Base Station Regulation Method Integrating Spatiotemporal Information and Traffic Characteristics

Abstract

This application relates to the field of communication technologies, and particularly to a base station regulation method that integrates spatio-temporal information and traffic characteristics. By integrating the spatio-temporal characteristics of the base station network, base station traffic is predicted, and then a regionalized base station energy-saving strategy is formulated according to the traffic prediction results. Specifically, it includes the regional division of a large-scale base station network; using a coding tree to extract the hierarchical structure characteristics of the base station network, and extracting the traffic correlation characteristics of the base station network based on the attention mechanism; generating a representation with spatio-temporal characteristics through feature fusion, using time-domain convolutional networks with multiple different dilation factors to capture the temporal characteristics of base station traffic and predict the changes in future base station traffic, and finally formulating a regionalized base station energy-saving strategy according to the traffic prediction results. It aims to solve the problem of how to reduce the energy consumption of base stations.

Description

Technical Field

[0001] This application relates to the field of communication technologies, and particularly to a base station regulation method that integrates spatio-temporal information and traffic characteristics. Background Art

[0002] With the popularization of intelligent mobile devices, the demand for wireless communication traffic and data-intensive services has grown rapidly. The communication network is developing towards ultra-large-scale and dense base station deployment, bringing high energy consumption and carbon emission problems. How to reduce the energy consumption of the base station network while ensuring the user experience has become a major challenge for operators. Dynamically adjusting the base station power helps reduce operating costs, but adjusting according to the real-time network traffic has a lag, which affects the user experience. Therefore, predicting the base station traffic and adjusting the base station operation strategy in advance is of great significance for improving the network service quality, reducing energy consumption, and optimizing resource allocation.

[0003] Currently, in order to effectively capture the internal dependence relationship between historical traffic, some well-known methods use deep learning methods such as recurrent neural networks to learn and predict the traffic change law. For example, Chinese Patent with application number CN202310000660.9 discloses an energy consumption optimization method and device for a 5G communication base station. It constructs an input set through the base station location, traffic average value, communication delay rate, and the number of access terminals, and uses the Grey Wolf Optimizer (GWO) and Bidirectional Long Short-Term Memory (BLSTM) to predict the reliability score of the base station. When the score is lower than a given threshold, it predicts the future traffic and the number of access terminals, and accordingly performs energy consumption control. However, the above methods only focus on the traffic change situation of a single base station in the time dimension, and do not fully consider the influence of multiple factors between base stations on traffic prediction, resulting in insufficient accuracy of base station traffic prediction and further affecting the effectiveness of energy-saving strategies.

[0004] On the other hand, in addition to time dependence, the traffic variation between adjacent base stations is also an important factor affecting the accuracy of traffic prediction. Some well-known methods consider both time and spatial information in traffic prediction. For example, Chinese Patent Application No. CN202010036123.6 discloses a base station sleep method based on mobile network traffic prediction, which uses a Temporal Convolutional Network and a Three-Dimensional Convolutional Neural Network to extract the time and spatial features of traffic. By obtaining the historical traffic information of base stations and the traffic information of base stations in the surrounding area, combined with external factors, the traffic of base stations is predicted. According to the prediction results, an appropriate base station is selected for sleep or wake-up through a scheduling objective function to save energy consumption and ensure service quality. However, the above method learns the artificially predefined spatial information in the original data, but the noise or errors in the original spatial data will affect the accuracy of traffic prediction, and thus have a negative impact on the optimization of base station energy consumption.

[0005] In view of this, there is a need for a regional energy-saving method that comprehensively considers multiple factors affecting base station traffic and formulates a regional energy-saving strategy for base stations based on accurate traffic prediction, so as to significantly reduce the energy consumption of the base station network while ensuring the quality of user services.

[0006] The above content is only used to assist in understanding the technical solution of the present application, and does not represent an admission that the above content is prior art. Summary of the Invention

[0007] The main object of the present application is to provide a base station regulation method that integrates spatio-temporal information and traffic characteristics, aiming to solve the problem of how to reduce the energy consumption of the base station network.

[0008] To achieve the above object, a base station regulation method that integrates spatio-temporal information and traffic characteristics provided by the present application includes:

[0009] S10, dividing the original base station network into multiple regional base station networks, constructing a graph representation of each regional base station network, and constructing a graph representation with regions as nodes by defining the connection relationship between regions;

[0010] S20, extracting the hierarchical structure features of the regional base station network based on an encoding tree, performing singular value decomposition on the traffic characteristics, adding learnable parameters to construct a directed weight matrix between base stations, and after updating the directed weight matrix to an attention matrix, capturing the traffic association features of the base station network based on the attention matrix;

[0011] S30. Embed three types of natural period markers into the base station traffic sequence to obtain a temporal feature representation of the traffic. Use convolution operations to fuse the temporal feature representation of the traffic with the associated features of the base station network traffic, obtaining a tensor representation that integrates spatio-temporal information. Employ a temporal convolutional network with different dilation factors to capture the traffic variation patterns of different periods in the tensor representation, resulting in a base station traffic representation that combines spatio-temporal features and features of different periods. Input the base station traffic representation into a fully connected layer to predict the base station traffic at the next moment;

[0012] S40. With the goal of minimizing the base station operation cost, determine the optimal regulation strategy for base station power adjustment in each region based on the predicted base station traffic.

[0013] Optionally, in S10, the step of dividing the original base station network into multiple regional base station networks specifically includes:

[0014] S11. Sort the traffic of each base station in the original base station network, and select the base station with the highest traffic density as the central base station;

[0015] S12. When it is determined that there are at least 3 target base stations with traffic greater than a preset traffic threshold within the circular neighborhood associated with the central base station, divide the circular neighborhood of the central base station and the circular neighborhoods associated with the target base stations into one of the regional base station networks;

[0016] S13. Repeat S11 - S12 until each base station is divided into the corresponding regional base station network, obtaining regional base station networks.

[0017] Optionally, in S10, constructing the graph representation of each regional base station network specifically includes:

[0018] S14. Use the base stations as graph nodes and the connection relationships between the base stations in the regional base station network as edges, and use to represent the base station network, where represents the set of base stations in this base station network, represents the number of base stations included, represents the connection relationship between base stations. If there is a connection relationship between base stations and , then the element value of the matrix in the th row and th column is 1, that is, , otherwise ;

[0019] S15. Successively construct the graph representations for each of the regional base station networks obtained in step S14, obtaining , where , represents the number of base stations included in, and N represents the sum of the number of base stations in the area;

[0020] S16. If there is a connection relationship between the base stations in two different areas in the original base station network, then connect these two areas to form a mesh structure as the area association network, and the area association network is denoted as , where is the set of all areas, represents the connection relationship between areas, the th row and th column element value of represents the number of base station pairs with all cross-area connections between the area base station networks numbered and .

[0021] Optionally, after the step S16, it further includes:

[0022] S17. Let the base station attribute feature be represented as , then the attribute feature matrix of is represented as , where represents the length of the base station attribute feature. The feature matrix of each area base station network is subjected to feature transformation through a fully connected layer to obtain the feature matrix of, and the continuously and equally spaced traffic at moments recorded by the base station constitute a traffic sequence . The overall traffic sequence of the area is obtained by summing the traffic of all base stations in the area at each moment, and the traffic sequence of all areas is spliced to obtain the traffic feature matrix of the area association network;

[0023] Optionally, in the S20, the step of extracting the hierarchical structure feature of the area base station network based on the coding tree specifically includes:

[0024] S21. Initialize the base stations in an area base station network sequentially into independent base station partitions, represents the number of base stations, and the base station partition after the base station network is initialized is , where represents the sub-partition of, represents the area base station network graph numbered ;

[0025] S22. Traverse all base station partitions. If the base station partitions and meet and the value is the largest and positive, then combine and to obtain a new base station partition . Among them, is the coding tree after partition merging, represents the base station partition and before merging on the current base station network diagram representation of the structure entropy, represents the structure entropy after merging;

[0026] S23. Traverse all base station partitions. If the base station partitions and meet and the value is the largest and positive, then combine and to create a new partition . Put and the two partitions into . and become sub - partitions of, where the height of the coding tree does not exceed 3;

[0027] S24. Repeat steps S22 and S23 until there are no base stations that can be merged or combined, and obtain the final base station partition set of this area. Among them, the 0th layer of the coding tree is the root node. Take all partitions in the base station partition set as the 1st - layer nodes. If a partition contains sub - partitions, then use the sub - partitions as sub - nodes to continue expanding until the base stations are leaf nodes, so as to obtain the optimal coding tree corresponding to the base station network of this area;

[0028] S25. Repeat steps S22 - S24 to construct the optimal coding tree of the entire area's associated network. Replace the root node of the optimal coding tree of each area's base station network with the leaf node of to obtain the hierarchical structure representation of the entire base station network. Perform information aggregation through graph convolution between the layers of . The node feature matrix of the layer is obtained through the graph convolution to get the node feature matrix , starting from the base station attribute features of the 6th layer of the coding tree are concatenated to obtain , and the information is iteratively passed upward, and finally a length of is obtained, which is the base station network hierarchy feature representation . Among them, the expression of the graph convolution is as follows:

[0029]

[0030] In the formula, represents the connection relationship between the layer nodes and the layer nodes of the coding tree , is the activation function, is the parameter matrix of the th layer.

[0031] Optionally, in the S20, the steps of performing singular value decomposition on the traffic features, adding learnable parameters to construct a directed weight matrix between base stations, and updating the directed weight matrix to an attention matrix, and then capturing the base station network traffic correlation features based on the attention matrix specifically include:

[0032] S26, let the traffic feature matrix , and use the equation to perform singular value decomposition on the traffic feature matrix to obtain the left singular value matrix representing the spatial distribution between base stations, and the right singular matrix where the traffic changes over time. The diagonal elements are the singular values. Select the first 10 singular values to obtain , select the first ten columns to obtain , select the first 10 rows to obtain ;

[0033] S27, add , and to the learnable parameters and , and use the function to constrain the range of the correlation weights between base stations, and use the function to ensure the non-negativity of the correlation weights between base stations, so as to obtain the learnable directed weight matrix :

[0034]

[0035] In the formula, Indicates the degree of association between each base station. The row represents the degree of association of other base station nodes with the current base station . The greater the weight value, the greater the correlation.

[0036] S28. Calculate the distance between any two base stations, and take 5 times the sum of their service radii as the distance threshold . If the distance between the base stations exceeds the set threshold , then set its weight to 0, and then use the function to retain the maximum values for each row vector, and set the rest to 0, to obtain an attention matrix that does not depend on the physical connection or communication path between base stations . The expression of

[0037]

[0038]

[0039] is as follows: is a piecewise function, represents the element in the row and column, that is, the association weight between base station and base station and .

[0040] S29. Aggregate and update the attribute feature vector of :

[0041]

[0042] is as follows: is the feature transformation matrix, is 's attribute feature vector, is the aggregated and updated feature vector.

[0043] S210. Repeat steps S26 to S29, calculate the attention matrices of each regional base station network and regional association network in turn, to obtain , ... and , and then perform feature aggregation based on the attention mechanism in turn, so as to obtain the traffic association features of the regional base station network, and the traffic association features of the regional association network 。

[0044] Optionally, in step S30, the step of embedding three types of natural period markers into the base station traffic sequence to obtain a temporal feature representation of the traffic and fusing the temporal feature representation of the traffic with the base station network traffic correlation feature by using a convolution operation includes:

[0045] S31, embedding daily markers, weekly markers, and monthly markers into the base station traffic sequence to obtain a temporal feature representation of the traffic ;

[0046] S32, concatenating the base station network hierarchy feature representation , the traffic correlation feature representation and the to obtain a concatenated base station network traffic correlation feature representation ;

[0047] S33, concatenating with and performing dimensionality expansion through a function:

[0048]

[0049] wherein, for an image with the number of channels being , height being 1, and width being , a 1×1 convolutional layer with an input channel of and an output channel of 32 is used for convolution operation to obtain a tensor representation integrating spatio-temporal information :

[0050]

[0051] Optionally, in step S30, the step of using a temporal convolutional network with different dilation factors to capture the traffic change rules of different periods in the tensor representation, obtaining a base station traffic representation integrating spatio-temporal features and different period features, and inputting the base station traffic representation into a fully connected layer to predict the base station traffic at the next moment includes:

[0052] S34, performing dimensionality conversion on the tensor representation integrating spatio-temporal information to obtain , and using as the input of the base station traffic temporal feature extraction model;

[0053] S35, respectively inputting the input sequences of the base station into 5 different hierarchical temporal convolutional network modules, and performing convolution with a stride of 1 through the following formula:

[0054]

[0055] In the formula, represents the flow rate at the th layer and the th moment. The size of the convolutional kernel is , is the weight of the convolutional kernel;

[0056] Obtain ;

[0057] S36. Obtain the results obtained after , , , and convolutions on the outputs of 5 time-domain convolutional network modules, respectively, to obtain , , , and ;

[0058] Among them, , , , , ;

[0059] S37. Add the convolution results according to the following formula to obtain the traffic time series feature representation of the base station :

[0060]

[0061] Among them, represents the last elements of the output result of the rd layer of the model with a dilation factor of ;

[0062] S38. Input the traffic time series feature representation into the fully connected layer to predict the base station traffic at the th moment, calculate the mean square error between its predicted value and the actual traffic of the base station at the th moment , and use the mean square error as the loss function to update the weight parameters of each module through the stochastic gradient descent algorithm. After the model training is completed, perform iterative prediction through the autoregressive process to achieve multi-time step base station traffic prediction.

[0063] Optionally, the step S40 specifically includes:

[0064] S41. Assume that the base station at The traffic at a certain moment is , and the base station energy consumption is positively correlated with the traffic. Under the premise of ensuring the user experience, it can be adjusted to the lowest power through the following formula : :

[0065]

[0066]

[0067] In the formula, is the maximum power that can be provided, is the base station the maximum traffic that can be processed simultaneously, is the static energy consumption proportionality coefficient, is at the predicted traffic at a certain moment, represents at a certain moment the switch state, indicating the off state when is the set of other base stations within the coverage area;

[0068] S42. Predict the base station traffic for the next moments, with an interval of minutes between adjacent moments. At each moment, the base station power and switch state are allowed to be adjusted once, and after adjustment, it runs continuously for time. When reaching the next moment, the next adjustment can be made. According to the operating states of the base stations at different time intervals, the total power of the base stations in the area for the next time period can be calculated. Through the unit energy consumption cost , the operating cost of the base stations in this area can be calculated:

[0069]

[0070]

[0071] Among them, represents the power of the base station in the th , is the number of startups of the base stations in the area during the time period, is the number of base stations in the area

[0072] This application has at least the following beneficial effects:

[0073] 1. Aiming at the problem that it is difficult to extract features due to the complex structure of large-scale base station networks, an efficient base station network partitioning method is designed, which can not only effectively identify areas with frequent traffic changes but also simplify the feature extraction of base station networks.

[0074] 2. Aiming at the noise or errors in the original spatial data of base stations, based on structural information theory and attention mechanism, a new method for extracting hierarchical structure features and traffic correlation features of base stations is proposed to capture effective features in the base station network structure, avoiding the interference of irrelevant information and the loss of effective information.

[0075] 3. By fusing the spatio-temporal features of the base station network, different periodic rules of base station traffic changes are effectively captured through a time-domain convolutional network with different dilation factors, realizing accurate prediction of base station traffic. And according to the traffic prediction results, an energy-saving strategy for base stations in each region is constructed, achieving effective energy saving and efficient regulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 is a schematic flow chart of the steps of the base station regulation method that fuses spatio-temporal information and traffic features in this application;

[0077] Figure 2 is a schematic diagram of the topological relationship of some regions involved in the embodiments of this application;

[0078] Figure 3 is a schematic diagram of the process of extracting traffic time series features involved in the embodiments of this application;

[0079] Figure 4 is a schematic diagram of the process of constructing the hierarchical structure representation of the base station network involved in the embodiments of this application.

[0080] The realization, functional features and advantages of the purpose of this application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0081] To better understand the above technical solutions, the exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be fully conveyed to those skilled in the art.

[0082] First Embodiment

[0083] Referring to Figure 1 , in this embodiment, the base station regulation method that fuses spatio-temporal information and traffic features includes the following steps:

[0084] S10. Divide the original base station network into multiple regional base station networks, construct a graph representation of each of the regional base station networks, and construct a graph representation with regions as nodes by defining the connection relationships between regions.

[0085] In this embodiment, referring to Figure 2 the schematic diagram of the topological relationship between base stations in the shown region, divide the large-scale base station network composed of the inherent topological structure between base stations into multiple smaller-scale regional base station networks; construct a graph representation of each regional base station network, and construct a graph representation with regions as nodes by defining the connection relationships between regions.

[0086] S20. Extract the hierarchical structure features of the regional base station network based on the coding tree, perform singular value decomposition on the traffic features, add learnable parameters to construct a directed weight matrix between base stations, update the directed weight matrix to an attention matrix, and then capture the traffic correlation features of the base station network based on the attention matrix.

[0087] In this embodiment, based on the structural information theory, construct a coding tree representing the hierarchical structure of the base station network, and extract the hierarchical structure features of the base station network based on the coding tree; perform singular value decomposition on the traffic features, add learnable parameters to construct a directed weight matrix between base stations, and then capture the traffic correlation features of the base station network based on the attention mechanism.

[0088] S30. Embed three types of natural period markers into the base station traffic sequence to obtain a temporal feature representation of the traffic, use convolution operations to fuse the temporal feature representation of the traffic with the traffic correlation features of the base station network to obtain a tensor representation integrating spatio-temporal information, use a temporal convolutional network with different dilation factors to capture the traffic change rules of different periods in the tensor representation, obtain a base station traffic representation integrating spatio-temporal features and different period features, and input the base station traffic representation into a fully connected layer to predict the base station traffic at the next moment.

[0089] In this embodiment, embed three types of natural period markers into the base station traffic sequence to obtain a temporal feature representation of the traffic, and use convolution operations to fuse it with the traffic correlation features of the base station network; then use a temporal convolutional network with different dilation factors to capture the different period rules of traffic changes; finally, predict the base station traffic through a fully connected layer (FCL).

[0090] S40. With the goal of minimizing the base station operation cost, determine the optimal regulation strategy for base station power adjustment in each region according to the predicted base station traffic.

[0091] In this embodiment, according to the base station traffic prediction result, a method for solving the minimum power of the base station is designed to ensure the user experience. By using the multi-layer coverage structure of the base station, an optimization objective of minimizing the operating cost of the base station network is constructed for each regional base station network divided in step S10, and the optimal strategy for adjusting the power of the base station in each region is obtained by solving the objective function.

[0092] In the technical solution provided in this embodiment, the base station network is constituted by using the inherent topological structure between base stations. According to the characteristics of traffic changes, a base station network division process is designed to divide the large-scale base station network into multiple regional base station networks with smaller scales. By using the characteristic of minimizing the structural entropy of the coding tree, the noise interference in the base station network structure is reduced, and the hierarchical structure representation of the base station network is constructed through the coding tree, so as to extract the hierarchical structure characteristics of the base station network. Through singular value decomposition of the historical traffic of the base station and adding learnable parameters to construct a directed weight matrix between base stations, the attention mechanism is used to capture the correlation characteristics of traffic changes between base stations. In addition, aiming at the periodicity of the change of base station traffic in the time dimension, time-domain convolutional networks with different dilation factors are designed to capture the temporal characteristics in the base station data. The above three types of characteristics are fused to predict the base station traffic, and then according to the traffic prediction result, an optimization objective of minimizing the operating cost of the base station is constructed for each regional base station network, and the optimal strategy for adjusting the power of the base station in each region is formulated under the constraint of ensuring the user experience.

[0093] Second Embodiment

[0094] In this embodiment, aiming at the problems of complex structure and difficult feature extraction existing in the large-scale base station network, a method for dividing the large-scale base station network into multiple regional base station networks is proposed in this embodiment. Specifically, in the step S10, the steps of dividing the original base station network into multiple regional base station networks specifically include:

[0095] S11, perform traffic sorting on each base station in the original base station network, and select the base station with the largest traffic density as the central base station;

[0096] S12, when it is determined that there are at least 3 target base stations with traffic greater than the preset traffic threshold in the circular neighborhood associated with the central base station, the circular neighborhood of the central base station and the circular neighborhoods associated with the target base stations are divided into one regional base station network;

[0097] S13, repeat S11 - S12 until each base station is divided into the corresponding regional base station network, and m regional base station networks are obtained.

[0098] In this embodiment, the number of mobile devices connected to a base station is referred to as the traffic of that base station. When a mobile device switches from one base station to another, it is said that there is a traffic variation relationship between these two base stations. In order to divide base stations with frequent traffic variation relationships into the same area, the present invention divides the base station network according to traffic density. The specific division steps are as follows:

[0099] Base station sorting based on traffic density. The base station traffic density is the base station traffic per unit service area. In order to adapt to different historical traffic sequence lengths of different base stations, the following formula is designed to calculate the base station traffic density, and based on this, the base stations in the base station network are sorted. The historical traffic data with a length of is cut into windows by a sliding window, and each window is sampled by the downsampling function . Then, the average of the sampled values is calculated, and the ratio of it to the base station service area is used as the traffic density of the base station.

[0100]

[0101] Among them, is the window size, is the base station service radius, means taking the maximum traffic value in the th window.

[0102] ② Base station selection. Let represent the set of selected base stations in a division area. First, the base station with the largest traffic density that has not been divided is put into .

[0103] ③ Neighboring base station inspection. Taking the base station selected in step ② as the center, a circular neighborhood with a radius of is constructed. The base station located at the center is called the central base station. Let the traffic density of the central base station be . If there are more than 3 base stations with a traffic density reaching in the circular neighborhood of the central base station, then this central base station is called a core base station. Considering the non-uniform characteristic of the base station spatial distribution, the following formula is used to dynamically calculate the neighborhood radius . Taking the selected base station as the center of a square with a length of 1 km and a width of 1 km, the number of base stations within this square is calculated to obtain , ensuring that the neighborhood radius has a negative correlation with the unit base station number .

[0104]

[0105] Among them, is the service radius of the selected base station.

[0106] ④ Region expansion. Add all the base stations within the radius neighborhood of the selected base station to and sequentially check the base stations within these neighborhoods. If there is a core base station, continue to add the base stations within its neighborhood to .

[0107] Continuously repeat steps ② to ④ until all base stations have their respective regions, thereby dividing the base station network into regions, retaining the original base station connection relationships within each region, and a regional base station network of regions can be obtained.

[0108] Further, in step S10, constructing the graph representation of each regional base station network specifically includes:

[0109] S14, taking the base stations as graph nodes and the connection relationships between the base stations in the regional base station network as edges, representing the base station network with , where represents the set of base stations in this base station network, represents the number of base stations included, represents the connection relationships between base stations. If there is a connection relationship between base stations and , then the element value of the th row and th column of the matrix is 1, that is , otherwise ;

[0110] S15, sequentially construct the graph representations for the regional base station networks obtained in step S14 to obtain , where , represents the number of base stations included in , and N represents the sum of the number of base stations in

[0111] S16, if there is a connection relationship between the base stations in two different regions in the original base station network, then connect these two regions to form a mesh structure as the regional association network, and the regional association network is denoted as , where is the set of all regions, represents the connection relationships between regions, and the element value of the th row and th column of represents that the base stations numbered and The number of base station pairs with all cross-region connections between regional base station networks.

[0112] In this embodiment, taking the base stations as graph nodes and the connection relationships between base stations in the base station network as edges, then can represent the base station network, where is the set of base stations in the base station network, represents the number of base stations it contains, represents the connection relationship between base stations. If there is a connection relationship between base stations and , then the element value of the th row and th column of the matrix is 1, that is , otherwise . Successively construct the graph representations of the regional base station networks obtained in step S13 to obtain , where , represents the number of base stations contained in

[0113] Furthermore, after step S16, it further includes:

[0114] S17. Let the base station attribute feature be represented as , then the attribute feature matrix of is represented as , where represents the length of the base station attribute feature. Perform feature transformation on the feature matrix of each regional base station network through a fully connected layer to obtain the feature matrix of . The traffic at successive and equally spaced moments recorded by the base stations constitutes a traffic sequence. Sum the traffic of all base stations in the region at each moment to obtain the overall traffic sequence of the region, and splice the traffic sequences of all regions to obtain the traffic feature matrix of the regional association network

[0115] Furthermore, in this embodiment, in order to represent the topological relationship between regions, if there is a connection relationship between base stations in two different regions in the original base station network, then connect these two regions to form a mesh structure, which is called the regional association network. The regional association network is denoted as , where is the set of all regions, represents the connection relationship between regions, the element value of the th row and th column of represents the number and the number of base station pairs for all cross - regional connections between the area base station networks. The base station number, service range, geographical location, etc. are used as base station attribute features , then the attribute feature matrix of is represented as where represents the length of the base station attribute features. The feature matrix of each area base station network is subjected to feature transformation through a fully - connected layer to obtain the feature matrix of . The traffic at consecutive and equally - spaced moments recorded by the base stations constitutes a traffic sequence . At each moment, the traffic of all base stations in the area is summed to obtain the overall traffic sequence of the area. The traffic sequences of all areas are concatenated to obtain the traffic feature matrix of the area - associated network

[0116] Third Embodiment

[0117] In this embodiment, in step S20, the step of extracting the hierarchical structure features of the area base station network based on the coding tree specifically includes:

[0118] S21, initialize the base stations in an area base station network sequentially as independent base station partitions, represents the number of base stations, and the base station partition after the base station network is initialized is , where represents the sub - partition, represents the area base station network diagram numbered ;

[0119] S22, traverse all base station partitions. If the base station partitions and satisfy that the value is the largest and positive, then merge and to obtain a new base station partition , where is the coding tree after partition merging, represents the structure entropy before merging of the base station partitions and on the current area base station network diagram , represents the structure entropy after merging;

[0120] S23, traverse all base station partitions. If the base station partitions and Meet If the value is the largest and positive, then and will be combined to create a new partition , and and the two partitions will be placed into , and will become sub - partitions, where the height of the coding tree does not exceed 3;

[0121] S24. Repeat steps S22 and S23 until there are no base stations that can be merged and combined, to obtain the final base station partition set of this area, where the 0th layer of the coding tree is the root node, and all partitions in the base station partition set are used as the 1st - layer nodes. If a partition contains sub - partitions, then the sub - partitions are used as child nodes to continue expanding until the base stations are leaf nodes, thus obtaining the optimal coding tree corresponding to the base station network of this area ;

[0122] S25. Repeat steps S22 - S24 to construct the optimal coding tree of the entire area's associated network . Replace the root nodes of the optimal coding trees of each area's base station network with the leaf nodes of , to obtain the hierarchical structure representation of the entire base station network . Information aggregation is performed through graph convolution between the layers of . The node feature matrix of the layer is obtained through the said graph convolution to get the node feature matrix of the layer. Starting from , the base station attribute features of the 6th layer of the coding tree are concatenated to obtain , and information is iteratively passed upward, finally obtaining the base station network hierarchical structure feature representation , where the expression of the graph convolution is as follows:

[0123]

[0124] In the formula, represents the connection relationship between the nodes of the layer and the nodes of the layer of the coding tree , is the activation function, is the parameter matrix of the layer.

[0125] In this embodiment, within a region, the hierarchical structure of the base station network is reflected in the coexistence and collaborative operation of base stations with different coverage ranges and operating frequency bands. For example, the macro base station operating in the low frequency band achieves wide signal coverage by virtue of its strong penetration power, and there are also micro base stations and pico base stations under the coverage of the macro base station, which play functions such as load balancing and signal enhancement. Based on structural information theory, the present invention generates an optimal coding tree of the base station network by minimizing the structural entropy of the base station network, and then extracts the hierarchical structure features of the base station network based on the coding tree. Among them, the structural entropy is an extension of the Shannon entropy, which is used to measure the uncertainty of a graph through hierarchical partitioning and is calculated based on the degree of nodes; the coding tree is the optimal hierarchical structure of the graph, representing the association between different levels in the data, and can be constructed based on minimizing the structural entropy.

[0126] The following formula gives the expression of the minimum structural entropy, by minimizing the graph representation of the base station network of the structural entropy , an optimal coding tree of a base station network is found , which can not only reduce the noise interfering with the base station traffic prediction in the base station network structure, but also effectively represent the hierarchical structure information of the base stations.

[0127]

[0128] Among them, represents the height of the coding tree, covers all coding trees with a height not exceeding , is the structural entropy of the base station network under the partition of the coding tree , and the calculation method is as follows: The calculation method is as follows:

[0129]

[0130]

[0131] Among them, represents a partition, that is, a set of base stations or sub - partitions, is the sum of the cut - edge weights of , is the parent - node partition of , calculates the sum of the weights of the nodes within the partition, is the root node of.

[0132] Considering the three different hierarchical base stations, namely macro base stations, micro base stations and pico base stations, in the base station network, an optimal coding tree with a height of 3 is constructed for the regional base station network , to represent the hierarchical structure of the regional base station network, which is constructed as follows:

[0133] The base stations in a regional base station network are initialized in sequence as independent base station partitions, representing the number of base stations. After the initialization of the base station network, the base station partition is , where represents the sub - partition.

[0134] Traverse all base station partitions. If the base station partitions and satisfy and the value is the largest and positive, then and are merged, that is, the sub - partitions or base stations included in are merged to obtain , and then is deleted, so as to obtain a new base station partition . In order to find the base station partition scheme that minimizes the structural entropy of the base station network, it should be ensured that the merging operation can minimize the structural entropy of the base station network to the greatest extent, that is, it is necessary to satisfy is the largest and positive, where is the coding tree after partition merging.

[0135] Traverse all base station partitions. If the base station partitions and satisfy and the value is the largest and positive, then and are combined, that is, a new partition is created, and and are put into , and become 's sub - partitions. Through the combination operation, it is ensured that the base stations with smaller service ranges are farther from the root node, thus reflecting the hierarchical structure of the base station network. The combination operation requires that the height of the combined coding tree does not exceed 3.

[0136] Repeat steps S22 and S23 until there are no partitions that can be merged or combined, and obtain the final base station partition set of this region. The 0th layer of the coding tree is the root node. All partitions in the base station partition set are used as the 1st - layer nodes. If a partition contains sub - partitions, its sub - partitions are used as sub - nodes to continue to expand until the base stations are used as leaf nodes, so as to obtain the optimal coding tree of the regional base station network 。

[0137] To represent the hierarchical structure information of the entire base station network, construct the optimal coding tree of the regional association network with reference to the above steps , The number of leaf nodes corresponds to the number of regions. Replace the root nodes of the optimal coding trees of the base station networks in each region with the leaf nodes to obtain the hierarchical structure representation of the entire base station network 。 The local and global hierarchical structure information of the base station network is retained. To effectively capture the hierarchical structure features of the base station network, graph convolution is used for information aggregation between the layers, and information is transmitted in a bottom-up manner. The following formula is the process of graph convolution operation. The node feature matrix of the layer is obtained through the following formula The node feature matrix of the layer, starting from the leaf nodes, that is , splice the base station attribute features of the 6th layer of the coding tree to obtain , and iteratively transmit information upward to finally obtain the base station network hierarchical structure feature representation with a length of 。

[0138]

[0139] Among them, represents the connection relationship between the nodes of the layer and the nodes of the layer of the coding tree , is the activation function is the parameter matrix of the layer

[0140] Furthermore, in the step S20, the steps of performing singular value decomposition on the traffic characteristics, adding learnable parameters to construct a directed weight matrix between base stations, and updating the directed weight matrix to an attention matrix, and then capturing the traffic association characteristics of the base station network based on the attention matrix specifically include

[0141] S26, set the traffic characteristic matrix , use the equation to perform singular value decomposition on the traffic characteristic matrix to obtain the left singular value matrix representing the spatial distribution between base stations, and the right singular matrix of the traffic change over time. The diagonal elements are the singular values. Select the first 10 singular values ​Obtain , select the first ten columns of the obtained , select the first 10 rows Obtain ;

[0142] S27. Incorporate , and into learnable parameters and . Constrain the range of the correlation weights between base stations through the function, and ensure the non-negativity of the correlation weights between base stations using the function, thereby obtaining the learnable directed weight matrix :

[0143]

[0144] In the formula, represents the degree of correlation between each pair of base stations. The th row represents the degree of correlation of other base station nodes with the current base station . The larger the weight value, the greater the correlation;

[0145] S28. Calculate the distance between any two base stations, and take five times the sum of their service radii as the distance threshold . If the distance between the base stations exceeds the set threshold , then set its weight to 0, and then use the function to retain the maximum values for each row vector and set the rest to 0, obtaining the attention matrix that does not depend on the physical connection or communication path between base stations , and its expression is as follows:

[0146]

[0147]

[0148] In the formula, is a piecewise function, represents the th row and th column element, that is, the correlation weight between base station and base station represents the distance between and

[0149] S29. For Aggregate and update the attribute feature vectors:

[0150]

[0151] In the formula, is the feature transformation matrix, is the attribute feature vector of is the feature vector after aggregate update.

[0152] S210. Repeat steps S26 to S29 to calculate the attention matrices of each regional base station network and regional association network in turn, and obtain , ... and , and then perform feature aggregation based on the attention mechanism in turn, so as to obtain the traffic association feature of the regional base station network, and the traffic association feature of the regional association network.

[0153] In this embodiment, the traffic changes between base stations are correlated and directional. In order to more accurately describe the potential traffic change correlation relationship between base stations, the present invention is based on the historical traffic data of base stations, breaks through the limitation of only relying on the existing base station topology, adds learnable parameters to construct a directed weight matrix between base stations, and effectively extracts the potential features that conform to traffic correlation in the base station network through the attention mechanism.

[0154] First, perform singular value decomposition on the traffic feature matrix through the following formula. Specifically, arrange the square roots of the eigenvalues of the matrix in descending order as the diagonal matrix , and use the equation to decompose the traffic feature matrix into the left singular value matrix representing the spatial distribution between base stations, and the right singular matrix where the traffic changes over time. The diagonal elements are the singular values, and the singular values represent the importance of the corresponding column vectors of and the row vectors of . Select the first 10 singular values to obtain , correspondingly select the first 10 columns of to obtain

[0155]

[0156] Then, for , and add learnable parameters and , and use the function to constrain the range of the correlation weights between base stations, and use the function to ensure the non-negativity of the correlation weights between base stations, so as to obtain a learnable directed weight matrix :

[0157]

[0158] represents the degree of correlation between each base station. The th row represents the degree of correlation of other base station nodes with the current base station . The greater the weight value, the greater the correlation. Then, improve the calculation efficiency of the model through the following formula. Specifically, first calculate the distance between any two base stations, and take 5 times the sum of their service radii as the distance threshold . If the distance between base stations exceeds the set threshold , then set its weight to 0, and then use the function to retain the maximum values for each row vector, and set the rest to 0, so as to ensure that the model only focuses on the neighboring base station nodes that are most relevant to the communication and traffic changes of the current base station, and prevent the misleading of long-distance base stations during traffic prediction.

[0159]

[0160] Among them, is a piecewise function that sets the correlation weight of two base stations with a geographical distance not less than the set threshold to 0. Its calculation method is as follows:

[0161]

[0162] Among them, represents the th row and column element, that is, the correlation weight between base station and base station . represents and the distance between.

[0163] It is an attention matrix that does not rely on the physical connection or communication path between base stations and can effectively reflect the potential correlation degree between base stations at the traffic level. Next, the attribute features of the base stations are updated through feature aggregation based on the attention mechanism. Specifically, The association weight with other base stations is of the row , where represents the association weight between and . The attribute feature vector of

[0164]

[0165] is aggregated and updated by the following formula: where is the feature transformation matrix, is the attribute feature vector of , and is the feature vector after aggregation and update of

[0166] Through the above steps, the attention matrices of each regional base station network and regional association network are calculated in turn to obtain , , and , and then the feature aggregation based on the attention mechanism is performed in turn to obtain the traffic association feature representations and of their respective networks.

[0167] Fourth Embodiment

[0168] In this embodiment, in step S30, the step of embedding the three types of natural period markers into the base station traffic sequence to obtain the temporal feature representation of the traffic and fusing the temporal feature representation of the traffic with the base station network traffic association feature by using the convolution operation includes:

[0169] S31, embedding the daily marker, weekly marker, and monthly marker into the base station traffic sequence to obtain the temporal feature representation of the traffic ;

[0170] S32, splicing the base station network hierarchical feature representation , the traffic association feature representation and the to obtain the spliced base station network traffic association feature representation ;

[0171] S33, splicing with and passing through The function performs dimensionality expansion:

[0172]

[0173] Among them, the number of channels is , the height is 1, and the width is For an image, a 1×1 convolutional layer with an input channel of and an output channel of 32 is used for convolution operation to obtain a tensor representation that fuses spatio-temporal information :

[0174]

[0175] In this embodiment, for the different periodicities of the base station traffic change in the time dimension, such as the alternation between weekdays and holidays, and the fluctuations during peak periods of different weekdays, three types of natural period markers are embedded in the traffic sequence. The daily, weekly, and monthly markers can be calculated through the following formula. The calculation of the daily marker starts from the time slice of 00:00 of a day , the current time slice minus the start time slice to obtain the position of this time slice in a day. To maintain numerical stability, the position of the current time slice is normalized by dividing by the total duration of a day . The marker of 00:00 is 0, the marker of 00:05 is , the marker of 00:10 is , and so on. Similarly, the weekly marker is the week where the current time slice is located divided by 7, and the monthly marker is the month where the current time slice is located divided by 12.

[0176]

[0177] The traffic sequence is concatenated with the above three types of natural period markers to obtain traffic time series features . Next, through convolution operation, the base station network traffic correlation features obtained in step S32 are incorporated into the base traffic time series features to achieve spatio-temporal feature fusion. The base station network hierarchical structure feature representation , the traffic correlation feature representation and are concatenated to obtain the base station network traffic correlation feature representation . Further concatenate and , and perform dimensionality expansion through the function. Consider as an image with a channel number of , a height of 1, and a width of . Use an input channel of Perform a convolution operation on a 1×1 convolutional layer with 32 output channels to obtain a tensor representation that fuses spatio-temporal information :

[0178]

[0179] Furthermore, in this embodiment, in step S30, a temporal convolutional network with different dilation factors is used to capture the traffic change rules of different periods in the tensor representation, obtaining a base station traffic representation that fuses spatio-temporal features and different period features. The step of inputting the base station traffic representation into a fully connected layer to predict the base station traffic at the next moment includes:

[0180] S34, perform a dimension transformation on the tensor representation that fuses spatio-temporal information to obtain , and use as the input of the base station traffic time series feature extraction model;

[0181] S35, input the input sequences of the base station into 5 different hierarchical temporal convolutional network modules respectively, and perform a convolution with a stride of 1 through the following formula:

[0182]

[0183] In the formula, represents the traffic representation at the th layer and the th moment, the convolution kernel size is , is the convolution kernel weight;

[0184] obtain ;

[0185] S36, obtain the results of 5, , , , , and convolutions respectively output by the 5 temporal convolutional network modules, obtaining , , , , and ;

[0186] Among them, , , , , ;

[0187] S37, add the convolution results according to the following formula to obtain the traffic time series feature representation of the base station ​ :

[0188]

[0189] Among them, represents the end -th element of the output result of the -th layer of the model with the dilation factor ;

[0190] S38, representing the traffic time series features is input to the fully connected layer to predict the base station traffic at time , calculating the mean square error between its predicted value and the true traffic of the base station at time

[0191] . After the model training is completed, iterative prediction is performed through an autoregressive process, so as to realize the base station traffic prediction of multiple time steps.

[0191] In this embodiment, the time-domain convolutional network combines one-dimensional convolution and dilated causal convolution, which can effectively capture the short-term and long-term time series dependencies of the base station traffic change while improving the calculation efficiency. By introducing dilated convolution, the time-domain convolutional network expands the receptive field for the base station traffic change without increasing the size of the convolutional kernel, so as to capture long-distance dependencies in a relatively shallow network. The following formula gives the convolution process of the time-domain convolutional network, and the dilation factor controls the receptive field of the convolution for the traffic change. The input traffic sequence of the -th layer obtains through one dilated causal convolution operation. The weighted sum of the dilated traffic representations at the corresponding time in is taken to obtain the traffic representation of the -th time of the -th layer:

[0192]

[0193] Among them, the size of the convolutional kernel is , and is the weight of the convolutional kernel.

[0194] The time-domain convolutional networks with different dilation factors are used to capture the dependence relationships of traffic changes at different time scales, so as to more effectively learn the different periodic laws of traffic changes. The common divisors 1, 2, 3, and 6 of the number of days 30 in a month and the number of hours 24 in a day, as well as the number of days 7 in a week, are used as the dilation factors of 5 time-domain convolutional networks.

[0195] First, perform dimensionality transformation on the fused representation of spatio-temporal information to obtain , and with reference to Figure 3 , use as the input to the base station traffic time series feature extraction model. Then, input the input sequences of the base station into 5 time-domain convolutional network modules respectively, and perform convolution with a stride of 1 through the following formula to obtain , and use it as the input of the second layer, and continue to perform the convolution operation of the following formula. The 5 time-domain convolutional network modules perform , , , and times of convolution respectively to obtain , , , and , where , , , , . Finally, add the convolution results in the following way to obtain the traffic time series feature representation of the base station :

[0196]

[0197] where, represents the last elements of the output result of the th layer of the model with a dilation factor of .

[0198] Through the above steps, the base station traffic representation fusing spatio-temporal features and different periodic features is obtained. Input into the fully connected layer to predict the base station traffic at time, and then calculate the mean square error between its predicted value and the actual traffic of the base station at time , and use it as the loss function, and update the weight parameters of each module through the stochastic gradient descent algorithm. After the model training is completed, iterative prediction can be performed through the autoregressive process to achieve multi-time step base station traffic prediction.

[0199] The Fifth Embodiment

[0200] In this embodiment, step S40 specifically includes:

[0201] S41, assume that the base station at The traffic at a moment is , and the base station energy consumption is positively correlated with the traffic. Under the premise of ensuring the user experience, it can be adjusted to the lowest power through the following formula : :

[0202]

[0203]

[0204] In the formula, is the maximum power that can be provided, is the base station the maximum traffic that can be processed simultaneously, is the static energy consumption proportionality coefficient, is at the predicted traffic at the moment, represents the moment of the switch state, indicating the off state when is the set of other base stations within the coverage area;

[0205] S42. Predict the base station traffic in the next moments. The interval between adjacent moments is minutes . At each moment, the base station power and switch state are allowed to be adjusted once. After adjustment, it runs continuously for time. When the next moment arrives, the next adjustment can be made. According to the operating states of the base stations at different time intervals, the total power of the base stations in the area in the next time period can be calculated. Through the unit energy consumption cost , the operating cost of the base stations in this area can be calculated:

[0206]

[0207]

[0208] Among them, represents the power of the th base station within , is the number of times the base stations in the area are started within the time period, is the number of base stations in the area

[0209] In this embodiment, increasing the base station power can enhance its signal and data processing capabilities, thereby providing high-quality communication and Internet services for more users. Reducing the base station power, on the other hand, will weaken the base station's data processing capabilities. Therefore, in order to ensure the user experience, the base station power is adjusted on the premise of not affecting the basic data processing capabilities of the base station. Assume that the base station at the traffic at time is , and the base station energy consumption is positively correlated with the traffic. Then, the minimum power that can be adjusted while ensuring the user experience can be expressed by the following formula can be adjusted to the lowest power :

[0210]

[0211] where, is the maximum power that can be provided, is the base station the maximum traffic that can be processed simultaneously, is the static energy consumption proportionality coefficient.

[0212] Adjusting the base station power in advance using the predicted base station traffic can solve the lag in adjusting the base station power according to the real-time monitored traffic. A region may have multi-layer base station signal coverage. The upper-layer base station has a larger coverage area, and within its coverage area, the lower-layer base station is responsible for a smaller coverage area to enhance the signal. The lower-layer base station is defined as a closable base station. When it is closed, its traffic will be transferred to the upper-layer base station. To further ensure the signal strength and user experience, a traffic threshold is set for the closable base station . When its traffic is less than , then the base station can be completely closed, and the traffic load is transferred to its upper-layer base station. Then the traffic at time can be expressed as:

[0213]

[0214] where, is the predicted traffic at time , represents the switch state at time , indicating the closed state when is the set of other base stations within the coverage area.

[0215] By predicting the base station traffic in the next time moments, with an adjacent time interval of minute The base station power and switch status are allowed to be adjusted once at each moment, and after the adjustment, it continues to operate for a period of time. At the next moment, the next adjustment can be made. According to the operating status of the base stations at different time intervals, the total power of the base stations in the area in the future time period can be calculated. Through the unit energy consumption cost the operating cost of the base stations in this area can be calculated. The following formula constructs an objective function by minimizing the cost. The startup process of the base station includes self-checking of equipment, initialization of signals, synchronization with the network, etc. To prevent the base station from being frequently turned off and on, which increases energy consumption, the startup cost of the base station is also considered in the objective function. The startup cost of the base station can be obtained according to the average startup power and startup duration of the base stations in the area . The following formula restricts the power adjustment range of the base station to ensure the user experience. Under its constraints, the objective function of the area is solved to obtain the base station power adjustment strategy in this area in the future time period.

[0216]

[0217]

[0218] Among them, represents the power of the th base station within , is the number of startups of the base stations in the area during the time period, and is the number of base stations in the area

[0219] As an implementation solution, in a specific embodiment, the base station traffic data segments in the embodiment are shown in Table 1. The length of the base station historical traffic data is 34272. The window size is set to 16, and the window moving step size is also set to 16. The traffic data is segmented into 2142 windows. The downsampling function samples the maximum traffic in the sampling window, calculates the traffic density of all base stations in this area, and sorts them from large to small according to the base station density. When selecting the base station with the largest traffic density that has not been visited, the selected base station is used as the center of a square with a length of 1 km and a width of 1 km, and the number of base stations within this square is calculated to obtain , and the radius of each area is calculated. According to the base station network division process of steps S11 - S13, base stations are divided into areas. The number of base stations in some areas and the average traffic density within the area are shown in Table 2.

[0220] Table 1 Fragment of Base Station Traffic Data

[0221]

[0222] Construct the graphical representation of 100 regions according to the steps S14 - S16 , For the topological relationship between base stations within the region, obtain the graphical representation of each region ,, and the graphical representation of the region association network For the region association network, the connection relationship between some regions is as Figure 2 shown. It can be known that the element values of the 1st row, 2nd column, 1st row, 3rd column, and 2nd row, 3rd column of the matrix representing the inter - region connection relationship are 2, 1, and 0 respectively.

[0223] Table 2 Number of Base Stations in Some Regions after Partitioning

[0224]

[0225] In Table 1, the base station records traffic every 5 minutes, and records the base station traffic at 288 moments every day. Set a window of size 288 ( ), with a moving step size of 1. Take the traffic sequence within the window as the training data, and the traffic at the next moment as the label. Thus, each base station can obtain 33985 groups of traffic sequences and corresponding labels. The traffic sequence of a region is the sum of the traffic of the base stations in the region at the corresponding moment. Part of the traffic data of the base stations in Region 2 is shown in Table 3, and the last column in the table represents the traffic sequence of this region. Base station attribute features include base station number, region number where it is located, service radius, service frequency band, maximum power, and the longitude and latitude of the base station. Concatenate the base station attribute features of base stations numbered 6, 9, 10, and 15 to obtain a feature vector with a length of 284, and reduce the dimension through a fully - connected layer with an input length of 284 and an output of 71 to obtain the attribute feature vector of Region 2 . Similarly, the traffic sequences and attribute feature vectors of all regions can be calculated, and after concatenation, we get , and then normalize the traffic sequences and attribute feature vectors of all base stations and regions in turn.

[0226] Table 3 Partial Traffic Data of Base Stations in

[0227]

[0228] 2: Extraction of Base Station Network Traffic Association Features

[0229] The generation method of the base station network hierarchical structure representation is asFigure 4 As shown in the figure. First, initialize the base stations in each regional base station network one by one. One base station is taken as a partition. The 0th layer of the initial coding tree is the root node, the 1st layer is the initial partition, the 2nd layer is the leaf node, and the base stations are used as leaf nodes. Let , and use the union and merge operations to construct an optimal coding tree with a height of for each regional base station network. The coding tree construction method in the previous steps obtains , , , . Construct the optimal coding tree for the regional association network in the same way , and then replace the root nodes of the coding tree , , , of the regional base station network with the leaf nodes of to obtain the hierarchical structure representation of the entire base station network .

[0230] Next, perform hierarchical convolution. The parameters of each layer are shown in Table 4. The 6th layer is the leaf node, that is, the base station. There are a total of 2,000 leaf nodes. Concatenate the attribute features of all base stations into a matrix , 71 is the length of the attribute feature. The number of nodes in the 5th layer is 940. Through , the features of the 6th layer can be aggregated to the 5th layer. By analogy, information transmission and data dimensionality reduction are achieved through bottom-up aggregation, and finally the feature representation of the root node is obtained, that is, the hierarchical structure features of the entire base station network . Through , control the feature length in the hierarchical convolution process. The finally output feature length 288 is equal to the traffic sequence length ( 288).

[0231] Table 4 Parameters of hierarchical convolution

[0232]

[0233] First, form a traffic matrix by sequentially combining the base station traffic sequences in each region . Perform singular value decomposition on the traffic matrix in sequence, and then sequentially select the first 10 columns from the obtained left singular matrix to get , select the first 10 rows and the first 10 columns from the singular value diagonal matrix to get , and select the first 10 rows from the right singular matrix to get . For example, after decomposition, it obtains , , , select the first 10 columns to obtain , select the first 10 rows and the first 10 columns of , select the first 10 rows of .

[0234] Add learnable parameters. The parameter dimensions of some regions are shown in Table 5. Construct a directed weight square matrix for each regional base station network to obtain . For example, if there are 4 base stations in Region 2, the finally constructed directed weight matrix is . Regard the nodes of the regional association network as base stations and construct the directed weight matrix in the same way . Set the distance threshold to 5 km ( ), and calculate the final attention matrix of each regional base station network . Take the position of the southernmost base station in each region as the representative position of the region, set the distance threshold of the regional association network to 50 km, and similarly calculate the attention matrix of the regional association network .

[0235] Table 5 Parameter dimensions of some regions

[0236]

[0237] Finally, perform attention aggregation on each regional base station network and regional association network in turn according to their respective attention matrices to obtain the traffic association features of their respective networks . 3: Base station traffic prediction

[0238] Calculate daily, weekly, and monthly markers based on the timestamps of the base station traffic sequence. The calculation results of the daily, weekly, and monthly markers for some timestamps are shown in Table 6. Let the starting time slices of the daily marker, weekly marker, and monthly marker be 0 ( = 0), and then calculate each type of marker. For example, "2023 / 3 / 1 0:05" corresponds to a daily marker of , since March 1, 2023 is Wednesday, the weekly marker is , and the monthly marker is . Concatenate the traffic sequence of each base station with the three types of markers corresponding to the corresponding timestamp to obtain the traffic time series features of all base stations .

[0239] Table 6 Marker calculation results for some timestamps

[0240]

[0241] Next, fuse the traffic association features of the base station network with the traffic time series features. The traffic association features of the base station network Each base station has traffic time series features with a dimension of 4×288. The time series features of 2000 base stations are sequentially concatenated with and the second dimension is expanded to obtain 2000 feature representations with a dimension of 2205×1×288, which are regarded as 2000 pictures with a height of 1 and a width of 288, and each picture has 2205 channels. A 1×1 convolution operation is performed to obtain a tensor representation that fuses spatio-temporal information .

[0242] The first and second dimensions are transformed to obtain , that is, the input channels remain 32, but the height becomes the number of base stations. The is simultaneously input into 5 temporal convolutional networks. Through dilated convolution, the parameters of each temporal convolutional network are shown in Table 7. The final outputs of each temporal convolutional network are added together to obtain the input of the fully connected layer . The traffic for the next moment is predicted through a fully connected layer with an input length of 8 and an output length of 1, and the mean square error between the label value and the predicted value of this group is calculated. The parameter weights of each module are updated through the gradient descent algorithm.

[0243] Table 7 Parameters of Each Temporal Convolutional Network

[0244]

[0245] The trained model is used for autoregressive iterative prediction to obtain the traffic prediction values for each base station in the next day. Some prediction results are shown in Table 8. For example, the traffic sequence on June 27, 2023 is input into the model, and the model predicts and outputs the traffic at 00:00 on June 28, 2023. Then, the traffic at 00:00 in the traffic sequence on June 27, 2023 is deleted, and the predicted traffic value at 00:00 on June 28, 2023 is added at the end as the model input, so as to obtain the traffic prediction value at 00:05 on June 28, 2023, and so on until the traffic prediction values for the next day are obtained.

[0246] Table 8 Partial Results of Base Station Traffic Prediction

[0247]

[0248] 4: Generation of Regional Energy Saving Strategies

[0249] The traffic for the next day has been predicted, and the traffic prediction interval is minutes, with a total of time segments. Table 9 gives the key base station parameters such as the maximum power and static power consumption ratio coefficient of each base station in Region 2. The base stations numbered 9 and 15 are closable base stations, and they form a multi-layer coverage structure with the base station numbered 6.

[0250] Key parameters of each base station in Area 2, Table 9

[0251]

[0252] For 200 areas, the objective function for the next day is constructed respectively. In each time period, the power of the base stations in the area can be any value under the following constraints:

[0253]

[0254]

[0255] If it is a base station that can be turned off, when the predicted number of connections is less than the turn-off threshold, the base station can be turned off and the energy consumption of the base station is reduced to 0. By solving the objective function of each area with a solver, the optimal on / off state of the base stations in each area and the optimal power of the non-turned-off base stations in each time period of the next day are obtained. The comparison of energy consumption before and after energy conservation in some areas is shown in Table 10.

[0256] Comparison of energy consumption before and after energy conservation in some areas, Table 10

[0257]

[0258] Although the preferred embodiments of the present application have been described, those skilled in the art can make additional changes and modifications once they know the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0259] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these changes and variations.

Claims

1. A base station control method integrating spatiotemporal information and traffic characteristics, characterized in that: The method comprises the following steps: S10, dividing the original base station network into a plurality of regional base station networks, and constructing a graph representation of each of the regional base station networks, and constructing a graph representation with regions as nodes by defining connection relationships between regions; In S10, constructing a graph representation of each of the regional base station networks specifically includes: S14, taking the base stations as graph nodes, and the connection relationships between the base stations in the regional base station network as edges, represents a base station network, where represents the set of base stations in the base station network, Indicates the number of base stations included. Indicates the connection relationship between base stations. If the base station and If there is a connection relationship between The matrix Line The column element value is 1, that is ,otherwise ; S15, which is obtained in step S14 Each regional base station network constructs its own graph representation and obtains ,in , express The number of base stations included in The sum of the number of base stations in each area; S16, if base stations in two different regions are connected in the original base station network, the two regions are connected to form a mesh structure as a regional association network, and the regional association network is recorded as ,in, is the set of all regions, Indicates the connection relationship between regions. No. Line Column element value Indicates the number and The number of all inter-regional connected base station pairs between regional base station networks; After step S16, the method further includes: S17, let the base station attribute feature be expressed as ,but The attribute feature matrix is ​​expressed as ,in Represents the feature length of base station attributes. The feature matrix of each regional base station network is transformed through the fully connected layer to obtain The characteristic matrix , continuous and equally spaced data recorded by the base station The flow at each moment constitutes the flow sequence At each moment, the traffic of all base stations in the area is summed up to get the overall traffic sequence of the area, and the traffic sequences of all areas are spliced ​​to get the traffic feature matrix of the regional association network ; S20, extracting hierarchical structural features of the regional base station network based on the coding tree, performing singular value decomposition on the traffic features, and adding learnable parameters to construct a directed weight matrix between base stations, and after updating the directed weight matrix to an attention matrix, capturing base station network traffic correlation features based on the attention matrix; In S20, the step of extracting the hierarchical structure features of the regional base station network based on the coding tree specifically includes: S21, a base station in a regional base station network Initialize them in sequence Independent base station division, express The number of base stations, the base stations after the base station network is initialized are divided into ,in express The subdivision of Indicates the number A regional base station network diagram is shown; S22, traverse all base station divisions, if the base station division and satisfy If the value is the largest and positive, and Merge to get new base station division ,in, To partition the merged coding tree, Indicates base station division and In the current base station network diagram The structural entropy before the merger, represents the structural entropy after merging; S23, traverse all base station divisions, if the base station division and satisfy If the value is the largest and positive, and Union, creating a new partition ,Will and Two partitions are placed , and become The sub-partition of The height of the stub does not exceed 3; S24, repeat steps S22 and S23 until there are no base stations that can be merged and combined, and obtain the final base station division set of the area , where the 0th layer of the coding tree is the root node, and the base stations are divided into sets All the partitions in are taken as the first-layer nodes. If a partition contains sub-partitions, the sub-partitions are continued to be expanded as child nodes until the base station is the leaf node, thereby obtaining the optimal coding tree corresponding to the base station network in the area. ; S25, repeat steps S22-S24 to construct the optimal coding tree of the entire regional association network , the optimal coding tree of the base station network in each region The root node is replaced by The leaf nodes of the whole base station network are obtained by ,exist The information is aggregated between layers through graph convolution. The node feature matrix of the layer Through the graph convolution, we get The node feature matrix of the layer ,from First, the base station attribute features of the 6th layer of the coding tree are concatenated to obtain , iteratively passing information upward, and finally obtaining a length of , Base station network hierarchical structure feature representation , where the expression of the graph convolution is as follows: ； In the formula, Represents the coding tree of Layer nodes and The connection relationship between layer nodes, is the activation function, For the The parameter matrix of the layer; In S20, the traffic features are subjected to singular value decomposition, and learnable parameters are added to construct a directed weight matrix between base stations. After the directed weight matrix is ​​updated to an attention matrix, the steps of capturing the base station network traffic correlation features based on the attention matrix specifically include: S26, let the flow characteristic matrix , using the equation Perform singular value decomposition on the traffic feature matrix to obtain the left singular value matrix representing the spatial distribution between base stations , and the right singular matrix of the flow rate over time , diagonal elements That is the singular value, select the first 10 singular values get , select the first ten columns Get , select the first 10 rows get ; S27, will , and Adding learnable parameters and ,pass The function constrains the range of association weights between base stations, using The function ensures the non-negativity of the association weights between base stations, thereby obtaining a learnable directed weight matrix : ； In the formula, Indicates the degree of association between base stations. The rows represent other base station nodes' The greater the weight value, the greater the correlation; S28, calculate the base station distance between any two base stations, and take 5 times the sum of the service radius of the two as the distance threshold , if the base station distance exceeds the set threshold , reset its weight to 0, and then use Function Each row vector retains its The maximum values ​​are set to 0, and the rest are set to 0, to obtain an attention matrix that does not depend on the physical connection or communication path between base stations. , The expression is as follows: ； ； In the formula, is a piecewise function, express No. OK The elements of the column, i.e. the base stations and base station The associated weight of express and The distance between S29, yes The attribute feature vector of is aggregated and updated: ； In the formula, is the feature transformation matrix, yes The attribute feature vector of for Aggregate the updated feature vectors; S210, repeat steps S26 to S29, calculate the attention matrix of each regional base station network and regional association network in turn, and obtain , ... and , and then perform feature aggregation based on the attention mechanism in turn to obtain the traffic correlation features of the regional base station network , and the traffic correlation characteristics of the regional correlation network ; In step S30, three types of natural periodic markers are embedded in the base station traffic sequence to obtain a time series feature representation of the traffic, and the time series feature representation of the traffic is fused with the base station network traffic association feature by using a convolution operation to obtain a tensor representation of fused spatiotemporal information, and a time domain convolutional network with different expansion factors is used to capture the traffic variation law of different periods in the tensor representation to obtain a base station traffic representation that integrates spatiotemporal features and different periodic features, and the base station traffic representation is input into a fully connected layer to predict the base station traffic at the next moment; In step S30, the steps of embedding three types of natural periodic markers into the base station traffic sequence to obtain a time series feature representation of the traffic, and fusing the time series feature representation of the traffic with the base station network traffic correlation feature by using a convolution operation include: S31, embedding a day mark, a week mark and a month mark into the base station traffic sequence to obtain a time series feature representation of the traffic ; S32, representing the base station network hierarchy characteristics The traffic correlation characteristics are represented by and stated After splicing, the associated characteristic representation of the base station network traffic after splicing is obtained: ； S33, will and Splice and pass Function to perform dimension expansion: ； The number of channels is , height is 1, width is The image is taken as , a 1×1 convolutional layer with 32 output channels is used for convolution operation to obtain a tensor representation of the fused spatiotemporal information : ； In step S30, a time domain convolution network with different expansion factors is used to capture the traffic change rules of different periods in the tensor representation, and a base station traffic representation that integrates spatiotemporal features and different period features is obtained. The base station traffic representation is input into a fully connected layer to predict the base station traffic at the next moment. The steps include: S34, representing the tensor of the fused spatiotemporal information Perform dimension conversion to obtain ,Will As the input of the base station traffic time series feature extraction model; S35, the base station The input sequence is input into 5 time-domain convolutional network modules of different levels, and the convolution with a stride of 1 is performed by the following formula: ； In the formula, Indicates Layer The flow at each moment is expressed as, and the convolution kernel size is , is the convolution kernel weight; get ; S36, obtain the outputs of 5 time domain convolutional network modules and perform , , , and After convolution, we get , , , and ; in, , , , , ; S37, add the convolution results according to the following formula to obtain the base station The traffic time series characteristics are expressed as : ； in, The expansion factor is The model of The end of the layer output elements; S38, represents the traffic time series characteristics Input to the fully connected layer to predict The base station traffic at time t is calculated and its predicted value is Time base station The mean square error of the real traffic is calculated, and the mean square error is used as the loss function to update the weight parameters of each module through the stochastic gradient descent algorithm. After the model training is completed, it is iteratively predicted through the autoregressive process to achieve multi-time step base station traffic prediction; S40, with the goal of minimizing the base station operation cost, determining the optimal control strategy for base station power regulation in each area according to the predicted base station traffic; The step S40 specifically includes: S41, set up base station exist The flow rate at the moment is , the energy consumption of the base station is positively correlated with the traffic, which can be expressed by the following formula under the premise of ensuring user experience Can be adjusted to the lowest power : ； ； In the formula, for The maximum power that can be provided, For base station The maximum flow that can be handled simultaneously, for Static energy consumption proportionality coefficient, yes exist Predict traffic flow at all times. express time The switch state, When it is closed, for A collection of other base stations within the coverage area; S42, predicting the future The base station traffic at each time point is minute , allowing the base station power and switch status to be adjusted once at each moment, and continuous operation after adjustment Time, the next adjustment can be made when the next moment arrives, and the future can be calculated according to the base station operation status at different time intervals The sum of the base station power in the area during the time period is calculated by the unit energy consumption cost The base station operating cost in this area can be calculated: ； ； in, Indicates indivual Internal base station The power, for The number of times the base station is started in the area during the time period, For Region The number of base stations in .

2. The base station control method integrating spatiotemporal information and traffic characteristics as claimed in claim 1, characterized in that: In S10, the step of dividing the original base station network into a plurality of regional base station networks specifically includes: S11, sorting the traffic of each base station in the original base station network, and selecting the base station with the largest traffic density as the central base station; S12, when it is determined that the circular neighborhood associated with the central base station includes at least three target base stations whose traffic is greater than a preset traffic threshold, the circular neighborhood of the central base station and the circular neighborhood associated with the target base station are divided into one of the regional base station networks; S13, repeat S11-S12 until each base station is divided into a corresponding regional base station network, and m regional base station networks are obtained.

Images