A method for expanding broadband capacity of a residential area

Abstract

The invention discloses a cell broadband capacity expansion method, aiming to overcome the problems of low capacity expansion efficiency and high cost in the prior art. The method comprises the following steps: a processor obtains a user onu rate time series and the onu rates of all users in the cell; the user onu rate time series is substituted into a time series prediction method to predict the user's future onu rate; feature learning is performed on the onu rates of all users in the cell to obtain initial derived features, and then the user's future onu rate and the initial derived features are combined to obtain derived features; an XGBOOST model is trained according to the derived features and the onu rates of all users in the cell, and a final predicted onu rate of the user is predicted according to the XGBOOST model; the cell expansion urgency is determined according to the final predicted onu rate of the user, and the cell is expanded according to the cell expansion urgency.

Description

Technical Field

[0001] The present invention relates to communication technology, and in particular to a method for expanding the broadband capacity of a cell. Background Art

[0002] With the development of the digital age, the number of network-connected devices owned by users has gradually increased. With the development of mobile content services such as short videos, live broadcasts, and mobile games, users' demands for broadband network quality have gradually increased, especially in terms of network width and network speed.

[0003] Broadband network is a communication bridge built between network service providers and users. It includes backbone computer rooms, community computer rooms and user terminals that are connected in sequence. User access devices access the network through user terminals. Community computer rooms serve as the connection point between users and backbone computer rooms. The broadband network quality requirements of user access devices in the same community are superimposed on the community computer rooms. Therefore, the broadband network quality requirements required by users in the same community are to identify and predict the traffic requirements of the community computer room, determine in advance whether to expand and optimize the network of the community computer room, and then make the network capacity planning of the community computer room keep up with the development of user needs, thereby ensuring and improving the user's perceived speed and experience effect.

[0004] To this end, for technical personnel in this field, in order to maximize user retention rate and meet users' needs for high-speed and uninterrupted service, how to identify and predict the traffic demand of community computer rooms through big data technology, and convert the traffic demand into a basis for expanding the capacity of community computer rooms, and analyze the root causes of the expansion of community computer rooms to achieve efficient and low-cost expansion are issues that need to be urgently addressed. Summary of the invention

[0005] In order to overcome the deficiencies and problems of the prior art, the present invention provides a method for expanding the broadband capacity of a cell.

[0006] To achieve the above object, the present invention adopts the following technical solution:

[0007] A method for expanding the capacity of a cell broadband comprises the following steps:

[0008] Step 1: Obtain the user onu rate time series and the onu rates of all users in the cell, where the user onu rate time series is represented by U = {u1,u2,…,u n}, u1 is the user onu rate obtained for the first time, u2 is the user onu rate obtained for the second time...u n is the user onu rate obtained for the nth time. The onu rate of all users in the cell adopts F = {U1, U2, ..., U v} indicates that U1 is the onu rate time series of the first user in the cell, and U2 is the onu rate time series of the second user in the cell...U v is the onu rate time series of the vth user in the cell, v is the total number of users in the cell;

[0009] Step 2: Substitute the user's onu rate time series into the time series prediction method to predict the user's future onu rate;

[0010] Step 3: Perform feature learning onu rates of all users in the cell to obtain initial derived features, and then combine the future onu rates of the users with the initial derived features to obtain derived features, where the derived features are X = {x1, x2, …, x v} indicates that x1 is the derived feature corresponding to U1, x2 is the derived feature corresponding to U2…x n For U v The corresponding derived features;

[0011] Step 4: Train the XGBOOST model based on the derived features and the onu rates of all users in the cell, and predict the final predicted onu rate of the user based on the XGBOOST model;

[0012] Step 5: Determine the urgency of cell expansion according to the final predicted ONU rate of the user, and expand the cell according to the urgency of cell expansion.

[0013] Preferably, the time series prediction method adopts the ARIMA prediction method, which specifically includes the following steps:

[0014] Step 21: determine whether the user onu rate time series is a stationary time series. If so, calculate the autocorrelation coefficient and the partial autocorrelation coefficient based on the user onu rate time series which is a stationary time series. If not, perform differential transformation on the user onu rate time series to obtain the user onu rate time series which is a stationary time series, and then calculate the autocorrelation coefficient and the partial autocorrelation coefficient.

[0015] Step 22: Determine the value range of the autoregressive coefficient according to the maximum lag period of the autocorrelation coefficient, and determine the value range of the moving average order according to the maximum lag period of the partial autocorrelation coefficient;

[0016] Step 23: Substitute the value range of the autoregressive coefficient and the value range of the moving average order into the minimum information criterion method to calculate the optimal autoregressive coefficient and moving average order, and establish the ARIMA model based on the optimal autoregressive coefficient and moving average order, where the autoregressive coefficient is represented by p and the moving average order is represented by q;

[0017] Step 24: Perform a validity test on the ARIMA model. If the validity test passes, substitute the user's onu rate time series into the ARIMA model to predict the user's future onu rate.

[0018] Preferably, the ARIMA model in step 23 is expressed by the following formula:

[0019] u rt =α1u r(t-1) +α2u r(t-2) +…+α p u r(t-p) +ε t +β1ε t-1 +β2ε t-2 +…+β q ε t-q ,

[0020] Where n=rt, r is a parameter, t is the time when the user onu rate is obtained, α1…α p u r(t-1) …u r(t-p) The parameters of β1…β q They are ε t-1 …ε t-q The parameter, ε t …ε t-q is the white noise sequence from time t to time tq.

[0021] Preferably, the step 21 specifically includes the following steps:

[0022] Substitute the user onu rate time series into the following formula to calculate the autocorrelation coefficient:

[0023]

[0024] Wherein, k is the lag period, AC(k) is the autocorrelation coefficient when the lag period is k, n=rt, r is a parameter, t is the acquisition time of the user onu rate, m is the number of acquisitions of the user onu rate between time tk and time t, It is the average onu rate of all users in the cell.

[0025] Preferably, the step 21 specifically includes the following steps:

[0026] Substitute the user onu rate time series into the following formula to calculate the partial autocorrelation coefficient:

[0027]

[0028] Wherein, k is the lag period, PAC(k) is the partial autocorrelation coefficient of the lag period k, cov is the covariance treatment, var is the variance treatment, n=rt, r is a parameter, t is the time when the user onu rate is obtained, is the average onu rate of all users in the cell at time t, It is the average onu rate of all users in the cell at time tk.

[0029] Preferably, the step 3 specifically includes the following steps:

[0030] Step 31: Determine the acquisition time of the user onu rate according to the user onu rate time series, judge whether the acquisition time of the user onu rate is a holiday, and obtain the holiday characteristics;

[0031] Step 32: Substitute the user's onu rate time series into the sliding time window method to calculate the trend preference feature;

[0032] Step 33: The processor substitutes the user onu rate time series into the Lag lag method to calculate the lag feature;

[0033] Step 34: Substitute the user onu time series into the trend learning method to calculate the trend feature;

[0034] Step 35: Combine holiday features, trend preference features, lag features and trend features to obtain preliminary derived features;

[0035] Step 36: Combine the preliminary derived features and the user's future onu rate to obtain the derived features.

[0036] Preferably, the step 4 specifically includes the following steps:

[0037] Step 41: Calculate the monthly maximum onu ​​rate set of users in the cell according to the onu rates of all users in the cell, where the monthly maximum onu ​​rate set of users in the cell adopts Y = {y1, y2, ..., y v} indicates that y1 is the monthly maximum ONU rate of the first user in the cell, and y2 is the monthly maximum ONU rate of the second user in the cell. v is the monthly maximum onu ​​rate of the vth user in the cell;

[0038] Step 42: Substitute the monthly maximum ONU rate set of users in the cell into the following model complexity leaf weight function to calculate the leaf weight of the optimal model complexity:

[0039]

[0040]

[0041]

[0042] In the formula, j is a leaf node, ω j is the weight of leaf node j, g i is the first-order partial derivative of y contained in leaf node j, h i is the second-order partial derivative of y contained in leaf node j, α is a hyperparameter, is the partial differential symbol, y i (τ-1) for y i The predicted value of the τ-1th iteration, for y i (τ-1) The first derivative of for y i (τ-1) The second-order partial derivative of is the loss degree of the i-th user, y i is the monthly maximum onu ​​rate of the ith user, is the predicted value of the monthly maximum onu ​​rate of the i-th user τ-1 iterations, where τ is the number of iterations;

[0043] Step 43: Substitute the monthly maximum ONU rate set of users in the cell into the following loss function to calculate the minimum loss:

[0044]

[0045] Where L is the loss degree, T is the number of leaf nodes, j is a leaf node, j is between 1 and T, α and γ are both hyperparameters, g i is the first-order partial derivative of the samples contained in leaf node j, h i is the second-order partial derivative of the samples contained in leaf node j;

[0046] Step 44: Substitute the leaf weights, minimum loss and derived features of the optimal model complexity into the following objective function to obtain the optimized XGBOOST model, and predict the user's final predicted ONU rate based on the optimized XGBOOST model:

[0047]

[0048] Where v is the total number of users in the cell, the sequence number of the user in cell i, and Obj (t) is the objective function of the τth iteration, f τ (x i ) is the x of the τth iteration i The predicted value, f τ 2 (x i ) is fτ (x i ) to the power of .

[0049] Preferably, the step 5 specifically includes the following steps:

[0050] Step 51: Set a user bandwidth configuration data set, wherein the user bandwidth configuration data set adopts B = {b1, b2, ..., b v}, b1 is the bandwidth configuration data of the first user in the cell, b2 is the bandwidth configuration data of the second user in the cell...b v is the bandwidth configuration data of the vth user in the cell, the user bandwidth configuration data is determined according to the bandwidth service subscribed by the user from the operator, and the user bandwidth excess rate set is determined according to the user's final predicted ONU rate and the user bandwidth configuration data set;

[0051] Step 52: Set a user broadband service quality score set, where the user broadband service quality score set adopts R = {r1, r2, ..., r v} indicates that r1 is the user broadband service quality score of the first user in the cell, r2 is the user broadband service quality score of the second user in the cell, and r v Score the user broadband service quality of the vth user in the cell;

[0052] Step 53: normalize the user's bandwidth excess rate feature and the user's broadband service quality score feature, and substitute them into the following expansion urgency formula to calculate the cell expansion urgency:

[0053]

[0054] Where D i is the urgency of cell expansion, i is the sequence number in the indicator sequence, d is the number of items in the indicator sequence, μ i is the weight value of the i-th indicator in the indicator sequence, z i is the ith indicator in the indicator sequence, and the ith indicator is the total value of the normalized user bandwidth excess rate set and the user broadband service quality score set;

[0055] Step 54: Expand the cell according to the urgency of cell expansion.

[0056] Preferably, the step 54 specifically includes the following steps:

[0057] Determine whether the cell expansion urgency is greater than the expansion threshold. If not, do not expand the cell. If yes, arrange all cell expansion urgency values ​​greater than the expansion threshold into an expansion sequence from large to small, and expand the cells in sequence according to the expansion sequence.

[0058] Compared with the prior art, the present invention has the following outstanding and beneficial technical effects:

[0059] In the present invention, the capacity expansion of a cell is based on the traffic demand of each user in the cell, which can more truly reflect the demand for cell capacity expansion.

[0060] In the present invention, the prediction of user onu rate adopts time series prediction method, feature learning and XGBOOST model, which has a higher convergence speed while ensuring the prediction accuracy, especially when using huge user onu rate time series data, it has better prediction accuracy and prediction efficiency.

[0061] In the present invention, the urgency of cell expansion takes into account the user's bandwidth overlimit rate and the user's broadband service quality score, and then expands different cells in stages. Compared with the traditional method of only considering the disorderly expansion of different cells when the cell broadband overlimit rate is too high, the urgency of cell expansion in the present invention can comprehensively consider the urgency of expansion under limited expansion resources and then carry out orderly, efficient and low-cost cell expansion.

[0062] In the present invention, the XGBOOST model is a supervised learning model, which takes the derived features as the supervised input and the user's final predicted onu rate as the supervised output, thereby converting the time series into supervised learning, preventing overfitting by determining the optimal attributes, and also using the first-order derivative and the second-order derivative to make the model loss and complexity controllable, thereby coordinating the computational amount of data processing and the accuracy of the prediction results, and taking into account both the efficiency and accuracy of data processing when predicting the user's onu rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a schematic diagram of the data processing flow of the present invention; DETAILED DESCRIPTION

[0064] To facilitate understanding by those skilled in the art, the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0065] It should be noted that in the field of network bandwidth, an area with an independent cell room and covered by the network bandwidth of the cell room is called a cell, and an individual using a broadband network and communicating connected to a node of an ONU device is called a user. For example, a cell may refer to an urban residential community or a residential community, in which a cell room may be built, which is used to provide network bandwidth to users in the urban residential community or the residential community, and a user may refer to a resident in the urban residential community or the residential community who accesses the network through the cell room. For another example, a cell may also refer to an office building or a commercial area, in which a cell room is built, which is used to provide network bandwidth to users in the office building or the commercial area, and a user may refer to a unit or shop using network broadband in the office building or the commercial area. The cell room includes an ONU device (Optical Network Unit), and the network service provider provides and configures network broadband to users in the cell through the ONU device. For example, if a user subscribes to a broadband service from an Internet service provider, the Internet service provider configures the user's IP address, MAC address, maximum upstream rate, and maximum downstream rate of the broadband through the ONU device.

[0066] This embodiment provides a method for expanding cell broadband, including the following steps:

[0067] Step 1: The processor obtains the user onu rate time series and the onu rates of all users in the cell, where the user onu rate time series is represented by U = {u1,u2,…,u n}, u1 is the user onu rate obtained for the first time, u2 is the user onu rate obtained for the second time...u n is the user onu rate obtained for the nth time. The onu rate of all users in the cell adopts F = {U1, U2, ..., U v} indicates that U1 is the onu rate time series of the first user in the cell, and U2 is the onu rate time series of the second user in the cell...U v is the onu rate time series of the vth user in the cell, v is the total number of users in the cell;

[0068] In the above steps, the user's onu rate time series is arranged according to the moment when the processor obtains the user's onu rate. In the present embodiment, the processor obtains the user's onu rate in an equal interval time mode. The interval between two adjacent acquisition times is set to 5 minutes. For example, the interval between the processor obtaining the user's onu rate for the 1st time and obtaining the user's onu rate for the 2nd time is 5 minutes.

[0069] The processor is set in the ONU device, and the user communication is connected to a node of the ONU device in the community. In this embodiment, the connection mode between the user and the ONU device includes fiber to the home, fiber to the building, and signal transmission through a coaxial cable. In fiber to the home (Fibre To The Home, FTTH), each user corresponds to an ONU device, and the ONU device only provides one node, and the user and the ONU device correspond one to one. In fiber to the building (Fiber To The Building, FTTB), multiple users share an ONU device, and the ONU device provides multiple nodes. The method of signal transmission by coaxial cable (Ethernet Over Cable, EOC) is mainly used in rural areas, and the bandwidth network quality provided is low. In order to facilitate the processor to determine the number and total number of users in the community, the processor numbers the users according to the node position of the ONU device to which the user is connected. The node position of the ONU device can be pre-set in the processor, and the node position of the ONU device is represented by 1, 2...n. The users in the community are connected to the nodes of the ONU device in sequence according to the node position order of the ONU device, and the network service provider configures the network broadband for the corresponding user according to the node position of the ONU device.

[0070] Step 2: The processor substitutes the user's onu rate time series into the time series prediction method to predict the user's future onu rate;

[0071] In the above steps, the time series prediction method refers to a method for predicting the future user's onu rate by performing statistical analysis, appropriate processing and trend prediction based on the user's onu rate time series, and the time series prediction method includes exponential smoothing method, seasonal trend prediction method, weighted moving average method, etc. The user's future onu rate refers to the user's onu rate at a certain point in the future. In the present embodiment, the processor sets the future point in time to be one month in the future.

[0072] Step 3: The processor performs feature learning on the onu rates of all users in the cell to obtain preliminary derived features, and then combines the future onu rates of the users with the initial derived features to obtain derived features, where the derived features are X = {x1, x2, …, x v} means that x1 is the derived feature of U1, x2 is the derived feature of U2…x v For U v Derivative features of

[0073] In the above steps, X represents the derived characteristics of all users in the cell, x1, x2, …, x vThey represent the derived features of the first user, the derived features of the second user, and the derived features of the vth user in sequence. Feature learning refers to a method of learning derived features based on the time series of the onu rates of all users in the cell and the actual features. The actual features include time attribute features, sliding window attribute features, hysteresis features, and the like. The processor can set the actual features in the processor in advance, and the actual features can be determined according to the actual situation. For example, if the time attribute features include holidays, if the time of predicting the user's onu rate is a holiday, then the actual user's future onu rate is likely to be higher than the predicted user's future onu rate. The use of time attribute features including holidays improves the accuracy of the prediction. The sliding window attribute feature refers to the statistical calculation of the user's onu rate in the same period time period based on the time window statistics to obtain its average value or median value, and the obtained rate is equidistantly binned, and the proportion of each gear is calculated. This method is used to facilitate understanding of the user's rate interval preference. The hysteresis feature refers to taking the user's onu rate at a certain moment in the past as an influencing factor of the current rate. Since the user's rate has a certain periodicity, the use of the hysteresis feature improves the accuracy of the user's onu rate that needs to be predicted.

[0074] Step 4: The processor obtains the XGBOOST model based on the user's future onu rate and derived feature training, and predicts the user's final predicted onu rate based on the XGBOOST model;

[0075] In the above steps, the XGBOOST model (eXtremeGradientBoosting) is an implementation of the boosting method, which can reduce the loss and limit the complexity of the model, taking into account both prediction accuracy and prediction efficiency.

[0076] Step 5: The processor determines the urgency of cell expansion according to the final predicted ONU rate of the user, and expands the cell according to the urgency of cell expansion;

[0077] In the above steps, the cell expansion urgency refers to the urgency of the cell to expand the network bandwidth. The cell expansion urgency is expressed in numerical values. The larger the cell expansion urgency value, the more serious the congestion of the cell network broadband and the lower the network service quality, the lower the user's perceived rate, and the worse the user's experience.

[0078] The time series prediction method in step 2 adopts the ARIMA prediction method. The ARIMA (Autoregressive Integrated Moving Average model) prediction method is one of the time series prediction and analysis methods, and specifically includes the following steps:

[0079] Step 21: The processor determines whether the user onu rate time series is a stationary time series. If so, the autocorrelation parameter (AC, AutocorrelationFunction) and the partial autocorrelation coefficient (PAC,

[0080] ParticalAutocorrelationCoefficient), if otherwise, the user onu rate time series is subjected to differential transformation to obtain the user onu rate time series which is a stationary time series, and then the autocorrelation parameter and partial autocorrelation coefficient are calculated according to the user onu rate time series which is a stationary time series;

[0081] In the above steps, a stationary time series refers to a time series in which the mean of the user's onu rate has no systematic change, the variance has no systematic change, and the periodic change is strictly eliminated. Differential transformation refers to a method of transforming the user's onu rate time series, which is used to eliminate the dependence of the user's onu rate time series on time and make the fluctuation curve more stable. The autocorrelation parameter refers to the degree of mutual correlation of the user's onu rate at different time points.

[0082] Step 22: The processor determines a value range of the autoregressive coefficient according to the maximum lag period of the autocorrelation coefficient, and determines a value range of the moving average order according to the maximum lag period of the partial autocorrelation coefficient;

[0083] In the above steps, the maximum lag period can be determined according to the lag order of the user onu rate time series.

[0084] Step 23: The processor substitutes the value range of the autoregressive coefficient and the value range of the moving average order into the minimum information criterion method to calculate the optimal autoregressive coefficient and moving average order, and establishes an ARIMA model according to the optimal autoregressive coefficient and moving average order, wherein the autoregressive coefficient is represented by p and the moving average order is represented by q;

[0085] In the above steps, the AIC (Akaike information criterion) method is a standard method for measuring the goodness of fit of the ARIMA model. Step 22 specifically includes the following steps:

[0086] Step 231: The processor substitutes the value range of the autoregressive coefficient and the value range of the moving average order into the following formula to calculate min(AIC):

[0087] AIC=-2In(L)+2k,

[0088] Where k is the number of parameters, L is the maximum likelihood function of the ARIMA model, and min(AIC) is the minimum value of AIC;

[0089] Step 232: The processor determines the optimal autoregressive coefficient and moving average order, and establishes an ARIMA model according to the optimal autoregressive coefficient and moving average order.

[0090] Step 24: The processor performs a validity check on the ARIMA model. If the validity check passes, the user's onu rate time series is substituted into the ARIMA model to predict the user's future onu rate.

[0091] In the above steps, the validity check refers to the comparison between the data calculated by the processor according to the ARIMA model and the corresponding actual user onu rate time series to obtain the residual, which can also be called white noise. The residual is used to determine whether the data calculated by the ARIMA model is valid and correct.

[0092] The step 24 specifically includes the following steps:

[0093] Substitute the autoregressive coefficient and the moving average order into the residual formula to calculate the residual:

[0094] ε t =u rt -α1u r(t-1) -α2u r(t-2) -…-α p u r(t-p) -β1ε t-1 -…-β q ε t-q ,

[0095] Where n = rt, r is a parameter, t is the time when the user onu rate is obtained, and u rt is the user onu rate time series obtained at time t, u r(t-1) is the user onu rate time series obtained at time t-1, ...u r(t-p) is the user onu rate time series obtained at time tp, α1, α2…α p u r(t-1) 、u r(t-2) …u r(t-p) The parameters of β1…β q They are β1…β q The parameter, ε t , ε t-1 …ε t-q is the white noise sequence from time t to time tq, ε t Also called residual.

[0096] The ARIMA model in step 23 is expressed by the following formula:

[0097] u rt =α1u r(t-1) +α2u r(t-2) +…+α p u r(t-p) +ε t +β1ε t-1 +…+β q εt-q,

[0098] Where n=rt, r is a parameter, t is the time when the user onu rate is obtained, α1…α p u r(t-1) …u r(t-p) The parameters of β1…β q They are ε t-1 …ε t-q The parameter, ε t …ε t-q is the white noise sequence from time t to time tq.

[0099] The step 21 specifically includes the following steps:

[0100] The processor substitutes the user onu rate time series into the following formula to calculate the autocorrelation coefficient:

[0101]

[0102] Wherein, k is the lag period, AC(k) is the autocorrelation coefficient when the lag period is k, n=rt, r is a parameter, t is the acquisition time of the user onu rate, m is the number of acquisitions of the user onu rate between time tk and time t, is the average onu rate of all users in the cell;

[0103] In the above steps, the above formula represents the correlation between the user's onu rate at time tk and time t. In actual use, the processor can use the above formula to represent the influence of the past user's onu rate on the current user's onu rate by setting the values ​​of t and k. The average onu rate of all users in the cell is equal to the sum of the onu rate time series from the first user to the onu rate time series of the vth user in the cell divided by v.

[0104] The step 21 further specifically includes the following steps:

[0105] The processor substitutes the user onu rate time series into the following formula to calculate the partial autocorrelation coefficient:

[0106]

[0107] Wherein, k is the lag period, PAC(k) is the partial autocorrelation coefficient of the lag period k, cov is the covariance treatment, var is the variance treatment, n=rt, r is a parameter, t is the time when the user onu rate is obtained, is the average onu rate of all users in the cell at time t, is the average onu rate of all users in the cell at time tk;

[0108] In the above steps, the average value of the onu rate of all users in the cell at time t is equal to the sum of the onu rate time series from the first user to the onu rate time series of the vth user in the cell at time t divided by v. The average value of the onu rate of all users in the cell at time tk is equal to the sum of the onu rate time series from the first user to the onu rate time series of the vth user in the cell at time tk divided by v.

[0109] In step 3, the feature learning method includes a time derivative method, a sliding time window method, a Lag lag method and a trend learning method, which specifically includes the following steps:

[0110] Step 31: The processor determines the acquisition time of the user onu rate according to the user onu rate time series, determines whether the acquisition time of the user onu rate is a holiday, and obtains the holiday feature;

[0111] The above step 31 adopts the time derivative method. When the processor obtains the user's onu rate, it may be a holiday time. The actual user's onu rate mutates, which does not conform to the trend of the user's onu rate time series. The actual future user's onu rate and the predicted corresponding user's onu rate have a huge deviation, which leads to a large prediction error. Holiday time includes statutory holidays and shopping promotion days. Statutory holidays include the Spring Festival, Dragon Boat Festival and Mid-Autumn Festival. Shopping promotion days include the JD 618 store anniversary and Tmall Double 11 shopping carnival defined by e-commerce. Therefore, the processor pre-sets a holiday list, which is determined in combination with the traffic data trend of historical holidays. The timing unit of the acquisition moment of the user's onu rate includes at least one of year, month, day and week.

[0112] Step 32: The processor substitutes the user's onu rate time series into the sliding time window method to calculate the trend preference feature;

[0113] The above step 32 adopts a sliding time window method, which specifically includes the following steps: the processor sets a time window, the time window has a certain preset time, the user onu rate time series enters the time window in chronological order, the processor counts the trend preference of the user onu rate time series in the time window, and then obtains the trend preference feature based on the trend preference.

[0114] Step 33: The processor substitutes the user onu rate time series into the Lag lag method to calculate the lag feature;

[0115] The above step 33 adopts the Lag hysteresis method, which specifically includes the following steps: the processor sets the hysteresis time M, and uses the user onu rate at time TM as the hysteresis rate of the user onu rate at time T, and uses this rate as the hysteresis feature.

[0116] Step 34: The processor substitutes the user onu time series into the trend learning method to calculate the trend feature;

[0117] The above step 34 adopts a trend learning method, which specifically includes the following steps: the processor counts the median of the user's onu rate in the past period of time in the user's onu time series, and the past period of time may be the past 1 month, 3 months, 6 months, 9 months, or 12 months. The processor then determines the user's onu trend according to the user's onu rate at the current moment and the median of the user's onu rate in the past period of time. The user's onu rate trend includes a downward trend, a stable trend, and an upward trend of the user's onu rate, and the above result is used as a trend feature;

[0118] Step 35: The processor optimizes the user onu time series according to the holiday features, trend preference features, lag features and trend features to obtain preliminary derived features.

[0119] Step 36: The processor combines the preliminary derived features and the user's future onu rate to obtain the derived features;

[0120] In the above steps, merging the preliminary derived features and the future onu rate of the user means that the processor performs a merge process on the preliminary derived features and the future onu rate of the user.

[0121] The step 4 specifically includes the following steps:

[0122] Step 41: The processor calculates the monthly maximum onu ​​rate set of users in the cell according to the onu rates of all users in the cell, wherein the monthly maximum onu ​​rate set of users in the cell adopts Y = {y1, y2, ..., y v} indicates that y1 is the monthly maximum ONU rate of the first user in the cell, and y2 is the monthly maximum ONU rate of the second user in the cell. v is the monthly maximum onu ​​rate of the vth user in the cell;

[0123] In the above steps, the user's monthly maximum onu ​​rate set is the true value of the user's onu rate, and the user's monthly maximum onu ​​rate set can be used to test subsequent formulas and models.

[0124] Step 42: The processor substitutes the monthly maximum onu ​​rate set of users in the cell into the following model complexity leaf weight function to calculate the leaf weight of the optimal model complexity:

[0125]

[0126]

[0127]

[0128] In the formula, j is a leaf node, ω j is the weight of leaf node j, g i is the first-order partial derivative of y contained in leaf node j, h i is the second-order partial derivative of y contained in leaf node j, α is a hyperparameter, is the partial differential symbol, for y i (τ-1) The first derivative of for y i (τ-1) The second-order partial derivative of is the loss degree of the i-th user, y i is the monthly maximum onu ​​rate of the ith user, is the predicted value of the monthly maximum onu ​​rate of the i-th user τ-1 iterations, where τ is the number of iterations;

[0129] Step 43: The processor substitutes the monthly maximum ONU rate set of users in the cell into the following loss function to calculate the minimum loss:

[0130]

[0131] Where L is the loss degree, T is the number of leaf nodes, j is a leaf node, j is between 1 and T, α and γ are both hyperparameters, g i is the first-order partial derivative of the samples contained in leaf node j, h i is the second-order partial derivative of the samples contained in leaf node j;

[0132] Step 44: The processor substitutes the leaf weights, minimum loss degree and derived features of the optimal model complexity into the following objective function to obtain the optimized XGBOOST model, and predicts the user's final predicted ONU rate based on the optimized XGBOOST model:

[0133]

[0134] Where v is the total number of users in the cell, the sequence number of the user in cell i, and Obj (t) is the objective function of the τth iteration, f τ (x i ) is the x of the τth iteration i The predicted value of Obj is generated by τ iterations and τ trees. τ 2 (x i ) is f τ (x i ) to the power of

[0135] In the above steps, the XGBOOST model is derived through the loss function and complexity function. The processor sets the loss function as follows:

[0136]

[0137] Where L(F) is the predicted loss of onu rate of all users in the cell, y i is the monthly maximum onu ​​rate of the ith user in the cell, is the predicted value of the monthly maximum onu ​​rate of the i-th user, i is the sequence number of the user in the cell, v is the total number of users in the cell, is the loss degree of the i-th user.

[0138] The processor sets the following complexity function:

[0139]

[0140]

[0141] Where, Ω(F) is the complexity of the onu rate prediction of all users in the cell, Ω(f i ) is the complexity of the time series prediction of the onu rate of the ith user in the cell, γ is a hyperparameter used to control the number of leaf nodes and thus the model complexity, α is also a hyperparameter used to control the leaf weight and thus the model complexity, ω j is the model complexity leaf weight.

[0142] The processor then linearly combines the loss function and the complexity function to obtain the following target formula:

[0143] Obj=L(F)+Ω(F),

[0144] Substituting Taylor's formula into the target formula, the target formula is converted to the following formula:

[0145]

[0146] This formula is the formula in step 44.

[0147] The step 5 specifically includes:

[0148] Step 51: The processor sets a user bandwidth configuration data set, wherein the user bandwidth configuration data set adopts B = {b1, b2, ..., b v}, b1 is the bandwidth configuration data of the first user in the cell, b2 is the bandwidth configuration data of the second user in the cell...b v is the bandwidth configuration data of the vth user in the cell, the user bandwidth configuration data is determined according to the bandwidth service subscribed by the user from the operator, and the user bandwidth excess rate set is determined according to the user's final predicted ONU rate and the user bandwidth configuration data set;

[0149] In the above steps, the bandwidth configuration data of the user includes the network bandwidth subscribed by the user, such as 100M, 500M, etc. The bandwidth overlimit rate of the user refers to the degree to which the ONU rate exceeds the upper limit of the network broadband rate subscribed by the user.

[0150] Step 52: The processor sets a user broadband service quality score set, wherein the user broadband service quality score set adopts R = {r1, r2, ..., r v} indicates that r1 is the user broadband service quality score of the first user in the cell, r2 is the user broadband service quality score of the second user in the cell, and r v Score the user broadband service quality of the vth user in the cell;

[0151] In the above steps, the user broadband service quality score can be determined based on indicators such as jitter, latency, packet loss rate, number of IP address switches, and number of fault reports within a period of time. The above period of time can be the past month.

[0152] Step 53: The processor normalizes the user's bandwidth excess rate set and the user's broadband service quality score set, and substitutes them into the following expansion urgency formula to calculate the cell expansion urgency:

[0153]

[0154] Where D i is the urgency of cell expansion, i is the sequence number in the indicator sequence, d is the number of items in the indicator sequence, μ iis the weight value of the i-th indicator in the indicator sequence, z i is the normalized value of the ith indicator;

[0155] In the above steps, since there is an inconsistency between the dimensions of the user's bandwidth excess rate set and the dimensions of the user's broadband service quality score, the processor normalizes the user's bandwidth excess rate set, that is, converts the user's bandwidth excess rate into a numerical value between 0-1, and also normalizes the user's broadband service quality score, that is, converts the user's broadband service quality score into a numerical value between 0-1.

[0156] Step 54: The processor expands the cell according to the urgency of cell expansion;

[0157] In the above steps, the processor expands the cell according to the urgency of cell expansion, specifically including: the processor determines whether the urgency of cell expansion is greater than the expansion threshold, if not, the cell is not expanded, and if so, all cell expansion urgency greater than the expansion threshold are arranged in order from large to small according to the numerical value into an expansion sequence, and the cells that need to be expanded are expanded in order according to the order of the expansion sequence.

[0158] In this embodiment, the expansion threshold can be pre-set in the processor. Since the network service provider serves multiple cells and the network service provider has limited expansion capacity, when expanding multiple cells, it is very likely that the expansion cost will be greatly increased. Therefore, the cells with expansion urgency greater than the expansion threshold are expanded one by one according to the cell expansion urgency value, which takes into account user satisfaction, improves expansion efficiency and reduces expansion cost.

[0159] The above embodiments are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for expanding cell broadband, characterized in that: The steps include: Step 1: Obtain the user onu rate time series and the onu rates of all users in the cell, where the user onu rate time series adopts U = {u1,u2,…,u n } indicates that u1 is the user onu rate obtained for the first time, u2 is the user onu rate obtained for the second time...u n is the user onu rate obtained for the nth time. The onu rate of all users in the cell adopts F = {U1, U2, ..., U v } indicates that U1 is the onu rate time series of the first user in the cell, and U2 is the onu rate time series of the second user in the cell...U v is the onu rate time series of the vth user in the cell, v is the total number of users in the cell; Step 2: Substitute the user's onu rate time series into the time series prediction method to predict the user's future onu rate; Step 3: Perform feature learning onu rates of all users in the cell to obtain initial derived features, and then combine the future onu rates of the users with the initial derived features to obtain derived features, where the derived features are X = {x1, x2, …, x v } indicates that x1 is the derived feature corresponding to U1, x2 is the derived feature corresponding to U2…x v For U v The corresponding derived features; The step 3 specifically includes the following steps: Step 31: Determine the acquisition time of the user onu rate according to the user onu rate time series, judge whether the acquisition time of the user onu rate is a holiday, and obtain the holiday characteristics; Step 32: Substitute the user's onu rate time series into the sliding time window method to calculate the trend preference feature; Step 33: Substitute the user onu rate time series into the Lag lag method to calculate the lag feature; Step 34: Substitute the user onu time series into the trend learning method to calculate the trend feature; Step 35: Combine holiday features, trend preference features, lag features and trend features to obtain preliminary derived features; Step 36: Combine the preliminary derived features and the user's future onu rate to obtain the derived features; Step 4: Train the XGBOOST model based on the derived features and the onu rates of all users in the cell, and predict the final predicted onu rate of the user based on the XGBOOST model; The step 4 specifically includes the following steps: Step 41: Calculate the monthly maximum onu ​​rate set of users in the cell according to the onu rates of all users in the cell, where the monthly maximum onu ​​rate set of users in the cell adopts Y = {y1, y2, ..., y v } indicates that y1 is the monthly maximum ONU rate of the first user in the cell, and y2 is the monthly maximum ONU rate of the second user in the cell. v is the monthly maximum onu ​​rate of the vth user in the cell; Step 42: Substitute the monthly maximum ONU rate set of users in the cell into the following model complexity leaf weight function to calculate the leaf weight of the optimal model complexity: In the formula, j is a leaf node, ω j is the weight of leaf node j, g i is the first-order partial derivative of y contained in leaf node j, h i is the second-order partial derivative of y contained in leaf node j, α is a hyperparameter, for y i (τ-1) The first derivative at the τ-1 iteration, for y i (τ-1) The second-order partial derivative of is the loss degree of the i-th user, y i is the monthly maximum onu ​​rate of the ith user, is the predicted value of the monthly maximum onu ​​rate of the i-th user τ-1 iterations, where τ is the number of iterations; Step 43: Substitute the monthly maximum ONU rate set of users in the cell into the following loss function to calculate the minimum loss: Where L is the loss degree, T is the number of leaf nodes, j is a leaf node, j is between 1 and T, α and γ are both hyperparameters, g i is the first-order partial derivative of the samples contained in leaf node j, h i is the second-order partial derivative of the samples contained in leaf node j; Step 44: Substitute the leaf weights, minimum loss and derived features of the optimal model complexity into the following objective function to obtain the optimized XGBOOST model, and predict the user's final predicted ONU rate based on the optimized XGBOOST model: Where v is the total number of users in the cell, the sequence number of the user in cell i, and Obj (τ) is the objective function of the τth iteration, f τ (x i ) is the x of the τth iteration i The predicted value of f τ (x i ) to the power of Step 5: Determine the urgency of cell expansion according to the final predicted ONU rate of the user, and expand the cell according to the urgency of cell expansion; The step 5 specifically includes the following steps: Step 51: Set a user bandwidth configuration data set, wherein the user bandwidth configuration data set adopts B = {b1, b2, ..., b v }, b1 is the bandwidth configuration data of the first user in the cell, b2 is the bandwidth configuration data of the second user in the cell...b v is the bandwidth configuration data of the vth user in the cell, the user bandwidth configuration data is determined according to the bandwidth service subscribed by the user from the operator, and the user bandwidth excess rate set is determined according to the user's final predicted ONU rate and the user bandwidth configuration data set; Step 52: Set a user broadband service quality score set, where the user broadband service quality score set adopts R = {r1, r2, ..., r v } indicates that r1 is the user broadband service quality score of the first user in the cell, r2 is the user broadband service quality score of the second user in the cell, and r v Score the user broadband service quality of the vth user in the cell; Step 53: normalize the user's bandwidth excess rate feature and the user's broadband service quality score feature, and substitute them into the following expansion urgency formula to calculate the cell expansion urgency: Where D i is the urgency of cell expansion, i is the index number, d is the number of index items, μ i is the weight value of the i-th indicator in the indicator, z i is the normalized value of the ith indicator; Step 54: Expand the cell according to the urgency of cell expansion.

2. A method for expanding cell broadband according to claim 1, characterized in that: The time series prediction method adopts the ARIMA prediction method, which specifically includes the following steps: Step 21: determine whether the user onu rate time series is a stationary time series. If so, calculate the autocorrelation coefficient and the partial autocorrelation coefficient based on the user onu rate time series which is a stationary time series. If not, perform differential transformation on the user onu rate time series to obtain the user onu rate time series which is a stationary time series, and then calculate the autocorrelation coefficient and the partial autocorrelation coefficient. Step 22: Determine the value range of the autoregressive coefficient according to the maximum lag period of the autocorrelation coefficient, and determine the value range of the moving average order according to the maximum lag period of the partial autocorrelation coefficient; Step 23: Substitute the value range of the autoregressive coefficient and the value range of the moving average order into the minimum information criterion method to calculate the optimal autoregressive coefficient and moving average order, and establish the ARIMA model based on the optimal autoregressive coefficient and moving average order, where the autoregressive coefficient is represented by p and the moving average order is represented by q; Step 24: Perform a validity test on the ARIMA model. If the validity test passes, substitute the user's onu rate time series into the ARIMA model to predict the user's future onu rate.

3. A method for expanding cell broadband according to claim 2, characterized in that: The ARIMA model in step 23 is expressed by the following formula: you rt =α1u r(t-1) +α2u r(t-2) +…+a p you r(t-p) +e t +β1ε t-1 +β2ε t-2 +…+b q e t-q , Where n=rt, r is a parameter, t is the time when the user onu rate is obtained, α1…α p u r(t-1) …u r(t-p) The parameters of β1…β q They are ε t-1 …ε t-q The parameter, ε t …ε t-q is the white noise sequence from time t to time tq.

4. A method for expanding cell broadband according to claim 2, characterized in that: The step 21 specifically includes the following steps: Substitute the user onu rate time series into the following formula to calculate the autocorrelation coefficient: Wherein, k is the lag period, AC(k) is the autocorrelation coefficient when the lag period is k, n=rt, r is a parameter, t is the acquisition time of the user onu rate, m is the number of acquisitions of the user onu rate between time tk and time t, It is the average onu rate of all users in the cell.

5. A method for expanding cell broadband according to claim 2, characterized in that: The step 21 specifically includes the following steps: Substitute the user onu rate time series into the following formula to calculate the partial autocorrelation coefficient: Wherein, k is the lag period, PAC(k) is the partial autocorrelation coefficient of the lag period k, cov is the covariance treatment, var is the variance treatment, n=rt, r is a parameter, t is the time when the user onu rate is obtained, is the average onu rate of all users in the cell at time t, It is the average onu rate of all users in the cell at time tk.

6. A method for expanding cell broadband according to claim 1, characterized in that: The step 54 specifically includes the following steps: Determine whether the cell expansion urgency is greater than the expansion threshold. If not, do not expand the cell. If yes, arrange all cell expansion urgency values ​​greater than the expansion threshold into an expansion sequence from large to small, and expand the cells in sequence according to the expansion sequence.

Images