A method and system for predicting 5G network slice traffic

Abstract

This invention discloses a method and system for predicting 5G network slice traffic, relating to the field of network traffic prediction technology. The method includes: mapping network slices to hypergraph nodes, mapping slice clusters sharing the same physical features to hyperedges, and constructing a hypergraph association matrix; constructing spatiotemporal feature tensors from the original features of each slice, inputting them into a hypergraph convolutional network to extract spatial features, and obtaining updated network slice features; concatenating the updated features with the previous hidden state, and adaptively generating a fractional order based on the concatenated features; using this fractional order to perform a weighted summation of the hidden states at multiple historical moments to obtain the updated hidden state; and finally outputting the traffic prediction result based on the updated hidden state. This invention captures high-order spatial coupling relationships between slices through hypergraph convolution and overcomes the long-sequence forgetting problem of traditional recurrent neural networks through an adaptive fractional-order memory mechanism, significantly improving the accuracy and robustness of network slice traffic prediction.

Description

Technical Field

[0001] This invention relates to the field of network traffic prediction technology, and specifically to a method and system for predicting 5G network slice traffic. Background Technology

[0002] 5G network slicing technology divides the physical network into multiple virtual logical networks, each serving one of three types of services: eMBB (enhanced Mobile Broadband), uRLLC (ultra-reliable low-latency communication), and mMTC (massive machine-type communication). To ensure service quality and maximize network resource utilization, operators need to dynamically adjust slice resources based on real-time changes in network traffic. Accurately predicting network traffic is the foundation for achieving intelligent network operation and maintenance.

[0003] There are three main types of existing traffic prediction technologies. Traditional statistical methods (such as ARIMA) are suitable for stationary linear sequences; deep learning methods (such as LSTM and GRU) can capture nonlinear temporal features; and graph neural network methods (such as GCN) attempt to fuse the spatial dependencies of base station topology. However, these technologies have the following key drawbacks in 5G network slicing scenarios:

[0004] (1) Long-term reliance on insufficient capture capability: Recurrent neural networks such as LSTM and GRU mainly rely on integer derivatives when performing gradient updates and state memory. However, real network traffic (especially mMTC and V2X services) often exhibits obvious fractal characteristics and self-similarity, and there is a weak long-memory connection between the current traffic state and the long-ago historical state. As the time step increases, the memory gradient of traditional integer-order RNN architecture decays rapidly, which cannot effectively utilize long-span historical time-series information, resulting in reduced prediction accuracy.

[0005] (2) Insufficient ability to model high-order spatial relationships: Traditional graph convolutional networks (GCNs) are based on simple graph structures and only consider the binary connection relationship between nodes. However, in 5G slicing networks, multiple slices (such as eMBB and uRLLC) often share the same set of physical infrastructure (such as base stations and edge computing nodes) or are affected by the same social activities, exhibiting complex many-to-many group characteristics. Relying solely on the binary graph topology structure makes it difficult to effectively aggregate high-order relational semantics, and the model is prone to ignoring the potential resource competition or cooperative fluctuations between slices, thus weakening the prediction accuracy.

[0006] To address the aforementioned issues, this invention proposes a 5G network slice traffic prediction method based on hypergraph convolution and adaptive fractional-order memory mechanism, which simultaneously solves the technical challenges of long-term reliance on capture and high-order spatial relationship modeling. Summary of the Invention

[0007] This invention provides a method and system for predicting 5G network slice traffic, in order to solve the problem of low accuracy in predicting 5G network slice traffic in existing traffic prediction technologies.

[0008] This invention is achieved through the following technical solution:

[0009] A first aspect of the present invention provides a method for predicting 5G network slice traffic, comprising:

[0010] Network slices are mapped to nodes in the hypergraph, clusters of network slices sharing the same physical features are mapped to hyperedges, and a hypergraph association matrix is ​​constructed based on the connection relationship between the nodes and the hyperedges.

[0011] The original features of each network slice in the hypergraph are constructed into a spatiotemporal feature tensor. The spatiotemporal feature tensor and the hypergraph association matrix are input into the hypergraph convolutional network for spatial feature extraction to obtain the updated features of the network slice.

[0012] The network slice update features are concatenated with the previous hidden state to obtain the concatenated features; wherein, the previous hidden state is obtained when the traffic prediction task was performed in the previous time step using the same temporal neural network model;

[0013] A fractional order is generated based on the splicing features. The fractional order is then used to perform a weighted summation of the hidden states at multiple historical moments recorded by the temporal neural network model to obtain the updated hidden states.

[0014] Based on the updated hidden state, a prediction result for network slice traffic is generated.

[0015] Furthermore, the physical characteristics include at least one of computing resource characteristics and transmission link characteristics.

[0016] Furthermore, the step of constructing a spatiotemporal feature tensor from the original features of each network slice in the hypergraph includes:

[0017] The historical traffic data of the network slices are subjected to max-min normalization to eliminate the dimensional differences between different network slices;

[0018] The original features are extracted from the normalized historical traffic data and integrated into a spatiotemporal feature tensor. , ;in, Represents the set of real numbers. This indicates the length of the sliding time window used when integrating historical traffic data. Indicates the number of network slices. The dimension represents the original feature, which includes traffic load and packet loss rate.

[0019] Furthermore, the hypergraph convolutional network includes a node-to-hyperedge aggregation module, a hyperedge feature weighting module, and a hyperedge-to-node backpropagation update module;

[0020] The node-to-hyperedge aggregation module is used to perform aggregation operations on the original features of each node in the same hyperedge to obtain hyperedge aggregated features;

[0021] The hyperedge feature weighting module is used to perform weighted calculation on the hyperedge aggregated features based on the weight of the hyperedge to obtain updated hyperedge aggregated features;

[0022] The hyperedge-to-node backhaul update module is used to backhaul the updated hyperedge aggregation features to each node to update the feature representation of each node and obtain the network slice update features.

[0023] Further, the step of generating the prediction result of network slice traffic based on the updated hidden state includes:

[0024] The spliced ​​features are input into the temporal neural network for state update to obtain state update features;

[0025] The state update feature is concatenated with the updated hidden state to obtain the hidden state at the current time.

[0026] The hidden state at the current moment is input into the output layer of the temporal neural network to generate the prediction result of the network slice traffic.

[0027] Further, the step of generating a fractional order based on the splicing features, and using the fractional order to perform a weighted summation of the hidden states at multiple historical moments recorded by the temporal neural network model to obtain the updated hidden states, includes:

[0028] The concatenated features are input into a fractional recurrent network to obtain the updated hidden state; wherein, the fractional recurrent network includes a fractional generation module and a fractional convolution calculation module;

[0029] The fractional order generation module is used to generate the fractional order number based on the splicing features;

[0030] The fractional convolution calculation module is used to perform a weighted summation of the hidden states at multiple historical moments based on the fractional order to obtain the updated hidden states.

[0031] Furthermore, the fractional order generation module includes a linear transformation layer, a sigmoid activation function, and a mapping layer connected in sequence;

[0032] The linear transformation layer is used to map the spliced ​​features into scalars;

[0033] The Sigmoid activation function is used to compress the scalar to the (0,1) interval;

[0034] The mapping layer is used to map the output value of the Sigmoid activation function to... Interval.

[0035] Furthermore, the step of using the fractional order to perform a weighted summation of the hidden states at multiple historical moments recorded by the temporal neural network model to obtain the updated hidden states includes:

[0036] Calculate the fractional order coefficients based on the fractional order number;

[0037] The fractional-order coefficients are convolved with the hidden states at multiple historical time points to obtain the updated hidden states.

[0038] Furthermore, the method also includes the step of training the network parameters of the hypergraph convolutional network and the fractional-order recurrent network using traffic sample data from network slices;

[0039] Among them, an asymmetric loss function is used. Conduct training:

[0040] ;

[0041] in, This represents the total number of traffic sample data. and The first The actual and predicted flow values ​​for each sample For asymmetric penalty weights; when hour, Take a value greater than 1, otherwise The value is 1.

[0042] A second aspect of the present invention provides a 5G network slice traffic prediction system, the prediction system being used to perform the 5G network slice traffic prediction method according to any one of the first aspects, comprising:

[0043] The hypergraph construction unit is used to map network slices to nodes in the hypergraph, map clusters of network slices sharing the same physical features to hyperedges, and construct a hypergraph association matrix based on the connection relationship between the nodes and the hyperedges.

[0044] The hypergraph feature extraction unit is used to construct a spatiotemporal feature tensor from the original features of each network slice in the hypergraph, and input the spatiotemporal feature tensor and the hypergraph association matrix into the hypergraph convolutional network for spatial feature extraction to obtain the updated features of the network slice.

[0045] The feature splicing unit is used to splice the network slice update features with the previous hidden state to obtain spliced ​​features; wherein, the previous hidden state is obtained when the traffic prediction task was performed in the previous time step using the same temporal neural network model;

[0046] A fractional-order memory unit is used to generate a fractional-order number based on the splicing features, and to use the fractional-order number to perform a weighted summation of the hidden states of multiple historical moments recorded by the temporal neural network model to obtain an updated hidden state.

[0047] The prediction output unit is used to generate prediction results for network slice traffic based on the updated hidden state.

[0048] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0049] This invention constructs a hypergraph structure and its association matrix, aggregating multiple slices sharing the same physical features to the same hyperedge. By performing convolution operations on the hypergraph, message passing between nodes, hyperedges, and nodes is achieved, and network slice update features are obtained. This effectively extracts high-order spatial coupling features between slices, solving the problem that traditional graph convolution cannot model "many-to-many" group relationships.

[0050] By concatenating the updated features of network slices with the previous hidden state and dynamically generating fractional order, and using the dynamic order to perform weighted summation on multiple historical hidden states, the model can adaptively adjust the memory depth and directly read long-term information from the historical buffer pool. This overcomes the problem of long-term dependency loss caused by gradient decay in integer-order RNNs and significantly improves the prediction accuracy of traffic sequences with fractal and self-similar features.

[0051] By combining hypergraph convolutional features with adaptive fractional-order features, the accuracy of traffic prediction is significantly improved. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0053] Figure 1 This is a flowchart of a 5G network slice traffic prediction method according to an embodiment of the present invention;

[0054] Figure 2 This is a schematic diagram of a hypergraph convolutional network according to an embodiment of the present invention;

[0055] Figure 3 This is a schematic diagram of a fractional-order cyclic network according to an embodiment of the present invention;

[0056] Figure 4 This is a schematic diagram of a training model according to an embodiment of the present invention;

[0057] Figure 5 This is a bar chart showing the quantitative comparison of the model of this invention with LSTM and GRU in terms of mean absolute error and mean square error.

[0058] Figure 6 This is a graph showing the fractional order change curves of the model of this invention adaptively learned on eMBB, uRLLC, and mMTC slices;

[0059] Figure 7 This is a comparison of the convergence curves of the loss function of the model of this invention with those of LSTM and GRU during the training process. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0061] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims, and accompanying drawings of this invention are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to other steps or units inherent in the device.

[0062] The terminology used in the various embodiments of the invention is for the purpose of describing particular embodiments only and is not intended to limit the various embodiments of the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of the invention pertain. The terms (such as those defined in commonly used dictionaries) are to be interpreted as having the same meaning as in the context of the relevant technical field and are not to be interpreted as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of the invention.

[0063] This invention addresses the shortcomings of existing 5G traffic prediction technologies by proposing a 5G network slice traffic prediction method based on an adaptive fractional-order hypergraph neural network. This method employs a hypergraph convolutional network (HGNN), using hyperedges to aggregate all slices sharing the same physical resource. Through a "node-hyperedge-node" message passing mechanism, it effectively extracts high-order spatial correlation features between slices. Simultaneously, it constructs a historical memory buffer and introduces an adaptive fractional-order calculus mechanism. By dynamically adjusting the fractional order, it forces the model to read more distant historical states, thereby accurately capturing the long-term dependency features of traffic data and significantly improving the accuracy of network slice traffic prediction.

[0064] Please see Figure 1 , Figure 1 This invention illustrates a method for predicting 5G network slice traffic, as detailed in the following steps. Specifically, this method analyzes resource competition among different slices by establishing a hypergraph model, and simultaneously utilizes fractional calculus to capture the long-term dependency characteristics of network traffic.

[0065] S101, Construct the hypergraph topology: Map network slices to nodes in the hypergraph, map clusters of network slices sharing the same physical characteristics to hyperedges, and construct the hypergraph association matrix based on the connection relationship between nodes and hyperedges.

[0066] Specifically, historical traffic records for different network slices (eMBB, uRLLC, mMTC) are first collected from the 5G network management component, and a hypergraph node is created for each network slice instance. For any type of physical feature, all slice instances sharing that resource are grouped into a set, which is mapped to a hyperedge, forming a hypergraph topology. The physical features can be one or more of the underlying physical infrastructure, such as computing resource features (computing resources of the same base station) and transmission link features (spectrum resources of the same transmission link).

[0067] Based on the connection relationships between each node and each hyperedge in the hypergraph topology, construct the hypergraph association moments. Where N is the total number of nodes and M is the total number of hyperedges. Represents the set of real numbers. If the slice node Occupied super edge The corresponding resources are the elements of the association matrix. ,otherwise This association matrix describes the attribution relationship between slice nodes and physical resource hyperedges, providing a topological structure for subsequent hypergraph convolution feature extraction.

[0068] S102, Hypergraph Convolutional Feature Extraction: The original features of each network slice in the hypergraph are constructed into a spatiotemporal feature tensor. The spatiotemporal feature tensor and the hypergraph association matrix are input into the hypergraph convolutional network for spatial feature extraction to obtain the updated features of the network slice.

[0069] In this step, the raw traffic data is first preprocessed. The collected historical traffic data contains raw characteristics such as traffic load and packet loss rate for each slice at different times. Since the traffic scales of different slices vary significantly (e.g., eMBB is at the GB level, while mMTC is only at the KB level), a normalization method is used to preprocess the raw traffic data to eliminate the differences in units.

[0070] Then, the normalized raw flow data is integrated into a spatiotemporal feature tensor. ,in Represents the set of real numbers. This indicates the length of the sliding time window used when integrating historical traffic data. Indicates the number of network slices. Dimensions representing the original characteristics (traffic load, packet loss rate, etc.).

[0071] Hypergraph Convolutional Networks (HGNNs) achieve spatial feature extraction through a two-stage aggregation of "node-hyperedge-node":

[0072] (1) Node-to-Hyperedge operation: Aggregate the features of all nodes within the same hyperedge to obtain the hyperedge features (reflecting the overall load status of the physical resource).

[0073] (2) Hyperedge-to-Node strategy: Learnable weighting is applied to the hyperedge features (different physical resources have different importance); finally, the weighted hyperedge features are back to each node to update the feature representation of the node.

[0074] Spatiotemporal feature tensor and the correlation matrix Input the hypergraph convolutional network and obtain network slices to update features.

[0075] The above extraction process can be repeated multiple times (by stacking multiple convolutional layers) to extract deeper, higher-order coupling information. After hypergraph convolution, the output network slices are used to update features. Its dimensions and input features The same, but this feature incorporates resource load information from other competing slices.

[0076] In one implementation, the hypergraph convolutional network stacks L (L=2~4) identical layers, each performing the same two-stage information transfer from node to hyperedge and from hyperedge to node, with the output of the previous layer serving as the input of the next layer. By repeatedly extracting information multiple times, the receptive field of each node continuously expands, thereby capturing higher-order, more global patterns in the graph.

[0077] like Figure 2 As shown, each layer includes a node-to-hyperedge aggregation module, a hyperedge feature weighting module, and a hyperedge-to-node backhaul update module.

[0078] Node-to-hyperedge aggregation module: Performs aggregation operations on the original features of each node in the same hyperedge to obtain the hyperedge aggregated features;

[0079] Hyperedge feature weighting module: The hyperedge aggregate features are weighted based on the weights of the hyperedges to obtain updated hyperedge aggregate features;

[0080] The hyperedge-to-node backhaul update module: backhauls the updated hyperedge aggregation features to each node to update the feature representation of each node and obtains the network slice update features.

[0081] Specifically, the calculation process for updating network slice features can be represented as follows:

[0082]

[0083] In the formula, This represents the output of the l-th layer (the network slice updated by the l-th layer updates the features). This represents the output of layer l+1 (the network slice updated by layer l+1 updates the features). Represents the hypergraph incidence matrix. express The transpose of the matrix; This represents the diagonal matrix of hyperedge weights; The degree matrix (diagonal matrix) represents the nodes and is used for normalization to prevent nodes with large degrees from having excessively large features; The degree matrix (diagonal matrix) representing the hyperedge is also used for normalization. The weight matrix represents the filter parameter matrix to be learned in the current layer (layer l+1); This represents a non-linear activation function.

[0084] The weight of a hyperedge quantitatively describes its importance or strength; a larger weight indicates a stronger and more important connection between the nodes represented by that hyperedge. The weights of all hyperedges are organized into an M×M diagonal matrix. This yields the hyperedge weight diagonal matrix, where M is the number of hyperedges. Indicates the superedge The weights. Those skilled in the art will understand that the hyperedge... The weights can be obtained from the hypergraph convolutional network. The general approach is to: As a trainable parameter, it participates in forward computation and backpropagation, initially... The values ​​are initialized to 1 (or a random small value). During training, the parameters are adjusted using gradient descent to minimize the loss function and output the final weights. .

[0085] S103, Feature Concatenation: Concatenate the updated features of the network slice with the previous hidden state to obtain the concatenated features; where the previous hidden state refers to the hidden state obtained when the same temporal neural network model was used to perform the traffic prediction task in the previous time step.

[0086] Let the current time be t, and the network slice update feature output in step S102 be: This refers to the output features of the hypergraph convolutional network at time t. The previous hidden state. This is the hidden state calculated by the model when it performed predictions at the previous time (t-1). For the initial time t=0, Let it be the zero vector.

[0087] In this step, the temporal neural network model can be LSTM, GRU, or other similar models. The hidden states are used by the output layer (such as a fully connected layer) of the temporal neural network model to perform classification or prediction. At the current time t, the network slices are updated with new features. Hidden state from the previous moment splicing features Input a temporal neural network model (the concatenated features simultaneously contain current spatial augmentation information and memory information from the previous time step), and the model extracts the new hidden state at the current time step. The model output layer performs traffic prediction based on this new hidden state, and... As input for the next moment.

[0088] S104, Fractional Order Memory Enhancement: Generate a fractional order based on the splicing features, and use the fractional order to perform a weighted summation of the hidden states of multiple historical moments recorded by the temporal neural network model to obtain the updated hidden states.

[0089] Unlike existing time series prediction models that rely on integer derivatives for gradient updates and state memory, this embodiment will concatenate features. This is used to generate a dynamic fractional order and to perform a weighted summation of the hidden states from multiple historical time points to obtain an updated hidden state, enabling the model to adaptively adjust the memory depth. Finally, the temporal neural network model performs traffic prediction based on this updated hidden state.

[0090] like Figure 3 As shown, this step is implemented using a fractional-order recurrent network, which consists of a fractional-order generation module and a fractional-order convolution calculation module. Simultaneously, a historical state buffer is maintained for the temporal neural network to store the hidden states from the past K time steps. , ... .

[0091] The internal computational logic of the fractional-order recurrent network is as follows: the output of the hypergraph convolutional network is concatenated with the previous hidden state, and then the features are concatenated. The input is fed into the fractional order generation module to generate the fractional order. The fractional-order convolution calculation module calls upon the historical hidden states in the historical state buffer pool, performs a weighted summation of the hidden states at K historical time points based on the fractional-order convolution, and outputs a fractional-order historical fusion term (i.e., the updated hidden state). This term is passed to the output layer (fully connected layer) of the temporal neural network model to predict the traffic of each network slice at future time points. Simultaneously, this updated hidden state is stored in the historical state buffer pool for use in the next prediction task.

[0092] In one specific implementation, the hidden states at multiple historical moments recorded by the temporal neural network model are weighted and summed using fractional order to obtain updated hidden states, including:

[0093] S401, Calculate the fractional order coefficients based on the fractional order;

[0094] S402, convolve the fractional coefficients with the hidden states at multiple historical time points to obtain the updated hidden states.

[0095] Specifically, the fractional-order coefficients are convolved with the hidden states at multiple historical time points, as follows:

[0096]

[0097] In the formula, Based on the current order The fractional coefficients are polynomial coefficients, consisting of K coefficients. It represents the Hadamah accumulation. This refers to the hidden state of historical moment tK. This represents the fractional-order historical fusion term, which is the updated hidden state.

[0098] The fractional order generation module dynamically generates a fractional order from the current input. It is used to control the weighted summation of historical hidden states of a fixed length K, through fractional order. Significantly non-zero bits adaptively adjust memory depth. The specific implementation steps are as follows: fractional order, order... (0<α<1) determines the decay rate of historical state weights, thus affecting the effective memory depth. A value close to 0 indicates long memory (uniform or slow decay). A value close to 1 indicates short memory (rapid decay).

[0099] The fractional order is dynamically generated by the current input (concatenated features) through the fractional order generation module. The model parameters learn how to output appropriate values ​​through end-to-end training (backpropagation). To minimize prediction loss.

[0100] The weighted summation formula is based on principles from fractional calculus, using binomial coefficients to normalize the sum of the weights of the historical states. The specific calculation is as follows:

[0101] For the fractional order at the current moment (The value usually ranges from 0.5 to 1.5), the first A history ( The corresponding coefficient is:

[0102]

[0103] in, It is the coefficient of the generalized binomial.

[0104] when When the coefficient is not an integer, it is a sequence of real numbers with alternating signs and gradually decaying.

[0105] In one specific implementation, the fractional order generation module is implemented using a sub-neural network. This sub-neural network consists of a linear transformation layer, a sigmoid activation function, and a mapping layer. The linear transformation layer maps the concatenated features to a scalar, the sigmoid function compresses them to the (0,1) interval, and the mapping layer constrains them to a specific value by adding a bias. The range, calculated using the following formula:

[0106] In the formula, This represents the network slice update feature at time t. Indicates the previous hidden state; and Here are the learnable parameters of the linear transformation layer, where... The weight matrix used to calculate the order. Indicates the bias term; This represents the activation function. Let be the fractional order at the current time t.

[0107] The sub-neural network parameters learn how to output appropriate values ​​through end-to-end training (backpropagation). To minimize prediction loss.

[0108] In another implementation, the concatenated features are split into two paths. One path is input into a fractional-order recurrent network to generate a fractional-order history fusion term, and the other path is input into a temporal neural network to update the hidden features at the current time step. The two paths are then concatenated to form an updated hidden state. Finally, the prediction result is output based on the updated hidden state to balance short-term and long-term memory.

[0109] S105, Traffic Prediction: Based on the updated hidden state, generate predictions of network slice traffic.

[0110] The updated hidden state from step S104 is input into the output layer of the temporal neural network to obtain the traffic prediction result.

[0111] Specifically, in this step, a fully connected layer is used as the output layer. The fully connected layer outputs the prediction result of the network slice traffic through linear transformation. The prediction result is a vector that includes the traffic values ​​of multiple slices.

[0112] In one specific implementation, based on the updated hidden state, a prediction result for network slice traffic is generated, including:

[0113] S201, Input the spliced ​​features into the temporal neural network to update the state and obtain the state update features;

[0114] S202, update the state features and the updated hidden states output by the fractional-order recurrent network. By concatenating the rows, we obtain the hidden state at the current moment, represented as:

[0115]

[0116] In the formula, This represents the hidden state at the current time t (after update). This represents the previous hidden state (the hidden state updated at time t-1). This represents the hidden state of history at time tk; This represents the update gate output vector in the temporal neural network (Gated Recurrent Neural Network, GRU, is used in this step); This represents the candidate hidden state at time t; This represents the Hadamard product, which is the product of corresponding elements of a matrix. This represents a regulatory factor used to balance short-term memory (GRU output features) and long-term memory ( The weight of ) ; K represents the length of the history memory buffer pool (i.e., the number of historical steps to backtrack); k represents the index of the current backtracking time step; Indicated based on order Calculated Fractional coefficients (weights) for each historical moment.

[0117] S203, input the hidden state at the current moment into the fully connected layer to generate the prediction result of the network slice traffic at the current moment.

[0118] In one specific implementation, a minimum-maximum normalization process is used to normalize historical traffic data, specifically including:

[0119] S301 performs min-max normalization on the historical traffic data of network slices to eliminate the dimensional differences between different network slices.

[0120] S302 extracts the original features from the normalized historical traffic data and integrates them into a spatiotemporal feature tensor. .

[0121] Embodiments of the present invention also provide a model training method for training the hypergraph convolutional network and fractional recurrent network described above. The training samples are historical traffic data from 5G network slices, and the entire model (including the hypergraph convolutional network and the fractional recurrent network) is trained end-to-end.

[0122] The hypergraph correlation matrix and spatiotemporal feature tensor are constructed using continuous time-time data of [tT, t] as input data, where T can be 20. For example... Figure 4 As shown, within the model, the hypergraph convolutional network receives the hypergraph correlation matrix and the spatiotemporal feature tensor as inputs at the current time t, and extracts network slices at time t to update features. , Compared to the previous hidden state The concatenation feature is divided into two paths. One path is input into the GRU unit to obtain the state update feature; the other path is input into the fractional recurrent network to obtain the updated hidden state. The state update feature is concatenated with the updated hidden state to obtain the current time step. , The fully connected layer is input to predict the flow rate at time t+1. The loss function is calculated based on the actual flow rate at time t+1 and the predicted flow rate. It is placed in the historical state buffer pool.

[0123] During training, an SLA-aware asymmetric loss function is used:

[0124] This represents the total number of traffic sample data. and The first The actual and predicted flow values ​​for each sample For asymmetric penalty weights; when hour, Take a value greater than 1, otherwise The value is set to 1. This loss function makes the model more inclined to avoid underestimation during training, thereby reducing SLA defaults caused by insufficient resource allocation in actual operation and maintenance.

[0125] The following uses publicly available 5G slice traffic data as an example to illustrate the training and application process of the prediction model of this invention. The slice traffic samples are from Kaggle's 5G_Traffic_Datasets.

[0126] Step 1: Build the topology of the slice resource hypergraph and construct a multidimensional data tensor.

[0127] This step aims to transform log data from the physical network into structured data that can be recognized by computer models. In the field of data analysis, converting traffic data into tensor form is a common and effective processing method. This is mainly to facilitate subsequent data analysis and model building. Through tensor processing, we can reorganize the raw traffic data into a structured tensor format, which not only preserves the original information of the data but also makes the data easier for machine learning algorithms to process.

[0128] (1) Tensor quantization processing of the flow data:

[0129] First, historical traffic records for different network slices (eMBB, uRLLC, mMTC) are collected from the 5G network management component. In this embodiment, the max-min normalization method is used to process the data to eliminate the dimensional differences between different network slices. Then, the processed data is integrated into a spatiotemporal feature tensor. , This represents the spatiotemporal feature tensor of the completed network slice traffic. Represents the set of real numbers. This indicates the length of the sliding time window, taking data from the past 20 moments. This represents the total number of nodes in a network slice. This represents the feature dimensions, which include traffic load and packet loss rate, among others.

[0130] (2) Construct a resource-sharing hypergraph:

[0131] Traditional graph structures can only describe simple point-to-point connections, which has certain limitations. To overcome this limitation, this invention uses a hypergraph to model the complex many-to-many relationships between slices. Each network slice is abstracted as a graph node; slice clusters that share the same type of underlying physical infrastructure (such as computing resources of the same base station or spectrum resources of the same transmission link) are aggregated into the same "hyperedge" to construct a higher-order association model. A hypergraph association matrix is ​​then constructed. ( (total number of superedges), if slice nodes Occupied super edge The corresponding resources are matrix elements. Set to 1 otherwise set to 0.

[0132] Step 2: Use Hypergraph Convolutional Network (HGNN) to extract spatial features.

[0133] This step primarily focuses on the coordinated changes or competition in traffic caused by resource sharing among different slices. Traditional graph convolution cannot handle hyperedge relationships; therefore, this step introduces a hypergraph convolution operator, a process involving two stages of feature aggregation:

[0134] Node-to-Hyperedge operation: integrates the features of all slices within the same hyperedge to reflect the overall load status of the resource;

[0135] Hyperedge-to-Node strategy: The load information on the hyperedge is passed back to each slice, so that the slices can know the status of other "competing" slices.

[0136] Step 3: Use adaptive fractional GRU units to extract time features.

[0137] This step primarily addresses the problem that traditional LSTM / GRU models struggle to remember long-term historical traffic trends. The adaptive fractional-order GRU unit mainly comprises two parallel processes: order generation and state update.

[0138] Regarding adaptive generation of order: current input State compared to the previous moment After concatenation, the output values ​​are processed through a linear transformation layer and a sigmoid activation function, and a mapping operation is used to strictly constrain the output values ​​to a specific value. Within the interval, obtain the current dynamic order. This means that the model can automatically determine the depth of its "memory" based on whether the current traffic is stable or bursty.

[0139] Regarding state updates: Unlike traditional GRUs that rely solely on the previous time step's state, this process introduces a "historical memory buffer." The system updates the state based on the calculated order. Calculate a set of fractional coefficients And use this set of coefficients to analyze the historical state of the buffer pool ( A weighted sum is performed. This sum is added to the result of the standard GRU gated update to generate a new state. And the new state Store it back in the buffer pool for calculations at subsequent times.

[0140] (1) Adaptive adjustment of fractional order in dynamic calculation:

[0141] Different traffic slices exhibit varying degrees of burstiness. For example, eMBB traffic is characterized by stability and continuity, while uRLLC traffic is bursty. Therefore, the model cannot use a fixed memory decay rate. This step designs a sub-neural network that can determine in real time how long it needs to review historical data based on the current traffic input.

[0142] Dynamically calculate the fractional order at the current time step. :

[0143]

[0144] In the formula: Indicates in The fractional order calculated adaptively at each time step has a range of values ​​constrained by... between; This represents the Sigmoid activation function, which maps values ​​to... interval; This represents the weight matrix used to calculate the order; express Real-time input flow data; express The hidden layer state at time (previous time); This indicates that the input data and the state from the previous time step will be concatenated. Indicates the bias term; A constant offset ensures that the lower bound of the order is not lower than... This formula ensures the order. Always It fluctuates between these values ​​to achieve adaptive adjustment of memory depth.

[0145] (2) Historical state update based on fractional derivative:

[0146] Traditional GRU updates the current state based solely on the state of the previous time step. This step, however, uses a fractional-order definition. When updating the current state, it not only references information from the previous time step but also directly utilizes historical information from multiple past time steps through a weighted summation method.

[0147] Hidden state The update formula is:

[0148]

[0149] In the formula, express The final hidden state after constant updates; express The hidden state at any given moment; This represents the update gate output vector in a GRU cell; express The candidate hidden state at each moment; This represents the Hadamard product, which is the product of corresponding elements of a matrix. This represents a regulatory factor used to balance the weights of short-term memory (the traditional GRU part) and long-term memory (the fractional part); This indicates the length of the historical memory buffer pool (i.e., the number of historical steps to be traced back). This indicates the index of the current backtracking time step; Indicates based on the current order Calculated Fractional coefficients for each historical moment; express The hidden state of history at any given moment.

[0150] Step 4: Employ an asymmetric decision-making method using the SLA-perceived loss function.

[0151] This step aims to address the risk control issues arising from prediction results in actual operation and maintenance. In 5G slicing operation and maintenance scenarios, the consequences of prediction errors are asymmetrical. Underestimating traffic (i.e., the predicted value is smaller than the actual value) will result in insufficient resource reservation, leading to serious Service Level Agreement (SLA) breaches, such as video stuttering and loss of control commands; while overestimating traffic will only result in slight resource idleness. Traditional mean squared error (MSE) treats these two types of errors equally, which is clearly unreasonable. Therefore, this invention designs a loss function with an asymmetric penalty mechanism. Constructing the SLA-aware loss function. The specific formula is as follows:

[0152]

[0153] In the formula: This represents the calculated SLA sensing loss value; This represents the total number of training samples; Indicates the sample's index number; Indicates the first The actual traffic value of each sample; Indicates the first The model predicts the flow value for each sample; Indicates the first The asymmetric penalty weights corresponding to each sample. When When (risk is underestimated), Take the larger value (e.g.) );otherwise The value is 1.

[0154] Compared with existing mainstream technologies (such as LSTM and GRU), the prediction model based on adaptive fractional hypergraph neural network (hereinafter referred to as FOD-HGNN) proposed in this invention exhibits significant technical advantages. Experimental verification on a real 5G network slice traffic dataset demonstrates its specific advantages in the following aspects:

[0155] (1) It significantly reduced the flow prediction error and greatly improved the prediction accuracy.

[0156] Figure 5 This is a comparison chart of the prediction errors of each model, such as... Figure 5 As shown, in the comprehensive test of the three slices eMBB, uRLLC, and mMTC, the mean absolute error (MAE) of the existing LSTM model (blue bar) and GRU model (green bar) is as high as about 0.45, and the mean squared error (MSE) is about 0.21. In contrast, the FOD-HGNN model proposed in this invention (red bar) shows extremely high accuracy, with an MAE of only about 0.02 and an MSE close to 0. This indicates that the present invention reduces the prediction error by an order of magnitude, and the prediction accuracy is substantially improved compared to the existing technology. This breakthrough improvement is mainly attributed to the effective capture of spatial correlation by hypergraph convolution and the accurate retention of long historical information by the fractional-order mechanism.

[0157] (2) It has adaptive memory capabilities and can automatically adjust the “memory depth” according to different business characteristics.

[0158] According to the appendix Figure 6 The analysis focuses on the adaptive learning curve of the fractional order alpha. Experimental results show that the model can automatically learn differentiated fractional order alpha for different slices. Figure 6As shown, for the uRLLC slice (red dashed line) with strong bursts, the model learns a higher Alpha value (approximately 1.38), indicating that it focuses more on capturing short-term burst features. For the eMBB slice (blue solid line) and mMTC slice (green dotted line), the Alpha value is relatively low, meaning the model automatically adjusts to rely on longer-term historical trends. This mechanism of "self-adjusting" the length of memory based on business characteristics is not present in traditional fixed-order RNN models.

[0159] (3) The model converges quickly and the training process is highly stable.

[0160] According to the appendix Figure 7 (Model training convergence comparison) Analysis is performed. During training, the FOD-HGNN model (red line) proposed in this invention exhibits excellent convergence characteristics. In the first few epochs of training, the loss value rapidly decreases and stabilizes at an extremely low level (close to 0). In contrast, the loss values ​​of LSTM (blue line) and GRU (green line) consistently oscillate between 0.2 and 0.3, failing to decrease further. This fully demonstrates that the algorithm of this invention possesses extremely strong robustness and efficient learning capabilities, enabling it to quickly adapt to complex 5G network traffic data.

[0161] Embodiments of the present invention provide a network slice traffic prediction system, which is used to execute the 5G network slice traffic prediction method of any embodiment of the present invention, including:

[0162] The hypergraph construction unit is used to map network slices to nodes in the hypergraph, map clusters of network slices sharing the same physical features to hyperedges, and construct a hypergraph association matrix based on the connection relationship between nodes and the hyperedges.

[0163] The hypergraph feature extraction unit is used to construct a spatiotemporal feature tensor from the original features of each network slice in the hypergraph, and input the spatiotemporal feature tensor and the hypergraph association matrix into the hypergraph convolutional network for spatial feature extraction to obtain the updated features of the network slice.

[0164] The feature splicing unit is used to splice the network slice update features with the previous hidden state to obtain spliced ​​features; wherein, the previous hidden state is obtained when the traffic prediction task was performed in the previous time step using the same temporal neural network model;

[0165] A fractional-order memory unit is used to generate a fractional-order number based on the splicing features, and to use the fractional-order number to perform a weighted summation of the hidden states of multiple historical moments recorded by the temporal neural network model to obtain an updated hidden state.

[0166] The prediction output unit is used to generate prediction results for network slice traffic based on the updated hidden state.

[0167] Embodiments of the present invention also provide an electronic device including a processor and a memory, wherein the number of processors may be one or more. The memory, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules. The processor executes various functional applications and data processing of the electronic device by running the software programs, instructions, and modules stored in the memory, thereby implementing the 5G network slice traffic prediction method of any of the above embodiments of the present invention.

[0168] The memory may primarily comprise a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function; the data storage area may store data created based on terminal usage. Furthermore, the memory may include high-speed random access memory (RAM) and non-volatile memory, such as at least one disk storage device, flash memory, or other non-volatile solid-state storage device. In some instances, the memory may further include memory remotely located relative to the processor, which can be connected to the electronic device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks (LANs), mobile communication networks, and combinations thereof.

[0169] Embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the 5G network slice traffic prediction method of any embodiment of the present invention.

[0170] The computer storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0171] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0172] Embodiments of the present invention also provide a computer program product that, when run on a computer, causes the computer to execute the 5G network slice traffic prediction method of any of the above embodiments of the present invention.

[0173] The above embodiments are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the appended claims.

Claims

1. A method for predicting 5G network slice traffic, characterized in that, include: Network slices are mapped to nodes in the hypergraph, clusters of network slices sharing the same physical features are mapped to hyperedges, and a hypergraph association matrix is ​​constructed based on the connection relationship between the nodes and the hyperedges. The original features of each network slice in the hypergraph are constructed into a spatiotemporal feature tensor. The spatiotemporal feature tensor and the hypergraph association matrix are input into the hypergraph convolutional network for spatial feature extraction to obtain the updated features of the network slice. The network slice update features are concatenated with the previous hidden state to obtain the concatenated features; wherein, the previous hidden state is obtained when the traffic prediction task was performed in the previous time step using the same temporal neural network model; A fractional order is generated based on the splicing features. The fractional order is then used to perform a weighted summation of the hidden states at multiple historical moments recorded by the temporal neural network model to obtain the updated hidden states. Based on the updated hidden state, a prediction result for network slice traffic is generated.

2. The method for predicting 5G network slice traffic according to claim 1, characterized in that, The physical characteristics include at least one of computing resource characteristics and transmission link characteristics.

3. The method for predicting 5G network slice traffic according to claim 1, characterized in that, The step of constructing a spatiotemporal feature tensor from the original features of each network slice in the hypergraph includes: The historical traffic data of the network slices are subjected to max-min normalization to eliminate the dimensional differences between different network slices; The original features are extracted from the normalized historical traffic data and integrated into a spatiotemporal feature tensor. , ;in, Represents the set of real numbers. This indicates the length of the sliding time window used when integrating historical traffic data. Indicates the number of network slices. The dimension represents the original feature, which includes traffic load and packet loss rate.

4. The method for predicting 5G network slice traffic according to claim 1, characterized in that, The hypergraph convolutional network includes a node-to-hyperedge aggregation module, a hyperedge feature weighting module, and a hyperedge-to-node backpropagation update module. The node-to-hyperedge aggregation module is used to perform aggregation operations on the original features of each node in the same hyperedge to obtain hyperedge aggregated features; The hyperedge feature weighting module is used to perform weighted calculation on the hyperedge aggregated features based on the weight of the hyperedge to obtain updated hyperedge aggregated features; The hyperedge-to-node backhaul update module is used to backhaul the updated hyperedge aggregation features to each node to update the feature representation of each node and obtain the network slice update features.

5. The method for predicting 5G network slice traffic according to claim 1, characterized in that, The generation of network slice traffic prediction results based on the updated hidden state includes: The spliced ​​features are input into the temporal neural network for state update to obtain state update features; The state update feature is concatenated with the updated hidden state to obtain the hidden state at the current time. The hidden state at the current moment is input into the output layer of the temporal neural network to generate the prediction result of the network slice traffic.

6. The method for predicting 5G network slice traffic according to claim 1, characterized in that, The step of generating a fractional order based on the splicing features, and then using the fractional order to perform a weighted summation of the hidden states at multiple historical moments recorded by the temporal neural network model to obtain the updated hidden states, includes: The concatenated features are input into a fractional recurrent network to obtain the updated hidden state; wherein, the fractional recurrent network includes a fractional generation module and a fractional convolution calculation module; The fractional order generation module is used to generate the fractional order number based on the splicing features; The fractional convolution calculation module is used to perform a weighted summation of the hidden states at multiple historical moments based on the fractional order to obtain the updated hidden states.

7. The method for predicting 5G network slice traffic according to claim 6, characterized in that, The fractional order generation module includes a linear transformation layer, a Sigmoid activation function, and a mapping layer connected in sequence. The linear transformation layer is used to map the spliced ​​features into scalars; The Sigmoid activation function is used to compress the scalar to the (0,1) interval; The mapping layer is used to map the output value of the Sigmoid activation function to... Interval.

8. The method for predicting 5G network slice traffic according to claim 1, characterized in that, The step of using the fractional order to perform a weighted summation of the hidden states at multiple historical moments recorded by the temporal neural network model to obtain the updated hidden states includes: Calculate the fractional order coefficients based on the fractional order number; The fractional-order coefficients are convolved with the hidden states at multiple historical time points to obtain the updated hidden states.

9. The method for predicting 5G network slice traffic according to claim 6, characterized in that, The method further includes the step of training the network parameters of the hypergraph convolutional network and the fractional-order recurrent network using traffic sample data from network slices; Among them, an asymmetric loss function is used. Conduct training: ； in, This represents the total number of traffic sample data. and The first The actual and predicted flow values ​​for each sample For asymmetric penalty weights; when hour, Take a value greater than 1, otherwise The value is 1.

10. A 5G network slice traffic prediction system, characterized in that, The prediction system is used to execute the 5G network slice traffic prediction method according to any one of claims 1-9, including: The hypergraph construction unit is used to map network slices to nodes in the hypergraph, map clusters of network slices sharing the same physical features to hyperedges, and construct a hypergraph association matrix based on the connection relationship between the nodes and the hyperedges. The hypergraph feature extraction unit is used to construct a spatiotemporal feature tensor from the original features of each network slice in the hypergraph, and input the spatiotemporal feature tensor and the hypergraph association matrix into the hypergraph convolutional network for spatial feature extraction to obtain the updated features of the network slice. The feature splicing unit is used to splice the network slice update features with the previous hidden state to obtain spliced ​​features; wherein, the previous hidden state is obtained when the traffic prediction task was performed in the previous time step using the same temporal neural network model; A fractional-order memory unit is used to generate a fractional-order number based on the splicing features, and to use the fractional-order number to perform a weighted summation of the hidden states of multiple historical moments recorded by the temporal neural network model to obtain an updated hidden state. The prediction output unit is used to generate prediction results for network slice traffic based on the updated hidden state.

Images