Base station load balancing method based on artificial intelligence

Abstract

The invention provides a base station load balancing method based on artificial intelligence, which comprises the following steps: determining a to-be-loaded area, determining to-be-distributed users according to the to-be-loaded area, determining a physical coverage range of a cell according to parameters configured by the cell so as to determine a same coverage area, and then reporting user information to a server. The reported user information is used for artificial intelligence training so as to obtain the user rate, load balancing is carried out through the user rate so as to realize balanced distribution of the users, and finally the artificial intelligence model is corrected so as to enable the artificial intelligence model to be optimal. According to the method, the users are balanced to the specific cell after artificial intelligence learning, the purposes of fast load balancing and optimization of the user rate and the cell flow are achieved, maximum utilization of resources is achieved, and the problem of long user migration time is solved.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communications, and in particular relates to a base station load balancing method based on artificial intelligence. Background Art

[0002] The existing base station connection state load balancing is generally a load balancing method based on the number of users or PRB. When the number of users or PRB reaches a certain set threshold, the load balancing mechanism is started. During balancing, it can be a blind switch or a measurement-based switching strategy to balance users in relatively high-load cells to low-load cells. However, this method cannot achieve user rate optimization, that is, the user experience is not necessarily optimal. The present invention mainly solves this type of problem. This solution is based on the idea of ​​artificial intelligence, performs dynamic learning, and then balances users to appropriate cells to achieve optimal user experience.

[0003] The current load balancing is based on the number of users or cell PRBs, that is, the base station will determine the number of users or PRBs in each cell. When the set threshold is met, the user is migrated to a cell with less load or less users by using measurement or blind addition. However, the user experience may not be optimal, and the cell spectrum resources may not be maximized. The existing technology cannot optimize the user experience in the cell, and the spectrum resources may not be maximized, and the user migration time is long. Summary of the invention

[0004] The purpose of the present invention is to solve the problem of spectrum resource utilization and the problem of long user migration time in the prior art.

[0005] To achieve the above object, the present invention provides a base station load balancing method based on artificial intelligence, comprising the following steps:

[0006] S1. Determine the area to be loaded

[0007] According to the load conditions of the cells, an area for load balancing is selected from the overlapping areas of multiple cells;

[0008] S2. Determine the user to be assigned

[0009] Determine the user to be allocated according to whether the user's terminal falls in the area to be load balanced in step S1;

[0010] S3. Determine the same coverage area

[0011] The physical coverage of the cell is determined according to the parameters configured in the cell. Once the parameters are determined, the actual coverage of the base station is constant. Then, according to the access location information of the user to be allocated in step S2, it can be determined whether the user to be allocated belongs to the same coverage area. For the user to be allocated in the same coverage area, it is placed in the queue A_list to be loaded and balanced.

[0012] S4. User information reporting

[0013] The user to be allocated enters the queue A_list, and the user to be allocated requires reporting the RSRP, RSRQ, CQI, DLMCS, BLER, BSR, cell bandwidth BW, and cell PRB of the area and the same coverage area in step S3 to the designated server;

[0014] S5. Use artificial intelligence model to learn user rate

[0015] S501. Clean the reported data, remove abnormal data, and standardize the data;

[0016] S502, classifying data using a feedforward neural network;

[0017] S503, obtaining the user rate under each combination, taking the arithmetic mean of all elements in the 100 classes obtained in step S502, and obtaining the theoretical rate value under the combination; using the artificial intelligence model to obtain the user rate set K1, K2, K3...Kn and the theoretical rate value under all combinations;

[0018] S504: Predicting the real user rate, based on the information of neighboring cells and the area reported by the users to be allocated in the same coverage area, and then obtaining the historical theoretical user rate through artificial intelligence;

[0019] S6. Load balancing method

[0020] First, a user to be allocated to be load balanced is assigned to obtain the user rate of each physical cell in the load balancing area by using the artificial intelligence model according to step S5, and the cell with the largest rate is selected for balancing;

[0021] S7. AI model correction and learning

[0022] Use artificial intelligence to balance users to M cells, and the theoretical rate is F m , after balancing to M, the actual user rate F, when |FF m |≤v, the artificial intelligence model is excellent, otherwise the artificial intelligence model needs to be corrected. The correction method is to re-select the load balancing cell based on the information reported by the UE.

[0023] Furthermore, the BW sizes of the cell broadbands are different and need to be uniformly converted to one broadband.

[0024] Furthermore, the input layer of the feedforward neural network is 8 features of rsrp, rsrq, cqi, mcs, bsr, bw, bler, and prb; the details are as follows:

[0025]

[0026] v i =u i +θ i

[0027] y i =f(v i )

[0028] x i For each input component, w ij is the weight parameter, θ i is the bias, f is the activation function, y i is the output result, v i To increase the bias output, u i is the output result of layer i;

[0029] Output the prediction results by forward propagation:

[0030]

[0031] in is the prediction result, w is the weight matrix, x is the vector, and b is the bias term;

[0032] Calculating the loss function Evaluate the size of the gap between the predicted value and the true value;

[0033] Update the value of the weight parameter w through the back-propagation neural network as follows:

[0034]

[0035] Where r is the learning rate, is the gradient, L is the loss function;

[0036] The final output layer has 100 categories, i.e. 100 neurons, and each output result After softmax regression, we get the probability value of the category to which the sample belongs. The larger the probability, the greater the possibility of belonging to a certain category. The specific calculation formula of softmax is:

[0037]

[0038] Beneficial effects:

[0039] The present invention proposes a new load balancing method, which performs artificial intelligence learning based on the historical information (mcs, cqi, user rate, etc.) reported by users of the entire network, and then balances the users to specific cells after artificial intelligence learning based on the information of each neighboring cell reported by the users to be load balanced in the same coverage area, thereby achieving rapid load balancing and optimizing the user rate and cell traffic. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is an overall flow chart provided by an embodiment of the present invention;

[0041] Figure 2 is a diagram of the same coverage area of ​​different cells provided by an embodiment of the present invention;

[0042] Figure 3 It is a structural diagram of a feedforward neural network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0044] The application principle of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0045] Embodiment 1:

[0046] like Figure 1 As shown, a base station load balancing method based on artificial intelligence includes the following steps:

[0047] S1. Determine the area to be loaded

[0048] The area to be load balanced refers to the overlapping area of ​​multiple cells. According to the load conditions of the cells, the area to be load balanced is selected from the overlapping areas of the multiple cells.

[0049] S2. Determine the user to be assigned

[0050] The users to be allocated refer to the set of all users whose terminals fall into the area to be load balanced. The number of users to be allocated is determined based on the same coverage area.

[0051] S3. Determine the same coverage area

[0052] like Figure 2 As shown in the figure, the same coverage area refers to different cells of a base station covering the same physical area. The specific method for calculating the same coverage area is as follows:

[0053] The physical coverage of the cell is determined according to the parameters of the cell configuration (power, antenna height, azimuth, downtilt, cell bandwidth). Once the parameters are determined, the actual coverage of the base station can be considered constant. For example: the signal configuration is 18.2db, the cell bandwidth is 20M, the antenna is a 4-port transmission, the azimuth and downtilt of the antenna are 20 degrees and 2 degrees respectively, and the height is 40 meters. At this time, the cell coverage is fixed, 3.4 kilometers away from the base station location, and the longitude and latitude information can be calculated in combination with the location of the base station.

[0054] Then, according to the access location information (longitude and latitude) of the user (UE) to be allocated in step S2, it can be determined whether the UE belongs to the same coverage area. If so, load balancing can be performed to determine the number of users to be allocated.

[0055] For UEs in the same coverage area, they are placed in a queue A_list to be loaded and balanced, that is, A_list is a set of users in the overlapping area between M cells.

[0056] S4. User information reporting

[0057] As shown in Table 1, the information of RSRP, RSRQ, CQI, DLMCS, BLER, BSR, cell broadband BW, cell PRB, and user rate reported by the global network UE in this area is stored in the server for the data source of artificial intelligence training. When the user enters the A_list queue, the UE requires to report the RSRP, RSRQ, CQI, DLMCS, BLER, BSR, cell broadband BW, and cell PRB of this area and the same coverage area in step S3 to the designated server, where this area refers to the area where the terminal initially accesses.

[0058] Table 1 User information table in historical community

[0059]

[0060] Please note that because the cell bandwidth sizes are different, the speed needs to be converted in the end, that is, converted to one broadband. For example, user A is in a 10M cell and the final speed is 8M. User B is under the same conditions, but the cell bandwidth is 20M and the final speed is 17M. In this case, the equivalent user efficiency of cell A is still 8M, and the equivalent user efficiency of cell B is 17 / 2=8.5M. If the conversion is not performed, the sizes cannot be directly compared.

[0061] When training the artificial intelligence model, it is required to use the information of all terminals for learning. However, once the UE enters the A_list queue, it needs to report all cells in the same coverage area (that is, the designated information of the cell that has been connected and all neighboring cells, specifically RSRP, RSRQ, CQI, DLMCS, BLER, BSR, cell broadband BW, cell PRB).

[0062] S5. Use artificial intelligence model to learn user rate

[0063] S501. Clean the reported data, remove abnormal data, and standardize the data. Delete the reported RSRP, RSRQ, CQI, DLMCS, BLER, BSR, cell bandwidth BW, and cell PRB that exceed the theoretical value. The following is the theoretical range of each parameter: -44>RSRP>-140, -19.5 <RSRQ<-3,1<=CQI<=15,0<=MCS<=31,0<=BLER<=0.2,BSR> 0,0 <PRB<=1。

[0064] S502, using a feedforward neural network to classify the data, such as Figure 3 shown.

[0065] The input layer consists of eight features: rsrp, rsrq, cqi, mcs, bsr, bw, bler, and prb. The output is initially determined to be 100 categories. The specific 100 categories are divided according to the rate from 0 to 100 with a step size of 1, such as 0-1 is the first category, 1-2 is the second category, 2-3 is the third category, and so on.

[0066] First, a set of weight matrices w is initialized by training 20,000 historical sample data with 8 features per sample ij , initialize x to 10, so that after vector x and weight matrix w ij The product calculation can get u i In order to make the result closer to reality, we add the bias term θ i ,θ i The initial implementation scheme uses random numbers to generate values ​​between 0 and 1. Then, in order to make the training values ​​of each layer of the neural network relatively reliable, the activation function f is used to change it to nonlinear. This paper uses the activation function Relu function of the feedforward neural network, that is, the output result v of taking 0 and adding bias i The maximum value, such as v i =-2, the final output is 0, otherwise it is v i , in order to train w ijThe matrix is ​​continuously iterated using the gradient descent method to update w. The last layer obtains a 100-dimensional vector through f, and then calculates a 100-dimensional vector through the softmax function. The sum of all vector elements is 1, and the largest value is the final classification to which this sample belongs.

[0067] The activation function of the feedforward neural network is the Relu function f(x)=max(0,x). The specific feedforward neural network is as follows:

[0068]

[0069] v i =u i +θ i (2)

[0070] y i =f(v i ) (3)

[0071] x j For each input component, w ij is the weight parameter, θ i is the bias, f is the activation function, y i is the output result, v i To increase the bias output, u i is the output result of layer i.

[0072] Output the prediction results by forward propagation:

[0073]

[0074] in is the prediction result, w is the weight matrix, x is the vector, and b is the bias term.

[0075] Calculating the loss function Evaluate the gap between the predicted value and the true value, that is, the gap between the predicted classification and the true classification.

[0076] The value of the weight parameter w is updated through the back-propagation neural network using the gradient descent principle, as shown in the following formula:

[0077]

[0078] Where r is the learning rate, is the gradient and L is the loss function.

[0079] The final output layer has 100 categories, that is, 100 neurons, and each output result After softmax regression, we get the probability value of the category to which the sample belongs. The larger the probability, the greater the possibility of belonging to a certain category. The specific calculation method of softmax is as follows, so that the output value is converted into a probability value.

[0080]

[0081] S503: Obtaining the user rates under each combination

[0082] As shown in Table 2, the arithmetic mean of all elements in the 100 classes obtained in step S502 is taken to obtain the theoretical rate value under the combination; the artificial intelligence model is used to obtain the user rate set K1, K2, K3...Kn and the theoretical rate value under all combinations; for example, after classification, a certain class A1 is obtained, which has 3 records, and the rates of these 3 records are as shown in the last column of the following table, then the final rate combination K1, so the average rate is (6+6.4+6.3) / 3=6.23.

[0083] Table 2 Similar user information table

[0084]

[0085] S504: Real User Rate Prediction

[0086] According to the information of neighboring cells and the current area reported by UE in the same coverage area, the historical theoretical user rate is obtained through the artificial intelligence model. Assuming there are 3 neighboring cells with the same coverage, the rates f1, f2, and f3 are learned respectively according to the reported information. That is to say, the artificial intelligence model can calculate the theoretical rate value that each terminal can achieve under the same conditions in history. Then, in actual use, according to the information of the current area and neighboring cells reported by the terminal, it is possible to find out which cell has the best load balancing, so that the balancing efficiency is also the fastest.

[0087] S6. Load balancing method

[0088] Step S5 has output how to use the artificial intelligence model to obtain the user rate under given conditions. Assuming that a UE has n neighboring cells with the same coverage at the same time, and the converted rates are f1, f2, ... fn, then take the cell M corresponding to max(f1, f2, ... fn), and balance the user to the M cell. That is, for a certain UE to be load balanced, first use the artificial intelligence model to obtain the rates of each physical cell in the area to be load balanced, and take the cell with the largest rate for balancing. For example, the public area where UE1 is located has two cells A and B. Through artificial intelligence, the rates of UE1 in A and B are 4M and 4.5M respectively, then the user is balanced to the B cell;

[0089] S7. AI model correction and learning

[0090] Because users change in real time, it is necessary to continuously learn and correct the AI ​​model. Assume that AI is used to balance users to M cells, and the theoretical rate is Fm. After balancing to M, the actual rate of users is F. When |FF m |≤v(error value), the artificial intelligence model is excellent, otherwise the artificial intelligence model needs to be corrected. The correction method is to re-select the load balancing cell based on the information reported by the UE.

[0091] The explanations of relevant professional terms used in this application are shown in Table 3.

[0092] Table 3 List of proper nouns

[0093]

[0094] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A base station load balancing method based on artificial intelligence, characterized in that: The steps include: S1. Determine the area to be loaded According to the load conditions of the cells, an area for load balancing is selected from the overlapping areas of multiple cells; S2. Determine the user to be assigned Determine the user to be allocated according to whether the user's terminal falls in the area to be load balanced in step S1; S3. Determine the same coverage area The physical coverage of the cell is determined according to the parameters configured in the cell. Once the parameters are determined, the actual coverage of the base station is constant. Then, according to the access location information of the user to be allocated in step S2, it can be determined whether the user to be allocated belongs to the same coverage area. For the user to be allocated in the same coverage area, it is placed in the queue A_list to be loaded and balanced. S4. User information reporting The user to be allocated enters the queue A_list, and the user to be allocated requires reporting the RSRP, RSRQ, CQI, DLMCS, BLER, BSR, cell bandwidth BW, and cell PRB of the area and the same coverage area in step S3 to the designated server; S5. Use artificial intelligence model to learn user rate S501. Clean the reported data, remove abnormal data, and standardize the data; S502, classifying data using a feedforward neural network; S503, obtaining the user rate under each combination, taking the arithmetic mean of all elements in the 100 classes obtained in step S502, and obtaining the theoretical rate value under the combination; using the artificial intelligence model to obtain the user rate set K1, K2, K3...Kn and the theoretical rate value under all combinations; S504: Predicting the real user rate, based on the information of neighboring cells and the area reported by the users to be allocated in the same coverage area, and then obtaining the historical theoretical user rate through artificial intelligence; S6. Load balancing method First, a user to be allocated to be load balanced is assigned to obtain the user rate of each physical cell in the load balancing area by using the artificial intelligence model according to step S5, and the cell with the largest rate is selected for balancing; S7. AI model correction and learning Use artificial intelligence to balance users to M cells, and the theoretical rate is F m , after balancing to M, the actual user rate F, when |FF m |≤v, the artificial intelligence model is excellent, otherwise the artificial intelligence model needs to be corrected. The correction method is to re-select the load balancing cell based on the information reported by the UE.

2. The base station load balancing method based on artificial intelligence according to claim 1, characterized in that: In step S4, the cell broadbands BW are different in size and need to be uniformly converted to one broadband.

3. The base station load balancing method based on artificial intelligence according to claim 1, characterized in that: The input layer of the feedforward neural network in step S502 is 8 features of rsrp, rsrq, cqi, mcs, bsr, bw, bler, and prb; specifically: v i =u i +θ i y i =f(v i ) x i For each input component, w ij is the weight parameter, θ i is the bias, f is the activation function, y i is the output result, v i To increase the bias output, u i is the output result of layer i; Output the prediction results by forward propagation: in is the prediction result, w is the weight matrix, x is the vector, and b is the bias term; Calculating the loss function Evaluate the size of the gap between the predicted value and the true value; Update the value of the weight parameter w through the back-propagation neural network as follows: Where r is the learning rate, is the gradient, L is the loss function; The final output layer has 100 categories, i.e. 100 neurons, and each output result After softmax regression, we get the probability value of the category to which the sample belongs. The larger the probability, the greater the possibility of belonging to a certain category. The specific calculation formula of softmax is:

Images