Disclosure of Invention
The invention aims to provide a management method and a management system for base station load balancing.
The technical scheme provided by the invention is as follows:
a management method for base station load balancing comprises the following steps: step S100, respectively acquiring training sample data of each base station; the training sample data of each base station is historical load data which is selected from a historical load database corresponding to the base station and meets a first preset condition; step S200, respectively training to obtain a load prediction model of each base station; the load prediction model of each base station is obtained by training by adopting a preset model according to training sample data of the base station; step S300, respectively acquiring current load data of each base station; step S400, respectively calculating the predicted load of each base station; the predicted load of each base station is the average load of the base station in a preset time period, which is obtained by calculation according to the current load data of the base station and a load prediction model corresponding to the base station; step S500, when the user to be accessed is to be accessed, selecting one base station according to the current load and the predicted load of each base station and the second preset condition, as the target base station to be accessed by the user to be accessed.
In the technical scheme, the current load of the base station and the historical load data of the base station are considered, and a load prediction model of the base station is obtained according to the historical load data and is used for calculating the predicted load of the base station; when a user to be accessed is available, the current load and the predicted load condition of all base stations covering the user are comprehensively considered, and overall distribution is realized, so that the load balance of the whole network is facilitated, and the overall network performance is improved.
Further, the process of training each base station to obtain the load prediction model in step S200 includes: step S210, obtaining a characteristic combination influencing the load of the base station according to the training sample data and the GBDT model of the base station; step S220, obtaining a regression function by adopting a logistic regression model according to the obtained feature combination, and taking each feature in the feature combination as each feature component in the regression function; step S230, obtaining values of all parameters in a regression function by adopting a gradient descent method according to training sample data of the base station; step S240 uses the regression function substituted for the value of each parameter as the load prediction model of the base station.
Further, the regression function in step S220 is:
hθ(x)=θTx=θ0x0+θ1x1+...+θnxn……………………………(1)
wherein x is0Is 1, x1,x2,...,xnRepresents each feature in the combination of features, x ═ x0,x1,x2,...,xn]T;θ0,θ1,θ2...,θnIs a parameter, theta ═ theta0,θ1,θ2,...,θn]T;
In step S230, a gradient descent method is used according to the training sample data of the base station, and the functions used for obtaining the values of the parameters in the regression function are as follows:
wherein i represents the ith training sample in the training sample data, and m is the total number of samples in the training sample data; h isθ(xi) For the predicted load, y, corresponding to the ith training sampleiRepresenting the actual load in the ith training sample; and when J (theta) is minimum, the theta value corresponding to the J (theta) is the value of each parameter in the regression function.
In the above technical solution, a process of obtaining a predicted load model of each base station through GBDT + LR model training is described, and it is further explained that the present solution not only considers the current load of the base station, but also considers the historical load data of the base station.
Further, the step S500 includes: step S510, when a user is to be accessed, all base stations covering the user are obtained according to the position of the user; step S520, respectively obtaining the current load and the predicted load corresponding to all base stations covering the user; step S530 selects one base station from all base stations covering the user according to a second preset condition, and the selected base station is used as a target base station to be accessed by the user to be accessed.
In the technical scheme, all base stations covering the user are obtained according to the position of the user to be accessed, the current load and the predicted load of the base stations are comprehensively considered, and overall distribution is performed, so that the load balance of the whole network is facilitated, and the performance of the whole network is improved.
Further, the step S530 further includes: and selecting one base station with the current load and the predicted load which are both smaller than a preset load threshold value and the predicted load is the smallest from all the base stations covering the user as a target base station to be accessed by the user to be accessed.
In the technical scheme, one base station with the current load and the predicted load both smaller than the preset load threshold and the smallest predicted load is selected as the target base station to be accessed by the user to be accessed, so that the load balance of the whole network is facilitated.
Further, after the step S500, the method further includes: in step S600, when there is a base station that reaches the update time of the load prediction model, the training sample data of the base station is updated, and the load prediction model of the base station is updated according to the updated training sample data.
In the technical scheme, the load prediction model of the base station is updated, the new model can accurately reflect the current situation of the base station, and the load predicted by using the new model is closer to the current situation of the base station.
The invention also provides a management system for load balancing of the base station, which comprises the following steps: the acquisition module is used for respectively acquiring training sample data of each base station, wherein the training sample data of each base station is historical load data which is selected from a historical load database corresponding to the base station and meets a first preset condition; the model training module is electrically connected with the acquisition module and used for respectively training to obtain a load prediction model of each base station, and the load prediction model of each base station is obtained by training by adopting a preset model according to training sample data of the base station; the obtaining module is further configured to obtain current load data of each base station respectively; the calculation module is electrically connected with the model training module and is used for calculating the predicted load of each base station respectively, and the predicted load of each base station is the average load of the base station in a preset time period, which is obtained by calculation according to the current load data of the base station and the load prediction model corresponding to the base station; the obtaining module is further configured to select one base station according to a second preset condition as a target base station to be accessed by the user to be accessed according to the current load and the predicted load of each base station when the user to be accessed is to be accessed.
In the technical scheme, the current load of the base station and the historical load data of the base station are considered, and a load prediction model of the base station is obtained according to the historical load data and is used for calculating the predicted load of the base station; when a user to be accessed is available, the current load and the predicted load condition of all base stations covering the user are comprehensively considered, and overall distribution is realized, so that the load balance of the whole network is facilitated, and the overall network performance is improved.
Further, the process of obtaining the load prediction model by training each base station in the model training module includes: obtaining a characteristic combination influencing the load of the base station according to the training sample data and the GBDT model of the base station; obtaining a regression function by adopting a logistic regression model according to the obtained feature combination, and taking each feature in the feature combination as each feature component in the regression function; obtaining values of all parameters in the regression function by adopting a gradient descent method according to training sample data of the base station; and taking the regression function substituted into the value of each parameter as a load prediction model of the base station.
In the above technical solution, a process of obtaining a predicted load model of each base station through GBDT + LR model training is described, and it is further explained that the present system considers not only the current load of the base station but also historical load data of the base station.
Further, the acquiring module is further configured to acquire, when a user is to be accessed, a base station covering the user according to the user location; respectively acquiring current loads and predicted loads corresponding to all base stations covering the user; and according to a second preset condition, selecting one base station from all base stations covering the user as a target base station to be accessed by the user to be accessed.
In the technical scheme, all base stations covering the user are obtained according to the position of the user to be accessed, the current load and the predicted load of the base stations are comprehensively considered, and overall distribution is performed, so that the load balance of the whole network is facilitated, and the performance of the whole network is improved.
Further, still include: the obtaining module is further used for updating training sample data of the base station when the base station reaching the update time of the load prediction model exists; the model training module is further used for updating the load prediction model of the base station according to the updated training sample data.
In the technical scheme, the load prediction model of the base station is updated, the new model can accurately reflect the current situation of the base station, and the load predicted by using the new model is closer to the current situation of the base station.
The management method and the system for base station load balancing provided by the invention can bring at least one of the following beneficial effects:
1. according to the load prediction method, not only the current load of the base station but also historical load data of the base station are considered, and a load prediction model of the base station is obtained according to the historical load data and is used for calculating the predicted load of the base station; when a user to be accessed is available, the current load and the predicted load condition of all base stations covering the user are comprehensively considered, and overall distribution is realized, so that the load balance of the whole network is facilitated, and the overall network performance is improved.
2. The invention updates the load prediction model of the base station to enable the predicted load to reflect the current situation of the base station more accurately, thereby enabling the management method of the base station load balance to be adaptive to the change of the environment.
Detailed Description
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following description will be made with reference to the accompanying drawings. It is obvious that the drawings in the following description are only some examples of the invention, and that for a person skilled in the art, other drawings and embodiments can be derived from them without inventive effort.
For the sake of simplicity, the drawings only schematically show the parts relevant to the present invention, and they do not represent the actual structure as a product. In addition, in order to make the drawings concise and understandable, components having the same structure or function in some of the drawings are only schematically illustrated or only labeled. In this document, "one" means not only "only one" but also a case of "more than one".
In an embodiment of the present invention, as shown in fig. 1, a method for managing load balancing of a base station includes:
step S100, respectively acquiring training sample data of each base station;
the training sample data of each base station is historical load data which is selected from a historical load database corresponding to the base station and meets a first preset condition;
step S200, respectively training to obtain a load prediction model of each base station;
the load prediction model of each base station is obtained by training by adopting a preset model according to training sample data of the base station;
step S300, respectively acquiring current load data of each base station;
step S400, respectively calculating the predicted load of each base station;
the predicted load of each base station is the average load of the base station in a preset time period, which is obtained by calculation according to the current load data of the base station and a load prediction model corresponding to the base station;
step S500, when the user to be accessed is to be accessed, selecting one base station according to the current load and the predicted load of each base station and the second preset condition, as the target base station to be accessed by the user to be accessed.
Specifically, the historical load database of the base station is composed of historical load data, that is, base station load data collected in historical time, wherein the base station load data includes values of various characteristics affecting the base station load. Factors influencing the base station load are many, such as dynamic variation of time, base station load, number of base station users, base station used power and the like, and well-defined static quantities during base station planning, such as location of the base station, maximum load capacity, allowed maximum transmitting power, frequency range, lobe width and the like, and all the factors can be used as characteristics influencing the base station load, and parts can also be selected according to experience, so that the training complexity is reduced.
The first preset condition may be that the selected data size meets the requirement of the preset training sample duration, and the load data occurring before the current time is the historical load data meeting the first preset condition. And each base station selects historical load data from the respective historical load database as training sample data. The training sample duration in the first preset condition of each base station may be the same or different, and for simplicity, the same is suggested. For example, the current time is 8 months and 31 days, the training sample duration of all base stations is set to 3 months, and in order to predict the base station load of 9 months and 1 day and later, base station load data of 5 months to 7 months can be selected from the historical load database of each base station, and base station load data of 6 months to 8 months can also be selected as the training sample data of each base station. The latter is better than the former, and the data selected closer to the current time can reflect the current situation of the base station. The longer the training sample time is, the better the training sample time is, the less the load prediction influence of the more long-term data on the current base station is, and the improvement on the prediction accuracy rate is not obvious. The first preset condition may also be that load data occurring before the current time is selected according to a preset data amount, that is, historical load data meeting the first preset condition. The preset data amount of different base stations can be the same or different, and for simplicity, the same is suggested. The preset data size is not larger, and the better, the reason is the setting of the training sample duration.
The model is preset, various classification and regression models in machine learning can be fully used for reference, for example, a combination of a gradient lifting tree (GBDT) and a Logistic Regression (LR) model is used, the GBDT model is used for extracting a characteristic combination which has the largest influence on the base station load, then the LR model is used for obtaining a regression function, the function contains a plurality of unknown parameters, the optimal solution of the parameters is estimated by learning through training sample data, the optimal solution is used as the value of the parameters and is substituted into the original regression function, and the model is successfully established; the method comprises the steps of presetting a model, or adopting a combination of a random forest and an LR model, firstly extracting a characteristic combination which has the largest influence on the base station load by using the random forest, then obtaining a regression function by using the LR model, then obtaining an optimal solution of parameters in the regression function by iteration through a gradient descent method, substituting the optimal solution as values of the parameters into an original regression function, and then successfully establishing the model; the aforementioned GBDT model may also be replaced by other models, such as XGBoost; the LR model in the foregoing can also be replaced by other regression models, such as linear regression, but the LR model is easier to implement than the LR model, so the LR model is recommended; the preset model used by each base station may be the same or different, and for simplicity, the same is suggested. Each base station obtains a load prediction model corresponding to the base station according to the training sample data and a preset model of the base station; and subsequently carrying out load prediction according to the load prediction model.
In the process of obtaining the load prediction model of each base station through training, multiple groups of window data can be adopted for parallel training, and a model with high accuracy is selected as a final model through cross validation, so that the bloom capability of the model can be improved. For example, the base station load of one day in the future is predicted according to historical load data of three months, a base station load of 9 month and 1 is supposed to be predicted, load data of the latest four months from 5 month and 1 to 8 months and 31 are selected, the data are divided into 3 windows, the time length of the sliding window is half a month, data windows from 5 month and 1 to 8 months and 1 (corresponding to window 1), data windows from 5 month and 15 to 8 months and 15 (corresponding to window 2), or data windows from 5 month and 30 to 8 months and 30 (corresponding to window 3) are sequentially obtained, each group of windows are respectively trained according to the data of three months to obtain corresponding temporary models, then the accuracy of the models is verified in a cross verification mode, for example, the model of the corresponding window 1 is verified by using all or part of the data of the window 2, the model of the data of the window 3 is verified by using the data of the window 1, and selecting a temporary model with high accuracy as a final load prediction model.
And respectively calculating the predicted load of each base station, wherein the predicted load of each base station is the average load of the base station in a preset time period, which is obtained by calculation according to the current load data of the base station and a load prediction model corresponding to the base station. For example, the preset time period is 3 days, data of 8 months and 31 days are input, the average load of 9 months and 1 day to 9 months and 3 days is expected to be predicted, if the load prediction model predicts the load of the next day according to the load data of the base station in one day, the load of 3 days needs to be continuously predicted, and the load of 3 days is averaged to obtain the required predicted load; and selecting a target base station according to the required predicted load, and if the second preset condition is that the base station with the current load and the predicted load both smaller than the preset load threshold and the predicted load being the smallest is selected, the base station with the current load being light and the average load of the future 3 days being lighter is expected to be selected as the target base station. The preset time periods of each base station can be the same or different, and the suggestions are the same, so that the referenced parameters are obtained according to a consistent mode when the target base station is selected subsequently.
And if the user is to be accessed at the current time, selecting one base station as a target base station to be accessed by the user to be accessed according to the current load and the predicted load of each base station and a second preset condition. Each base station refers to all base stations covering the user; in a heterogeneous network, there is an overlapping coverage case; in a conventional cellular network consisting of macro base stations, there is also overlapping coverage; the present embodiment is applicable to scenarios where there is overlapping coverage. The second preset condition may be set as needed, for example, one way is to select a base station whose current load and predicted load are both smaller than a preset load threshold and whose predicted load is the smallest; the other mode is that a base station with the current load and the predicted load both smaller than a preset load threshold value is selected first, and then the base station with the best user channel condition is selected from the base stations; the former is the principle of minimum load, and the latter is a base station with good priority channel condition and load within a certain range. The user may be a new access user or a user migrated from another base station.
Respectively calculating the predicted load of each base station, wherein the predicted load can be executed when a user is to be accessed, and can also be executed only aiming at all base stations covering the user; for example, the user 1 to be accessed is covered by the base stations 1, 2, 3, and 4, and when the user to be accessed is detected, the current loads of the base stations 1, 2, 3, and 4 are obtained and the corresponding predicted loads are calculated, and then the target base station is selected according to the second preset condition; if the user 2 is still to be accessed at the current time, the base stations 1, 2, 5, and 6 cover the user, and considering that the current loads of the base stations 1 and 2 are already obtained and the corresponding predicted loads are already calculated, only the base stations 5 and 6 need to obtain the current loads and calculate the corresponding predicted loads.
The embodiment is described in a mode of firstly calculating the predicted load of all base stations and then using the predicted load when the user is to be accessed, so that the advantage of directly using the calculated predicted load when the user is to be accessed is that the target base station can be quickly selected for the user, and the user experience is possibly better.
In another embodiment of the present invention, as shown in fig. 2, a method for managing load balancing of a base station includes:
step S100, respectively acquiring training sample data of each base station;
the training sample data of each base station is historical load data which is selected from a historical load database corresponding to the base station and meets a first preset condition;
step S200, respectively training to obtain a load prediction model of each base station;
the load prediction model of each base station is obtained by training by adopting a preset model according to training sample data of the base station;
step S300, respectively acquiring current load data of each base station;
step S400, respectively calculating the predicted load of each base station;
the predicted load of each base station is the average load of the base station in a preset time period, which is obtained by calculation according to the current load data of the base station and a load prediction model corresponding to the base station;
step S500, when a user to be accessed is to be accessed, selecting one base station according to the current load and the predicted load of each base station and a second preset condition as a target base station to be accessed by the user to be accessed;
the process of training each base station to obtain the load prediction model in the step S200 includes:
step S210, obtaining a characteristic combination influencing the load of the base station according to the training sample data and the GBDT model of the base station;
step S220, obtaining a regression function by adopting a logistic regression model according to the obtained feature combination, and taking each feature in the feature combination as each feature component in the regression function;
step S230, obtaining values of all parameters in a regression function by adopting a gradient descent method according to training sample data of the base station;
step S240, taking the regression function substituted into the value of each parameter as a load prediction model of the base station;
the regression function in step S220 is:
hθ(x)=θTx=θ0x0+θ1x1+...+θnxn……………………………(1)
wherein x is0Is 1, x1,x2,...,xnRepresents each feature in the combination of features, x ═ x0,x1,x2,...,xn]T;θ0,θ1,θ2...,θnIs a parameter, theta ═ theta0,θ1,θ2,...,θn]T;
In step S230, a gradient descent method is used according to the training sample data of the base station, and the functions used for obtaining the values of the parameters in the regression function are as follows:
wherein i represents the ith training sample in the training sample data, and m is the total number of samples in the training sample data; h isθ(xi) For the predicted load, y, corresponding to the ith training sampleiRepresenting the actual load in the ith training sample; when J (theta) is minimum, the theta value corresponding to the J (theta) is the value of each parameter in the regression function;
the step S500 includes:
step S510, when a user is to be accessed, all base stations covering the user are obtained according to the position of the user; step S520, respectively obtaining the current load and the predicted load corresponding to all base stations covering the user; step S530, according to a second preset condition, one base station is selected from all base stations covering the user and is used as a target base station to be accessed by the user to be accessed;
the step S530 further includes:
and selecting one base station with the current load and the predicted load which are both smaller than a preset load threshold value and the predicted load is the smallest from all the base stations covering the user as a target base station to be accessed by the user to be accessed.
Specifically, compared with the previous embodiment, steps S210 to S240 are added, and a combination of GBDT and LR models is respectively adopted for each base station, and training is performed to obtain respective corresponding load prediction models. The following description is only for the process of training a single base station to obtain a load prediction model, and other base stations operate similarly.
GBDT (gradient Boost Decision Tree) is a common nonlinear model, which is based on boosting thought in ensemble learning, and a new Decision tree is established in the gradient direction for reducing residual error in each iteration, and how many Decision trees are generated by iteration. The GBDT concept has natural advantages that various distinctive features and feature combinations can be found, the path of the decision tree can be directly used as an LR (Logistic Regression model) input feature, and the step of manually searching the features and the feature combinations is omitted. LR is a linear fitting model that can be transformed into a classifier using a logistic function.
Firstly, extracting a characteristic combination which has the largest influence on the base station load by using a GBDT model, inputting the characteristic combination into a logistic regression model, and obtaining a regression function reflecting the base station load, wherein the function contains a plurality of unknown parameters, namely theta; finding out a theta value which enables the function of the formula (2) to be minimum, namely an optimal solution of theta according to training sample data of the base station by a gradient descent method; substituting the optimal solution of theta as the value of each parameter into the original regression function, and successfully establishing a load prediction model of the base station; the regression function with the determined parameters is used as a load prediction model of the base station.
Steps S510-S530 are added, when a user to be accessed is available, all base stations covering the user are obtained according to the position of the user; respectively acquiring current loads and predicted loads corresponding to the base stations; and selecting a target base station to be accessed by the user from the base stations according to a second preset condition.
The second preset condition in this embodiment is to select a base station whose current load and predicted load are both smaller than the preset load threshold and whose predicted load is the smallest. The preset load threshold is set for dividing the base station into a heavy load base station and a light load base station. If the current load of the base station is smaller than the preset load threshold value, indicating that the base station belongs to light load currently; if the predicted load of the base station is smaller than the preset load threshold value, indicating that the base station belongs to light load for the next period of time; and preferably selecting the base station with light load at the current time and the next period of time, so that the load balance of the whole network is facilitated, and the utilization rate of resources of the whole network is improved.
In another embodiment of the present invention, as shown in fig. 3, a method for managing load balancing of a base station includes:
step S100, respectively acquiring training sample data of each base station;
the training sample data of each base station is historical load data which is selected from a historical load database corresponding to the base station and meets a first preset condition;
step S200, respectively training to obtain a load prediction model of each base station;
the load prediction model of each base station is obtained by training by adopting a preset model according to training sample data of the base station;
step S300, respectively acquiring current load data of each base station;
step S400, respectively calculating the predicted load of each base station;
the predicted load of each base station is the average load of the base station in a preset time period, which is obtained by calculation according to the current load data of the base station and a load prediction model corresponding to the base station;
step S500, when a user to be accessed is to be accessed, selecting one base station according to the current load and the predicted load of each base station and a second preset condition as a target base station to be accessed by the user to be accessed;
the process of training each base station to obtain the load prediction model in the step S200 includes:
step S210, obtaining a characteristic combination influencing the load of the base station according to the training sample data and the GBDT model of the base station;
step S220, obtaining a regression function by adopting a logistic regression model according to the obtained feature combination, and taking each feature in the feature combination as each feature component in the regression function;
step S230, obtaining values of all parameters in a regression function by adopting a gradient descent method according to training sample data of the base station;
step S240, taking the regression function substituted into the value of each parameter as a load prediction model of the base station;
the regression function in step S220 is:
hθ(x)=θTx=θ0x0+θ1x1+...+θnxn……………………………(1)
wherein x is0Is 1, x1,x2,...,xnRepresents each feature in the combination of features, x ═ x0,x1,x2,...,xn]T;θ0,θ1,θ2...,θnIs a parameter, theta ═ theta0,θ1,θ2,...,θn]T;
In step S230, a gradient descent method is used according to the training sample data of the base station, and the functions used for obtaining the values of the parameters in the regression function are as follows:
wherein i represents the ith training sample in the training sample data, and m is the total number of samples in the training sample data; h isθ(xi) For the ith training sampleCorresponding predicted load, yiRepresenting the actual load in the ith training sample; when J (theta) is minimum, the theta value corresponding to the J (theta) is the value of each parameter in the regression function;
the step S500 includes:
step S510, when a user is to be accessed, all base stations covering the user are obtained according to the position of the user;
step S520, respectively obtaining the current load and the predicted load corresponding to all base stations covering the user;
step S530, according to a second preset condition, one base station is selected from all base stations covering the user and is used as a target base station to be accessed by the user to be accessed;
the step S530 further includes:
selecting one base station with the current load and the predicted load both smaller than a preset load threshold value and the minimum predicted load from all base stations covering the user as a target base station to be accessed by the user to be accessed;
the step S500 is followed by:
in step S600, when there is a base station that reaches the update time of the load prediction model, the training sample data of the base station is updated, and the load prediction model of the base station is updated according to the updated training sample data.
Specifically, compared with the previous embodiment, the present embodiment adds step S600 and introduces an update of the load prediction model.
And updating the base station load data acquired at each time into a historical load database of the base station, wherein the historical load database of the base station contains more or less new data along with the time, so that when the base station reaching the update time of the load prediction model exists, the training sample data of the base station can be updated, and the load prediction model of the base station can be updated according to the updated training sample data. The load prediction model update time may be different or the same for each base station, and some base stations may not be provided.
The load prediction model is updated relatively infrequently, but if the environment surrounding the base station changes, the model needs to be refreshed. For example, in a new residential living area, the load of the base station changes greatly from the initial living period of few people living to the gradual living period of people living to the cell maturation period of people living, and the corresponding base station load prediction model should be updated in time.
In another embodiment of the present invention, as shown in fig. 4, a management system for load balancing of base stations includes:
an obtaining module 10, configured to obtain training sample data of each base station, where the training sample data of each base station is historical load data that meets a first preset condition and is selected from a historical load database corresponding to the base station;
the model training module 20 is electrically connected with the acquisition module 10 and is used for respectively training to obtain a load prediction model of each base station, and the load prediction model of each base station is obtained by training by adopting a preset model according to training sample data of the base station;
the obtaining module 10 is further configured to obtain current load data of each base station respectively;
a calculating module 30, electrically connected to the model training module 20, configured to calculate a predicted load of each base station, where the predicted load of each base station is an average load of the base station in a preset time period, which is calculated according to current load data of the base station and a load prediction model corresponding to the base station;
the obtaining module 10 is further configured to select one base station according to a second preset condition and according to the current load and the predicted load of each base station when the user to be accessed is to be accessed, and use the selected base station as a target base station to be accessed by the user to be accessed.
Specifically, the historical load database of the base station is composed of historical load data, that is, base station load data collected in historical time, wherein the base station load data includes values of various characteristics affecting the base station load. Factors influencing the base station load are many, such as dynamic variation of time, base station load, number of base station users, base station used power and the like, and well-defined static quantities during base station planning, such as location of the base station, maximum load capacity, allowed maximum transmitting power, frequency range, lobe width and the like, and all the factors can be used as characteristics influencing the base station load, and parts can also be selected according to experience, so that the training complexity is reduced.
The first preset condition may be that the selected data size meets the requirement of the preset training sample duration, and the load data occurring before the current time is the historical load data meeting the first preset condition. And each base station selects historical load data from the respective historical load database as training sample data. The training sample duration in the first preset condition of each base station may be the same or different, and for simplicity, the same is suggested. For example, the current time is 8 months and 31 days, the training sample duration of all base stations is set to 3 months, and in order to predict the base station load of 9 months and 1 day and later, base station load data of 5 months to 7 months can be selected from the historical load database of each base station, and base station load data of 6 months to 8 months can also be selected as the training sample data of each base station. The latter is better than the former, and the data selected closer to the current time can reflect the current situation of the base station. The longer the training sample time is, the better the training sample time is, the less the load prediction influence of the more long-term data on the current base station is, and the improvement on the prediction accuracy rate is not obvious. The first preset condition may also be that load data occurring before the current time is selected according to a preset data amount, that is, historical load data meeting the first preset condition. The preset data amount of different base stations can be the same or different, and for simplicity, the same is suggested. The preset data size is not larger, and the better, the reason is the setting of the training sample duration.
The model is preset, various classification and regression models in machine learning can be fully used for reference, for example, a combination of a gradient lifting tree (GBDT) and a Logistic Regression (LR) model is used, the GBDT model is used for extracting a characteristic combination which has the largest influence on the base station load, then the LR model is used for obtaining a regression function, the function contains a plurality of unknown parameters, the optimal solution of the parameters is estimated by learning through training sample data, the optimal solution is used as a value of the parameters and is substituted into the original regression function, and the model is successfully established; the method comprises the steps of presetting a model, or adopting a combination of a random forest and an LR model, firstly extracting a characteristic combination which has the largest influence on the base station load by using the random forest, then obtaining a regression function by using the LR model, then obtaining an optimal solution of parameters in the regression function by iteration through a gradient descent method, substituting the optimal solution as values of the parameters into an original regression function, and then successfully establishing the model; the aforementioned GBDT model may also be replaced by other models, such as XGBoost; the LR model in the foregoing can also be replaced by other regression models, such as linear regression, but the LR model is easier to implement than the LR model, so the LR model is recommended; the preset model used by each base station may be the same or different, and for simplicity, the same is suggested. Each base station obtains a load prediction model corresponding to the base station according to the training sample data and a preset model of the base station; and subsequently carrying out load prediction according to the load prediction model.
In the process of obtaining the load prediction model of each base station through training, multiple groups of window data can be adopted for parallel training, and a model with high accuracy is selected as a final model through cross validation, so that the bloom capability of the model can be improved. For example, the base station load of one day in the future is predicted according to historical load data of three months, a base station load of 9 month and 1 is supposed to be predicted, load data of the latest four months from 5 month and 1 to 8 months and 31 are selected, the data are divided into 3 windows, the time length of the sliding window is half a month, data windows from 5 month and 1 to 8 months and 1 (corresponding to window 1), data windows from 5 month and 15 to 8 months and 15 (corresponding to window 2), or data windows from 5 month and 30 to 8 months and 30 (corresponding to window 3) are sequentially obtained, each group of windows are respectively trained according to the data of three months to obtain corresponding temporary models, then the accuracy of the models is verified in a cross verification mode, for example, the model of the corresponding window 1 is verified by using all or part of the data of the window 2, the model of the data of the window 3 is verified by using the data of the window 1, and selecting a temporary model with high accuracy as a final load prediction model.
And respectively calculating the predicted load of each base station, wherein the predicted load of each base station is the average load of the base station in a preset time period, which is obtained by calculation according to the current load data of the base station and a load prediction model corresponding to the base station. For example, the preset time period is 3 days, data of 8 months and 31 days are input, the average load of 9 months and 1 day to 9 months and 3 days is expected to be predicted, if the load prediction model predicts the load of the next day according to the load data of the base station in one day, the load of 3 days needs to be continuously predicted, and the load of 3 days is averaged to obtain the required predicted load; and selecting a target base station according to the required predicted load, and if the second preset condition is that the base station with the current load and the predicted load both smaller than the preset load threshold and the predicted load being the smallest is selected, the base station with the current load being light and the average load of the future 3 days being lighter is expected to be selected as the target base station. The preset time periods of each base station can be the same or different, and the suggestions are the same, so that the referenced parameters are obtained according to a consistent mode when the target base station is selected subsequently.
And if the user is to be accessed at the current time, selecting one base station as a target base station to be accessed by the user to be accessed according to the current load and the predicted load of each base station and a second preset condition. Each base station refers to all base stations covering the user; in a heterogeneous network, there is an overlapping coverage case; in a conventional cellular network consisting of macro base stations, there is also overlapping coverage; the present embodiment is applicable to scenarios where there is overlapping coverage. The second preset condition may be set as needed, for example, one way is to select a base station whose current load and predicted load are both smaller than a preset load threshold and whose predicted load is the smallest; the other mode is that a base station with the current load and the predicted load both smaller than a preset load threshold value is selected first, and then the base station with the best user channel condition is selected from the base stations; the former is the principle of minimum load, and the latter is a base station with good priority channel condition and load within a certain range. The user may be a new access user or a user migrated from another base station.
Respectively calculating the predicted load of each base station, wherein the predicted load can be executed when a user is to be accessed, and can also be executed only aiming at all base stations covering the user; for example, the user 1 to be accessed is covered by the base stations 1, 2, 3, and 4, and when the user to be accessed is detected, the current loads of the base stations 1, 2, 3, and 4 are obtained and the corresponding predicted loads are calculated, and then the target base station is selected according to the second preset condition; if the user 2 is still to be accessed at the current time, the base stations 1, 2, 5, and 6 cover the user, and considering that the current loads of the base stations 1 and 2 are already obtained and the corresponding predicted loads are already calculated, only the base stations 5 and 6 need to obtain the current loads and calculate the corresponding predicted loads.
The embodiment is described in a mode of firstly calculating the predicted load of all base stations and then using the predicted load when the user is to be accessed, so that the advantage of directly using the calculated predicted load when the user is to be accessed is that the target base station can be quickly selected for the user, and the user experience is possibly better.
In another embodiment of the present invention, as shown in fig. 4, a management system for load balancing of base stations includes:
an obtaining module 10, configured to obtain training sample data of each base station, where the training sample data of each base station is historical load data that meets a first preset condition and is selected from a historical load database corresponding to the base station;
the model training module 20 is electrically connected with the acquisition module 10 and is used for respectively training to obtain a load prediction model of each base station, and the load prediction model of each base station is obtained by training by adopting a preset model according to training sample data of the base station;
the obtaining module 10 is further configured to obtain current load data of each base station respectively;
a calculating module 30, electrically connected to the model training module 20, configured to calculate a predicted load of each base station, where the predicted load of each base station is an average load of the base station in a preset time period, which is calculated according to current load data of the base station and a load prediction model corresponding to the base station;
the obtaining module 10 is further configured to, when a user to be accessed is detected, select one base station according to a second preset condition and the current load and the predicted load of each base station, and use the selected base station as a target base station to be accessed by the user to be accessed;
the process of obtaining the load prediction model by training each base station in the model training module 20 includes:
obtaining a characteristic combination influencing the load of the base station according to the training sample data and the GBDT model of the base station; obtaining a regression function by adopting a logistic regression model according to the obtained feature combination, and taking each feature in the feature combination as each feature component in the regression function; obtaining values of all parameters in the regression function by adopting a gradient descent method according to training sample data of the base station; and taking the regression function substituted into the value of each parameter as a load prediction model of the base station;
the obtaining module 10 is further configured to, when a user is to be accessed, obtain a base station covering the user according to the user location; respectively acquiring current loads and predicted loads corresponding to all base stations covering the user; according to a second preset condition, selecting one base station from all base stations covering the user as a target base station to be accessed by the user to be accessed;
the obtaining module 10 is further configured to select, from all base stations covering the user, a base station with a current load and a predicted load both smaller than a preset load threshold and with a smallest predicted load as a target base station to be accessed by the user to be accessed;
the obtaining module 10 is further configured to update training sample data of a base station when the base station arrives at the update time of the load prediction model;
the model training module 20 is further configured to update the load prediction model of the base station according to the updated training sample data.
Specifically, the functions of the acquisition module 10 and the model training module 20 are further enhanced compared to the previous embodiment.
The model training module 20 is further configured to train each base station to obtain a corresponding load prediction model by using a combination of GBDT and LR models;
the obtaining module 10 is further configured to, when a user is to be accessed, obtain all base stations covering the user according to the user location; respectively acquiring current loads and predicted loads corresponding to the base stations; and selecting a target base station to be accessed by the user from the base stations according to a second preset condition. The second preset condition in this embodiment is to select a base station whose current load and predicted load are both smaller than the preset load threshold and whose predicted load is the smallest. The preset load threshold is set for dividing the base station into a heavy load base station and a light load base station. If the current load of the base station is smaller than the preset load threshold value, indicating that the base station belongs to light load currently; if the predicted load of the base station is smaller than the preset load threshold value, indicating that the base station belongs to light load for the next period of time; and preferably selecting the base station with light load at the current time and the next period of time, so that the load balance of the whole network is facilitated, and the utilization rate of resources of the whole network is improved.
The obtaining module 10 and the model training module 20 are further configured to update the load prediction model; and updating the base station load data acquired at each time into a historical load database of the base station, wherein the historical load database of the base station contains more or less new data along with the time, so that when the base station reaching the update time of the load prediction model exists, the training sample data of the base station can be updated, and the load prediction model of the base station can be updated according to the updated training sample data. The load prediction model update time may be different or the same for each base station, and some base stations may not be provided.
The load prediction model is updated relatively infrequently, but if the environment surrounding the base station changes, the model needs to be refreshed. For example, in a new residential living area, the load of the base station changes greatly from the initial living period of few people living to the gradual living period of people living to the cell maturation period of people living, and the corresponding base station load prediction model should be updated in time.
It should be noted that the above embodiments can be freely combined as necessary. The foregoing is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, various modifications and decorations can be made without departing from the principle of the present invention, and these modifications and decorations should also be regarded as the protection scope of the present invention.