Network resource optimization method for network slicing

Abstract

The invention discloses a network resource optimization method for network slices, which comprises the following steps of: acquiring historical flow data, user behavior data and network topology information of each network slice from each network node in real time by utilizing a distributed data acquisition system, and performing data preprocessing on the historical flow data, the user behavior data and the network topology information to form a feature data set; inputting the feature data set into a prediction model constructed based on an attention mechanism and a mixed density network algorithm, optimizing the network weight through a gradient descent algorithm, and training the prediction model; inputting feature data of the current moment into the trained prediction model, and outputting prediction values of probability distribution of different network slices for future traffic demands; and formulating a resource pre-allocation strategy according to the predicted probability distribution of different network slices for future traffic demands, and dynamically optimizing and adjusting network resources. According to the invention, through accurate resource pre-allocation, network resources are utilized to the maximum extent, the operation cost is reduced, and the resource utilization efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of communication network resource optimization and intelligent scheduling, and in particular to a network resource optimization method for network slicing. Background Art

[0002] With the rapid development of mobile communication technology, especially the popularization of 5G and future networks, network slicing technology has become a key means to meet the personalized needs of different services. However, existing technologies still have many shortcomings in network resource management and optimization, and it is difficult to cope with complex and changing network environments and user needs.

[0003] First, traditional network resource allocation methods are usually based on static or empirical strategies and lack accurate perception of real-time network status and user behavior, which results in resource allocation being unable to flexibly adapt to the dynamic changes of network slices and is prone to resource waste or shortage. Second, traditional traffic prediction methods mainly rely on simple statistical models or basic machine learning algorithms, such as autoregressive models (AR), moving average models (MA), and support vector machines (SVM). These methods can usually only provide a single-point estimate of future traffic, and cannot capture the probability distribution of traffic demand, and it is difficult to reflect the randomness and uncertainty of network traffic. In addition, with the increasing complexity of network topology and the increased correlation between network nodes, traditional methods find it difficult to effectively utilize network topology information and fully mine the spatiotemporal characteristics of data.

[0004] In practical applications, due to the high nonlinearity and non-stationarity of network traffic, the accuracy of traditional prediction methods drops significantly when faced with sudden traffic or abnormal events, which not only reduces the efficiency of resource allocation, but also may affect service quality and even cause network congestion. On the other hand, the existing technology lacks in-depth analysis of user behavior data and cannot capture the changing trends of user needs in a timely manner. In addition, traditional methods often have problems such as difficulty in data fusion and insufficient feature extraction when dealing with multi-source heterogeneous data, which limits the prediction performance of the model.

[0005] In summary, existing technologies have problems in network resource management and optimization, such as insufficient prediction accuracy, slow response speed, and poor adaptability, which makes it difficult to meet the needs of future networks for efficient and intelligent resource management. Summary of the invention

[0006] The purpose of this section is to summarize some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract of the specification and the title of the invention of this application to avoid blurring the purpose of this section, the abstract of the specification and the title of the invention, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0007] In view of the above existing problems, the present invention is proposed.

[0008] Therefore, the technical problem solved by the present invention is that the prior art cannot accurately predict the probability distribution of network traffic demand.

[0009] In order to solve the above technical problems, the present invention provides the following technical solutions: using a distributed data collection system, historical traffic data, user behavior data and network topology information of each network slice are collected from each network node in real time, and data preprocessing is performed to form a feature data set;

[0010] Inputting the feature data set into a prediction model constructed based on an attention mechanism and a hybrid density network algorithm, optimizing network weights through a gradient descent algorithm, and training the prediction model;

[0011] Input the feature data at the current moment into the prediction model after training, and output the predicted value of the probability distribution of future traffic demand of different network slices;

[0012] Based on the predicted probability distribution of future traffic demand for different network slices, a resource pre-allocation strategy is formulated to dynamically optimize and adjust network resources.

[0013] As a preferred solution of the network resource optimization method for network slicing described in the present invention, the historical traffic data at least includes bandwidth utilization, data packet transmission rate and data packet loss rate;

[0014] The user behavior data at least includes the number of active users, average session duration and business type distribution;

[0015] The network topology information at least includes a node set, an edge set, and link capacity or bandwidth.

[0016] As a preferred solution of the network resource optimization method for network slicing described in the present invention, feature advance technology is used to extract time series features, user behavior features and network topology features from the historical traffic data, the user behavior data and the network topology information to form a complete feature vector:

[0017] f i (t)=[F AR,i (t),F user,i (t),F topo,i (t)]

[0018] Among them, f i (t) is the complete feature vector of the ith network slice at time t, F AR,i (t) is the time series feature vector, F user,i (t) is the user behavior feature vector, F topo,i(t) is the network topology feature vector.

[0019] As a preferred solution of the network resource optimization method for network slicing described in the present invention, the complete feature vector is used to form a feature data set X, which is expressed in matrix form as follows:

[0020]

[0021] Among them, X is the feature dataset matrix, each row represents the feature vector of a network slice at a certain time point, n is the number of network slices, and r is the number of time points.

[0022] As a preferred solution of the network resource optimization method for network slicing described in the present invention, the constructed prediction model at least includes an input layer, an attention mechanism layer, a mixed density network layer and an output layer.

[0023] As a preferred solution of the network resource optimization method for network slicing according to the present invention, training the prediction model includes:

[0024] Taking the feature data set as input and the real future traffic demand as target output;

[0025] Define the negative log-likelihood loss function to measure how well the probability distribution predicted by the model matches the real data;

[0026] The gradient descent algorithm is used to optimize the model parameters, minimize the loss function, and complete the training.

[0027] As a preferred solution of the network resource optimization method for network slicing according to the present invention, outputting predicted values ​​of probability distribution of future traffic demands of different network slices includes:

[0028] Inputting the feature vector at the current moment into the input layer of the prediction model;

[0029] Set the number of components of the prediction model to M and calculate the mixing weight π m , mean μ m and standard deviation σ m ;

[0030] For the i-th network slice, the predicted traffic demand probability density function is:

[0031]

[0032] Among them, P i (y) is the probability distribution prediction value of the future traffic demand of the ith network slice, y is the random variable of the future traffic demand, The mean is μ m ,variance Gaussian distribution.

[0033] As a preferred solution of the network resource optimization method for network slicing described in the present invention, it includes:

[0034] According to the predicted value P i (y), based on the confidence level γ, the flow demand threshold y is obtained γ , calculate the resource requirement R of each network slice i =y γ,i ;

[0035] Total resource requirement R total =∑ i R i ;

[0036] Determine whether the total resource demand exceeds the total resource capacity C total ,in:

[0037] If R total ≤C total , then it is allocated to each network slice according to the calculated resource requirements;

[0038] If R total >C total , then resource scheduling and optimization are performed.

[0039] Beneficial effects of the present invention:

[0040] 1. The distributed data acquisition system can efficiently obtain various types of data from each network node, ensuring the comprehensiveness and timeliness of the data. Through data preprocessing, it enhances the validity of the data and improves the accuracy and robustness of the prediction;

[0041] 2. The integrated attention mechanism enables the model to automatically focus on important features. The mixed density network can output the probability distribution of future traffic demand, capture the uncertainty of data, and improve the accuracy and reliability of prediction;

[0042] 3. The model can output the prediction results of future traffic demand in a timely manner based on the latest feature data, ensuring that network resources can be dynamically adjusted to meet actual needs, thereby improving the elasticity and adaptability of the network;

[0043] 4. According to the prediction results, resources are pre-allocated reasonably to avoid resource waste or shortage, and the dynamic optimization allocation of network resources is realized. Through accurate resource pre-allocation, the network resources are utilized to the maximum extent, the operating costs are reduced, and the resource utilization efficiency is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them:

[0045] Figure 1 A schematic diagram of a flow chart of a network resource optimization method for network slicing shown in the present invention;

[0046] Figure 2 It is a schematic diagram of prediction accuracy comparison shown in the present invention;

[0047] Figure 3 A schematic diagram showing resource utilization comparison according to the present invention;

[0048] Figure 4 A schematic diagram of service quality comparison shown in the present invention;

[0049] Figure 5 It is a schematic diagram of bandwidth requirement prediction comparison of slice 1 shown in the present invention. DETAILED DESCRIPTION

[0050] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, but not all of the embodiments.

[0051] Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without making any creative work should fall within the scope of protection of the present invention.

[0052] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0053] In view of the deficiencies of the above-mentioned prior art, an embodiment of the present invention proposes a network resource optimization method for network slicing based on an attention mechanism and a hybrid density network: using a distributed data collection system, historical traffic data, user behavior data and network topology information of each network slice are collected from each network node in real time, and a high-quality feature data set is formed by preprocessing and feature extraction of multi-source data, laying the foundation for model training; the feature data set is input into the constructed prediction model, which integrates the attention mechanism and the hybrid density network algorithm. The attention mechanism can automatically focus on the most useful features for prediction, thereby improving the prediction accuracy of the model; the hybrid density network can output the probability distribution of future traffic demand, fully capturing the randomness and uncertainty of the traffic; the network weights of the model are optimized by the gradient descent algorithm to improve the generalization ability of the model. After the model training is completed, the feature data at the current moment is input, and the model can output the probability distribution prediction value of different network slices for future traffic demand; according to the predicted probability distribution, a resource pre-allocation strategy is formulated, and network resources are dynamically optimized and adjusted to achieve refined management of network resources.

[0054] According to an embodiment of the present invention, Figure 1 The flowchart shown is a network resource optimization method for network slicing, which specifically includes the following steps:

[0055] S1. Use the distributed data collection system to collect historical traffic data, user behavior data, and network topology information of each network slice from each network node in real time, perform data preprocessing on it, and form a feature data set. Among them, what needs to be explained in this step is:

[0056] Deploy data collection agents at each network node to form a distributed data collection system to collect multi-source data from each network slice in real time, including historical traffic data, user behavior data, and network topology information;

[0057] In an optional implementation, a traffic monitoring module is deployed on the network node to record the traffic information of each network slice in different time periods to obtain historical traffic data T i (t):

[0058] T i (t) = {B i (t),R i (t),L i (t)}

[0059] Among them, T i (t) is the traffic data set of the ith network slice at time t, B i (t) is the bandwidth utilization, R i (t) is the data packet transmission rate, L i (t) is the packet loss rate;

[0060] In an optional implementation, user session information is recorded at the network node access layer, the number of active users, session duration, and service type are obtained, and user behavior data U is obtained. i (t):

[0061] U i (t) = {N i (t),D i (t),S i (t)}

[0062] Among them, U i (t) is the user behavior data set of the ith network slice at time t, N i (t) is the number of active users, D i (t) is the average session duration, S i (t) is the business type distribution;

[0063] In an optional implementation, a network management protocol (such as SNMP) is used to periodically obtain network topology and performance indicators, collect network node connection relationships and link bandwidths, and obtain network topology information G(t):

[0064] G(t)=(V(t),E(t),C(t))

[0065] Where G(t) is the network topology at time t, V(t) is the node set, E(t) is the edge set (link information), and C(t) is the link capacity or bandwidth;

[0066] Furthermore, the collected data is cleaned by interpolation, 3σ principle, timestamp alignment, and normalization, and then feature extraction technology is used to extract time series features, user behavior features, and network topology features;

[0067] Specifically, the extracted time series features are the autoregressive features of historical traffic data:

[0068] FF AR,i (t) = [T i (t-1),T i (t-2),,T i (tp)]

[0069] Among them, FF AR,i (t) is the autoregressive eigenvector of the ith network slice at time t, T i (t-1) is the traffic data of the ith network slice at time tk, where t is the current time point, p is the lag order, i.e. the number of historical time steps considered, and k is the lag index, which ranges from 1 to p;

[0070] The extracted user behavior feature is the proportion of active users:

[0071]

[0072] Among them, ActiveUserRatio i (t) is the proportion of active users of the ith network slice at time t, N active,i (t) is the number of active users of the ith network slice at time t, N total,i (t) is the total number of users of the ith network slice at time t;

[0073] The extracted network topology feature is node degree centrality:

[0074] C degree (v i )=deg(v i )

[0075] Among them, C degree (v i ) is the node v i The degree centrality of i ) is the node v i The degree of is the number of edges directly connected to it, v i is the network node associated with the i-th network slice;

[0076] Statistical characteristics (mean and variance):

[0077]

[0078] Among them, μ is the mean value of the data set, σ 2 is the variance of the data set, indicating the degree of dispersion of the data, x k is the kth data point, n is the total number of data points;

[0079] Furthermore, various eigenvectors are concatenated into a complete eigenvector f i (t), the feature vectors of all samples are combined into a matrix to form a feature data set X, and the principal component analysis method is used to reduce the dimension of the feature data set X to obtain the reduced-dimensional matrix Z for subsequent prediction model training; including:

[0080] Time series feature vector:

[0081] F AR,i (t) = [T i (t-1),T i (t-2),,T i (tp)]

[0082] Among them, F AR,i(t) is the autoregressive feature vector of the ith network slice at time t;

[0083] User behavior feature vector:

[0084] F user,i (t) = [N i (t),D i (t),S i (t),ActiveUserRatio i (t)]

[0085] Among them, F user,i (t) is the user behavior feature vector of the ith network slice at time t, N i (t) is the number of active users, D i (t) is the average session duration, S i (t) is the business type distribution, ActiveUserRatio i (t) is the proportion of active users;

[0086] Network topology feature vector:

[0087] F topo,i (t) = [C degree (v i ),…]

[0088] Among them, F topo,i (t) is the network topology feature vector of the ith network slice at time t, C degree (v i ) is the node v i The degree centrality of , … are other possible topological features, such as the shortest path length between nodes and the clustering coefficient;

[0089] Concatenate the above feature vectors to form a complete feature vector:

[0090] f i (t)=[F AR,i (t),F user,i (t),F topo,i (t)]

[0091] Among them, f i (t) is the complete feature vector of the ith network slice at time t;

[0092] The feature data set X is formed, and its matrix form is expressed as:

[0093]

[0094] Where X is the feature dataset matrix, each row represents a sample (the feature vector of a network slice at a certain time point), n is the number of network slices, and r is the number of time points;

[0095] In a preferred embodiment, principal component analysis (PCA) is used to reduce the dimension of X to obtain a feature matrix Z after dimension reduction, thereby reducing the dimension of features, reducing computational complexity, avoiding model overfitting, and improving training efficiency while retaining the main information of the data; including:

[0096] Compute the covariance matrix:

[0097]

[0098] Where C is the covariance matrix, n is the number of samples, is the transpose of matrix X;

[0099] Perform eigenvalue decomposition on the covariance matrix C:

[0100] Ce i =λ i e i

[0101] Among them, λ i is the i-th eigenvalue, representing the variance of the principal component, e i is the corresponding eigenvector, indicating the direction of the principal component;

[0102] Sort the eigenvalues ​​from large to small, select the first d largest eigenvalues ​​and their corresponding eigenvectors, and form the projection matrix W:

[0103] W=[e 1 ,e 2 ,,e d ]

[0104] Project the data matrix X into the new subspace using the projection matrix W:

[0105] Z=XW

[0106] Among them, Z is the data matrix after dimensionality reduction, and each row represents the dimensionality reduction feature vector of a sample;

[0107] Preferably, the dimensionally reduced matrix Z can be used as input for subsequent prediction models to improve model efficiency and performance.

[0108] S2. Input the feature data set into the prediction model constructed based on the attention mechanism and the mixed density network algorithm, optimize the network weights through the gradient descent algorithm, and train the prediction model. It should be noted that in this step, the constructed prediction model at least includes an input layer, an attention mechanism layer, a mixed density network layer, and an output layer, where:

[0109] (1) Input layer

[0110] Input feature vector: feature dataset f obtained in step S1 i (t):

[0111] f i (t)=[F AR,i (t),F user,i (t),F topo,i (t)]

[0112] Among them, F AR,i (t) is the time series feature vector, F user,i (t) is the user behavior feature vector, F topo,i (t) is the network topology feature vector;

[0113] (2) Attention mechanism layer

[0114] Calculate the attention weights:

[0115]

[0116] Among them, α j is the attention weight of the jth feature, e j is the attention score of the j-th feature, is the learnable parameter vector of attention weights, h j is the hidden representation of the jth feature, b a is the bias term;

[0117] The weighted summation gives the attention representation:

[0118]

[0119] Among them, h att Feature representation output by the attention mechanism;

[0120] (3) Mixed Density Network Layer

[0121] Output the parameters of the mixture model:

[0122] The number of hybrid components is set to M;

[0123] For each mixture component, predict its weight π m , mean μ m and standard deviation σ m , its mathematical expression formula is as follows:

[0124]

[0125] m=1,2,,M

[0126] Among them, π m is the weight of the mth mixture component, μ m is the mean of the mth mixture component, σ m is the standard deviation of the mth mixture component, exponentially guaranteed to be positive, is the learnable weight parameter of the model, b π,m , b μ,m , b σ,m is the bias term;

[0127] (4) Output layer

[0128] Get the probability distribution of future traffic demand:

[0129] Probability density function of the prediction:

[0130]

[0131] Among them, y is the random variable of future traffic demand, Indicates that the mean is μ m , the variance is Gaussian distribution of

[0132] Furthermore, the model training includes: taking the feature data set obtained in step S1 as input and the real future traffic demand as target output; defining a negative log-likelihood loss function to measure the degree of match between the probability distribution predicted by the model and the real data; optimizing the model parameters using a gradient descent algorithm to minimize the loss function and complete the training;

[0133] In a preferred embodiment, the specific implementation of the model training includes the following steps:

[0134] Input features: feature dataset X, each sample is f i (t);

[0135] Target output: corresponding real future traffic demand y true ;

[0136] The loss function is defined as negative log-likelihood loss:

[0137]

[0138] Where N is the number of samples, is the mixed component parameter corresponding to the i-th sample, is the actual traffic demand of the i-th sample;

[0139] Select the stochastic gradient descent (SGD) algorithm to optimize the network weights and update the rules:

[0140]

[0141] Among them, θ is the set of all learnable parameters of the model, including weights and biases, η is the learning rate, is the gradient of the loss function with respect to the parameter θ;

[0142] Start initializing model parameters;

[0143] Set the number of training epochs or until the loss function converges;

[0144] Each iteration consists of the following steps:

[0145] The feature vector f i (t) Input model; calculate attention weight α j And get the attention representation h att ; Calculate the mixed component parameter π through the mixed density network layer m , μ m , σ m ; Calculate the predicted probability density function P(y|f i (t));

[0146] Use negative log-likelihood loss to calculate the loss of the current model on the training data;

[0147] Calculate the gradient of the loss function with respect to the model parameters θ;

[0148] Update model parameters using gradient descent algorithm;

[0149] At the end of each training epoch, the validation set is used to calculate the loss, where:

[0150] Set a patience value P, which means that training will be stopped when the validation loss does not improve for P consecutive epochs;

[0151] Defining the optimal validation loss NoImproveCount=0;

[0152] At the end of each epoch:

[0153] If the current validation loss Then update the best validation loss Reset the unimproved counter NoImproveCount = 0, and save the current model parameters as the best model;

[0154] Otherwise, increase the unimproved counter NoImproveCount=NoImproveCount+1;

[0155] Check if the no-improvement counter reaches the patience value P. If NoImproveCount ≥ P, stop training.

[0156] Finally, a trained prediction model is obtained, which can be used for subsequent calculations in steps S3 and S4.

[0157] S3. Input the feature data at the current moment into the trained prediction model, and output the predicted value of the probability distribution of future traffic demand of different network slices. Among them, what needs to be explained in this step is:

[0158] At the current time t, the latest feature data is obtained from the network node, and after processing in step S1, the feature vector is obtained:

[0159] f i (t)=[F AR,i (t),F user,i (t),F topo,i (t)]

[0160] Among them, f i (t) is the complete feature vector of the ith network slice at the current time t, F AR,i (t) is the time series feature, F user,i (t) is the user behavior feature, F topo,i (t) is the network topology feature;

[0161] Furthermore, the feature vector at the current moment is input into the input layer of the prediction model, and the mapping function Get the hidden representation of each feature and calculate the attention score e j , attention weight α j , generate the attention representation h att ;

[0162] In an optional implementation, the number of components in the prediction model is set to M, and the hybrid weight π is calculated m , mean μ m and standard deviation σ m , for the i-th network slice, the predicted traffic demand probability density function is:

[0163]

[0164] Among them, P i (y) is the probability density function of the future traffic demand of the ith network slice, y is the random variable of the future traffic demand, The mean is μ m ,variance Gaussian distribution of

[0165] Output mixture model parameters: mixture weight π m , mean μ m , and standard deviation σm ;

[0166] The above prediction model (predicted traffic demand probability density function) predicts P i (y), and then calculate the expected value, variance and confidence interval to obtain the probability distribution of future traffic demand:

[0167] The expected value E[Y] is calculated as:

[0168]

[0169] Among them, E[Y] represents the average predicted value of future traffic demand, that is, the expected value, and M is the number of hybrid components;

[0170] The calculation formula of variance Var[Y] is:

[0171]

[0172] Among them, Var[Y] represents the degree of dispersion of the predicted value, that is, the variance, and M is the number of mixed components;

[0173] Calculate the confidence interval to determine the range of flow demand under a certain probability:

[0174] Define the confidence level γ: for example, 95%, which means that there is a 95% probability that the traffic demand falls within this interval;

[0175] Cumulative Distribution Function (CDF):

[0176]

[0177] Among them, F i (y) is the CDF of the i-th network slice, indicating the probability that the traffic demand is less than or equal to y;

[0178] Compute confidence interval bounds:

[0179] Lower limit y lower :satisfy

[0180] upper limit y upper :satisfy

[0181] For example, when the confidence level γ is 95%, y lower Corresponding to F i (y) = 2.5%, y upper Corresponding to F i (y) = 97.5%;

[0182] It should be noted that the confidence interval provides a range of traffic demand, reflecting the uncertainty of the prediction, and helps network managers understand the maximum and minimum values ​​that traffic demand may reach under a certain probability. upper , ensuring that sufficient resources are reserved to meet high traffic demands;

[0183] In an optional embodiment, by comparing the threshold value y γ With available resource capacity C available , it can be concluded whether the current resources need to be reserved:

[0184] If y γ ≤C available : If the predicted traffic demand is lower than the available resource capacity, the current resource configuration is maintained without adjustment;

[0185] If y γ >C available :The predicted traffic demand is higher than the available resource capacity, and resource allocation needs to be increased by at least y γ -C available capacity.

[0186] As an example, define the model prediction to get two Gaussian mixture components:

[0187] Component 1:

[0188] π 1 =0.8

[0189] μ 1 =100Mbps

[0190] σ 1 =10Mbps

[0191] Component 2:

[0192] π 2 =0.2

[0193] μ 2 =200Mbps

[0194] σ 2 =20Mbps

[0195] Expected value:

[0196] E[Y]=0.8×100+0.2×200=120Mbps

[0197] It indicates that the average level of future traffic demand will be 120Mbps;

[0198] variance:

[0199] Var[Y]=[0.8(102 +100 2 )+0.2(20 2 +200 2 )-120 2 ]-120 2

[0200] =[0.8(100+10000)+0.2(400+40000)]-14400

[0201] =[8080+8080]-14400=16160-14400=1760Mbps 2

[0202] Variance is 1760Mbps 2 , the standard deviation is

[0203] Setting the confidence level γ to 95%, we obtain:

[0204] y lower ≈70Mbps

[0205] y upper ≈190Mbps

[0206] Should consider upper =190Mbps is used as the basis for resource allocation to ensure that the network can carry peak traffic.

[0207] S4: According to the predicted probability distribution of future traffic demand for different network slices, formulate a resource pre-allocation strategy and dynamically optimize and adjust network resources. It should be noted that the traffic demand threshold y obtained in step S3 is γ , determine the amount of resources that need to be pre-allocated, considering the total capacity of network resources C total According to the needs of different network slices, resources are allocated reasonably and resource allocation plans are formulated, including increasing or decreasing resource configuration of certain network slices; among which:

[0208] According to the prediction results of step S3, the traffic demand threshold is obtained based on the confidence level γ, and the resource demand R of each network slice is calculated. i =y γ,i ;

[0209] Total resource requirement R total =∑ i R i ;

[0210] Determine whether the total resource demand exceeds the total resource capacity C total :

[0211] If R total ≤Ctotal , then it is allocated to each network slice according to the calculated resource requirements;

[0212] If R total >C total , resource scheduling and optimization are required, including priority sorting and resource preemption strategies. For example, according to the service level agreement (SLA) information, each network slice is assigned a priority. The larger the value, the higher the priority and the minimum resource requirement. Indicates the minimum amount of resources and maximum resource requirements required for the normal operation of each network slice Indicates the maximum amount of resources that can be used by each network slice;

[0213] In an optional implementation, the priority sorting strategy is to sort by priority P i Sort from high to low to get the sorted network slice list {S 1 ,S 2 ,,S N}; The specific implementation process includes the following steps:

[0214] Traverse the sorted list, for each network slice S i ;

[0215] If R i ≤C available , assign R i Resources, Update C available =C available -R i ;

[0216] if distribute Setting C available =0;

[0217] Otherwise, record the network slice S that cannot be satisfied i , marked as requiring resource preemption;

[0218] In an optional implementation, the resource preemption strategy includes:

[0219] Starting from the low-priority network slice, traverse from low to high priority and calculate the amount of resources that can be preempted:

[0220]

[0221] If ΔR j >0, ΔR j Resources are allocated to high priority network slices S i ;

[0222] Update resource allocation for low priority network slices:

[0223]

[0224] Update resource allocation for high priority network slices:

[0225]

[0226] Update the remaining requirements. If the resource requirements of the high-priority network slices have been met, stop preemption.

[0227] In an optional implementation manner, resource scheduling and optimization also includes the following implementation manners:

[0228] Calculate the total resource gap:

[0229] ΔR total =R total -C total

[0230] Calculate the reduction ratio δ:

[0231]

[0232] Update resource allocation for each network slice:

[0233]

[0234] Check minimum resource requirements: If

[0235] Adjust the reduction ratio δ, or make greater resource reductions for low-priority network slices;

[0236] Furthermore, in the case of extremely insufficient resources, service requests from low-priority network slices may need to be rejected:

[0237] Allocate its resources

[0238] Allocate the released resources to high-priority network slices;

[0239] Record the final resource allocation for each network slice

[0240] Through the network management system, the new resource allocation plan is sent to each network node (for example, adjusting bandwidth allocation, computing resources, and storage resources), and notifications are sent to the affected network slices, including information on resource adjustments and possible service impacts.

[0241] It should also be noted in this embodiment that the traditional prediction and resource allocation method based on the long short-term memory network (LSTM) model usually outputs a single-point estimate of future traffic when performing traffic prediction, and cannot provide a confidence interval or probability distribution of the prediction result. LSTM cannot quantify the uncertainty of the prediction, resulting in resource allocation that may be too conservative or aggressive; LSTM models tend to predict the average value of future traffic and cannot effectively capture the multimodal characteristics of traffic demand. Network traffic is affected by many factors, including user behavior, network topology, etc. LSTM is mainly good at processing time series data and it is difficult to effectively integrate non-sequence features. The LSTM model treats input features equally and cannot identify and highlight the most important features for prediction, which limits the model's prediction ability; the LSTM's loop structure leads to a long training time, which is not conducive to real-time prediction and rapid response to needs. Network traffic data often contains noise and outliers. The LSTM model is sensitive to these data, which may lead to unstable prediction results. Network operators cannot obtain information about the causes or patterns of traffic changes from the model, which is not conducive to strategy formulation;

[0242] The effectiveness of the method of the present invention in network resource optimization is verified by conducting comparative experiments with the traditional prediction and resource allocation method based on the long short-term memory network (LSTM) model, aiming to verify the effectiveness of the method of the present invention in the following aspects:

[0243] Improve prediction accuracy: By integrating the attention mechanism and hybrid density network, the prediction accuracy of future traffic demand can be improved;

[0244] Provides forecast uncertainty quantification: outputs a probability distribution of flow demand rather than a single point estimate;

[0245] Optimize resource utilization: formulate dynamic resource pre-allocation strategies based on predicted probability distribution to improve the utilization efficiency of network resources;

[0246] Ensure service quality: avoid the decline of service quality due to insufficient resource allocation;

[0247]

Experimental subjects

[0248] Network environment: simulates a 5G network environment with 6 network slices, including enhanced mobile broadband (eMBB), ultra-reliable low-latency communication (URLLC) and massive machine-type communication (mMTC);

[0249] Table 1. Network slice list

[0250]

[0251]

[0252]

Experimental data

[0253] Data collection:

[0254] Historical traffic data: traffic data for each slice for the past 7 days, with a sampling interval of 10 minutes, for a total of 1008 data points;

[0255] User behavior data: including the number of active users, average session duration, and business type ratio;

[0256] Network topology information: node connection relationship, link bandwidth, node load;

[0257] Data preprocessing:

[0258] Data cleaning: remove missing values ​​and outliers;

[0259] Data alignment: Time series synchronization to ensure that timestamps of different data sources are consistent;

[0260] Data normalization: Min-Max normalization is used to scale the data to the [0,1] interval;

[0261] Feature extraction:

[0262] Time series features: lagged values ​​of historical traffic (6 data points in the past hour);

[0263] User behavior characteristics: percentage of active users, average session duration;

[0264] Network topology characteristics: node degree centrality, network average load;

[0265]

Experimental environment conditions

[0266] Server: Intel Xeon E5-2650 v4 @ 2.20 GHz (16 cores), 128 GB RAM, NVIDIA Tesla V100 GPU;

[0267] Operating system: Ubuntu 20.04LTS;

[0268] Development platform: Python 3.8, TensorFlow 2.5, Keras;

[0269] MATLAB R2021b: for data visualization and results analysis;

[0270]

Experimental process

[0271] Traditional methods

[0272] Model construction: Use LSTM model for traffic prediction. The model structure is a single-layer LSTM with 64 hidden units.

[0273] Model training:

[0274] Input: time series features (historical traffic data);

[0275] Output: Single-point forecast of traffic demand in the next 10 minutes;

[0276] Loss function: mean square error (MSE);

[0277] Optimizer: Adam, learning rate 0.001;

[0278] Training set and validation set division: 80% for training and 20% for validation;

[0279] Number of training epochs: 50;

[0280] Resource allocation strategy: Allocate equal amounts of network resources based on predicted single-point traffic demand;

[0281] Method of the present invention

[0282] Model construction: prediction model based on attention mechanism and mixture density network (MDN);

[0283] Attention mechanism: used to weight the importance of different features;

[0284] MDN layer: outputs the probability distribution parameters of traffic demand (mean, variance and weight of mixed Gaussian distribution);

[0285] Model training:

[0286] Input: comprehensive feature vector (time series features, user behavior features, network topology features);

[0287] Output: Probability distribution of traffic demand in the next 10 minutes;

[0288] Loss function: Negative Log-Likelihood;

[0289] Optimizer: Adam, learning rate 0.001;

[0290] Training set and validation set division: 80% for training and 20% for validation;

[0291] Number of training epochs: 50;

[0292] Resource allocation strategy: Based on the predicted probability distribution, the traffic demand threshold is calculated at a 95% confidence level and the corresponding amount of resources is allocated.

[0293] Reference Figure 2It can be seen intuitively that the MAPE value of the traditional method is relatively high, ranging from 10.7% to 15.3%. Since LSTM cannot quantify prediction uncertainty and has insufficient ability to fuse multi-source features, the prediction error is large; the MAPE value of the method of the present invention is significantly reduced, ranging from 5.4% to 6.2%, and the prediction accuracy is improved by about 50%. Since the method of the present invention can automatically focus on the most important features for prediction by integrating the attention mechanism, the ability to capture complex patterns is improved, and the mixed density network can output the probability distribution of future traffic demand, quantify the prediction uncertainty, and improve the prediction accuracy.

[0294] Reference Figure 3 It can be seen intuitively that the resource utilization rate of the traditional method is between 85% and 90%, which is low. Due to the large prediction error, the traditional method needs to reserve more redundant resources to avoid insufficient resources, resulting in resource waste; the resource utilization rate of the method of the present invention is between 93% and 96%, which is close to the ideal value. Since the method of the present invention calculates resource requirements at a confidence level of 95% by predicting the probability distribution of traffic requirements, resource allocation is more accurate, resource waste is reduced, resource utilization efficiency is improved, and operating costs are reduced.

[0295] Reference Figure 4 It can be seen intuitively that the traditional method has 4-8 service interruptions on each network slice. Due to inaccurate prediction, especially underestimation of traffic during peak hours, insufficient resource allocation leads to service interruption. Service interruption affects user experience, resulting in user churn and service complaints. The method of the present invention has no service interruption on all network slices, and the number of times is 0. Since the method of the present invention quantifies prediction uncertainty and takes possible traffic fluctuations into account when allocating resources, it avoids resource shortages, improves service reliability, and meets the requirements of the service level agreement (SLA).

[0296] Reference Figure 5 The actual observed demand curve represents the actual measured traffic demand of Slice 1. The curve shows periodic fluctuations, reflecting the changing trend of traffic demand. Random noise is added to simulate the uncertainty in the actual network environment. The prediction curve of the traditional method cannot keep up with the actual changes at the inflection point of traffic increase or decrease, showing lag. During the peak period, the predicted value is significantly lower than the actual demand, resulting in insufficient resource allocation. In addition, the prediction curve is relatively smooth and cannot capture the rapid changes in traffic.

[0297] The prediction curve of the method of the present invention closely follows the actual demand, that is, the predicted value is highly consistent with the actual demand, accurately capturing the fluctuation of flow rate. At the inflection point of flow rate change, the predicted value can be adjusted quickly, showing good sensitivity. Even in the presence of random noise, the predicted value is still stable and reliable.

[0298] It should also be noted that the LSTM model has limited ability to capture sudden and nonlinear traffic changes, resulting in prediction lags and errors, and is unable to quantify prediction uncertainties, leading to deficiencies in resource allocation. The method of the present invention uses an attention mechanism to enable the model to focus on key features and improve sensitivity to traffic changes. The mixed density network can capture the probability distribution of traffic demand and provide more accurate and robust predictions.

[0299] like Figure 5 As shown, the superiority of the method of the present invention in traffic demand prediction is intuitively demonstrated, and its effectiveness is verified. Through accurate traffic prediction, a reliable basis is provided for the optimal allocation of network resources, resource utilization is improved, and service quality is guaranteed.

[0300] The data preprocessing method of the aforementioned collected data can be performed using methods and means in the prior art, and will not be described in detail in this example.

[0301] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A network resource optimization method for network slicing, characterized in that: include: Using the distributed data collection system, historical traffic data, user behavior data, and network topology information of each network slice are collected from each network node in real time, and data preprocessing is performed to form a feature data set; Inputting the feature data set into a prediction model constructed based on an attention mechanism and a hybrid density network algorithm, optimizing network weights through a gradient descent algorithm, and training the prediction model; Input the feature data at the current moment into the prediction model after training, and output the predicted value of the probability distribution of future traffic demand of different network slices; Based on the predicted probability distribution of future traffic demand for different network slices, a resource pre-allocation strategy is formulated to dynamically optimize and adjust network resources.

2. The network resource optimization method for network slicing according to claim 1, characterized in that: The historical traffic data at least includes bandwidth utilization, data packet transmission rate and data packet loss rate; The user behavior data at least includes the number of active users, average session duration and business type distribution; The network topology information at least includes a node set, an edge set, and link capacity or bandwidth.

3. The network resource optimization method for network slicing according to claim 1 or 2, characterized in that: The feature advance technology is used to extract time series features, user behavior features and network topology features from the historical traffic data, the user behavior data and the network topology information to form a complete feature vector: f i (t)=[F AR,i (t),F user,i (t),F topo,i (t)] Among them, f i (t) is the complete feature vector of the ith network slice at time t, F AR,i (t) is the time series feature vector, F user,i (t) is the user behavior feature vector, F topo,i (t) is the network topology feature vector.

4. The network resource optimization method for network slicing according to claim 3, characterized in that: The complete feature vector is used to form a feature data set X, which is expressed in matrix form as follows: Among them, X is the feature dataset matrix, each row represents the feature vector of a network slice at a certain time point, n is the number of network slices, and r is the number of time points.

5. The network resource optimization method for network slicing according to claim 1, characterized in that: The constructed prediction model includes at least an input layer, an attention mechanism layer, a mixed density network layer and an output layer.

6. The network resource optimization method for network slicing according to claim 1, characterized in that: Training the prediction model comprises: Taking the feature data set as input and the real future traffic demand as target output; Define the negative log-likelihood loss function to measure how well the probability distribution predicted by the model matches the real data; The gradient descent algorithm is used to optimize the model parameters, minimize the loss function, and complete the training.

7. The network resource optimization method for network slicing according to claim 1, characterized in that: Output the predicted values ​​of the probability distribution of future traffic demand for different network slices, including: Inputting the feature vector at the current moment into the input layer of the prediction model; Set the number of components of the prediction model to M and calculate the mixing weight π m , mean μ m and standard deviation σ m ; For the i-th network slice, the predicted traffic demand probability density function is: Among them, P i (y) is the probability distribution prediction value of the future traffic demand of the ith network slice, y is the random variable of the future traffic demand, The mean is μ m ,variance Gaussian distribution.

8. The network resource optimization method for network slicing according to claim 7, characterized in that: include: According to the predicted value P i (y), based on the confidence level γ, the flow demand threshold y is obtained γ , calculate the resource requirement R of each network slice i =y γ,i ; Total resource requirement R total =∑ i R i ; Determine whether the total resource demand exceeds the total resource capacity C total ,in: If R total ≤C total , then it is allocated to each network slice according to the calculated resource requirements; If R total >C total , then resource scheduling and optimization are performed.

Images