changes of wireless network traffic data, analyzes the amplitude difference and phase difference of the low data validity sub-region set, and clusters the low data validity sub-region set based on the scale and translation changes of the wireless network traffic data and the similarity of the distribution, thereby achieving deep clustering and sub-clustering of the low data validity sub-region set, improving the sustained effectiveness of clustering and sub-clustering, and facilitating the subsequent construction of different traffic prediction models to perform accurate traffic prediction on each low data validity sub-region under different clusters.

[0127] Specifically, a communication network traffic prediction model for each low data validity sub-region in the low data validity sub-region set is constructed, including steps S1041 and S1042:

[0128] S1041, based on the width learning algorithm, constructs the first The mapping nodes, enhancement nodes and model parameters of each cell under the cluster are obtained in the set of low data validity sub-regions. The initial prediction model of each cell under the cluster, where N represents the number of clusters in the set of low data validity sub-regions.

[0129] In this embodiment, Figure 4 As shown, first C i Wireless network traffic data After linear transformation, mapping nodes are generated to obtain a mapping layer mapping node set. Then the mapping nodes are transformed nonlinearly to generate enhanced nodes, and the enhanced node set ε of the enhanced layer is obtained. i(*) (t), where C i ∈C s (T), the mapping node group set of the mapping layer is The set of mapping nodes in the kth group of the mapping layer is The set of enhanced node groups in the enhanced layer is ε i (t), the set of enhancement nodes of the jth group in the enhancement layer is ε i(j) (t), such as: the mapping layer has 3 sets of mapping node sets, the enhancement layer has 3 sets of enhancement node sets, and the mapping node sets of the mapping layer are

[0130] Enhancement layer enhancement node collection

[0131]

[0132] For example For example, the number of mapping nodes in the mapping layer is as follows: As shown, the number of enhanced nodes in the enhanced layer is as follows: Therefore, the number of mapping nodes in the mapping layer is The number of enhanced nodes in the enhanced layer is 3*|ε i(3) (t)|. Connect the mapping nodes of the mapping layer and the enhancement nodes of the enhancement layer to the output layer nodes, and i Communication network traffic data at time t in time scale s Construct and train an initial prediction model, and obtain model parameters of the initial prediction model.

[0133] Based on the width learning algorithm, we construct the first Cluster C i Mapping nodes, enhancement nodes, and model parameters, such as Figure 6 As shown, specifically: (1) Constructing the initial model parameters of the mapping layer and the enhancement layer: (1-1) C i The sparse autoencoder is used to generate the mapping layer model parameters, and the random weight matrix in the mapping layer model parameters is recorded as The bias vector in the mapping layer model parameters is (1-2)C i Use sparse autoencoder to generate enhancement layer model parameters, and record the random weight matrix in the enhancement layer model parameters The bias vector in the enhancement layer model parameters is (2) According to the formula Generate the kth set of mapping nodes, where It is a nonlinear activation function. Nonlinear activation functions include but are not limited to ReLu function, Sigmoid function, and Tanh function. The ReLu function formula is as follows: Sigmoid function such as: formula Functions like formulas According to the formula Splice the mapping nodes, where (4) According to the formula Generate the jth set of enhanced nodes, where is a nonlinear activation function, such as ReLu function, Sigmoid function, Tanh function, etc. For example. (5) According to the formula Among them, the splicing enhancement node. (6) According to the formula Splice the mapping node and the enhancement node, where (7) According to the formula and C i Communication network traffic data at time t in time scale s Construct and train the initial prediction model, where As the initial model parameters of the initial prediction model. There are irreversible situations, so the ridge regression method is used to calculate the model parameters of the initial prediction model (7-1) Note for The pseudo-inverse matrix of calculate Among them, λ can take a minimum value, including but not limited to: 2 -50 , is the unit matrix. (7-2) According to the formula Calculate the model parameters of the initial prediction model (8) According to the formula Calculate C i The root mean square logarithmic error is in, Predict C for the initial prediction model i The predicted communication network traffic data at time t in the time scale s, C i The communication network traffic data at time t in the time scale s. This embodiment is based on the width learning algorithm to construct the first The mapping nodes, enhanced nodes and model parameters of each cell under the cluster can more effectively utilize the existing data to optimize the nodes and parameters and improve the learning effect of the model.

[0134] It should be noted that Each low data validity sub-region under the cluster, namely Each cell in the cluster has a corresponding edge server, and each cluster has a corresponding central server, such as: Cluster exists in Central Server, The central server is responsible for The data of all edge servers under the cluster are centrally processed, managed and stored.

[0135] S1042, based on reinforcement learning algorithm, optimize the The initial prediction model of each cell under the cluster is obtained. Communication network traffic prediction model for each low data validity sub-area under the cluster.

[0136] In this embodiment, the prior art generally consists of Each edge server in the cluster locally builds and trains The initial prediction model of each cell in the cluster is constructed and trained locally, and the model parameters of the initial prediction model are uploaded to the Central server; The central server receives the After all edge servers in the cluster send model parameters, based on the reinforcement learning algorithm, the model parameters sent by all edge servers are integrated and sent to the All edge servers in the cluster are optimized The initial prediction model for all low data validity sub-regions under the cluster. The central server needs to select the The model parameters of all edge servers in the cluster participate in the fusion, which not only increases the The central server load will also increase Communication overhead to the central server.

[0137] Specifically, S1042: Based on the reinforcement learning algorithm, optimize the first The initial prediction model of each cell under the cluster is obtained in the set of low data validity sub-regions. The communication network traffic prediction model of the cluster includes: constructing the first Cluster reinforcement learning environment; Based on the reinforcement learning environment and deep Q learning algorithm, optimize the first The initial prediction model of each cell under the cluster is obtained in the set of low data validity sub-regions. The optimized prediction model of each cell under the cluster is calculated, and the The long-term prediction error of the optimized prediction model of each cell under the cluster; the optimized prediction model with the smallest long-term prediction error is taken as the first in the low data validity sub-region set Cluster communication network traffic prediction model.

[0138] In this embodiment, On the edge servers of each cell under the cluster, the first The mapping nodes, enhanced nodes and model parameters of each cell under the cluster are obtained. The initial prediction model of each cell in the cluster. The root mean square error of the initial prediction model of each cell under the cluster, and according to the The root mean square error and the first The data validity of each cell under the cluster is calculated. The root mean square error level and data validity level of each cell under the cluster are determined The required selection actions for each cell under the cluster are obtained. The status of each cell in the cluster. Cluster reinforcement learning environment, deep Q learning algorithm, determine the The best action among all the required selected actions under the cluster is taken as the first The optimal action of the cluster, such as: The cluster includes cell 1 and cell 2. The action required for cell 1 is a, and the action required for cell 2 is b. The long-term prediction error after cell 1 and cell 2 select action a to optimize the initial prediction model of cell 1 and cell 2 is calculated, and the long-term prediction error after cell 1 and cell 2 select action b to optimize the initial prediction model of cell 1 and cell 2 is calculated. The action with the smallest long-term prediction error is taken as the first The optimal action of the cluster and The required selection actions for each cell under the cluster are converted into The optimal action of the cluster is based on The optimal action of the cluster, optimizing The initial prediction model of each cell under the cluster is used to obtain the communication network traffic prediction model of each low data validity sub-area.

[0139] Specifically, the reinforcement learning environment includes an agent, a state space, an action space, a reward function, and a state transition probability. The agent is the The virtual servers of the cluster, The cluster's virtual server acts as the The central server corresponding to all edge servers of the cluster,

[0140] State Space

[0141] Action Space

[0142] Reward Function

[0143] State transition probability

[0144] in, Indicates Cluster, T represents the current time, C i Indicates cluster The i-th cell under represents the Cartesian product, Represents C i The state at time T, Represents C i The data validity level at time T, Represents C i The root mean square logarithmic error level at time T, Represents C i The root mean square logarithmic error at time T, D i (T) means C i The action to be chosen at time T, D i (t)∈{TR,+E,-E,-M,+M,FL},D i (t) represents C i The action chosen at time t is i When (t) = TR, it means self-optimization C i The model parameters, when D i When (t) = +E, it means increasing C i Enhanced nodes and optimized C i The model parameters, when D i When (t) = -E, it means reducing C i Enhanced nodes and optimized C i The model parameters, when D i When (t) = -M, it means reducing C i Mapping nodes and optimize C i The model parameters, when D i When (t) = +M, it means increasing C i Mapping nodes and optimize C i The model parameters, when D i When (t) = FL, it means that C is optimized based on the federated learning algorithm. i The model parameters.

[0145] In this embodiment, C i The state at time T is T time cluster Status Remember C i Data validity level at time T Among them, L E C is the data validity level. i The root mean square logarithmic error level at time T Among them, L R is the root level of the mean square logarithmic error. Therefore, C i Status at time T Moment cluster Status Among them, the status D i (T) indicates status The action you want to select.

[0146] Remember the action space Among them, D i (t)∈{TR,+E,-E,+M,-M,+D,FL}, such as Figure 7 As shown, the actions can be converted. The action space at time t is

[0147] The strategy is Right now arrive 's mapping.

[0148] Remember T moment All edge servers in the policy π, that is, all edge servers in the state , select action The rewards after

[0149] Remember T moment Select Action hour, The status of Transfer to The state transition probability is

[0150] As time T = 0, 1, 2, ..., ∞ increases, the root mean square logarithm error of the long-term prediction of wireless network data traffic of the intelligent agent is

[0151] The goal of this embodiment is to maximize the long-term prediction accuracy of edge server wireless network data traffic and find the optimal strategy π* So that each agent can be in each state A valid action can be taken Make Minimum, that is

[0152] Solve the optimal strategy π through deep Q learning algorithm * Specifically, according to the Bellman equation, the minimum long-term prediction logarithmic root mean square error of the wireless network data traffic of the intelligent agent is calculated.

[0153] Where γ∈[0,1] is the discount factor. The optimal strategy under According to the action value Q function, that is, the formula

[0154]

[0155] When the Q function converges, the optimal strategy learned by the agent Wherein, η∈(0,1] represents the learning rate, including but not limited to 0.1, 0.01, and 0.001.

[0156] It should be noted that in order to solve the optimal strategy with lower computational complexity, the experience after each training of the deep Q learning algorithm is Stored in the experience pool, the action value Q function is learned using the experience replay mechanism. Using the value function approximation method, the optimal action value Q function is converted to Approximately the action value Q function Right now In order to determine the optimization target of the deep Q network, the target action value Q function value is recorded Among them, θ - Represents the parameters of the previous iteration. In the process of training the deep Q learning network, the mean square error of the target Q function value is reduced by iteratively adjusting θ, and according to the loss function, that is, Calculate the loss value. According to the formula

[0157] and formula Use stochastic gradient descent to optimize and update θ.

[0158] The trade-off parameter between exploration and reward is ε. Based on the reinforcement learning environment and deep Q learning, the initial prediction model of all edge servers is optimized, and the long-term prediction error of the optimized initial prediction model of all edge servers is calculated. Specifically, (1) The number of rounds is make (2) Initialization and That is, determine. (3) Let t ← 0. (4) Generate a random number r ∈ [0, 1]. (5) According to the formula Select an action (6) Execute Obtain and (7) Obtain a small batch of data from the experience pool, such as when T = 32, 64, 128 (8) According to the formula Calculate the Q function and calculate the loss of θ. (9) Every L steps, such as letting L = 4, update θ - ← θ. (10) If t < T, such as letting T = 1000, then let t ← t + 1 and go to (4). If Such as letting Then let Go to (2).

[0159] This embodiment is based on the Markov decision process and constructs a reinforcement learning environment for the central server in the cluster, that is, determine all possible states, all selectable actions, the rewards obtained by the actions selected by the states, and the state transition probabilities of the cluster, and based on the reinforcement learning environment and the deep Q learning algorithm, enable the central server to determine and learn the optimal actions of the cells in different states under the cluster, so as to optimize the initial prediction models of the cells in different states, maximize the long-term rewards while effectively improving the decision-making efficiency of resource scheduling and load balancing, enabling the edge server to respond to user requests faster and optimizing the service quality.

[0160] Specifically, based on the reinforcement learning environment and the deep Q learning algorithm, optimize the initial prediction models of each cell under the cluster to obtain the optimized prediction models of each cell under the cluster, including: obtaining the state of C i at time T According to the state of C i at time T Judge the action D i that C needs to select at time T i (T); in response to D i (T) = FL, determine the initial prediction model of C i as the optimized federated learning model parameter fusion participant at time T, and based on the federated learning algorithm, determine the optimized model parameters of C i ; based on the mapping node, enhancement node and optimized model parameters of C i , generate C iOptimized prediction model.

[0161] In this embodiment, when D i When (t) = FL, C i The initial prediction model of is selected as the federated learning model parameter fusion participant, and based on the federated learning algorithm, C is determined i The optimized model parameters of the initial prediction model are: C i The edge server of C stops training the initial prediction model locally; i The edge server uses cryptographic algorithms to encrypt the model parameters of the initial prediction model The encryption result is recorded as C i Will Upload to C i The central server of C, namely the virtual server; i The central server determines C based on the federated learning algorithm i Optimized model parameters of the initial prediction model Cryptography algorithms are also used to encrypt and optimize model parameters get and will Send to C i Edge server of C i The edge server uses cryptographic algorithms to decrypt And according to the optimization model parameters Optimizing C i The initial prediction model is obtained by i The optimized prediction model; similarly, when constructing the initial prediction model, C i The root mean square logarithmic error calculation process is used to calculate C under FL action. i The root mean square logarithmic error of .

[0162] The same is true for the calculation of model parameters when constructing the initial prediction model, C i The root mean square logarithmic error calculation process, when D i When (t) = TR, repeat the local training of the initial prediction model and calculate C under the action at TR i The root mean square logarithmic error of .

[0163] The same is true for the initial model parameter calculation of the enhancement layer, generation of enhancement node set, splicing of enhancement nodes, splicing of enhancement nodes and mapping nodes, model parameter calculation, C i The root mean square logarithmic error calculation process, when D i When (t) = +E, add an enhanced node group, specifically: calculate the model parameters of the newly added enhanced node group and represents the random weight matrix in the model parameters of the newly added enhanced node group, Represents the bias vector in the model parameters of the newly added enhanced node group, According to the formula Generate the j+1th set of enhanced nodes; according to the formula Splicing enhancement nodes, where According to the formula Splice the mapping node and the enhancement node, where According to the formula Calculate Q1; according to the formula Calculate Q2; according to the formula Calculate Q3; according to the formula calculate According to the formula calculate in, Prediction C for the initial prediction model after adding an enhanced node group i The predicted communication network traffic data at time t in the time scale s; C is predicted based on the initial prediction model after adding an enhanced node group i The predicted communication network traffic data at time t in the time scale s, calculate the C under +E action i The root mean square logarithmic error of .

[0164] The same is true for enhancing node splicing, enhancing node and mapping node splicing, model parameter calculation, C i The root mean square logarithmic error calculation process, when D i When (t) = -E, one enhanced node group is reduced, specifically: C i Deleting an Enhanced Node Group Splice the remaining enhanced node sets; Splice the mapping node set and the remaining enhanced node sets; Calculate the model parameters of the initial prediction model after reducing one enhanced node group; Calculate C under -E action i The root mean square logarithmic error of .

[0165] The same is true for the initial model parameter calculation of the mapping layer, generation of mapping node set, mapping node splicing, splicing of enhanced nodes and mapping nodes, model parameter calculation, C i The root mean square logarithmic error calculation process, when D i When (t)=+M, a mapping node group is added. Specifically, the model parameters of the newly added mapping node group are calculated, and the random weight matrix in the model parameters of the newly added mapping node group is recorded as The bias vector in the model parameters of the newly added mapping node group is According to the formula Generate the k+1th group of mapping nodes; according to the formula Splice the mapping nodes, where According to the formula As shown, update the j-th group of enhanced nodes; according to the formula Splice the mapping node and the enhancement node, where According to the formula Calculate Q1; according to the formula Calculate Q2; according to the formula Calculate Q3; according to the formula calculate According to the formula calculate in, Prediction C for the initial prediction model after adding a mapping node group i The predicted communication network traffic data at time t in the time scale s; according to the initial prediction model after adding a mapping node group, C i The predicted communication network traffic data at time t in the time scale s, calculates C under +M actions i The root mean square logarithmic error of .

[0166] The same is true for mapping node splicing, enhancing node and mapping node splicing, model parameter calculation, C i The root mean square logarithmic error calculation process, when D i When (t) = -M, one mapping node group is reduced, specifically: C i Deleting a Mapping Node Group Splice the remaining mapping node set; Splice the remaining mapping node set and the enhanced node set; Calculate the model parameters of the initial prediction model after reducing one mapping node group; Calculate C under -M action i The root mean square logarithmic error of .

[0167] The same is true for the initial model parameter calculation of the enhancement layer, generation of enhancement node set, splicing of enhancement nodes, splicing of enhancement nodes and mapping nodes, model parameter calculation, C i The root mean square logarithmic error calculation process, when D i When (t) = +D, increase C i The sample set of multivariate cross-domain data, the sample capacity after adding the sample set is The added sample set is Specifically: According to the formula Update the kth group of mapping nodes; according to the formula Splice the mapping nodes, where According to the formula Update the jth group of enhanced nodes; according to the formula Splicing enhancement nodes, where According to the formula Splice the mapping node and the enhancement node, where According to the formula Splicing C i The original sample set and the increased sample set of multivariate cross-domain data, where According to the formula Calculate Q1; according to the formula Calculate Q2; according to the formula Calculate Q3; according to the formula calculate According to the formula calculate Calculate + D action C i The root mean square logarithmic error of .

[0168] It should be noted that in order to maintain the The model parameter matrix dimensions of each cell under the cluster are unified. Cluster C i It is necessary to add an enhanced node group, namely D i When (t) = +E, let Each of the remaining cells in the cluster adds an enhanced node group, i.e. Similarly, for D i (t) = -E, or D i (t) = +M, or D i When (t) = -M, let Or +M or -M.

[0169] Specifically, based on the federated learning algorithm, C i The optimization model parameters include: cluster All optimized federated learning model parameter fusion participants at time T are obtained to obtain the set of federated learning model parameter fusion participants Calculate the model parameter acceptance rate, where the model parameter acceptance rate refers to the sum of the data validity of all optimized federated learning model parameter fusion participants at time T and the sum of the data validity of the first cluster The ratio of the sum of the data validity of all cells at the next time T; determine whether the model parameter acceptance rate is greater than or equal to the model parameter acceptance rate threshold; in response to the model parameter acceptance rate being greater than or equal to the model parameter acceptance rate threshold, based on the preset weights, fuse the federated learning model parameter fusion participant set The model parameters of all optimized federated learning model parameter fusion participants are fused, and the fusion results are used as the optimized model parameters of all optimized federated learning model parameter fusion participants at time T.

[0170] In this embodiment, the The central server receives Inside The edge server set is And let the initial The model parameter acceptance rate is ω θ is the model parameter acceptance rate threshold, .

[0171] Based on the federated learning algorithm, determine C i The optimization model parameters are as follows: (1) The central server receives C i Model parameters (2) Order (3) According to the formula Calculate the model parameter acceptance rate. (4) If In this embodiment, θ = 0.6, go to (1). (5) The central server will All model parameters in the system are fused with preset weights to obtain optimized model parameters. This embodiment determines the optimized federated learning model parameter fusion participants, and based on the federated learning algorithm, determines the optimized model parameters of the optimized federated learning model parameter fusion participants, improves the selection of optimized federated learning model parameter fusion participants and the parameter fusion process, strengthens the focus on important participants, reduces redundant calculations, thereby improving the overall efficiency of model training, shortening the training time, and reducing the first Central server load and communication overhead.

[0172] The present embodiment provides a method for predicting communication network traffic, by obtaining auxiliary feature data of all cells in a target area, calculating the data quality of the auxiliary feature data of each cell that changes with time as the data validity of each cell in the target area, determining a set of low data validity sub-areas in the target area, and constructing a communication network traffic prediction model for each low data validity sub-area in the set of low data validity sub-areas, so as to predict the communication network traffic of each low data validity sub-area in the set of low data validity sub-areas based on the communication network traffic prediction model, realize accurate prediction of communication network traffic, improve the accuracy of selecting base station addresses and equipment types in the subsequent planning stage and designing intelligent base station sleep strategies, and reduce the construction and maintenance costs of communication networks. In addition, according to the weight, sample capacity and data integrity of the auxiliary feature data of the cell at different times, the data quality of the auxiliary feature data of the cell that changes with time is dynamically calculated, and then the data validity of the auxiliary feature data of the cell that changes with time is used as the data validity of the cell, comprehensively considering the time-varying nature reflected by the changes in the collection cycle, acquisition conditions and number of users of multivariate cross-domain data, improving the reliability and effectiveness of data quality, and real-time monitoring of data quality changes, which helps to timely discover and solve potential problems and ensure the continuous effectiveness of data. According to the spatial feature data and data validity of the cell, and the spatial feature data and data validity of the cell's neighboring areas, the Voronoi boundary points between the cell and the cell's neighboring areas are calculated to determine the target sub-area under the cell. The Voronoi boundary between the cell and the cell's neighboring areas is adjusted in combination with the cell's data validity, and the situations where the cells in the target area have neighboring area relationships or do not have neighboring area relationships are fully considered to reduce the division error of the wireless network area and improve the accuracy of the wireless network area division. Based on statistics, the target sub-area set after the wireless network area is divided is further divided into a low data validity sub-area set and a non-low data validity sub-area, achieving a clear division of low data validity and non-low data validity sub-areas, which is helpful for the subsequent construction of traffic prediction models for low data validity and non-low data validity sub-areas, and for in-depth traffic prediction analysis, thereby improving the modeling efficiency and prediction accuracy of the traffic prediction model. Based on the time series analysis algorithm, the low data validity sub-regions are decomposed and fitted in time series to better capture the scale and translation changes of wireless network traffic data, analyze the amplitude differences and phase differences of the low data validity sub-region set, and cluster the low data validity sub-region set based on the scale and translation changes of wireless network traffic data and the similarity of distribution, so as to achieve deep clustering and clustering of the low data validity sub-region set, improve the continuous effectiveness of clustering and clustering, so as to facilitate the subsequent construction of different traffic prediction models for each low data validity sub-region under different clusters. Based on the width learning algorithm, the low data validity sub-region set is constructed. The mapping nodes, enhanced nodes and model parameters of each cell under the cluster can more effectively use the existing data to optimize the nodes and parameters and improve the learning effect of the model. Based on the Markov decision process, Build the first The reinforcement learning environment of the cluster, that is, to determine the All possible states of the cluster, all selectable actions, rewards obtained by the selected actions of the state, and state transition probabilities, and based on the reinforcement learning environment and deep Q learning algorithm, the The central server can determine and learn the The optimal actions of cells in different states under the cluster are used to optimize the initial prediction model of cells in different states, maximize long-term returns, and effectively improve the decision-making efficiency of resource scheduling and load balancing, so that edge servers can respond to user requests faster and optimize service quality. By determining the participants in the optimization of federated learning model parameter fusion, and based on the federated learning algorithm, the optimization model parameters of the participants in the optimization of federated learning model parameter fusion are determined, the selection of participants in the optimization of federated learning model parameter fusion and the parameter fusion process are improved, the focus on important participants is strengthened, and redundant calculations are reduced, thereby improving the overall efficiency of model training, shortening training time, and reducing the first Central server load and communication overhead.

[0173] Embodiment 2:

[0174] like Figure 8 As shown, this embodiment provides a method for predicting communication network traffic. The method for predicting communication network traffic includes:

[0175] S201, collecting multivariate cross-domain data of all cells in the target area, wherein the multivariate cross-domain data includes but is not limited to: auxiliary feature data and spatial feature data.

[0176] In this embodiment, the target cell is Figure 8For the wireless network area in the data center, the auxiliary feature data includes one or more of the following combinations: natural attribute feature data and weather attribute feature data. The natural attribute feature data includes weather factor feature data and air quality feature data. Among them, the weather factor feature data includes: air pressure, wind direction, wind speed, temperature, relative humidity, and precipitation in the time and space of the community. The air quality feature data refers to the air quality index in the time and space of the community, which can be calculated by the concentration of six major pollutants (i.e., fine particulate matter, inhalable particulate matter, sulfur dioxide, nitrogen dioxide, ozone, and carbon monoxide). The weather factor feature data can be obtained through the website of the National Meteorological Science Data Center; the AQI of the air quality feature data can be obtained through the ecological environment website. The social attribute feature data includes: day type, scene type, and scenario type. Among them, the day type includes: working day or holiday. The scene type refers to the scene classification of the time and space of the community, such as: residential area, hospital, school, office area, scenic spot. The scenario type refers to the scenario classification of the time and space of the community, such as: daily, cultural and sports activities, and major meetings. Day types can be obtained through government and enterprise websites; scene types can be obtained through the communication operator's engineering parameter system; scenario types can be obtained through the communication operator's major event support system.

[0177] The spatial feature data may be the location of the cell in two-dimensional space or three-dimensional space, such as: in two-dimensional space, the spatial feature data refers to the longitude and latitude of the location of the cell; in three-dimensional space, the spatial feature data refers to the longitude, latitude, and altitude of the location of the cell. The spatial feature data may be obtained through the engineering parameter system of the communication operator. This embodiment takes the location of the cell in two-dimensional space as an example of spatial feature data.

[0178] In addition to auxiliary feature data and spatial feature data, multivariate cross-domain data also includes time feature data and wireless network traffic data. Among them, time feature data includes time scale feature data and moment feature data. The time scale features include: year (YR), quarter (QT), month (MO), week (WK), day (DY), hour (HR), minute (MN), second (SE). Time feature data can be obtained through the network management system of communication operators.

[0179] S202, determining the data validity of each cell according to the auxiliary feature data of each cell, and dividing the target area based on the data validity of each cell, the cell-neighborhood relationship and the Voronoi diagram method to obtain the target sub-area under each cell.

[0180] In this embodiment, the target sub-area under each cell is Figure 2 Sub-area 1, sub-area 2, … in .

[0181] S203, based on statistics and K-means clustering algorithm, cluster the target sub-regions under each cell to obtain a set of sub-regions with low data validity.

[0182] In this embodiment, the data validity of the target sub-region is sorted, and the set of low data validity sub-regions in the target region is determined according to the interquartile range method. Then the communication network traffic data of each low data validity sub-region in the low data validity sub-region set Δ is extracted. Based on the time series analysis algorithm, the communication network traffic data of each low data validity sub-region is respectively decomposed and fitted to obtain the communication network traffic time series corresponding to each low data validity sub-region. Based on the K-means clustering algorithm, the distance between the communication network traffic time series is calculated, and the low data validity sub-region set is clustered based on the distance.

[0183] 204. Based on the width learning algorithm, the reinforcement learning algorithm and the federated learning algorithm, a communication network traffic prediction model of each low data validity sub-region in the low data validity sub-region set is constructed, and based on the communication network traffic prediction model, the communication network traffic of each low data validity sub-region in the low data validity sub-region set is predicted.

[0184] In this embodiment, Each low data validity sub-region under the cluster, namely Each cell in the cluster has a corresponding edge server, and each cluster has a corresponding central server, such as: Cluster exists in Central Server, The central server is responsible for The data of all edge servers under the cluster are centrally processed, managed and stored.

[0185] In the On the edge servers of each cell under the cluster, the first The mapping nodes, enhanced nodes and model parameters of each cell under the cluster are obtained. The initial prediction model of each cell in the cluster. The root mean square error of the initial prediction model of each cell under the cluster, and according to the The root mean square error and the first The data validity of each cell under the cluster is calculated. The root mean square error level and data validity level of each cell under the cluster are determined The required selection actions for each cell under the cluster are obtained. The status of each cell in the cluster. Cluster reinforcement learning environment, deep Q learning algorithm, determine the The best action among all the required selected actions under the cluster is taken as the first The optimal action of the cluster, such as: The cluster includes cell 1 and cell 2. The action required for cell 1 is a, and the action required for cell 2 is b. The long-term prediction error after cell 1 and cell 2 select action a to optimize the initial prediction model of cell 1 and cell 2 is calculated, and the long-term prediction error after cell 1 and cell 2 select action b to optimize the initial prediction model of cell 1 and cell 2 is calculated. The action with the smallest long-term prediction error is taken as the first The optimal action of the cluster and The required selection actions for each cell under the cluster are converted into The optimal action of the cluster is based on The optimal action of the cluster, optimizing The initial prediction model of each cell under the cluster is used to obtain the communication network traffic prediction model of each low data validity sub-area.

[0186] Jordi The i-th cell C under the cluster i The required selection action is based on the federated learning algorithm to optimize C i The model parameters of C i The initial prediction model is the optimized federated learning model parameter fusion participant at time T, and based on the federated learning algorithm, determines C i The optimization model parameters are based on C i Mapping nodes, enhancement nodes and optimized model parameters to generate C i Optimized prediction model.

[0187] The present embodiment provides a method for predicting communication network traffic, which obtains auxiliary feature data of all cells in a target area, calculates the data quality of the auxiliary feature data of each cell that changes with time as the data validity of each cell in the target area, determines a set of low data validity sub-areas in the target area, and constructs a communication network traffic prediction model for each low data validity sub-area in the set of low data validity sub-areas, so as to predict the communication network traffic of each low data validity sub-area in the set of low data validity sub-areas based on the communication network traffic prediction model, thereby achieving accurate prediction of communication network traffic, improving the accuracy of selecting base station addresses and equipment types and designing intelligent base station sleep strategies in subsequent planning stages, and reducing the cost of communication network construction and maintenance.

[0188] Embodiment 3:

[0189] like Fig. 9As shown, this embodiment also provides a communication network traffic prediction device, including an acquisition module 31, a first determination module 32, a second determination module 33 and a construction module 34, the acquisition module 31, used to acquire multivariate cross-domain data of all cells in the target area, wherein the multivariate cross-domain data includes auxiliary feature data and spatial feature data, the first determination module 32, connected to the acquisition module 31, used to determine the data validity of each cell in the target area according to the auxiliary feature data, wherein the data validity is used to characterize the data quality of the multivariate cross-domain data of the cell that changes over time, the second determination module 33, connected to the first determination module 32, used to determine a set of low data validity sub-areas in the target area according to the spatial feature data and the data validity, wherein the low data validity sub-area is used to characterize the sub-area under the cell with poor communication network traffic data quality, the construction module 34, connected to the second determination module 33, used to construct a communication network traffic prediction model for each low data validity sub-area in the low data validity sub-area set, and based on the communication network traffic prediction model, predict the communication network traffic of each low data validity sub-area in the low data validity sub-area set.

[0190] Specifically, the first determination module 32 includes: an acquisition unit 321, an evaluation unit 322 and a calculation unit 323. The acquisition unit 321 is used to obtain the weight and sample size of the auxiliary feature data of each cell in the target area. The evaluation unit 322 is used to evaluate the data completeness of the auxiliary feature data of each cell in the target area, wherein the auxiliary feature data includes one or more combinations of the following: natural attribute feature data, social attribute feature data, and the data completeness is used to characterize the proportion of non-empty feature data of the auxiliary feature data. The calculation unit 323 is used to calculate the data validity of each cell in the target area based on the weight, sample size and data completeness of the auxiliary feature data of each cell in the target area.

[0191] Specifically, the calculation unit 323 includes: a first calculation subunit, configured to calculate the data validity of each cell in the target area according to formula (1):

[0192]

[0193] Where T represents the current time, represents the sample capacity of the auxiliary feature data of the i-th cell in the target area at time t on the time scale S, represents the weight of the auxiliary feature data of the i-th cell in the target area at time t on the time scale S, Indicates the data completeness of the auxiliary feature data of the i-th cell in the target area at time t on the time scale S, It represents the data validity of the i-th cell in the target area at time T on the time scale S.

[0194] Specifically, the second determination module 33 includes: a division unit 331 and a sorting unit 332, the division unit 331 is used to divide the target area based on the Voronoi diagram method, the spatial feature data and the data validity, and obtain the target sub-area under each cell, wherein the spatial feature data is used to characterize the physical location of the cell, and the sorting unit 332 is used to sort the data validity of the target sub-area and determine the set of low data validity sub-areas in the target area according to the interquartile range method.

[0195] Specifically, the division unit 331 includes: an acquisition subunit, a second calculation subunit, a connection subunit, a judgment subunit, a first response subunit and a second response subunit, the acquisition subunit is used to obtain the neighbor relationship information of all cells in the target area, and determine the neighboring area of ​​the i-th cell in the target area according to the neighbor relationship information, the second calculation subunit is used to calculate the Voronoi boundary point between the i-th cell and the neighboring area of ​​the i-th cell according to the spatial feature data and data validity of the i-th cell, and the spatial feature data and data validity of the neighboring area of ​​the i-th cell, the connection subunit an element for connecting the Voronoi boundary points to obtain the sub-region under the i-th cell; a judging sub-unit for judging whether the i-th cell is within the sub-region under the i-th cell; a first responding sub-unit for, in response to the i-th cell not being within the sub-region under the i-th cell, connecting the Voronoi boundary points and the i-th cell to obtain a target sub-region under the i-th cell; and a second responding sub-unit for, in response to the i-th cell being within the sub-region under the i-th cell, determining that the sub-region under the i-th cell is the target sub-region under the i-cell.

[0196] Specifically, the second calculation subunit includes: a calculation microunit, which is used to calculate the Voronoi boundary points between the i-th cell and the neighboring cells of the i-th cell according to formula (2):

[0197]

[0198] Among them, P i represents the Voronoi boundary point between the ith cell and the neighboring cell of the ith cell, (x i ,y i )、(x j ,y j ) represent the spatial feature data of the i-th cell and the j-th cell respectively, They respectively represent the data validity of the i-th cell and the j-th cell at time t in the S time scale, and the j-th cell refers to the neighboring cell of the i-th cell.

[0199] Specifically, the communication network traffic prediction device further includes: a clustering module 35, which is used to cluster the low data validity sub-region set according to the communication network traffic data and a clustering algorithm.

[0200] Specifically, the clustering module 35 includes: an extraction unit 351, a decomposition and fitting unit 352 and a clustering unit 353. The extraction unit 351 is used to extract the communication network traffic data of each low data validity sub-region in the low data validity sub-region set. The decomposition and fitting unit 352 is used to perform time series decomposition and fitting on the communication network traffic data of each low data validity sub-region based on a time series analysis algorithm to obtain the communication network traffic time series corresponding to each low data validity sub-region. The clustering unit 353 is used to calculate the distance between the communication network traffic time series based on the K-means clustering algorithm, and cluster the low data validity sub-region set based on the distance.

[0201] Specifically, the construction module 34 includes: a construction unit 341 and an optimization unit 342, the construction unit 341 is used to construct the first in the low data validity sub-region set based on the width learning algorithm. The mapping nodes, enhancement nodes and model parameters of each cell under the cluster are obtained to obtain the first The initial prediction model of each cell under the cluster, where N represents the number of clusters in the low data validity sub-region set, and the optimization unit 342 is used to optimize the first The initial prediction model of each cell under the cluster is obtained in the low data validity sub-region set. Communication network traffic prediction model for each low data validity sub-area under the cluster.

[0202] Specifically, the optimization unit 342 includes: a construction subunit, an optimization subunit and a determination subunit, wherein the construction subunit is used to construct the first The reinforcement learning environment of the cluster, the optimization subunit, is used to optimize the first The initial prediction model of each cell under the cluster is obtained The optimized prediction model of each cell in the cluster is used to calculate the long-term prediction error of the optimized prediction model of each cell, and determine the subunit, which is used to take the optimized prediction model with the smallest long-term prediction error as the first subunit in the low data validity sub-region set. Communication network traffic prediction model for each low data validity sub-area under the cluster.

[0203] Specifically, the optimization subunit includes: an acquisition microunit, a judgment microunit, a response microunit and a generation microunit. The acquisition microunit is used to obtain C i Status at time T The judgment micro unit is used to determine the i Status at time T Judgment C i Action D i (T), optimized microcell for response to D i (T) = FL, determine C i The initial prediction model is the optimized federated learning model parameter fusion participant at time T, and based on the federated learning algorithm, determines C i Optimize model parameters to generate microcells for C-based i Mapping nodes, enhancement nodes and optimized model parameters to generate C i Optimized prediction model.

[0204] Specifically, the response micro-unit includes: a summary minimum unit, a calculation minimum unit, a judgment minimum unit and a response minimum unit, and the summary minimum unit is used to summarize the first cluster All optimized federated learning model parameter fusion participants at time T are obtained to obtain the set of federated learning model parameter fusion participants The minimum calculation unit is used to calculate the model parameter acceptance rate, where the model parameter acceptance rate refers to the sum of the data validity of all optimized federated learning model parameter fusion participants at time T and the cluster The ratio of the sum of the data validity of all cells at the next time T, the minimum judgment unit, is used to judge whether the model parameter acceptance rate is greater than or equal to the model parameter acceptance rate threshold, the minimum response unit, is used to respond to the model parameter acceptance rate is greater than or equal to the model parameter acceptance rate threshold, based on the preset weights, the federated learning model parameter fusion participant set The model parameters of all optimized federated learning model parameter fusion participants are fused, and the fusion results are used as the optimized model parameters of all optimized federated learning model parameter fusion participants at time T.

[0205] It can be understood that the above-mentioned communication network traffic prediction device executes the communication network traffic prediction method corresponding to the embodiment 1 provided above. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects of the scheme corresponding to the communication network traffic prediction method of the above embodiment 1, and will not be repeated here.

[0206] Embodiment 4:

[0207] This embodiment further provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to implement the method for predicting communication network traffic in the above-mentioned embodiment 1.

[0208] Embodiment 5:

[0209] This embodiment further provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method for predicting communication network traffic in the above-mentioned embodiment 1 is implemented.

[0210] It is to be understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present invention, but the present invention is not limited thereto. For those of ordinary skill in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for predicting communication network traffic, characterized in that: include: Acquire multivariate cross-domain data of all cells in the target area, wherein the multivariate cross-domain data includes auxiliary feature data and spatial feature data; Determine the data validity of each cell in the target area based on the auxiliary feature data, where the data validity is used to characterize the data quality of the multivariate cross-domain data of the cell that changes over time; Determine a set of low data validity sub-regions in the target area according to the spatial feature data and the data validity, wherein the low data validity sub-regions are used to characterize sub-regions under cells with poor communication network traffic data quality; A communication network traffic prediction model for each low data validity sub-region in the low data validity sub-region set is constructed, and based on the communication network traffic prediction model, the communication network traffic of each low data validity sub-region in the low data validity sub-region set is predicted.

2. The method for predicting communication network traffic according to claim 1, characterized in that: Determining the data validity of each cell in the target area according to the auxiliary feature data specifically includes: Obtain the weight and sample size of auxiliary feature data of each cell in the target area; Evaluate the data completeness of the auxiliary feature data of each community in the target area, where the auxiliary feature data includes one or more of the following combinations: natural attribute feature data, social attribute feature data, and the data completeness is used to characterize the proportion of non-empty feature data in the auxiliary feature data; The data validity of each cell in the target area is calculated based on the weight, sample size and data completeness of the auxiliary feature data of each cell in the target area.

3. The method for predicting communication network traffic according to claim 2, characterized in that: The method of calculating the data validity of each cell in the target area according to the weight, sample size and data completeness of the auxiliary feature data of each cell in the target area specifically includes: According to formula (1), the data validity of each cell in the target area is calculated: Where T represents the current time, represents the sample capacity of the auxiliary feature data of the i-th cell in the target area at time t on the time scale S, represents the weight of the auxiliary feature data of the i-th cell in the target area at time t on the time scale S, Indicates the data completeness of the auxiliary feature data of the i-th cell in the target area at time t on the time scale S, It represents the data validity of the i-th cell in the target area at time T on the time scale S.

4. The method for predicting communication network traffic according to claim 1, characterized in that: The step of determining a set of low data validity sub-regions in the target region according to the spatial feature data and the data validity specifically includes: Based on the Voronoi diagram method, spatial feature data and the data validity, the target area is divided to obtain the target sub-areas under each cell, wherein the spatial feature data is used to characterize the physical location of the cell: The data validity of the target sub-regions is ranked, and a set of sub-regions with low data validity in the target region is determined according to the interquartile range method.

5. The method for predicting communication network traffic according to claim 4, characterized in that: The method of dividing the target area based on the Voronoi diagram method, the spatial feature data and the data validity to obtain the target sub-areas under each cell specifically includes: Obtaining neighboring cell relationship information of all cells in the target area, and determining neighboring cells of the i-th cell in the target area according to the neighboring cell relationship information; Calculate the Voronoi boundary points between the i-th cell and the neighboring area of ​​the i-th cell according to the spatial feature data and data validity of the i-th cell and the spatial feature data and data validity of the neighboring area of ​​the i-th cell; Connect the Voronoi boundary points to obtain the sub-region under the i-th cell; Determine whether the i-th cell is within a sub-area under the i-th cell; In response to the i-th cell not being in the sub-region under the i-th cell, connecting the Voronoi boundary point and the i-th cell to obtain a target sub-region under the i-th cell; In response to the i-th cell being within the sub-area under the i-th cell, determining the sub-area under the i-th cell as a target sub-area under the i-th cell.

6. The method for predicting communication network traffic according to claim 5, characterized in that: The calculating, according to the spatial feature data and data validity of the i-th cell and the spatial feature data and data validity of the neighboring area of ​​the i-th cell, the Voronoi boundary point between the i-th cell and the neighboring area of ​​the i-th cell specifically includes: According to formula (2), the Voronoi boundary points between the i-th cell and the neighboring cells of the i-th cell are calculated: Among them, P i represents the Voronoi boundary point between the ith cell and the neighboring cell of the ith cell, (x i ,y i )、(x j ,y j ) represent the spatial feature data of the i-th cell and the j-th cell respectively, They respectively represent the data validity of the i-th cell and the j-th cell at time t in the S time scale, and the j-th cell refers to the neighboring cell of the i-th cell.

7. The method for predicting communication network traffic according to claim 1, characterized in that: The multivariate cross-domain data also includes communication network traffic data, After determining a set of low data validity sub-regions in the target area according to the spatial feature data and the data validity, and before constructing a communication network traffic prediction model for each low data validity sub-region in the set of low data validity sub-regions, the method further includes: The low data validity sub-region set is clustered according to the communication network traffic data and a clustering algorithm.

8. The method for predicting communication network traffic according to claim 7, characterized in that: The clustering of the low data validity sub-region set according to the communication network traffic data and the clustering algorithm specifically includes: Extracting communication network traffic data of each low data validity sub-region in the low data validity sub-region set; Based on the time series analysis algorithm, the communication network traffic data of each low data validity sub-area is respectively decomposed and fitted in time series to obtain the communication network traffic time series corresponding to each low data validity sub-area; Based on the K-means clustering algorithm, the distance between the communication network traffic time series is calculated, and the low data validity sub-region set is clustered based on the distance.

9. The method for predicting communication network traffic according to claim 7, characterized in that: The constructing of a communication network traffic prediction model for each low data validity sub-region in the low data validity sub-region set specifically includes: Based on the width learning algorithm, the first The mapping nodes, enhancement nodes and model parameters of each cell under the cluster are obtained to obtain the first The initial prediction model of each cell under the cluster, where N represents the number of clusters in the low data validity sub-region set; Based on the reinforcement learning algorithm, optimize the The initial prediction model of each cell under the cluster is obtained, and the first Communication network traffic prediction model for each low data validity sub-area under the cluster.

10. The method for predicting communication network traffic according to claim 9, characterized in that: Based on the reinforcement learning algorithm, optimizing the The initial prediction model of each cell under the cluster is obtained in the low data validity sub-region set. The communication network traffic prediction model for each low data validity sub-area under the cluster includes: Based on the Markov decision process, the Cluster reinforcement learning environment; Based on the reinforcement learning environment and deep Q learning algorithm, optimize the The initial prediction model of each cell under the cluster is obtained The optimized prediction model of each cell in the cluster is used to calculate the long-term prediction error of the optimized prediction model of each cell; Take the optimized prediction model with the smallest long-term prediction error as the first in the low data validity sub-region set. Communication network traffic prediction model for each low data validity sub-area under the cluster.

11. The method for predicting communication network traffic according to claim 10, characterized in that: The reinforcement learning environment includes an agent, a state space, an action space, a reward function and a state transition probability. The agent is the A virtual server of a cluster, wherein the The cluster's virtual servers act as The central server corresponding to all edge servers of the cluster, State Space Action Space Reward Function State transition probability in, Indicates Cluster, T represents the current time, C i Indicates cluster The i-th cell under represents the Cartesian product, Represents C i The state at time T, Represents C i The data validity level at time T, Represents C i The root mean square logarithmic error level at time T, Represents C i The root mean square logarithmic error at time T, D i (T) means C i The action to be chosen at time T, D i (t)∈{TR,+E,-E,-M,+M,FL},D i (t) represents C i The action chosen at time t is i When (t) = TR, it means self-optimization C i The model parameters, when D i When (t) = +E, it means increasing C i Enhanced nodes and optimized C i The model parameters, when D i When (t) = -E, it means reducing C i Enhanced nodes and optimized C i The model parameters, when D i When (t) = -M, it means reducing C i Mapping nodes and optimize C i The model parameters, when D i When (t) = +M, it means increasing C i Mapping nodes and optimize C i The model parameters, when D i When (t) = FL, it means that C is optimized based on the federated learning algorithm. i The model parameters.

12. The method for predicting communication network traffic according to claim 11, characterized in that: Based on the reinforcement learning environment and the deep Q learning algorithm, optimizing the The initial prediction model of each cell under the cluster is obtained The optimized prediction model of each cell under the cluster includes: Get C i Status at time T According to C i Status at time T Judgment C i Action D i (T); In response to D i (T) = FL, determine C i The initial prediction model is the optimized federated learning model parameter fusion participant at time T, and based on the federated learning algorithm, determines C i Optimization model parameters of C-based i Mapping nodes, enhancement nodes and optimized model parameters to generate C i Optimized prediction model.

13. The method for predicting communication network traffic according to claim 12, characterized in that: Based on the federated learning algorithm, C i The optimization model parameters include: Summarize the above cluster All optimized federated learning model parameter fusion participants at time T are obtained to obtain the set of federated learning model parameter fusion participants Calculate the model parameter acceptance rate, where the model parameter acceptance rate refers to the sum of the data validity of all optimized federated learning model parameter fusion participants at time T and the cluster The ratio of the total data validity of all cells at the next time T; Determine whether the model parameter acceptance rate is greater than or equal to the model parameter acceptance rate threshold; In response to the model parameter acceptance rate being greater than or equal to the model parameter acceptance rate threshold, based on the preset weights, the federated learning model parameter fusion participant set is fused The model parameters of all optimized federated learning model parameter fusion participants are fused, and the fusion results are used as the optimized model parameters of all optimized federated learning model parameter fusion participants at time T.

14. A communication network traffic prediction device, characterized in that: include: an acquisition module, a first determination module, a second determination module and a construction module, An acquisition module is used to acquire multivariate cross-domain data of all cells in the target area, wherein the multivariate cross-domain data includes auxiliary feature data and spatial feature data. The first determination module is connected to the acquisition module and is used to determine the data validity of each cell in the target area based on the auxiliary feature data, wherein the data validity is used to characterize the data quality of the multivariate cross-domain data of the cell that changes over time. The second determination module is connected to the first determination module and is used to determine a set of low data validity sub-regions in the target area according to the spatial feature data and the data validity, wherein the low data validity sub-regions are used to characterize sub-regions under cells with poor communication network traffic data quality. A construction module is connected to the second determination module, and is used to construct a communication network traffic prediction model for each low data validity sub-region in the low data validity sub-region set, and predict the communication network traffic of each low data validity sub-region in the low data validity sub-region set based on the communication network traffic prediction model.

15. An electronic device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to implement a method for predicting communication network traffic as claimed in any one of claims 1 to 13.

16. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, a method for predicting communication network traffic as described in any one of claims 1 to 13 is implemented.

Images