Prediction method of base station coverage range based on XGBoost regression algorithm

Abstract

The invention discloses a prediction method for base station coverage based on an XGBoost regression algorithm, and the specific steps are as follows: S1 data collection: collecting base station data and sampling point MDT data in a test area, S2 data cleaning; S3 feature engineering; S4 model training: selecting a point as a preselected point of a newly built base station, screening out base stations whose distance from the preselected point is less than a set threshold, integrating the sampling points associated with each base station, and dividing the integrated base station data and sampling point MDT data into a training set and a test set, using the XGBoost algorithm to train the data set to obtain an XGBoost model; S5 model prediction: using the trained XGBoost model to obtain the field strength of the sampling point, if the field strength is greater than the original field strength, switching the base station connected to the sampling point to the preselected point base station of the newly built base station, traversing all sampling points, and presenting all sampling points cut into the newly built base station on a map.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning, and in particular relates to a method for predicting base station coverage based on an XGBoost regression algorithm. Technical Background

[0002] In recent years, as my country's communications market has become saturated, operators have faced the problem of difficulty in increasing users. Competition has shifted from the incremental market to the stock market. The maturity of 5G technology has also become a new competitive point for operators, and all operators have actively deployed 5G base stations. Since the band used by 5G is shorter and the coverage is smaller than that of 4G, in order to achieve full coverage of the 5G network, multiple 5G base stations need to be added. The construction cost of 5G base stations is high, and it is inevitable that they cannot be built on a large scale. The primary goal of 5G base station site selection is the area where users gather, such as shopping malls, factories, and transportation hubs. Geographical location will also have a great impact on the coverage of 5G base stations. Therefore, the site selection of 5G base stations and whether the site selection and construction can meet the user's Internet needs. This has become an urgent problem for operators to solve. After the construction is completed, it is also necessary to consider the user access problem and whether the quality of the access network can meet normal needs. In this case, we need to make a judgment on the newly built base station to determine whether the coverage can achieve the expected effect. Usually, manual analysis is used to determine the location of the final base station. This method is not only inefficient and costly, but also the experience value is not accurate enough, and the final result will also be biased.

[0003] Chinese patent document (application number: 201911308145.7) discloses a carrier shutoff method and system based on time series prediction; the traffic prediction method includes: obtaining historical data of the target base station cell, extracting features based on time series through associated fields; activating the multi-model fusion algorithm, selecting multiple models as the first layer of the stack according to the extracted features for multi-fold cross-validation, and merging the output results; based on the prediction algorithm model of the second layer of the specified stack, retraining the output results after multi-fold cross-validation to obtain the target prediction value and accuracy index. The carrier shutoff method is based on the traffic prediction step, using the prediction data of a certain period of time to make a specific cell base station perform carrier shutoff or station closure operation at the predicted time node. The system in the scheme of the present invention includes a computing device and a module for implementing the above method. However, the method and system mainly predict the traffic prediction of existing base stations in different periods of time, and are applied in the field of base station energy saving; it does not use the data of existing base stations to predict the coverage of base stations, nor is it applied in the scenario of base station coverage construction planning; at the same time, the algorithm used in the technical solution is a time series prediction algorithm, which predicts the fluctuation of traffic within a day, and the algorithms used in it, such as XGBoost, are used for data processing, not for prediction.

[0004] The Chinese patent document (application number: 202110862707.3) discloses a cellular network fault diagnosis method, comprising the following steps: step 1, determining a network fault data set; step 2, obtaining a network fault data set after dimensionality reduction; step 3, representing the network fault data set after dimensionality reduction in step 2 in the form of a feature matrix; representing the label information of the network fault data set after dimensionality reduction in step 2 in the form of a label matrix; converting the introduced weight matrix into an adjacency matrix with only 0 and 1 matrix elements; step 4, fault diagnosis based on graph convolutional neural network. The XGBoost regression algorithm used in this new cellular network fault diagnosis method only uses the function of ranking the feature importance, while the core of realizing network diagnosis is realized by convolutional neural network.

[0005] Chinese patent document (application number: 201911363066.6) discloses a method for multi-access edge computing offloading and frame time slot resource allocation in the Internet of Vehicles environment. The application scenario of this method is a single base station with an edge server deployed and all high-speed moving vehicle nodes within its coverage. The steps are as follows: model the multi-node time slot resource allocation and computing offloading problem as a mixed integer programming problem, and the system objective function is to maximize the sum of the throughput of all nodes within the coverage of the base station; perform channel preprocessing on the wireless time-varying channels generated by the vehicle moving nodes; use deep reinforcement learning to perceive the preprocessing channels of all vehicle nodes to obtain a set of candidate offloading vectors; each set of offloading vectors obtained can convert the system objective function into a convex function, and use the Lagrange multiplier method to solve the time slot allocation strategy, and finally select the maximum system throughput; establish an ordered offloading queue and unload in order. This technical solution is applied to the Internet of Vehicles. The main algorithm used is to solve the problem of maximizing the throughput in the multi-access edge computing system under low time complexity. Therefore, this method is to solve the problem of real-time performance. The algorithms used are all to solve the problems of real-time performance and throughput, not to solve the problem of base station coverage prediction.

[0006] The Chinese patent document (application number: 202010892797.6) discloses a network planning method based on big data mining. The method first converts actual big data into useful information through data mining technology, conducts in-depth analysis of the collected cellular network data, discovers and extracts the information and rules hidden in it, and the acquired information will serve as the basis for base station deployment; then, based on the particle swarm algorithm, the requirements of weak coverage and the minimum number of base stations are met, and the network planning model in the cellular network is established by using python software for simulation to obtain the location and number of base stations; the network performance can be optimized and the cost of operators can be greatly reduced. This technical solution uses the particle swarm algorithm for network planning, and the problem it solves is the weak coverage problem of existing base stations. It can only be applied in the scenario of existing base stations, not the problem of planning new base stations, and the data it uses contains hourly granularity.

[0007] Therefore, the problem that the site selection of base stations can only rely on manual experience at this stage needs to be solved urgently. Summary of the invention

[0008] The purpose of the present invention is to provide a prediction method for base station coverage based on the XGBoost regression algorithm, which can predict the coverage of a new base station and the network quality under the coverage without the help of manual experience; at the same time, the network quality within the predicted coverage is analyzed to ensure normal use, so that 5G planning is more reasonable, more load problems can be solved, and users can obtain a better 5G user experience, while saving 5G construction costs.

[0009] In order to solve the above technical problems, the technical solution adopted by the present invention is: the prediction method of base station coverage based on XGBoost regression algorithm specifically includes the following steps:

[0010] S1 Data Collection: Collect base station data and sampling point MDT data in the test area to obtain full data in the area;

[0011] S2 data cleaning: perform data cleaning on all data in the test area;

[0012] S3 feature engineering: perform feature engineering on the raw data cleaned in step S2;

[0013] S4 model training: Select a point as the pre-selected point for the new base station, screen out base stations whose distance from the pre-selected point is less than the set threshold, integrate the sampling points associated with each base station, and divide the integrated base station data and sampling point MDT data into training sets and test sets. Use the XGBoost algorithm to train the data set to obtain the XGBoost model.

[0014] S5 model prediction: Use the trained XGBoost model to obtain the field strength of the sampling point. If the field strength is greater than the original field strength, the base station connected to the sampling point is switched to the pre-selected base station of the new base station. All sampling points are traversed and all sampling points cut into the new base station are presented on the map, thereby obtaining the predicted range of the new base station and the coverage rate of the base station.

[0015] By adopting the above technical solution, the height, azimuth, downtilt angle, etc. of the base station are used, and the modeling method based on machine learning is used to obtain the associated data of the base station and the sampling point from the MDT data of the sampling point as features. Feature processing and feature analysis are performed on a large amount of original data to train a prediction model, and the coverage range of the base station is predicted to help the planning department to reasonably plan the base station construction plan and achieve cost reduction and efficiency improvement. At the same time, the network quality within the predicted coverage range is analyzed to ensure that it can be used normally, making the 5G planning more reasonable, solving more load problems, and enabling users to obtain a better 5G user experience while saving 5G construction costs.

[0016] As a preferred technical solution of the present invention, the sampling point MDT data in step S1 includes cellid, longitude, latitude and field strength; the base station data includes base station name, nci, longitude, latitude, base station height, base station azimuth and base station downtilt, where cellid and nci are both base station IDs, which are the same data and are used to associate base station data with MDT data during training. The coverage rate of the base station is determined by the field strength, and the field strength of the sampling point accessing the base station is less than -100dm as the boundary; if the field strength of the sampling point is less than -100dm, it is considered that the base station cannot cover the sampling point, and the coverage rate of the base station is calculated by the ratio of the field strength of the sampling point accessed by the base station greater than -100dm to the total field strength accessed by the base station; therefore, the collected data includes base station data, such as the number of the base station, the longitude and latitude of the base station, the type of base station equipment, the height of the base station, the azimuth of the base station (the angle with the true north direction), the mechanical downtilt, and the electronic downtilt; the sampling point data are the longitude and latitude of the sampling point, the base station associated with the sampling point, and the field strength of the sampling point.

[0017] As a preferred technical solution of the present invention, the data cleaning described in step S2 includes missing value processing, erroneous data deletion, string feature conversion and relative deviation. The relative deviation is calculated based on the latitude and longitude information of the sampling point and the base station and the base station azimuth; the original data needs to be cleaned before use, and some of the original data has erroneous values ​​and missing values. Before training, this part of the data needs to be removed; first, the missing values ​​are processed, and the data with missing values ​​are first screened out from the full amount of data, and the data is filled with mean filling and mode filling. If the missing value is greater than 15%, the filling of the data will affect the authenticity of the data; therefore, this part of the data will be deleted. The erroneous values ​​in the data will also affect the subsequent training. Therefore, it is necessary to delete the erroneous values, for example, if the field strength value in the data is negative, if the field strength in the data is found to be positive, it is directly deleted. XGBoost does not receive non-numeric data, so it is necessary to convert string data into numeric data. For example, the model of the base station equipment needs to be converted into a numeric type through feature conversion. For example, convert the three different models in this article into three categories: 0, 1, and 2.

[0018] As a preferred technical solution of the present invention, the feature engineering process in step S3 uses feature dimension expansion, wherein the feature dimension expansion is based on the original features in the collected user information, including calculating the distance between the base station and the sampling point and the relative azimuth between the base station and the sampling point; the specific steps are:

[0019] S31: According to the longitude and latitude of the base station and the longitude and latitude of the user, the distance between the base station and the sampling point is calculated using the spherical distance formula; the formula is:

[0020] D = R × arccos[cosβ1 cosβ2 cos(α1-α2)+sinβ1sinβ2]; where D is the straight-line distance between the user and the base station; α1 and β1 are the longitude and latitude of the base station respectively; α2 and β2 are the longitude and latitude of the user respectively; R is the radius of the earth;

[0021] S32: The azimuth of the base station is based on the north direction and increases clockwise. The value range is 0-360°. The azimuth of the sampling point relative to the base station is calculated using the latitude and longitude of the sampling point and the base station. The formula is:

[0022]

[0023] Where angle is the azimuth of the sampling point relative to the base station; lon and lat are the longitude and latitude of the sampling point respectively; lon cell and lat cell are the longitude and latitude of the base station respectively. Each base station is usually composed of 3 cells.

[0024] Each cell covers an area of ​​180 degrees and is spread out in a fan-shaped manner. The closer to the center of the fan, the stronger the signal is, and the actual signal strength is similar to an ellipse. Therefore, if you want to ensure the accuracy of the prediction, in addition to predicting the distance between the sampling point and the cell, the relative azimuth between the two is also very important. The accuracy of the relative azimuth will have a great impact on the accuracy of the predicted coverage, so the relative azimuth needs to be accurately calculated. The azimuth can be understood as the angle experienced by the plane in the north direction rotating clockwise to coincide with the plane where the antenna is located. Therefore, to find the relative azimuth between the cell and the sampling point, you only need to find the angle between the sampling point and the cell and the north direction, and subtract the azimuth of the cell to obtain the relative azimuth.

[0025] As a preferred technical solution of the present invention, if the relative azimuth obtained in step S32 is a negative value, indicating that the sampling point is on the right side of the cell azimuth, the negative value is converted into a positive value, thereby obtaining the relative azimuth of the base station relative to the sampling point. The coverage of the base station is based on the antenna main lobe azimuth as the central axis, with a deflection of 60 degrees to the left and right. Therefore, no matter whether the calculated relative azimuth is a positive value or a negative value, as long as the absolute value of the relative azimuth is less than 60 degrees, the sampling point is considered to be under the coverage of the base station.

[0026] As a preferred technical solution of the present invention, the specific steps of step S4 are:

[0027] S41 selects base stations close to the pre-selected point: selects a point on the map as the pre-selected point for the new base station, extracts the longitude and latitude values ​​of the pre-selected point, takes the longitude and longitude of the selected pre-selected point as the center point, calculates the distance between each base station and the pre-selected point, selects base stations whose distance is less than a set threshold, evaluates the geographical environment of the area based on the selected base stations, and uses the average coverage rate of the base stations as the coverage rate of the new base station;

[0028] S42 associates the base station and the cell: the sampling points associated with each screened base station are integrated, and the base station coverage rate of the geographical location and the distance and azimuth between the base station and the sampling point are obtained through the association between the base station and the sampling point;

[0029] S43 Data Fitting: Integrate base station data and sampling point MDT data, associate base station data and sampling point MDT data into a table through base station ID, establish a mapping relationship between the distance, height, downtilt angle, relative azimuth angle, base station type, base station power and sampling point field strength; then use the XGBoost algorithm to fit the relationship between the field strength of the sampling point and the distance, height, downtilt angle, and relative azimuth angle of the base station, and establish a regression model;

[0030] S44 trains the XGBoost model: the integrated data is divided into a training set and a test set in a ratio of 8:2, the XGBoost algorithm is used to train the data set, the feature importance is sorted, the feature importance is obtained, and the features with high importance weights are selected. After multiple rounds of iterative training, the trained model is applied to the validation set, and the results are tested on the validation set to determine whether the trained model is usable; if not, the model parameters are adjusted, and after multiple rounds of iterative training, a usable XGBoost model is finally fitted.

[0031] As a preferred technical solution of the present invention, the formula for obtaining the XGBoost model by training with the XGBoost algorithm in step S4 is:

[0032]

[0033]

[0034]

[0035] …

[0036]

[0037] Among them, x i represents sample i; Represents the prediction result of sample i after the tth iteration; represents the prediction results of the first t-1 trees; f t (x i ) represents the function of the t-th tree.

[0038] XGBoost continuously splits features to grow a tree. Each round of learning a tree is actually to fit the residual between the predicted value and the actual value of the previous round of model. Model optimization is to make the residual smaller and the model more accurate through continuous training. The XGBoost model is a type of gradient boosting tree model. It also generates models serially and takes the sum of all models as output. XGBoost performs a second-order Taylor expansion on the loss function, optimizes the loss function using the second-order derivative information of the loss function, and greedily chooses whether to split the node based on whether the loss function is reduced. At the same time, the XGBoost algorithm adds regularization, learning rate, column sampling, and approximate optimal split points to prevent overfitting.

[0039] As a preferred technical solution of the present invention, the specific steps of step S5 are:

[0040] S51: Calculate the distance between the pre-selected point of the base station and all the sampling points by using the latitude and longitude of the base station and the latitude and longitude of the sampling point, sort the points according to the distances obtained, select the data with a distance less than a threshold as a test set, and calculate the characteristic data of the data in the test set;

[0041] S52: Use the XGBoost model trained in step S4 to obtain the field strength of the sampling point. If the obtained field strength is greater than the original field strength, switch the base station connected to the sampling point to the pre-selected base station of the newly-built base station.

[0042] S53: Mark all sampling points that are connected to the new base station on the map, that is, obtain the predicted range of the new base station and the coverage rate of the base station, and then compare the field strength of the sampling points before and after the new base station is built to obtain the display result.

[0043] Compared with the prior art, the present invention has the following beneficial effects: using the XGBoost regression algorithm to train a coverage prediction model based on the MDT data of existing base stations and base station sampling points, so as to determine whether the construction of 5G base stations in this area can solve the network coverage problem in the current area through the predicted coverage range and coverage rate; solving the problem that the site selection of base stations can only rely on manual experience, the coverage prediction algorithm can predict the location of base station construction, without relying on experience judgment, and select the most suitable site for operators to reduce network signal problems and construction cost problems caused by site selection. It also helps telecom operators improve 5G services, enhance users' Internet experience, and improve user satisfaction, and helps telecom operators reduce costs and increase efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of a method for predicting base station coverage based on XGBoost regression algorithm of the present invention;

[0045] Figure 2 A schematic diagram of MDT data samples at sampling points of a method for predicting base station coverage based on an XGBoost regression algorithm of the present invention;

[0046] Figure 3 A schematic diagram of a base station data sample of a base station coverage prediction method based on an XGBoost regression algorithm of the present invention;

[0047] Figure 4 An example diagram of the base station azimuth angle of the base station coverage prediction method based on the XGBoost regression algorithm of the present invention;

[0048] Figure 5 A schematic diagram of the relative azimuth angle of the base station coverage prediction method based on the XGBoost regression algorithm of the present invention;

[0049] Figure 6 A schematic diagram of base station signal strength of a method for predicting base station coverage based on an XGBoost regression algorithm of the present invention;

[0050] Figure 7 A base station coverage map before prediction of the base station coverage range prediction method based on the XGBoost regression algorithm of the present invention;

[0051] Figure 8 It is a base station coverage map after prediction by the base station coverage prediction method based on XGBoost regression algorithm of the present invention. DETAILED DESCRIPTION

[0052] Example: Figure 1 As shown, the prediction method of base station coverage based on XGBoost regression algorithm specifically includes the following steps:

[0053] S1 Data Collection: Collect base station data and sampling point MDT data in the test area to obtain full data in the area;

[0054] The sampling point MDT data in step S1 includes cellid, longitude, latitude and field strength; the base station data includes base station name, nci, longitude, latitude, base station height, base station azimuth and base station downtilt, wherein cellid and nci are both base station IDs, which are the same data and are used to associate base station data with MDT data during training; the coverage rate of the base station is determined by the field strength, and the field strength of the sampling point accessing the base station is less than -100dm as the boundary; if the field strength of the sampling point is less than -100dm, it is considered that the base station cannot cover the sampling point, and the coverage rate of the base station is calculated by the ratio of the field strength of the sampling point accessed by the base station greater than -100dm to the total field strength accessed by the base station; therefore, the collected data includes base station data, such as the number of the base station, the longitude and latitude of the base station, the type of base station equipment, the height of the base station, the azimuth of the base station (the angle with the true north direction), the mechanical downtilt, and the electronic downtilt; the sampling point data are the longitude and latitude of the sampling point, the base station associated with the sampling point, and the field strength of the sampling point;

[0055] S2 Data Cleaning: Data cleaning is performed on the entire data of the test area; the data cleaning in step S2 includes missing value processing, erroneous data deletion, string feature conversion and relative deviation; wherein the relative deviation is calculated based on the latitude and longitude information of the sampling point and the base station and the azimuth of the base station; the original data needs to be cleaned before use, and some of the original data has erroneous values ​​and missing values. Before training, this part of the data needs to be removed; first, the missing values ​​are processed, and the data with missing values ​​are first screened out from the entire data, and the data is filled with mean filling and mode filling. If the missing value is greater than 15%, the filling of the data will affect the authenticity of the data; therefore, this part of the data will be deleted; the erroneous values ​​in the data will also affect the subsequent training; therefore, the erroneous values ​​need to be deleted, for example, the field strength value in the data is a negative value, if the field strength in the data is found to be a positive value, it is directly deleted; XGBoost does not receive non-numeric data, so the string type data needs to be converted into numerical type data; for example, the model of the base station equipment needs to be converted into a numerical type through feature conversion. For example, convert the three different models in this article into three categories: 0, 1, and 2;

[0056] S3 feature engineering: perform feature engineering on the raw data cleaned in step S2;

[0057] The feature engineering process in step S3 uses feature dimension expansion, wherein the feature dimension expansion is performed based on the original features in the collected user information, including calculating the distance between the base station and the sampling point and the relative azimuth between the base station and the sampling point; the specific steps are:

[0058] S31: According to the longitude and latitude of the base station and the longitude and latitude of the user, the distance between the base station and the sampling point is calculated using the spherical distance formula; the formula is:

[0059] D = R × arccos[cosβ1 cosβ2 cos(α1-α2)+sinβ1 sinβ2]; where D is the straight-line distance between the user and the base station; α1 and β1 are the longitude and latitude of the base station; α2 and β2 are the longitude and latitude of the user; R is the radius of the earth;

[0060] S32: The azimuth of the base station is based on the north direction and increases clockwise. The value range is 0-360°. The azimuth of the sampling point relative to the base station is calculated using the latitude and longitude of the sampling point and the base station. The formula is:

[0061]

[0062] Where angle is the azimuth of the sampling point relative to the base station; lon and lat are the longitude and latitude of the sampling point respectively; lon celland lat cell are the longitude and latitude of the base station respectively; each base station is usually composed of 3 cells; each cell covers an area of ​​180 degrees and is spread out in a fan-shaped manner; the closer to the center of the fan, the stronger the signal is, and the actual signal strength is similar to an ellipse. Therefore, if the accuracy of the prediction is to be guaranteed, in addition to predicting the distance between the sampling point and the cell, the relative azimuth between the two is also very important. The accuracy of the relative azimuth will have a great impact on the accuracy of the predicted coverage, so it is necessary to accurately calculate the relative azimuth. The azimuth can be understood as the angle experienced by the plane in the north direction rotating clockwise to coincide with the plane where the antenna is located. Therefore, to obtain the relative azimuth between the cell and the sampling point, it is only necessary to obtain the angle between the sampling point and the cell and the north direction, and subtract the azimuth of the cell to obtain the relative azimuth; if the relative azimuth obtained in step S32 is a negative value, indicating that the sampling point is on the right side of the cell azimuth, the negative value is converted to a positive value, thereby obtaining the relative azimuth of the base station relative to the sampling point. The coverage range of the base station is based on the antenna main lobe azimuth as the central axis, with a deflection of 60 degrees to the left and right. Therefore, no matter whether the calculated relative azimuth is positive or negative, as long as the absolute value of the relative azimuth is less than 60 degrees, the sampling point is considered to be within the coverage range of the base station.

[0063] S4 model training: select a point as the pre-selected point of the new base station, screen out base stations whose distance from the pre-selected point is less than the set threshold, integrate the sampling points associated with each base station, and divide the integrated base station data and sampling point MDT data into training set and test set, use XGBoost algorithm to train the data set, and obtain the XGBoost model; the specific steps of step S4 are:

[0064] S41 selects base stations close to the pre-selected point: selects a point on the map as the pre-selected point for the new base station, extracts the longitude and latitude values ​​of the pre-selected point, takes the longitude and longitude of the selected pre-selected point as the center point, calculates the distance between each base station and the pre-selected point, selects base stations whose distance is less than a set threshold, evaluates the geographical environment of the area based on the selected base stations, and uses the average coverage rate of the base stations as the coverage rate of the new base station;

[0065] S42 associates the base station and the cell: the sampling points associated with each screened base station are integrated, and the base station coverage rate of the geographical location and the distance and azimuth between the base station and the sampling point are obtained through the association between the base station and the sampling point;

[0066] S43 Data Fitting: Integrate base station data and sampling point MDT data, associate base station data and sampling point MDT data into a table through base station ID, establish a mapping relationship between the distance, height, downtilt angle, relative azimuth angle, base station type, base station power and sampling point field strength; then use the XGBoost algorithm to fit the relationship between the field strength of the sampling point and the distance, height, downtilt angle, and relative azimuth angle of the base station, and establish a regression model;

[0067] S44 trains the XGBoost model: the integrated data is divided into a training set and a test set in a ratio of 8:2, the XGBoost algorithm is used to train the data set, the importance of the features is sorted, the importance of the features is obtained, and the features with high importance weights are selected. After multiple rounds of iterative training, the trained model is applied to the validation set, and the results of the validation set are used to determine whether the trained model is usable; if not, the parameters of the model are adjusted, and multiple rounds of iterative training are performed to finally fit a usable XGBoost model;

[0068] The formula for obtaining the XGBoost model by training with the XGBoost algorithm in step S4 is:

[0069]

[0070]

[0071] …

[0072]

[0073] Among them, x i represents sample i; Represents the prediction result of sample i after the tth iteration; represents the prediction results of the first t-1 trees; f t (x i ) represents the function of the tth tree; XGBoost continuously performs feature splitting to grow a tree. Each round of learning a tree is actually to fit the residual between the predicted value and the actual value of the previous round of model. The optimization of the model is to make the residual smaller and the model more accurate through continuous training; the XGBoost model is a kind of gradient boosting tree model, which also generates models serially and takes the sum of all models as output; XGBoost performs a second-order Taylor expansion on the loss function, optimizes the loss function using the second-order derivative information of the loss function, and greedily chooses whether to split the node based on whether the loss function is reduced; at the same time, the XGBoost algorithm adds regularization, learning rate, column sampling, approximate optimal split point and other means to prevent overfitting. The technical solution has also made certain optimizations in dealing with missing values;

[0074] S5 model prediction: Use the trained XGBoost model to obtain the field strength of the sampling point. If the field strength is greater than the original field strength, switch the base station connected to the sampling point to the pre-selected base station of the new base station, traverse all sampling points, and present all sampling points cut into the new base station on the map, so as to obtain the predicted range of the new base station and the coverage rate of the base station; the specific steps of step S5 are:

[0075] S51: Calculate the distance between the pre-selected point of the base station and all the sampling points by using the latitude and longitude of the base station and the latitude and longitude of the sampling point, sort the points according to the distances obtained, select the data with a distance less than a threshold as a test set, and calculate the characteristic data of the data in the test set;

[0076] S52: Use the XGBoost model trained in step S4 to obtain the field strength of the sampling point. If the obtained field strength is greater than the original field strength, switch the base station connected to the sampling point to the pre-selected base station of the newly-built base station.

[0077] S53: Mark all sampling points that are connected to the new base station on the map, that is, obtain the predicted range of the new base station and the coverage rate of the base station, and then compare the field strength of the sampling points before and after the new base station is built to obtain the display result.

[0078] Specific implementation example: The data comes from base station data and sampling point data collected by a local mobile communication company in June 2022; the prediction method of base station coverage based on the XGBoost regression algorithm specifically includes the following steps:

[0079] S1 Data Collection: Collect base station data and sampling point MDT data in the test area to obtain full data in the area;

[0080] The sampling point MDT data in step S1 includes cellid, longitude, latitude and field strength; the base station data includes base station name, nci, longitude, latitude, base station height, base station azimuth and base station downtilt, where cellid and nci are both base station IDs, which are the same data and are used to associate base station data with MDT data during training; the data samples of the sampling point MDT data are as follows Figure 2 As shown, the data dimensions are cellid, longitude1, latitude1, and scrsrp; they are base station ID, longitude, latitude, and field strength respectively; the sample of base station data is as follows Figure 3As shown in the figure, the data dimensions are cell_name, nci, longitude, latitude, ant_height, ant_azimuth, cell_all_pitching, which respectively represent the base station name, base station ID, longitude, latitude, base station height, base station azimuth, and base station downtilt angle; preliminary statistical analysis shows that there are 68 base stations in the test area, with a total of 21,246 sampling points. According to the MDT sampling point data, the base station with the most sampling points has 1,930 sampling points, and the base station with the least sampling points has only 12 sampling points. The sampling points of most base stations are in the range of 100-400.

[0081] S2 Data Cleaning: Data cleaning is performed on the entire data of the test area; the data cleaning in step S2 includes missing value processing, erroneous data deletion, string feature conversion and relative deviation; wherein the relative deviation is calculated based on the latitude and longitude information of the sampling point and the base station and the azimuth of the base station; the original data needs to be cleaned before use, and some of the original data has erroneous values ​​and missing values. Before training, this part of the data needs to be removed; first, the missing values ​​are processed, and the data with missing values ​​are first screened out from the entire data, and the data is filled with mean filling and mode filling. If the missing value is greater than 15%, the filling of the data will affect the authenticity of the data; therefore, this part of the data will be deleted; the erroneous values ​​in the data will also affect the subsequent training; therefore, the erroneous values ​​need to be deleted, for example, the field strength value in the data is a negative value, if the field strength in the data is found to be a positive value, it is directly deleted; XGBoost does not receive non-numeric data, so the string type data needs to be converted into numerical type data; for example, the model of the base station equipment needs to be converted into a numerical type through feature conversion. For example, convert the three different models in this article into three categories: 0, 1, and 2;

[0082] S3 feature engineering: performing feature engineering processing on the raw data cleaned in step S2; the feature engineering processing in step S3 uses feature dimension expansion, wherein the feature dimension expansion is performed based on the original features in the collected user information, including calculating the distance between the base station and the sampling point and the relative azimuth between the base station and the sampling point; the specific steps are:

[0083] S31: According to the longitude and latitude of the base station and the longitude and latitude of the user, the distance between the base station and the sampling point is calculated using the spherical distance formula; the formula is:

[0084] D = R × arccos[cosβ1 cosβ2 cos(α1-α2)+sinβ1 sinβ2]; where D is the straight-line distance between the user and the base station; α1 and β1 are the longitude and latitude of the base station; α2 and β2 are the longitude and latitude of the user; R is the radius of the earth;

[0085] S32: Each base station is usually composed of three cells; each cell covers an area of ​​180 degrees, spreading out in a fan-shaped manner; the closer to the center of the fan, the stronger the signal, and the actual signal strength is similar to an ellipse. Therefore, if you want to ensure the accuracy of the prediction, in addition to predicting the distance between the sampling point and the cell, the relative azimuth between the two is also very important; the accuracy of the relative azimuth will have a great impact on the accuracy of the predicted coverage, so it is necessary to accurately calculate the relative azimuth; such as Figure 4 As shown, the azimuth can be understood as the angle experienced by rotating the plane in the north direction clockwise to coincide with the plane where the antenna is located; therefore, to obtain the relative azimuth between the cell and the sampling point, it is only necessary to obtain the angle between the sampling point and the cell and the north direction, and subtract the azimuth of the cell to obtain the relative azimuth; the azimuth of the base station takes the north direction as the reference direction, increases clockwise, and the value range is 0-360°; the azimuth of the sampling point relative to the base station is calculated using the latitude and longitude of the sampling point and the base station; the formula is:

[0086]

[0087] Where angle is the azimuth of the sampling point relative to the base station; lon and lat are the longitude and latitude of the sampling point respectively; lon cell and lat cell are the longitude and latitude of the base station respectively; if the relative azimuth obtained in step S32 is a negative value, it means that the sampling point is on the right side of the cell azimuth, then the negative value is converted to a positive value, thereby obtaining the relative azimuth of the base station relative to the sampling point. The coverage of the base station is based on the antenna main lobe azimuth as the central axis, with a deflection of 60 degrees to the left and right. Therefore, no matter whether the calculated relative azimuth is a positive value or a negative value, as long as the absolute value of the relative azimuth is less than 60 degrees, it is considered that the sampling point is under the coverage of the base station; the schematic diagram of the relative azimuth is as follows Figure 5 As shown;

[0088] S4 model training: select a point as the pre-selected point of the new base station, screen out base stations whose distance from the pre-selected point is less than the set threshold, integrate the sampling points associated with each base station, and divide the integrated base station data and sampling point MDT data into training set and test set, use XGBoost algorithm to train the data set, and obtain the XGBoost model; the specific steps of step S4 are:

[0089] S41 selects base stations close to the pre-selected point: selects a point on the map as the pre-selected point for the new base station, extracts the longitude and latitude values ​​of the pre-selected point, takes the longitude and longitude of the selected pre-selected point as the center point, calculates the distance between each base station and the pre-selected point, selects base stations whose distance is less than a set threshold, evaluates the geographical environment of the area based on the selected base stations, and uses the average coverage rate of the base stations as the coverage rate of the new base station;

[0090] S42 Associate base stations and cells: Integrate the sampling points associated with each screened base station, and through the association between the base station and the sampling point, obtain the base station coverage rate of the geographical location and the distance and azimuth between the base station and the sampling point. S43 Data fitting: Integrate the base station data and the sampling point MDT data, associate the base station data and the sampling point MDT data into a table through the base station ID, and establish a mapping relationship between the distance, height, downtilt angle, relative azimuth angle, base station type, base station power and sampling point field strength; then use the XGBoost algorithm to fit the relationship between the field strength of the sampling point and the distance, height, downtilt angle, and relative azimuth angle of the base station, and establish a regression model;

[0091] S44 trains the XGBoost model: the integrated data is divided into a training set and a test set in a ratio of 8:2, the XGBoost algorithm is used to train the data set, the importance of the features is sorted, the importance of the features is obtained, and the features with high importance weights are selected. After multiple rounds of iterative training, the trained model is applied to the validation set, and the results of the validation set are used to determine whether the trained model is usable; if not, the parameters of the model are adjusted, and multiple rounds of iterative training are performed to finally fit a usable XGBoost model;

[0092] The formula for obtaining the XGBoost model by training with the XGBoost algorithm in step S4 is:

[0093]

[0094]

[0095]

[0096] …

[0097]

[0098] Among them, x i represents sample i; Represents the prediction result of sample i after the tth iteration; represents the prediction results of the first t-1 trees; f t (x i) represents the function of the tth tree; XGBoost continuously splits features to grow a tree. Each round of learning a tree is actually to fit the residual between the predicted value and the actual value of the previous round of model; the optimization of the model is to make the residual smaller through continuous training, and the accuracy of the model is higher; the XGBoost model is a kind of gradient boosting tree model, which also generates models serially and takes the sum of all models as output; XGBoost performs a second-order Taylor expansion on the loss function, optimizes the loss function using the second-order derivative information of the loss function, and greedily chooses whether to split the node according to whether the loss function is reduced; at the same time, the XGBoost algorithm adds regularization, learning rate, column sampling, approximate optimal split point and other means to prevent overfitting. The technical solution has also made certain optimizations in dealing with missing values;

[0099] S5 model prediction: Use the trained XGBoost model to obtain the field strength of the sampling point. If the field strength is greater than the original field strength, switch the base station connected to the sampling point to the pre-selected base station of the new base station, traverse all sampling points, and present all sampling points cut into the new base station on the map, so as to obtain the predicted range of the new base station and the coverage rate of the base station; the specific steps of step S5 are:

[0100] S51: Calculate the distance between the pre-selected point of the base station and all the sampling points by using the latitude and longitude of the base station and the latitude and longitude of the sampling point, sort the points according to the distances obtained, select the data with a distance less than a threshold as a test set, and calculate the characteristic data of the data in the test set;

[0101] S52: Use the XGBoost model trained in step S4 to obtain the field strength of the sampling point. If the obtained field strength is greater than the original field strength, switch the base station connected to the sampling point to the pre-selected base station of the new base station. S53: Mark all the sampling points that are connected to the new base station on the map, that is, obtain the predicted range of the new base station and the coverage rate of the base station, and then compare the field strength of the sampling points before and after the new base station to obtain the display results, such as Figures 6 to 8 As shown; Figure 7 and Figure 8 The boxed area shows the comparison effect before and after the new base station planning. Figure 7 In the figure, the color of the sampling point in the box is darker, which means the field strength of the sampling point is between -110 and -100db, and the signal is poor. Figure 8 After the new base station is planned, the color of the sampling point in the box is lighter, which means that the field strength of the sampling point is between -80 and -70 db, and the signal is strong. Therefore, it can be proved that our coverage prediction method can solve the actual coverage prediction problem.

[0102] The specific embodiments described in the present invention are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications, supplements, explanations, or substitutes to the specific embodiments described, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A method for predicting base station coverage based on XGBoost regression algorithm, characterized in that: The specific steps include: S1 Data Collection: Collect base station data and sampling point MDT data in the test area to obtain full data in the area; S2 data cleaning: perform data cleaning on all data in the test area; S3 feature engineering: perform feature engineering on the raw data cleaned in step S2; S4 model training: Select a point as the pre-selected point for the new base station, screen out base stations whose distance from the pre-selected point is less than the set threshold, integrate the sampling points associated with each base station, and divide the integrated base station data and sampling point MDT data into training sets and test sets. Use the XGBoost algorithm to train the data set to obtain the XGBoost model. S5 model prediction: Use the trained XGBoost model to obtain the field strength of the sampling point. If the field strength is greater than the original field strength, the base station connected to the sampling point is switched to the pre-selected base station of the new base station. All sampling points are traversed and all sampling points connected to the new base station are presented on the map, thereby obtaining the predicted range of the new base station and the coverage rate of the base station. The feature engineering process in step S3 uses feature dimension expansion, wherein the feature dimension expansion is performed based on the original features in the collected user information, including calculating the distance between the base station and the sampling point and the relative azimuth between the base station and the sampling point; the specific steps are: S31: According to the longitude and latitude of the base station and the longitude and latitude of the user, the distance between the base station and the sampling point is calculated using the spherical distance formula; the formula is: D = R × arccos[cosβ1cosβ2cos(α1-α2)+sinβ1sinβ2]; where D is the straight-line distance between the user and the base station; α1 and β1 are the longitude and latitude of the base station respectively; α2 and β2 are the longitude and latitude of the user respectively; R is the radius of the earth; S32: The azimuth of the base station is based on the north direction and increases clockwise. The value range is 0-360°. The azimuth of the sampling point relative to the base station is calculated using the latitude and longitude of the sampling point and the base station. The formula is: Where angle is the azimuth of the sampling point relative to the base station; lon and lat are the longitude and latitude of the sampling point respectively; lon cell and lat cell are the longitude and latitude of the base station respectively; The specific steps of step S4 are: S41 selects base stations close to the pre-selected point: selects a point on the map as the pre-selected point for the new base station, extracts the longitude and latitude values ​​of the pre-selected point, takes the longitude and longitude of the selected pre-selected point as the center point, calculates the distance between each base station and the pre-selected point, selects base stations whose distance is less than a set threshold, evaluates the geographical environment of the area based on the selected base stations, and uses the average coverage rate of the base stations as the coverage rate of the new base station; S42 associates the base station and the cell: the sampling points associated with each screened base station are integrated, and the base station coverage rate of the geographical location and the distance and azimuth between the base station and the sampling point are obtained through the association between the base station and the sampling point; S43 Data Fitting: Integrate base station data and sampling point MDT data, associate base station data and sampling point MDT data into a table through base station ID, establish a mapping relationship between the distance, height, downtilt angle, relative azimuth angle, base station type, base station power and sampling point field strength; then use the XGBoost algorithm to fit the relationship between the field strength of the sampling point and the distance, height, downtilt angle, and relative azimuth angle of the base station, and establish a regression model; S44 trains the XGBoost model: the integrated data is divided into a training set and a test set in a ratio of 8:2, the XGBoost algorithm is used to train the data set, the importance of the features is sorted, the importance of the features is obtained, and the features with high importance weights are selected. After multiple rounds of iterative training, the trained model is applied to the validation set, and the results of the validation set are used to determine whether the trained model is usable; if not, the parameters of the model are adjusted, and multiple rounds of iterative training are performed to finally fit a usable XGBoost model; The specific steps of step S5 are: S51: Calculate the distance between the pre-selected point of the base station and all the sampling points by using the latitude and longitude of the base station and the latitude and longitude of the sampling point, sort the points according to the distances obtained, select the data with a distance less than a threshold as a test set, and calculate the characteristic data of the data in the test set; S52: The XGBoost model trained in step S4 is used to obtain the field strength of the sampling point. If the obtained field strength is greater than the original field strength, the base station connected to the sampling point is switched to the pre-selected base station of the newly-built base station. S53: Mark all sampling points that are connected to the new base station on the map, that is, obtain the predicted range of the new base station and the coverage rate of the base station, and then compare the field strength of the sampling points before and after the new base station is built to obtain the display result.

2. The method for predicting base station coverage based on XGBoost regression algorithm according to claim 1, characterized in that: The sampling point MDT data in step S1 includes cellid, longitude, latitude and field strength; the base station data includes base station name, nci, longitude, latitude, base station height, base station azimuth and base station downtilt, wherein cellid and nci are both base station IDs, which are the same data and are used to associate base station data with MDT data during training.

3. The method for predicting base station coverage based on XGBoost regression algorithm according to claim 2, characterized in that: The data cleaning in step S2 includes missing value processing, erroneous data deletion, string feature conversion and relative deviation.

4. The method for predicting base station coverage based on XGBoost regression algorithm according to claim 1, characterized in that: If the relative azimuth angle obtained in step S32 is a negative value, indicating that the sampling point is on the right side of the cell azimuth angle, the negative value is converted into a positive value, thereby obtaining the relative azimuth angle of the base station relative to the sampling point.

5. The method for predicting base station coverage based on XGBoost regression algorithm according to claim 1, characterized in that: The formula for obtaining the XGBoost model by training with the XGBoost algorithm in step S4 is: Among them, x i represents sample i; Represents the prediction result of sample i after the tth iteration; represents the prediction results of the first t-1 trees; f t (x i ) represents the function of the t-th tree.

Images