Invention content
To solve the above problems, based in City scenarios, crowd moves with existing road, crowd's future time point
Position this thinking can be predicted, the present invention provides the small base station switch control methods based on deep neural network.
In order to achieve the above object, the present invention provides following technical solution:
Small base station switch control method based on deep neural network, includes the following steps:
Step 1:Acquire the user information in base station
Sampling is primary at regular intervals, records Customs Assigned Number, access moment and the customer position information of access base station,
It is put into sample set L={ (ui,ti,pi), wherein uiFor the number of accessing user, tiAt the time of for record data, piFor user
Geographical location, including longitude coordinate xiWith latitude coordinate yi;
Step 2:Data Integration
The user data that base station in step 1 is collected arranges, is merged into path data for model training, merges institute
There is the data sample of user, obtain eventually for trained sample set Ltrain={ (p1,i,p2,i,p3,i,p4,i,p5,i,p6,i)};
Step 3:Build neural network model
Full Connection Neural Network is selected as training pattern, training error is calculated using Mean Square Error, is used when training
It is that common forward-propagating method is once trained as a result, using reverse transmission method update neural network in parameter;
Step 4:Input data and training pattern
1) the set L that selecting step two obtainstrainIn a sample, sample is denoted as dj=(p1,j,p2,j,p3,j,p4,j,
p5,j,p6,j), djIn preceding 5 data (p1,j,p2,j,p3,j,p4,j,p5,j) for inputting, last 1 data p6,jFor with prediction
Results contrast calculates error;
2) preceding 5 data (p of a sample are inputted into neural network1,j,p2,j,p3,j,p4,j,p5,j), using nerve net
It is that forward-propagating method in network is once trained as a result, i.e. prediction position coordinates
3) actual value of the prediction coordinate in sample isIt calculates and misses using actual coordinate and prediction coordinate
DifferenceParameter in neural network is updated using reverse transmission method, completes a sample instruction
Practice;
4) by entire sample set LtraiN is substituted into and is trained primary, referred to as one wheel sample set training, the so more wheel samples of progress
Collection training calculates the wheel training error after often wheel trainingWhen | Ei+1-Ei| < ecWhen, training stops;At this point, model
Parameter update finish, model training finishes;
Wherein, EiFor the error of the i-th wheel training, ei,jThe training error of j-th of sample in the sample set trained for the i-th wheel,
ecFor minimal error constant;
Step 5:Collect user data to be predicted, prediction user's subsequent time position
1) moment to be predicted is set as tpredict, user data is acquired according to step 1, is denoted as L'={ (ui',ti',pi')};
2) according to the step integral data in step 2, it is denoted as set Lpredict={ (p1',i,p'2,i,p'3,i,p'4,i,
p'5,i)};
It 3) will set LpredictIn sample input model, you can obtain the position prediction of user as a result, prediction result collection
Conjunction is denoted as
Step 6:The number of calculation base station future service user, control base station switch
1) information of base station is denoted as setWherein,For the longitude coordinate of base station location,For
The latitude coordinate of base station location, numiNumber for base station future service user;
2) the prediction result set obtained from step 5In, sample is chosen successivelyCalculate its distance with base station each in set C
Obtain the number of the base station nearest with the sampleThe corresponding num in set CiIt is upper to add 1;
3) threshold value of control base station switch is set as numc, all base station datas in set C are traversed, work as numi≥numcWhen, it opens
Open respective base station i;Work as numi≤numc, close respective base station i.
Further, the process that user data is arranged, merged in the step 2 specifically comprises the following steps:
1) in set L, a fixed Customs Assigned Number c is chosen, the data of the user is collected, is denoted as set
2) by LcIn sample temporallyIt is ranked up;
3) sample in set is grouped in chronological order with six one group, each group is a new sample, i.e.,
Each new sample isK is grouping group number;
4) time data in new samples is removedRetention position coordinate dataI.e.
To user (ui=c) for trained sample set, it is denoted as
5) sample of other users is arranged according to above-mentioned steps, and the sample set of all users is merged, is obtained
Eventually for trained sample set, it is denoted as Ltrain={ (p1,i,p2,i,p3,i,p4,i,p5,i,p6,i)}。
Further, the neural network selected in the step 3 is of five storeys neuron in total, and first layer has 10 input god
Through member, last layer has 2 output neurons, for predicting unknown coordinate;To neural network inputs in the step 4
Each data share 10 numbers for inputting including latitude coordinates in sample data, with the input neuron of neural network one by one
It is corresponding.
Further, at the time of prediction time and collecting sample are set in the step 5 within the scope of certain time.
Compared with prior art, the invention has the advantages that and advantageous effect:
The method of the present invention controls the switch of small base station in super-intensive network, reaches by predicting number to be serviced in base station
Reduction base station power consumption reduces the interference between base station, the purpose of optimization super-intensive resources in network distribution;In founding mathematical models
During, this method combines data mining and machine learning, improves the accuracy rate of prediction and the practicability of system.
Specific embodiment
Technical solution provided by the invention is described in detail below with reference to specific embodiment, it should be understood that following specific
Embodiment is only illustrative of the invention and is not intended to limit the scope of the invention.
The present invention is established the prediction model of crowd position, is predicted in following small base station using deep neural network model
Number to be serviced.Specifically, the small base station switch control method provided by the invention based on deep neural network, including as follows
Step:
The first step:Acquire the user information in base station.Sampling in each minute is primary, record access base station Customs Assigned Number,
Moment and customer position information are accessed, is put into sample set L={ (ui,ti,pi), wherein uiFor the number of accessing user, tiFor
At the time of recording data, the moment is accurate to minute, piFor the geographical location of user, including longitude coordinate xiWith latitude coordinate yi。
Second step:Data Integration, the user data that base station in the first step is collected are arranged, are merged into for model training
Path data.
1) in set L, a fixed Customs Assigned Number (u is choseni=c), the data of the user are collected, are denoted as set
2) by LcIn sample temporallyIt is ranked up.
3) sample in set is grouped in chronological order with six one group, each group is a new sample, i.e.,
Each new sample isK is grouping group number.
4) time data in new samples is removedRetention position coordinate dataI.e.This
When obtain user (ui=c) for trained sample set, it is denoted as
5) sample of other users is arranged according to above-mentioned steps, and the sample set of all users is merged, is obtained
Eventually for trained sample set, it is denoted as
Third walks:Build neural network model.
Full Connection Neural Network is selected as training pattern.The model is of five storeys neuron in total, and first layer has 10 input god
Through member, last layer has 2 output neurons, and to predict unknown coordinate, intermediate three layers respectively have 50 hidden neuron (values
It can voluntarily be adjusted according to real network complexity by operator).Training error is calculated using Mean Square Error, training pace
It is set as 0.001 (value can voluntarily be adjusted by operator according to real network complexity).Using common positive biography during training
It is that broadcasting method is once trained as a result, using reverse transmission method update neural network in parameter.
4th step:Input data and training pattern.
1) the set L that second step obtains is chosentrainIn a sample, sample is denoted as dj=(p1,j,p2,j,p3,j,p4,j,
p5,j,p6,j), djIn preceding 5 data (p1,j,p2,j,p3,j,p4,j,p5,j) for inputting, last 1 data p6,jFor with prediction
Results contrast calculates error.
2) preceding 5 data (p of a sample are inputted into neural network1,j,p2,j,p3,j,p4,j,p5,j), each data
It is position data, including two number of latitude coordinates, a total of 10 numbers are for inputting, with the input neuron of neural network one by one
It is corresponding, using the forward-propagating method in neural network once trained as a result, being denoted asThe position predicted
Coordinate.
3) actual value of the prediction coordinate in sample isIt calculates and misses using actual coordinate and prediction coordinate
DifferenceParameter in neural network is updated using reverse transmission method, completes a sample instruction
Practice.
4) by entire sample set LtrainIt substitutes into and trains primary, referred to as one wheel sample set training.In a wheel sample set training
Afterwards, the wheel training error is calculated(EiFor the error of the i-th wheel training, ei,jIt is j-th in the sample set of the i-th wheel training
The training error of sample).More wheel sample set training are carried out, when | Ei+1-Ei| < ec(ecFor minimal error constant, which can be by transporting
Battalion quotient voluntarily adjusts according to real network operating condition) when, training stops.At this point, the parameter update of model finishes, model training
It finishes.
5th step:Collect user data to be predicted, prediction user's subsequent time position.
1) moment to be predicted is set as tpredict, user data is acquired according to step 1, is denoted as L'={ (ui',ti',pi')}
(1≤tpredict-ti'≤5) represent collecting sample at the time of (value can be by operator according to reality within 5 minutes with prediction time
Internet operating condition voluntarily adjusts).
2) according to the step integral data in second step, it is denoted as set Lpredict={ (p1',i,p'2,i,p'3,i,p'4,i,
p'5,i), sample set at this time is only inputted, is not exported, therefore each sample only has 5 position datas for predicting.
6th step:The number of calculation base station future service user, control base station switch.
1) information of base station is denoted as setWherein,For the longitude coordinate of base station location, yi cFor
The latitude coordinate of base station location, numiFor the number of base station future service user, it is initialized as 0.
2) the prediction result set obtained from the 5th stepIn, sample is chosen successivelyCalculate its distance with base station each in set C
Obtain the number of the base station nearest with the sampleThe corresponding num in set CiIt is upper to add 1.
3) threshold value of control base station switch is set as numc(value can be voluntarily true according to real network operating condition by operator
It is fixed), all base station datas in set C are traversed, work as numi≥numcWhen, open respective base station i;Work as numi≤numc, close and correspond to
Base station i.
The technical means disclosed in the embodiments of the present invention is not limited only to the technological means disclosed in the above embodiment, further includes
By more than technical characteristic arbitrarily the formed technical solution of combination.It should be pointed out that for those skilled in the art
For, various improvements and modifications may be made without departing from the principle of the present invention, these improvements and modifications are also considered as
Protection scope of the present invention.