Cellular network flow prediction method based on Mama model and graph convolutional network

Abstract

The invention discloses a cellular network traffic prediction method based on a Mama model and a graph convolutional network, which belongs to the field of deep learning and cellular network traffic prediction, and comprises the following steps: (1) carrying out fast Fourier transform on original traffic data, converting a time domain into a frequency domain, screening a main frequency component and estimating a period; (2) constructing a graph structure by taking the preprocessed features as nodes through a periodic graph convolutional network, and extracting periodic features through aggregation of the graph convolutional network; (3) dividing regional homogeneous / heterogeneous neighbors by using a Pearson coefficient through a variable correlation graph convolutional network, and aggregating and extracting regional correlation features through the graph convolutional network; (4) a bidirectional Mama module performs time-dependent modeling on the flow data according to positive and negative time sequences, and fuses bidirectional features; and (5) the output layer outputs a prediction result through linear transformation. According to the method, flow periodicity, regional association and long-term dependence are considered, accurate long-term prediction is realized, and support is provided for network resource optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning and cellular communication network, and particularly relates to a cellular network traffic prediction method based on Mamba model and graph convolution network. BACKGROUND

[0002] With the rapid rise of the fifth generation mobile communication (5G) technology, the global cellular network user scale continues to expand. According to the prediction of Ericsson, the global 5G users will reach 2.8 billion by 2025, which directly drives the cellular network traffic demand to grow exponentially. Such traffic data presents significant complex periodicity characteristics, including both short-term cycles of daily peak-valley alternation and medium-long-term cycles of traffic fluctuation difference between weekdays and holidays, and is also affected by user behavior, regional function and other factors, forming a strong nonlinear time-dependent relationship. For telecom operators, accurate long-term cellular network traffic prediction is the core support for optimizing network resource allocation - by predicting traffic peaks and troughs, high-load and low-load areas in advance, precise base station resource scheduling, bandwidth dynamic adjustment and network expansion planning can be realized, thereby improving network service quality and reducing operating costs. Therefore, long-term cellular network traffic prediction has become a key technical requirement in the field of mobile communication.

[0003] Current cellular network traffic prediction methods mainly include statistical methods and deep learning methods, but both have obvious limitations and challenges. Among them, statistical methods such as ARIMA model and covariance function method need to assume that traffic data follows a specific distribution, which is difficult to adapt to the complex nonlinear patterns of cellular traffic in actual scenarios, and the prediction accuracy is greatly reduced when facing peak-valley mutations, periodic fluctuations and other situations.

[0004] Traditional deep learning methods such as CNN, RNN and LSTM can capture some spatio-temporal features, but they are prone to gradient vanishing or gradient explosion problems when processing long sequence traffic data, and cannot effectively model long-term time-dependent relationships. The Transformer architecture alleviates the long dependence modeling problem through the self-attention mechanism, but its quadratic computational complexity leads to low computational efficiency on high-dimensional and long-time sequence cellular traffic data, making it difficult to meet real-time prediction needs. The emerging Mamba architecture has linear complexity and excellent long sequence modeling capability, but its existing applications do not fully emphasize the periodicity feature capture of cellular traffic and do not deeply explore the homogeneous and heterogeneous relationship between different monitoring areas, resulting in that the existing technology cannot meet the actual needs of long-term accurate prediction of cellular networks. SUMMARY

[0005] In view of the deficiencies in the prior art, the application provides a cellular network traffic prediction method based on a Mamba model and a graph convolution network, which accurately predicts long-term cellular network traffic by combining periodic features, regional correlation features and time-dependent features, thereby providing support for network resource optimization.

[0006] The object of the application is achieved in that a cellular network traffic prediction method based on a Mamba model and a graph convolution network comprises the following steps:

[0007] Step 1) data preprocessing step: performing fast Fourier transform (FFT) on the original cellular network traffic time series data to convert the time domain data into frequency domain data to highlight the periodic features of the traffic; screening the main frequency components in the frequency domain data and estimating the period length corresponding to each main frequency component to determine the core period mode of the traffic data; performing maximum pooling operation on the transformed data, and concatenating the pooled data in each dimension according to the feature dimension to obtain preprocessed features suitable for subsequent graph network processing;

[0008] Step 2) periodic graph convolution network (PGCN) feature extraction step: taking the preprocessed features obtained in step 1) as the node set of the graph structure, initializing the adjacency matrix containing learnable vectors to describe the periodic correlation between nodes; converting the adjacency matrix elements to positive values through an activation function and normalizing them through a softmax function to balance the node weights; constructing a periodic graph structure based on the node set and the normalized adjacency matrix; using the node aggregation operation of the graph convolution network to iteratively update the node features in the periodic graph structure, strengthening the extraction of periodic features in the traffic data, and outputting a periodic feature matrix;

[0009] Step 3) variable correlation graph convolution network (VCGCN) feature extraction step: calculating the Pearson correlation coefficients between the cellular network traffic data of different monitoring regions to quantify the degree of linear correlation between the traffic changes between regions; introducing a learnable threshold parameter θ to divide the homogeneous neighbors and heterogeneous neighbors of each region; normalizing the correlation weights of the homogeneous neighbors and the heterogeneous neighbors, respectively; constructing a variable correlation graph structure based on the normalized weights, updating the node features through the node aggregation operation of the graph convolution network, extracting the homogeneous / heterogeneous correlation features between different regions, and outputting a regional correlation feature matrix;

[0010] Step 4) Bidirectional Mamba module feature extraction step: splicing the periodic feature matrix of step 2) and the region correlation feature matrix of step 3) to form an input feature sequence of the bidirectional Mamba module; using a pair of Mamba blocks to process the input sequence in forward and reverse time sequence respectively: the forward Mamba block is processed in the original time sequence, modeling the long sequence time dependence; the reverse Mamba block is processed after reversing the time sequence dimension of the input sequence, and then the output result is flipped back to the original time sequence to align the dimension; performing summation on the output features of the two Mamba blocks, performing layer normalization to stabilize the training process, and then compressing the high-dimensional features through the projection layer to obtain a feature matrix that fuses the bidirectional time dependence;

[0011] Step 5) Prediction output step: mapping the fused high-dimensional features to the target prediction length of the cellular network traffic prediction result through linear transformation, the dimension of the prediction result matches the number of monitoring areas and the prediction length, realizing long-term traffic accurate prediction.

[0012] Further, the step 1) specifically comprises:

[0013] Step 1.1) original data for fast Fourier transform FFT: converting the original cellular network traffic time domain sequence into frequency domain data, wherein B is the batch size, L is the historical traffic length, and N is the number of base stations, and the calculation method is:

[0014]

[0015] wherein, is the kth frequency component, , is the nth time domain sample, and the exponential term is the basis function of the Fourier transform, which is used to decompose the frequency components of the time domain signal;

[0016] Step 1.2) main frequency component screening: calculating the amplitude of each frequency component , selecting the top frequency components with the largest amplitude as the main frequency components, is a preset hyperparameter, which is set according to the periodicity of the traffic data, and the screening formula is:

[0017]

[0018] to ensure that the frequency components with the largest contribution to the traffic periodicity are screened out;

[0019] Step 1.3) period length estimation: calculating the corresponding period based on the main frequency component, the formula is:

[0020]

[0021] wherein, For the first The period duration corresponding to each dominant frequency component The total time span of the original data. The frequency value of the i-th dominant frequency component , Index of the main frequency component, symbol This indicates rounding to the nearest integer, ensuring that the cycle length is an integer number of time steps.

[0022] Step 1.4) Max Pooling and Concatenation: For the transformed data corresponding to each main frequency component, max pooling is performed to calculate the maximum value within the pooling window as the pooling result. The formula is as follows:

[0023]

[0024] The pooled data is used to reduce the estimation mean bias caused by convolutional layer parameter errors; all principal frequency components are corresponding to... Preprocessed features are obtained by concatenating them according to feature dimensions.

[0025] Furthermore, step 2) specifically includes:

[0026] Step 2.1) Adjacency Matrix Initialization: Construct the graph structure of PGCN:

[0027]

[0028] Among them, the node set For the preprocessed features obtained in step 1), each node corresponds to a feature vector at one time step; the learnable vectors are randomly initialized. The initial incidence matrix is ​​calculated by matrix multiplication. ; Use activation function The elements of the initial correlation matrix are converted to positive values ​​to avoid negative correlations interfering with the extraction of periodic features. Then, the softmax function is used to normalize each row of the matrix to obtain the adjacency matrix of the periodic graph. The formula is:

[0029]

[0030] Ensure that the sum of the weights of each node with respect to other nodes is 1;

[0031] Step 2.2) Node aggregation operation: Iteratively update node features using hierarchical graph convolution, in the l-th layer, The node feature matrix is initial layer The feature update formula for the nodes in the (l+1)th layer is:

[0032]

[0033] in, Here is the learnable weight matrix for the l-th layer, used for feature transformation; activation function. ReLU is employed to enhance the model's nonlinear expressive power; after iterative updates, the final periodic feature matrix is ​​output. Where B is the batch size, L is the historical traffic length, and D is the model feature dimension.

[0034] Furthermore, step 3) specifically includes:

[0035] Step 3.1) Pearson correlation coefficient calculation: For any two monitoring areas i and j, The total number of regions is N. The preprocessed flow characteristic sequence from step 1) is used:

[0036]

[0037] Where L is the time step;

[0038] Calculate the Pearson correlation coefficient between the two. The formula is:

[0039]

[0040] in, Let r be the mean of the feature sequences of regions i and j, respectively. ij The larger the absolute value, the stronger the correlation between regions;

[0041] Step 3.2) Neighbor Partitioning and Weight Normalization: Initialize the learnable threshold parameter θ, and partition the neighbors into heterogeneous neighbor sets and homogeneous neighbor sets, using the following formula:

[0042]

[0043] in, Heterogeneous neighbors For homogeneous neighbors, normalized weights are then calculated for both types of neighbors separately:

[0044]

[0045] in, For heterogeneous neighbor weights, The weights are those of homogeneous neighbors;

[0046] Step 3.3) Node Aggregation and Feature Output: Construct the graph structure of VCGCN, using the following formula:

[0047]

[0048] Among them, the node set For each monitoring area, the adjacency matrix of the variable correlation plot is given. The elements are determined by the weights of their homogeneous neighbors. Or heterogeneous neighbor weight Fill in the gaps, setting the weights of non-neighboring nodes to 0; employ a hierarchical graph convolutional node aggregation operation similar to PGCN, with the following formula:

[0049]

[0050] in, Given the region association feature matrix of layer l, after iterative updates, output the region association feature matrix of layer (l+1). .

[0051] Furthermore, step 4) specifically includes:

[0052] Step 4.1) Input Feature Preparation: Prepare the periodic feature matrix from Step 2). The region association feature matrix of step 3) Expanding along the region dimension, it is transformed into a unified input feature sequence, as shown in the formula:

[0053]

[0054] Where each x t This represents the fusion characteristics of all regions at time step t;

[0055] Step 4.2) Forward Mamba block processing: Input in the original timing order The state vector is calculated step-by-step using Mamba blocks. and output vector The formula is:

[0056]

[0057] After processing, the positive output sequence is obtained:

[0058]

[0059] Modeling time dependencies from the past to the future;

[0060] Step 4.3) Reverse Mamba block processing: Process the input feature sequence Perform time reversal to obtain the reverse sequence. Processed in the same way as forward Mamba blocks. The reverse output sequence is obtained. The forward output sequence is then reversed again according to the timing sequence to obtain the reverse output sequence aligned with the original timing sequence. :

[0061]

[0062] Modeling the time dependency from the future to the past;

[0063] Step 4.4) Bidirectional Feature Fusion: For and Perform element-wise summation to obtain the feature matrix that incorporates bidirectional dependencies. :

[0064]

[0065] The Y layer is normalized to stabilize the training process, and then bidirectional time-dependent features are obtained through a feedforward network. .

[0066] Furthermore, step 5) specifically includes: assigning bidirectional time-dependent features... The input is fed into a linear transformation layer, where high-dimensional fusion features are compressed through linear transformation and mapped to an output dimension adapted to the target prediction task. This output dimension is matched with a preset prediction duration S and the number of monitored areas N, resulting in a final output dimension of [missing value]. Cellular network traffic prediction results This enables long-term cellular network traffic prediction.

[0067] Compared with existing technologies, the advantages of this invention are as follows: 1) This invention proposes PGCN and VCGCN to capture the periodicity of cellular network traffic data, as well as the homogeneity and heterogeneity between different regions, extracting more effective spatiotemporal features. 2) This invention extracts more effective temporal dependency features of traffic data through a bidirectional Mamba module. 3) Experimental results on real cellular network traffic datasets show that this scheme improves prediction accuracy compared with the latest prediction models. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0069] Figure 1 This is a schematic diagram of the cellular network traffic prediction method based on the Mamba model and graph convolutional networks of the present invention.

[0070] Figure 2 This is a diagram illustrating the architecture of the cellular network traffic prediction method based on the Mamba model and graph convolutional networks of this invention.

[0071] Figure 3 This is a bar chart showing the average performance comparison between the method of this invention and other methods on the Milan dataset for various prediction lengths.

[0072] Figure 4 This is a bar chart comparing the average performance of the method of this invention with other methods on the Trentino dataset for various prediction lengths. Detailed Implementation

[0073] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0074] like Figures 1-2 The method for predicting cellular network traffic based on the Mamba model and graph convolutional networks, as shown, includes the following steps:

[0075] Step 1) Data Preprocessing Steps: Perform Fast Fourier Transform (FFT) on the original cellular network traffic time-series data to convert the time-domain data into frequency-domain data to highlight the periodic characteristics of the traffic; filter the main frequency components in the frequency-domain data and estimate the period duration corresponding to each main frequency component to determine the core periodic pattern of the traffic data; perform max pooling on the transformed data to reduce the estimation mean bias caused by convolutional layer parameter errors; concatenate the pooled data of each dimension according to the feature dimension to obtain preprocessed features suitable for subsequent graph network processing;

[0076] Step 1.1) The original data is a Fast Fourier Transform (FFT): the original cellular network traffic time-domain sequence is transformed. Convert to frequency domain data, where B is the batch size, L is the historical traffic length, and N is the number of base stations. The calculation method is as follows:

[0077]

[0078] in, For the k-th frequency component, , For the nth time-domain sample, the exponent term is the basis function of the Fourier transform, used to decompose the frequency components of the time-domain signal;

[0079] Step 1.2) Filtering of dominant frequency components: Calculate the amplitude of each frequency component. (Amplitude is the magnitude of a frequency component, reflecting the signal strength at the corresponding frequency). The top M frequency components with the largest amplitudes are selected as the main frequency components, where M is a preset hyperparameter set according to the periodic characteristics of the flow data. The selection formula is as follows:

[0080]

[0081] Ensure that the frequency components that contribute the most to the periodicity of the flow are selected; in this embodiment, M equals 3;

[0082] Step 1.3) Period Duration Estimation: Calculate the corresponding period based on the main frequency component, using the following formula:

[0083]

[0084] in, For the first The period duration corresponding to each dominant frequency component The total time span of the original data. The frequency value of the i-th dominant frequency component , Index of the main frequency component, symbol This indicates rounding to the nearest integer, ensuring that the cycle length is an integer number of time steps.

[0085] Step 1.4) Max Pooling and Concatenation: For the transformed data corresponding to each main frequency component, max pooling is performed to calculate the maximum value within the pooling window as the pooling result. The formula is as follows:

[0086]

[0087] The pooled data is used to reduce the estimation mean bias caused by convolutional layer parameter errors; all principal frequency components are corresponding to... Preprocessed features are obtained by concatenating them according to feature dimensions. In this example, B equals 32, N equals 400, and L equals 720.

[0088] Step 2) Periodic Graph Convolutional Network (PGCN) Feature Extraction Steps: Using the preprocessed features obtained in Step 1) as the node set of the graph structure, initialize an adjacency matrix containing learnable vectors to characterize the periodic relationships between nodes; convert the elements of the adjacency matrix to positive values ​​through an activation function, and normalize them through a softmax function to balance the node weights; construct a periodic graph structure based on the node set and the normalized adjacency matrix; use the node aggregation operation of the graph convolutional network to iteratively update the node features in the periodic graph structure, enhance the extraction of periodic features such as peak-valley cycles and weekday / holiday cycles in traffic data, and output a periodic feature matrix;

[0089] Step 2.1) Adjacency Matrix Initialization: Construct the graph structure of PGCN:

[0090]

[0091] Among them, the node set For the preprocessed features obtained in step 1), each node corresponds to a feature vector at one time step; the learnable vectors are randomly initialized. The initial incidence matrix is ​​calculated by matrix multiplication. ; Use activation function The elements of the initial correlation matrix are converted to positive values ​​to avoid negative correlations interfering with the extraction of periodic features. Then, the softmax function is used to normalize each row of the matrix to obtain the adjacency matrix of the periodic graph. The formula is:

[0092]

[0093] Ensure that the sum of the weights of each node with respect to other nodes is 1;

[0094] Step 2.2) Node aggregation operation: Iteratively update node features using hierarchical graph convolution, in the l-th layer, The node feature matrix is initial layer The feature update formula for the nodes in the (l+1)th layer is:

[0095]

[0096] in, Here is the learnable weight matrix for the l-th layer, used for feature transformation; activation function. ReLU is employed to enhance the model's nonlinear expressive power; after iterative updates, the final periodic feature matrix is ​​output. Where B is the batch size, L is the historical traffic length, and D is the model feature dimension; in the embodiment, D equals 64.

[0097] Step 3) Feature extraction steps of Variable Correlation Graph Convolutional Network (VCGCN): Calculate the Pearson correlation coefficient between cellular network traffic data in different monitoring areas to quantify the linear correlation between traffic changes in different areas; introduce a learnable threshold parameter θ to divide each area into homogeneous and heterogeneous neighbors; normalize the correlation weights of homogeneous and heterogeneous neighbors respectively to highlight the influence of strongly correlated areas; construct a variable correlation graph structure based on the normalized weights, update node features through node aggregation operation of graph convolutional network, extract homogeneous / heterogeneous correlation features between different areas, and output the regional correlation feature matrix;

[0098] Step 3.1) Pearson correlation coefficient calculation: For any two monitoring areas i and j, The total number of regions is N. The preprocessed flow characteristic sequence from step 1) is used:

[0099]

[0100] Where L is the time step;

[0101] Calculate the Pearson correlation coefficient between the two. The formula is: ;

[0102] in Let r be the mean of the feature sequences of regions i and j, respectively. ij The larger the absolute value, the stronger the correlation between regions;

[0103] Step 3.2) Neighbor Partitioning and Weight Normalization: Initialize the learnable threshold parameter θ (updated iteratively during model training; initialized to 0.5 in this example), and divide the neighbors into heterogeneous neighbor sets and homogeneous neighbor sets, using the following formula:

[0104]

[0105] in, Heterogeneous neighbors For homogeneous neighbors, normalized weights are then calculated for both types of neighbors separately:

[0106]

[0107] in, For heterogeneous neighbor weights, The weights are those of homogeneous neighbors;

[0108] Step 3.3) Node Aggregation and Feature Output: Construct the graph structure of VCGCN, using the following formula:

[0109]

[0110] Among them, the node set For each monitoring area, the adjacency matrix of the variable correlation plot is given. The elements are determined by the weights of their homogeneous neighbors. Or heterogeneous neighbor weight Fill in the gaps, setting the weights of non-neighboring nodes to 0; employ a hierarchical graph convolutional node aggregation operation similar to PGCN, with the following formula:

[0111]

[0112] in, Given the region association feature matrix of layer l, after iterative updates, output the region association feature matrix of layer (l+1). .

[0113] Step 4) Bidirectional Mamba module feature extraction steps: Concatenate the periodic feature matrix from Step 2) and the region association feature matrix from Step 3) to form the input feature sequence of the bidirectional Mamba module; use a pair of Mamba blocks to process the input sequence in forward and reverse time order respectively: the forward Mamba block processes the original time order and models the time dependency of the long sequence; the reverse Mamba block first reverses the time dimension of the input sequence and then processes it, and then flips the output result back to the original time order to align the dimensions; perform summation on the output features of the two Mamba blocks, stabilize the training process through layer normalization, and then compress the high-dimensional features through the projection layer to obtain the feature matrix that fuses the bidirectional time dependency;

[0114] Step 4.1) Input Feature Preparation: Prepare the periodic feature matrix from Step 2). The region association feature matrix of step 3) Expanding along the region dimension, it is transformed into a unified input feature sequence, as shown in the formula:

[0115]

[0116] Where each x t This represents the fusion characteristics of all regions at time step t;

[0117] Step 4.2) Forward Mamba block processing: Input in the original timing order The state vector is calculated step-by-step using Mamba blocks. and output vector The formula is:

[0118]

[0119] After processing, the positive output sequence is obtained:

[0120]

[0121] Modeling time dependencies from the past to the future;

[0122] Step 4.3) Reverse Mamba block processing: Process the input feature sequence Perform time reversal to obtain the reverse sequence. Processed in the same way as forward Mamba blocks. The reverse output sequence is obtained. The forward output sequence is then reversed again according to the timing sequence to obtain the reverse output sequence aligned with the original timing sequence. :

[0123]

[0124] Modeling the time dependency from the future to the past;

[0125] Step 4.4) Bidirectional Feature Fusion: For and Perform element-wise summation to obtain the feature matrix that incorporates bidirectional dependencies. : ; The Y layer is normalized to stabilize the training process, and then bidirectional time-dependent features are obtained through a feedforward network. .

[0126] Step 5) Prediction output step: The fused high-dimensional features are mapped to a cellular network traffic prediction result of the target prediction length through linear transformation. The dimension of the prediction result is matched with the number of monitoring areas and the prediction duration to achieve accurate long-term traffic prediction.

[0127] Bidirectional time dependence features The input is fed into a linear transformation layer, where high-dimensional fusion features are compressed through linear transformation and mapped to an output dimension adapted to the target prediction task. This output dimension is matched with a preset prediction duration S and the number of monitored areas N, resulting in a final output dimension of [missing value]. Cellular network traffic prediction results This enables long-term cellular network traffic prediction.

[0128] Two evaluation metrics, MAE and MSE, are used to assess the accuracy of the final predictions made by the system model.

[0129]

[0130]

[0131] Where n is the total sample size. pr i and tr i These represent the predicted value and the actual value at the i-th time point, respectively.

[0132] Comparison: After training, the proposed solution was validated on two real-world datasets: the Milan dataset (containing 400 base stations) and the Trentino dataset (containing 100 base stations). The prediction window lengths for both datasets were set to 96, 192, 336, and 720. The comparison of prediction results between this solution and LSTM, CrossGNN, iTransformer, PatchTST, Dlinear, Crossformer, TiDE, S-Mamba, and TimeMachine models on the cellular network traffic data of the two datasets under these prediction window lengths is shown in the following figures.Figure 3 , Figure 4 As shown.

[0133] Therefore, this invention employs the aforementioned cellular network traffic prediction method based on the Mamba model and graph convolutional networks, which achieves accurate long-term prediction of cellular network traffic by fully extracting the periodic features, regional correlation features, and time-dependent features of traffic data.

[0134] The above description of the embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A method for cellular network traffic prediction based on Mamba model and graph convolution network, characterized in that, The method comprises the following steps: Step 1) data preprocessing step: performing fast Fourier transform (FFT) on the original cellular network traffic time series data to convert time domain data into frequency domain data to highlight the periodicity characteristics of the traffic; screening the main frequency components in the frequency domain data and estimating the period length corresponding to each main frequency component to determine the core period pattern of the traffic data; performing maximum pooling operation on the transformed data, and concatenating the pooled data in each dimension according to the feature dimension to obtain preprocessed features suitable for subsequent graph network processing; Step 2) periodic graph convolution network (PGCN) feature extraction step: taking the preprocessed features obtained in step 1) as the node set of the graph structure, initializing the adjacency matrix containing learnable vectors to depict the periodic association between nodes; converting the elements of the adjacency matrix to positive values through an activation function and normalizing them through a softmax function to balance the node weights; constructing a periodic graph structure based on the node set and the normalized adjacency matrix; using the node aggregation operation of the graph convolution network to iteratively update the node features in the periodic graph structure, strengthening the extraction of periodic features in the traffic data, and outputting a periodic feature matrix; Step 3) variable correlation graph convolution network (VCGCN) feature extraction step: calculating the Pearson correlation coefficients between the cellular network traffic data of different monitoring areas to quantify the degree of linear correlation between the traffic changes of different areas; introducing a learnable threshold parameter θ to divide the homogeneous neighbors and heterogeneous neighbors of each area; normalizing the correlation weights of homogeneous neighbors and heterogeneous neighbors respectively; constructing a variable correlation graph structure based on the normalized weights, updating the node features through the node aggregation operation of the graph convolution network, extracting homogeneous / heterogeneous correlation features between different areas, and outputting a regional correlation feature matrix; Step 4) bidirectional Mamba module feature extraction step: concatenating the periodic feature matrix of step 2) and the regional correlation feature matrix of step 3) to form an input feature sequence of the bidirectional Mamba module; using a pair of Mamba blocks to process the input sequence in forward and reverse time sequences respectively: the forward Mamba block processes in the original time sequence to model the long sequence time dependence; the reverse Mamba block first processes the input sequence after reversing the time sequence, and then flips the output back to the original time sequence to align the dimensions; performing summation on the output features of the two Mamba blocks, performing layer normalization to stabilize the training process, and then compressing the high-dimensional features through a projection layer to obtain a feature matrix that fuses the bidirectional time dependence; Step 5) prediction output step: mapping the fused high-dimensional features to the cellular network traffic prediction results of the target prediction length through linear transformation, the dimension of the prediction results matches the number of monitoring areas and the prediction length, realizing accurate long-term traffic prediction.

2. The method according to claim 1, wherein, The step 1) specifically comprises: Step 1.1) Raw data is Fast Fourier Transform (FFT): convert raw cellular network traffic time-domain sequence to frequency-domain data, where B is batch size, L is history traffic length, N is the number of base stations, and the calculation is as follows: ； wherein is the kth frequency component, , is the nth time domain sample, the exponential term is a basis function of the Fourier transform used to decompose the frequency components of the time domain signal; Step 1.2) Main frequency component screening: calculate the amplitude of each frequency component , select the top frequency components with the largest amplitude as the main frequency components, is a preset hyperparameter, which is set according to the periodic characteristics of the flow data, and the screening formula is: ； Ensure that the frequency component that contributes most to the periodicity of the traffic is screened out; Step 1.3) period length estimation: calculate the corresponding period based on the main frequency component, the formula is: ； wherein, is the period length corresponding to the i-th dominant frequency component, T is the total time span of the raw data, is the frequency value of the i-th dominant frequency component , , is the index of the dominant frequency component, the symbol denotes rounding to the nearest integer, ensuring that the period length is an integer number of time steps. Step 1.4) maximum pooling and concatenation: for the transformed data corresponding to each main frequency component, use the maximum pooling operation to calculate the maximum value in the pooling window as the pooling result, the formula is: ； For the pooled data, to reduce the estimated mean deviation caused by the error of the convolution layer parameters; all the corresponding main frequency components of Splicing according to the feature dimension, get the preprocessed features.

3. The method according to claim 2, wherein, The step 2) specifically comprises: Step 2.1) Adjacency matrix initialization: construct the graph structure of PGCN: ； wherein the node set For the pre-processed features obtained in step 1), each node corresponds to a feature vector of a time step; randomly initialize the learnable vector Calculate the initial correlation matrix by matrix multiplication ; use the activation function Convert the initial correlation matrix elements to positive values to avoid negative correlation interference period feature extraction, and then normalize each row of the matrix by the softmax function to obtain the adjacency matrix of the period graph , the formula is: ； Ensure that the weight sum of each node to other nodes is 1; Step 2.2) Node aggregation operation: update the node feature iteratively using hierarchical graph convolution, the l-th layer, The node feature matrix of the l-th layer is The initial layer The node feature update formula of the l+1-th layer is: ； wherein, is the learnable weight matrix of the lth layer, used for feature transformation; the activation function ReLU is used to enhance the nonlinear expression ability of the model; after iterative updating, the final periodic feature matrix is output , wherein B is the batch size, L is the historical traffic length, and D is the feature dimension of the model.

4. The method according to claim 3, wherein, The step 3) specifically includes: Step 3.1) Pearson correlation coefficient calculation: for any two monitoring regions i, j, , the total number of regions is N, take the flow feature sequence after preprocessing in step 1): ； Where L is the time step; The Pearson correlation coefficient of both is calculated with the formula: ； wherein, are the mean values of the region i, j characteristic sequence, respectively, r ij The greater the absolute value, the stronger the correlation between regions. Step 3.2) Neighbor division and weight normalization: initialize the learnable threshold parameter θ, divide the neighbor type into a heterogeneous neighbor set and a homogeneous neighbor set, and the formula is: ； where, heterogeneous neighbors, are homogeneous neighbors, then the normalized weights are computed separately for the two types of neighbors: ； wherein, is a heterogeneous neighbor weight, is a homogeneous neighbor weight; Step 3.3) Node aggregation and feature output: construct the graph structure of VCGCN, and the formula is: ； where the node set is the characteristic sequence of each monitoring area, and the adjacency matrix of the variable correlation graph is filled by homogeneous neighbor weights or heterogeneous neighbor weights , and the non-neighbor node weights are set to 0; a hierarchical graph convolution node aggregation operation similar to PGCN is adopted, and the formula is: ； wherein, is the region associated feature matrix of the lth layer, which is iteratively updated and outputted as the region associated feature matrix of the (l+1)th layer .

5. The method according to claim 4, wherein, The step 4) specifically includes: Step 4.1) Input feature preparation: The periodic feature matrix of step 2) is unfolded along the region dimension, transformed into a uniform input feature sequence, formula is: The region-associated feature matrix of step 3) is unfolded along the region dimension, transformed into a uniform input feature sequence, formula is: The region-associated feature matrix of step 3) is unfolded along the region dimension, transformed into a uniform input feature sequence, formula is: ； wherein each x t represents the fusion feature of all regions at the t-th time step; Step 4.2) Forward Mamba block processing: input in original time sequence order Compute state vector and output vector at each time step through the Mamba block, formula is: ； After processing, the forward output sequence is obtained: ； Modeling the time dependence from the past to the future; Step 4.3) Reverse Mamba block processing: the input feature sequence is time-reversed to get a reversed sequence ; the same way as the forward Mamba block is processed to get a reversed output sequence ; the above forward output sequence is time-reversed again to get a reversed output sequence aligned with the original time sequence : ； Modeling the time dependence from the future to the past; Step 4.4) Bidirectional feature fusion: For each element in the feature matrix and perform element-wise summation to get the fused bidirectional dependent feature matrix : ； The Y is executed layer normalization to stabilize the training process, and then through the feedforward network, get the bidirectional time-dependent features .

6. The method according to claim 5, wherein, The step 5) specifically comprises: inputting the bidirectional time-dependent feature to a linear transformation layer, compressing the high-dimensional fusion feature through linear transformation, and mapping it to an output dimension adapted to the target prediction task; the output dimension matches a preset prediction length S and a preset number of monitoring regions N, and the final output dimension is a cellular network traffic prediction result , and long-term cellular network traffic prediction is realized.

Images