A centralized cloud-edge collaborative wireless communication traffic prediction method

Abstract

The present invention provides a centralized cloud-edge collaborative wireless communication traffic prediction method, including: describing mobile traffic prediction as a discrete wavelet transform process and implementing linear part prediction through machine learning; constructing a residual network, taking community users with similar characteristics within a region as a cluster, and the edge server quickly aggregates data of the same cluster to form a global model of the cluster; decomposing the complex characteristics of multi-source and cross-domain model data through the Lasso algorithm, calculating the correlation between spatio-temporal features, and combining XGboost to construct an optimal weight allocation mechanism for community traffic features; XGBoost processes the non-linear relationships and the mutual dependencies between variables in the model through CART trees, captures deep features, and performs non-linear prediction on mobile traffic; reconstructing the prediction results of the linear and non-linear parts through wavelet reconstruction. The present invention can improve the performance of the global model for predicting wireless communication traffic obtained through machine learning training, and improve the accuracy and real-time performance of wireless communication traffic prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic prediction, and more particularly, to a centralized cloud-edge collaborative wireless communication traffic prediction method. Background Art

[0002] Current traffic prediction technologies can be regarded as time series prediction problems, and the distribution law in the past time period is analyzed through modeling and then extended to the future development trend. The distribution of traffic data is affected by factors such as human activity habits and time seasons, and usually shows typical time series characteristics. Accordingly, the existing literature can be roughly divided into two categories, namely statistical-based methods and machine learning-based methods.

[0003] Statistical-based methods mainly rely on using methods such as signal processing and time series analysis to model the base station load time series. The model expression form used in the prediction of the load time series is fixed, and when the time series is relatively complex and difficult to be characterized by a known specific model, the prediction effect is not ideal.

[0004] The accuracy of mobile traffic modeling and prediction has been greatly improved through machine learning solutions. However, as more and more indicators are referred to in deep learning, some indicators when the model is put into use, such as the convergence time during the operation of the model, are slow due to the relatively complex model design and the training method, which makes it difficult to meet the real-time business prediction requirements. Therefore, it is also a difficult problem to accurately screen features and achieve higher prediction accuracy and faster convergence speed of wireless communication services considering time and space factors and combining cross-domain data;

[0005] With the development of technological innovation, the development of smart phones in the past decade has accelerated the generation and explosion of data, thus accelerating the era of big data. How to efficiently and reasonably allocate limited spectrum resources has become a widely concerned issue for relevant researchers. Among them, mobile traffic prediction is one of the research hotspots. In recent years, artificial intelligence technology has developed rapidly, and machine learning is more suitable for mobile traffic prediction applications due to its good perception and decision-making capabilities. Essentially, mobile service traffic can be regarded as a time series prediction problem, and the distribution law in the past time period is analyzed through modeling and then extended to the future development trend. The distribution of traffic data is affected by factors such as human activity habits and time seasons, and usually shows typical time series characteristics. Through the action of integrated learning on the interaction between the agent and the environment, when the environment changes, the agent affects the global model weight parameters through the strategy of integrated learning and the reward of the corresponding actions to quickly adjust the agent's actions on the environment. Summary of the Invention

[0006] According to the above-mentioned technical problem of how to improve the accuracy of mobile traffic prediction, a centralized cloud-edge collaborative wireless communication traffic prediction method is provided. The present invention uses the advantages of deep learning to propose a hybrid model combining the optimal feature weighting method based on Lasso, DWT decomposition ARIMA linear prediction, and XGboost residual network prediction. Considering the nonlinearity of mobile traffic data, the data set is decomposed into a linear part and a residual part by using discrete wavelet transform, and the ARIMA model processes the linear part. The optimal feature variables of the residual data are constructed by Lasso to reduce the model complexity and prevent overfitting. At the same time, community users with similar characteristics in the region are taken as a cluster, and their own training models are uploaded to the same edge parameter server for neural network model parameter aggregation, so as to quickly form a global model of the cluster, and the residual part is predicted by XGboost. Finally, wavelet reconstruction is used to combine the prediction results. The present invention has good approximation ability and generalization ability in mobile network traffic prediction, greatly improving the prediction accuracy. It can help developers make project decisions and allocate resources to achieve the goal of maximizing benefits.

[0007] The technical means adopted by the present invention are as follows:

[0008] A centralized cloud-edge collaborative wireless communication traffic prediction method includes:

[0009] Describe mobile traffic prediction as a process of discrete wavelet transform, and realize linear part prediction through machine learning;

[0010] Construct a residual network, take community users with similar characteristics in the region as a cluster, and the edge server quickly aggregates the data of the same cluster to form a global model of the cluster;

[0011] Decompose the complex characteristics of multi-source and cross-domain model data through the Lasso algorithm, calculate the correlation between spatio-temporal features, and combine XGboost to construct an optimal weight allocation mechanism for community traffic features;

[0012] XGBoost processes the nonlinear relationships and the mutual dependencies between variables in the model through CART trees, captures deep features, and performs nonlinear prediction on mobile traffic;

[0013] Reconstruct the prediction results of the linear and nonlinear parts through wavelet reconstruction.

[0014] Furthermore, the process of describing mobile traffic prediction as a discrete wavelet transform and realizing linear part prediction through machine learning includes:

[0015] Perform discrete wavelet decomposition on the data set in the data preprocessing stage. Given a square-integrable signal f(t), that is, f(t) ∈ L 2(R), the corresponding discrete wavelet transform equation is as follows:

[0016]

[0017] where a represents the offset factor, b represents the displacement coefficient, a and b are constants, and a > 0. is the basis function which is first shifted and then scaled; if a and b vary at each time point of the traffic flow, a family of functions can be obtained 〈*, *〉 represents the inner product; * represents the complex conjugate, m and n are the results of the discretization of a and b; the sequence is decomposed into decaying orthogonal bases to express the mutations and non-stationary parts in the sequence.

[0018] The traffic flow dataset is divided into an approximation part and a residual part by the discrete wavelet transform with the basis function db4, that is:

[0019] Y t = L t + N t

[0020] where Y t represents the traffic flow dataset, L t represents the linear part, and N t represents the residual part.

[0021] The linear part L of the moving traffic flow will be stripped out t and input into the ARIMA(p, d, q) differential autoregressive moving average model for linear part modeling and prediction. AR represents autoregression, MA represents moving average, p and q are the corresponding orders, and d is the number of differences to make the time series stationary.

[0022] Furthermore, the ARIMA(p, d, q) differential autoregressive moving average model is a linear regression model used to track the linear trend in stationary time series data, where the future values of the time series are generated by a linear function observed in the past; specifically:

[0023] The ARIMA(p, d, q) differential autoregressive moving average model performs d-order differential processing on the non-stationary community traffic flow historical data L t to obtain a new stable traffic flow historical sequence X t , and fitting X t into the ARMA(p, q) model and restoring according to the original d-order difference can obtain the predicted data of L t , where the expression of ARMA(p, q) is as follows:

[0024]

[0025] Among them, the first half of the expression is the autoregressive part, where p is the autoregressive order and a p is the autoregressive coefficient; the second half of the expression is the moving average part, where q is the moving average order and b q is the moving average coefficient; L t is the relevant sequence of community mobile traffic, and w t is the random error.

[0026] Furthermore, when constructing the residual network, community users with similar characteristics within the region are taken as a cluster, and the edge server quickly aggregates the data of the same cluster to form a global model of the cluster, including:

[0027] Randomly initialize k community points as cluster centroids;

[0028] Assign each point in the sample set to a cluster; calculate the distance between each point and the centroid, and assign it to the cluster corresponding to the nearest centroid;

[0029] Update the centroid of the cluster, and update the centroid of each cluster to the average value of all points in the cluster;

[0030] Repeatedly iterate the steps of assigning each point in the sample set to a cluster and the steps of updating the centroid of the cluster until the cluster center no longer changes significantly, that is:

[0031]

[0032] where x i represents the i-th sample, c i represents the cluster to which x i belongs, μ i represents the center point corresponding to the cluster, and M represents the total number of samples; let t = 0, 1, 2… be the iteration steps, and repeatedly execute the following two expressions until J converges. The expressions are as follows:

[0033]

[0034] This expression means that for each sample x i it is assigned to the nearest cluster;

[0035]

[0036] This expression means that for each cluster class k, recalculate the center of the cluster until it is stable.

[0037] Furthermore, when decomposing the complex characteristics of multi-source and cross-domain model data through the Lasso algorithm, calculating the correlation between spatio-temporal features, and constructing an optimal weight allocation mechanism for community traffic features in combination with XGboost, it includes:

[0038] Use ensemble learning to build a tree model for the residual part of mobile traffic. Through continuous interaction and trial-and-error between the residual terms of community users and feature variables, find the maximum splitting point of the tree model; in order to avoid feature redundancy in the residual network and reduce the model complexity, introduce Lasso to construct a penalty function to obtain a refined model, which can accurately compress and calculate the estimated values of the data when performing compressed estimation on important parameter variables, and at the same time generate a sparse solution. Specifically:

[0039] Construct a linear regression model as follows:

[0040] Y = Xβ + ε

[0041] where the response variable is Y = (y1, y2,......, y n ) T , the covariate is X = (X (1) , X (2) ,......, X (d) ). For each X (j) there is

[0042] Design the random error term as follows:

[0043] ε i ~N(0, σ 2 ),(i = 1, 2..., n), ε = (ε1, ε2,..., ε n ) T

[0044] Design the regression coefficients as follows:

[0045] β = (β1, β2,......, β n ) T

[0046] Adopt the Lasso method to add the penalty term into the multiple linear model, and take the minimum value of the likelihood function as the regression coefficient to obtain the following expression:

[0047]

[0048] where t and λ correspond one-to-one as the adjustment coefficients; set the threshold t0. When t < t0, reduce the model complexity by reducing the X dimension to avoid overfitting of the residual network model for traffic prediction.

[0049] Furthermore, the XGBoost processes the non-linear relationships and the mutual dependencies between variables in the model through CART trees, captures deep features, and performs non-linear prediction on mobile traffic, including:

[0050] Construct an XGBoost model, and the expression is as follows:

[0051]

[0052] where x i is the sample variable, the predicted value, T is the number of trees, F is all possible CART trees, and f t is a specific CART tree;

[0053] The objective function for designing the XGBoost model is expressed as follows:

[0054]

[0055] where represents the model prediction in the previous t - 1 rounds, Ω(f t ) represents the model complexity, and C represents a constant term;

[0056] Perform a second - order Taylor expansion on the objective function and define two variables in the original objective function for easy calculation. The scoring function is expressed as follows:

[0057]

[0058] where γ is the adjustment function, and w j is the number of leaves;

[0059] Use the scoring function to select the best split point to build a CART tree, determine all cut points of the sample features, and divide each determined cut point. The criteria for good or bad are shown in the following formula:

[0060]

[0061] where Gain represents the difference between the single - node obj(t) after splitting and the two - node tree obj(t). By traversing all split points of the features, the largest split point is the best split point.

[0062] Furthermore, the reconstruction of the prediction results for the linear and non - linear parts through wavelet reconstruction includes:

[0063] Divide the community mobile traffic dataset into a linear part and a residual part through discrete wavelet transform with the basis function db4;

[0064] For the separated linear - part dataset L t , use the ARIMA(p, d, q) model to predict the training and obtain the linear - part prediction dataset L T ;

[0065] Use Lasso to screen the features of the residual part of the traffic to avoid overfitting. At the same time, use the XGboost model to predict the dataset N t and obtain the residual prediction dataset N T ;

[0066] Reconstruct the linear part prediction dataset L T and the residual prediction dataset N T through wavelet reconstruction to obtain the final prediction result as follows:

[0067] Y T = L T + N T

[0068] where Y T represents the final prediction result.

[0069] Compared with the prior art, the present invention has the following advantages:

[0070] 1. The centralized cloud-edge collaborative wireless communication traffic prediction method provided by the present invention performs dual modeling from the perspectives of time and space. First, it expands the scope of external features, such as base station information and third-party data of cells (such as the distribution of some popular places, public facilities, and social activity levels), weather, and whether it is a holiday, etc. It effectively integrates the problem of multi-source and cross-domain data directly related to external features and wireless service traffic. Secondly, it considers the correlation between features and screens the features to avoid problems such as feature redundancy. Finally, it considers some indicators when the model is put on the line, such as the convergence time during the operation of the model. The model of the present invention effectively reduces the model complexity, improves the model convergence speed, and meets the real-time business prediction requirements. Under the condition of considering time and space factors and combining cross-domain data, it realizes higher prediction accuracy and faster convergence speed of wireless cellular services.

[0071] 2. The centralized cloud-edge collaborative wireless communication traffic prediction method provided by the present invention proposes a new prediction model (DALXG) DWT-ARIMA-Lasso-XGboost based on Lasso optimal feature weighting for the problems faced by traffic prediction. The model includes a DWT algorithm with a decomposed linear part and a residual part, an ARIMA part for linear autoregressive prediction, a Lasso model for optimal feature screening and optimal weight allocation, and an XGboost ensemble learning model for integrating non-linear relationships in the residual network.

[0072] 3. The centralized cloud-edge collaborative wireless communication traffic prediction method provided by the present invention proposes to construct a DWT preprocessing model in the concept of traditional time series traffic prediction. The time series is decomposed into a linear part and a non-linear part through the DWT model by discrete wavelet transform. After processing these two parts, wavelet reconstruction is used to fuse the prediction results.

[0073] 4. The centralized cloud-edge collaborative wireless communication traffic prediction method provided by the present invention performs autoregressive prediction of the linear part on mobile traffic data through ARIMA; mines the hidden periodic rules of the traffic in the target area, analyzes the stationarity of the residual part of the data at the same time, and constructs a residual network; decomposes the complex characteristics of multi-source and cross-domain data through the Lasso regression algorithm, and calculates the correlation between spatio-temporal features. Feature variables are screened according to the degree of correlation, and an optimal weight allocation mechanism for community traffic features is constructed in combination with XGboost to avoid feature redundancy and reduce the model complexity.

[0074] 5. When the centralized cloud-edge collaborative wireless communication traffic prediction method provided by the present invention predicts the non-linear behavior of community traffic through XGboost, the CART tree is used to process the non-linear relationship and the mutual dependence between variables in the model to capture deep features. At the same time, regularization reduces the model variance and prevents overfitting, effectively improving the prediction accuracy and convergence speed.

[0075] For the above reasons, the present invention can be widely promoted in the fields of traffic prediction and the like. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0077] Figure 1 It is a framework diagram of the centralized cloud-edge collaborative wireless communication traffic prediction method of the present invention.

[0078] Figure 2 It is a flow chart of the present invention and its learning algorithm.

[0079] Figure 3 It is a diagram of the proportion of feature variables provided by the embodiment of the present invention.

[0080] Figure 4 It is a diagram of the weight ratio of feature variables provided by the embodiment of the present invention.

[0081] Figure 5The prediction effect diagram of traffic by the machine learning framework provided by the embodiments of the present invention. Detailed implementation manners

[0082] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments may be combined with each other. The present invention will be described in detail below with reference to the drawings and in conjunction with the embodiments.

[0083] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part rather than all of the embodiments of the present invention. The following description of at least one exemplary embodiment is actually illustrative only and in no way limits the present invention and its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0084] It should be noted that the terms used herein are only for describing the specific implementation manners and are not intended to limit the exemplary embodiments of the present invention. As used herein, unless otherwise clearly specified in the context, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or their combinations.

[0085] Unless otherwise specifically stated, the relative arrangements of components and steps, numerical expressions and values set forth in these embodiments do not limit the scope of the present invention. At the same time, it should be clear that, for the convenience of description, the dimensions of the various parts shown in the drawings are not drawn in actual proportional relationships. Technologies, methods and devices known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods and devices should be regarded as part of the authorized specification. In all the examples shown and discussed here, any specific values should be construed as merely exemplary and not as a limitation. Therefore, other examples of the exemplary embodiments may have different values. It should be noted that: like reference numerals and letters denote like items in the following drawings, and thus, once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0086] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by orientation words such as "front, back, up, down, left, right", "lateral, vertical, perpendicular, horizontal" and "top, bottom", etc. is usually based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description. Without contrary explanation, these orientation words do not indicate and imply that the device or element referred to must have a specific orientation or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation on the protection scope of the present invention: the orientation words "inside, outside" refer to the inside and outside relative to the contour of each component itself.

[0087] For convenience of description, spatial relative terms such as "above...", "over...", "on the upper surface of...", "above-mentioned", etc. can be used here to describe the spatial positional relationship of a device or feature shown in the drawings with other devices or features. It should be understood that the spatial relative terms are intended to include different orientations in use or operation in addition to the orientation described in the drawings for the device. For example, if the device in the drawing is inverted, the device described as "above other devices or structures" or "over other devices or structures" will then be positioned as "below other devices or structures" or "under other devices or structures". Thus, the exemplary term "above..." can include both orientations of "above..." and "below...". The device can also be positioned in other different ways (rotated 90 degrees or in other orientations), and corresponding interpretations should be made for the spatial relative descriptions used here.

[0088] In addition, it should be noted that using words such as "first", "second", etc. to limit the components is only for the convenience of distinguishing the corresponding components. Without separate statement, the above words have no special meaning. Therefore, it should not be construed as a limitation on the protection scope of the present invention.

[0089] Such as Figure 1 , 2As shown in the figure, in order to improve the accuracy of mobile traffic prediction and maximize the utilization rate of communication network resources, the present invention provides a centralized cloud-edge collaborative wireless communication traffic prediction method. First, the known regional traffic data is decomposed into a linear part and a non-linear part through discrete wavelet transform; the autoregressive modeling and prediction of the linear part data of mobile traffic are carried out through the ARIMA model; for the characteristic variables of mobile traffic non-linearity, the present invention constructs a penalty function by introducing Lasso to obtain a relatively refined model, which can accurately compress and calculate the estimated value of the data when estimating and compressing important parameter variables, and at the same time generate a sparse solution. On this basis, the tree model of the mobile traffic residual part is constructed through the integrated learning XGBoost, and the global model is aggregated through the continuous interaction and trial-and-error of the community user residual term and the characteristic variables to find the maximum splitting point of the tree model. Avoiding the feature redundancy of the residual network, reducing the model complexity, and improving the prediction accuracy.

[0090] The centralized cloud-edge collaborative wireless communication traffic prediction method provided by the present invention includes:

[0091] S1. Describe the mobile traffic prediction as a process of discrete wavelet transform, and realize the linear part prediction through machine learning;

[0092] S2. Construct a residual network, take community users with similar characteristics in the region as a cluster, and the edge server quickly aggregates the data of the same cluster to form a global model of the cluster;

[0093] S3. Decompose the complex characteristics of multi-source and cross-domain model data through the Lasso algorithm, calculate the correlation between spatio-temporal features, and combine XGboost to construct an optimal weight allocation mechanism for community traffic features;

[0094] S4. XGBoost processes the non-linear relationship in the model and the mutual dependence between variables through CART trees, captures deep features, and performs non-linear prediction on mobile traffic;

[0095] S5. Reconstruct the prediction results of the linear and non-linear parts through wavelet reconstruction.

[0096] Specifically, as a preferred embodiment of the present invention, in step S1, describing the mobile traffic prediction as a process of discrete wavelet transform and realizing the linear part prediction through machine learning includes:

[0097] S11. Perform discrete wavelet decomposition on the data set in the data preprocessing stage. Given the square-integrable signal f(t), that is, f(t) ∈ L 2 (R), the corresponding discrete wavelet transform equation is as follows:

[0098]

[0099] Among them, a represents the offset factor, b represents the displacement coefficient, a and b are constants, and a > 0. is the basis function First, it is shifted and then scaled; if a and b at each time point of the mobile traffic are constantly changing, a cluster of functions can be obtained 〈*, *〉 represents the inner product; * represents the complex conjugate, m and n are the results of the discretization processing of a and b; the sequence is decomposed into decaying orthogonal bases to express the mutations and non-stationary parts in the sequence.

[0100] S12. The traffic data set is divided into an approximation part and a residual part through discrete wavelet transform with the basis function db4, that is:

[0101] Y t = L t + N t

[0102] Among them, Y t represents the traffic data set, L t represents the linear part, and N t represents the residual part;

[0103] S13. The linear part L of the mobile traffic is stripped out t and input into the ARIMA(p, d, q) differential autoregressive moving average model for linear part modeling and prediction. AR represents autoregression, MA represents moving average, p and q are the corresponding orders, and d is the number of differences to make the time series stationary.

[0104] In specific implementation, as a preferred implementation manner of the present invention, in the step S13, the ARIMA(p, d, q) differential autoregressive moving average model is a linear regression model for tracking the linear trend in stationary time series data, where the future value of the time series is generated by a linear function observed in the past; specifically:

[0105] The ARIMA(p, d, q) differential autoregressive moving average model performs d-order differential processing on the non-stationary community traffic historical data L t to obtain a new stable traffic historical sequence X t , fit X t into the ARMA(p, q) model, and the predicted data of L t can be obtained according to the original d-order difference recovery. Among them, the expression of ARMA(p, q) is as follows:

[0106]

[0107] Among them, the first half of the expression is the autoregressive part, p is the autoregressive order, ap is the autoregressive coefficient; the second half of the expression is the moving average part, where q is the order of the moving average and b q is the moving average coefficient; L t is the relevant sequence of the community mobile traffic, and w t is the random error.

[0108] In specific implementation, as a preferred implementation manner of the present invention, in the step S2, a residual network is constructed, and community users with similar characteristics in the area are taken as a cluster. The edge server quickly aggregates the data of the same cluster to form a global model of the cluster; since users in the same regional space proximity or communities with the same attributes have approximate operation conditions, we aggregate the mobile traffic trends of community users to obtain a mutually correlated community. The mutually correlated community can be used as the characteristics of each other's communities to realize that the cluster will aggregate the global model. The specific implementation steps are as follows:

[0109] S21. Randomly initialize k community points as cluster centroids;

[0110] S22. Assign each point in the sample set to a cluster; calculate the distance between each point and the centroid, and assign it to the cluster corresponding to the centroid with the closest distance;

[0111] S23. Update the centroid of the cluster, and update the centroid of each cluster to the average value of all points in the cluster;

[0112] S24. Repeatedly iterate steps S22 and S23 until the cluster centers no longer change significantly, that is:

[0113]

[0114] where x i represents the i-th sample, c i represents the cluster to which x i belongs, μ i represents the center point corresponding to the cluster, and M represents the total number of samples; let t = 0, 1, 2... be the iteration steps, and repeat the following two expressions until J converges. The expressions are as follows:

[0115]

[0116] This expression means that for each sample x i it is assigned to the closest cluster;

[0117]

[0118] This expression means that for each cluster class k, recalculate the center of the cluster until it is stable. When the K-means algorithm is iterating, assume that the current J has not reached the minimum value. Then first fix the cluster centers {μ k}, adjust each sample x i to the category c it belongs to i , to minimize the J function, and then fix {c i}, adjust the cluster centers {μ k}, to minimize J. These two steps alternate in a cycle, J monotonically decreases, and when J decreases to the minimum value, {μ k} and {c i} also converge simultaneously.

[0119] In specific implementation, as a preferred implementation manner of the present invention, in the step S3, the Lasso algorithm is used to decompose the complex characteristics of multi-source and cross-domain model data, calculate the correlation between spatio-temporal features, and combine XGboost to construct an optimal weight allocation mechanism for community traffic features. In this embodiment, in order to avoid overfitting of the ensemble learning model to mobile traffic features, the present invention constructs a penalty function by introducing Lasso to obtain a more refined model, which can accurately compress and calculate the estimated value of the data when estimating important parameter variables, and at the same time generate a sparse solution. Specifically, it includes:

[0120] Use ensemble learning to construct a tree model for the residual part of mobile traffic. Through continuous interaction and trial and error between the residual terms of community users and feature variables, find the maximum splitting point of the tree model; in order to avoid feature redundancy in the residual network and reduce the model complexity, introduce Lasso to construct a penalty function to obtain a refined model, which can accurately compress and calculate the estimated value of the data when estimating important parameter variables, and at the same time generate a sparse solution. Specifically, it is:

[0121] S31. Construct a linear regression model as follows:

[0122] Y = Xβ + ε

[0123] where the response variable is Y = (y1, y2,......, y n ) T , the covariate is X = (X (1) , X (2) ,......, X (d) ). For each X (j) there is

[0124] S32. Design the random error term as follows:

[0125] ε i ~N(0, σ 2 ),(i = 1, 2..., n), ε = (ε1, ε2,..., ε n ) T

[0126] S33. Design the regression coefficients as follows:

[0127] β = (β1, β2,......, β n ) T

[0128] S34. Add the penalty term into the multiple linear model by using the Lasso method, and take the minimum value of the likelihood function as the regression coefficient to obtain the following expression:

[0129]

[0130] where t and λ correspond one-to-one as adjustment coefficients; set the threshold t0. When t < t0, reduce the model complexity by reducing the X dimension to avoid overfitting of the residual network model for traffic prediction. The Lasso algorithm of the present invention is as follows after screening the model features Figure 3 Figure 4 As shown, in the simulation experiment, the system parameter λ = -1.5, and it can be seen that the model can converge quickly and the performance is stable at this time.

[0131] In specific implementation, as a preferred implementation manner of the present invention, in the step S4, XGBoost processes the non-linear relationship and the mutual dependence between variables in the model through CART trees, captures deep features, and performs non-linear prediction on mobile traffic, including:

[0132] S41. Construct the XGBoost model, and the expression is as follows:

[0133]

[0134] where x i is the sample variable, is the predicted value, T is the number of trees, F is all possible CART trees, and f t is a specific CART tree;

[0135] S42. Design the objective function of the XGBoost model, and the expression is as follows:

[0136]

[0137] where, represents the model prediction of the previous t - 1 rounds, Ω(f t ) represents the model complexity, and C represents a constant term;

[0138] S43. Perform a second-order Taylor expansion on the objective function, and define two variables in the original objective function for easy calculation. The scoring function expression is as follows:

[0139]

[0140] Among them, γ is a regulation function, and w j is the number of leaves;

[0141] S44. Use a scoring function to select the best splitting point to build a CART tree, determine all splitting points of the sample features, and divide each determined splitting point. The criteria for good or bad are shown in the following formula:

[0142]

[0143] Among them, Gain represents the difference between the single node obj(t) after splitting and the two-node tree obj(t). By traversing the splitting points of all features, the largest splitting point found is the best splitting point. According to this method, the nodes are continuously split to obtain a CART tree; when XGboost predicts the non-linear behavior of the data set, the CART tree can explain the non-linear relationship in the model and the interdependence between variables, and has good performance in non-linear data prediction.

[0144] In specific implementation, as a preferred implementation manner of the present invention, in the step S5, the prediction results of the linear and non-linear parts are reconstructed through wavelet reconstruction, including:

[0145] S51. Divide the community mobile traffic data set into a linear part and a residual part through discrete wavelet transform with the basis function db4;

[0146] S52. For the separated linear part data set L t , use the ARIMA(p, d, q) model to predict and train, and obtain the linear part prediction data set L T ;

[0147] S53. Screen the features of the residual part of the traffic through Lasso to avoid overfitting. At the same time, use the XGboost model to predict the data set N t , and obtain the residual prediction data set N T ;

[0148] S54. Recombine the linear part prediction data set L T and the residual prediction data set N T through wavelet reconstruction to obtain the final prediction result, as follows:

[0149] Y T = L T + N T

[0150] Among them, Y TIndicates the final prediction result. After the centralized cloud-edge collaborative wireless communication traffic prediction algorithm framework in the present invention is simulated, the prediction effect is as Figure 5 shown.

[0151] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A centralized cloud-edge collaborative wireless communication traffic prediction method, characterized in that Including: Describe the mobile traffic prediction as a process of discrete wavelet transform, and achieve linear part prediction through machine learning; Construct a residual network, take community users with similar characteristics in the region as a cluster, and the edge server quickly aggregates the data of the same cluster to form a global model of the cluster; Decompose the complex characteristics of multi-source and cross-domain model data through the Lasso algorithm, calculate the correlation between spatio-temporal features, and combine XGboost to construct an optimal weight allocation mechanism for community traffic features; XGBoost processes the non-linear relationships and the mutual dependencies between variables in the model through CART trees, captures deep features, and performs non-linear prediction on mobile traffic, including: Construct an XGBoost model, the expression is as follows: Among them, is a sample variable, the predicted value, T is the number of trees, F are all possible CART trees, and is a specific CART tree; Design the objective function of the XGBoost model, the expression is as follows: Among them, represents retaining the model prediction of the previous iteration at e- round 1, represents the model complexity, C represents the constant term; Perform a second-order Taylor expansion on the objective function, and define two variables in the original objective function for easy calculation. The scoring function expression is as follows: Among them, , is an adjustment function, is the number of leaves; Use the scoring function to select the best splitting point to build a CART tree, determine all the cutting points of the sample features, and divide each determined cutting point. The criteria for good or bad are shown in the following formula: Among them, Gain represents a single node after splitting obj ( t ) and the difference between the tree of two nodes obj ( t ). By traversing the splitting points of all features, the largest splitting point is found as the best splitting point; Reconstruct the prediction results of the linear and non-linear parts through wavelet reconstruction, including: Through discrete wavelet transform with the basis function db4, divide the community mobile traffic data set into a linear part and a residual part; For the separated linear part data set , use the ARIMA( p, d, q ) model to predict and train, and obtain the linear part prediction data set ; The characteristics of the residual part of the flow are screened by Lasso to avoid overfitting. At the same time, the XGboost model is used to N t predict the residual part and obtain the residual prediction data set N T ; Reconstruct the linear part prediction dataset through wavelet reconstruction and the residual prediction dataset N T , and obtain the final prediction result as follows: Among them, represents the final prediction result.

2. The centralized cloud-edge collaborative wireless communication traffic prediction method according to claim 1, wherein The process of describing the mobile traffic prediction as a discrete wavelet transform and achieving linear part prediction through machine learning includes: Perform discrete wavelet decomposition on the dataset in the data preprocessing stage, given a square-integrable signal , that is , and the corresponding discrete wavelet transform equation is as follows: Among them, a represents the offset factor, b represents the displacement coefficient, a, b is a constant, and a>0 , is the basis function which is first shifted and then scaled; the a, b at each time point of the moving traffic is constantly changing, then a cluster of functions can be obtained; 〈*, *〉 represents the inner product; * represents the complex conjugate, m、 is a and b the result of the discretization process; the sequence is decomposed into decaying orthogonal bases to express the mutation and non-stationary parts in the sequence; The traffic data set is divided into an approximation part and a residual part through discrete wavelet transform with the basis function db4, that is: Among them, represents the flow data set, represents the separated linear part data set, represents the residual part; Strip out the linear part of the mobile traffic Input the ARIMA( p, d, q ) differential autoregressive moving average model to perform linear part modeling and prediction. AR represents autoregression, MA represents moving average, p and q are the corresponding orders, d is the number of differences to make the time series stationary.

3. The centralized cloud-edge collaborative wireless communication traffic prediction method according to claim 2, wherein The ARIMA( p, d, q ) differential autoregressive moving average model is a linear regression model used to track the linear trend in stationary time series data, where the future values of the time series are generated by a linear function of past observations; specifically: The ARIMA( p, d, q ) differential autoregressive moving average model performs differential processing on the non-stationary historical community traffic data d to obtain a new stable traffic historical sequence X t . Then, X t is fitted into the ARMA( p,q ) model, and the prediction data d of can be obtained by restoring according to the original -order difference. Among them, the expression of ARMA( p,q ) is as follows: Among them, the first half of the expression is the autoregressive part, p is the autoregressive order, a p are the autoregressive coefficients; the second half of the expression is the moving average part, q is the moving average order, b q are the moving average coefficients; is the non-stationary historical data of community traffic, w t is the random error.

4. The centralized cloud-edge collaborative wireless communication traffic prediction method according to claim 1, characterized in that The construction of the residual network, taking community users with similar characteristics in the region as a cluster, and the edge server quickly aggregating the data of the same cluster to form a global model of the cluster includes: Random initialization k The points in the community are used as cluster centroids; Assign each point in the sample set to a cluster; calculate the distance between each point and the centroid, and assign it to the cluster corresponding to the centroid with the closest distance; Update the centroid of the cluster, and update the centroid of each cluster to the average value of all points in the cluster; Iteratively repeat the steps of assigning each point in the sample set to a cluster and updating the centroid of the cluster until the centroid no longer changes significantly, that is: Among them, represents the i-th sample, represents the cluster to which it belongs, represents the center point corresponding to the cluster, M represents the total number of samples; let e = 0, 1, 2… be the iteration step number, and repeat the following two expressions until J converges. The expressions are as follows: This expression means that for each sample it is assigned to the nearest cluster; This expression means that for each cluster k , recalculate the center of the cluster until it stabilizes.

5. The centralized cloud-edge collaborative wireless communication traffic prediction method according to claim 1, wherein The decomposition of the complex characteristics of multi-source and cross-domain model data through the Lasso algorithm, calculating the correlation between spatio-temporal features, and combining XGboost to construct an optimal weight allocation mechanism for community traffic features includes: Use ensemble learning to construct a tree model for the residual part of mobile traffic. Through continuous interaction and trial and error between the residual terms of community users and feature variables, find the maximum splitting point of the tree model; in order to avoid feature redundancy in the residual network and reduce the model complexity, introduce Lasso to construct a penalty function to obtain a refined model, which can accurately compress and calculate the estimated value of the data when performing compressed estimation on important parameter variables, and at the same time generate a sparse solution. Specifically: Construct a linear regression model as follows: Among them, the response variable is , and the covariate is For each , there is ; Design the random error term as follows: Design the regression coefficient as follows: The Lasso method is used to add a penalty term to the multiple linear model, and the minimum value of the likelihood function is used as the regression coefficient to obtain the following expression: Among them, and correspond to adjustment coefficients one by one; set a threshold t 0, when < t is 0, the complexity of the model is reduced by adjusting the X dimension, avoiding overfitting of the residual network model for traffic prediction.

Images