Cellular network traffic spatio-temporal prediction method and system based on hybrid neural network

Abstract

The application provides a kind of based on hybrid neural network's cellular network flow space-time prediction method and system, belongs to intelligent communication technical field;The application adopts layered deep neural network architecture, including data embedding layer, space-time coding layer, feature fusion layer and output layer.Data embedding layer maps historical traffic sequence, cross-domain external data and metadata into high-dimensional features;Space-time coding layer extracts multi-scale time features by fusing one-dimensional causal convolution and Mamba neural network through time and space modeling branches, and captures multi-scale space features by densely connected convolution and multi-head attention mechanism;Feature fusion layer realizes adaptive weighted fusion of space-time features, cross-domain features and metadata features using gated fusion mechanism;The output layer linearly transforms the fused features to generate the final prediction results.The application accurately captures the space-time dynamics of traffic flow, and the predicted curve closely matches the true value, achieving accurate prediction of cellular network multi-service traffic.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent communication, and particularly relates to a cellular network traffic spatiotemporal prediction method and system based on a hybrid neural network. BACKGROUND

[0002] With the large-scale deployment of 5G / 6G networks and the popularity of Internet of Things devices, cellular networks are facing the challenges of traffic explosion, business diversification, and the continuous improvement of user experience. Under this background, network traffic prediction, as a core technology for realizing intelligent network operation, dynamic resource allocation, and guaranteeing service quality, its accuracy is directly related to network operation efficiency and user satisfaction. Accurate short-term and long-term traffic prediction can provide crucial decision-making basis for key tasks such as network slice resource allocation, base station energy management, and network congestion prevention.

[0003] A network traffic prediction method and system based on spatiotemporal sequences is disclosed in Chinese patent application No. CN202410499848.7, which uses a diffusion convolution gated recurrent unit and a convolutional neural network to extract time and spatial features, respectively, and performs convolution on the time features, spatial features, and influence features to obtain a combination matrix. After multiplying the combination matrix with a weight matrix and activating it through a Sigmoid function, the final prediction result is output. This method can realize end-to-end prediction of cellular network traffic, and its advantages lie in its simple model structure, high computational efficiency, and suitability for resource-constrained scenarios. However, it has the following limitations: first, it is difficult to effectively capture long-term temporal dependencies in traffic. The diffusion convolution GRU model used is limited by the recursive structure of recurrent neural networks, and there is a gradient decay problem, which cannot accurately learn the long-period patterns of traffic data; second, the model does not introduce key metadata such as hours and holidays, making it difficult to effectively model the periodicity of traffic and lacking the ability to respond to traffic peaks caused by sudden events; third, it uses a static weight fusion mechanism that cannot adaptively adjust the contribution of multiple source features, which restricts the prediction accuracy and generalization performance in complex spatiotemporal scenarios.

[0004] In addition, existing network traffic prediction methods still face systemic technical bottlenecks when dealing with complex spatiotemporal dynamics, specifically manifested in the following three aspects: First, at the level of basic time series modeling, mainstream models suffer from an inherent contradiction between efficiency and performance. Models based on recurrent neural networks (RNN / LSTM) are limited by their recursive computation mode, resulting in the vanishing gradient problem when processing long sequences, making it difficult to effectively capture long-term dependencies, and their training efficiency is low. While Transformer-based models improve their ability to model long-term dependencies through self-attention mechanisms, their computational complexity increases quadratically with sequence length, leading to unbearable computational overhead in large-scale fine-grained traffic prediction scenarios. Furthermore, while models such as convolutional neural networks (CNN) can process sequences through causal convolution, their receptive field is limited, making it inherently difficult to model long-term dependencies beyond the size of the convolution kernel. Second, at the level of architectural design, existing methods generally lack effective spatiotemporal coordination mechanisms, deep cross-domain capabilities, and metadata fusion capabilities. Most models separate or simply superimpose temporal and spatial features, failing to achieve true spatiotemporal joint modeling. Especially in utilizing cross-domain external data and metadata, existing strategies mostly involve simple early concatenation or linear weighting. These methods cannot dynamically adjust the fusion weights according to the context of the data itself, resulting in coarse information fusion and potentially introducing noise, weakening the model's ability to extract deep, discriminative features from diverse data. Third, at the overall performance level, the above limitations collectively lead to insufficient prediction accuracy and generalization ability of existing technologies when dealing with sudden traffic fluctuations and complex spatial relationships in real-world network environments. Existing solutions often struggle to achieve a good balance between computational efficiency, long-term reliance on model accuracy, and the ability to fusion complex features.

[0005] Therefore, there is an urgent need in this field for an integrated solution that can fundamentally improve the efficiency of time series modeling, achieve deep fusion of spatiotemporal features, and support adaptive fusion of cross-domain and metadata information, so as to overcome the comprehensive deficiencies of existing technologies in terms of accuracy, efficiency and practicality. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, this invention proposes a spatiotemporal prediction method and system for cellular network traffic based on hybrid neural networks. This solves the problems of limited prediction accuracy and low computational efficiency caused by the long-term reliance on insufficient capture capabilities, inefficient spatiotemporal feature coordination, and coarse cross-domain information fusion in existing methods.

[0007] To achieve the above objectives, the first aspect of the present invention provides a spatiotemporal prediction method for cellular network traffic based on a hybrid neural network, comprising the following processes:

[0008] S1. Obtain historical cellular network traffic data, cross-domain external data, and metadata, and preprocess them to obtain historical traffic sequences, cross-domain data matrices, and metadata vectors;

[0009] S2. Input the processed historical traffic sequence, cross-domain data matrix and metadata vector into the data embedding layer to obtain high-dimensional spatiotemporal sequence features, high-dimensional cross-domain features and high-dimensional metadata features respectively.

[0010] S3. Input the high-dimensional spatiotemporal sequence features into the time modeling branch, and extract and fuse short-term and long-term time features through a one-dimensional causal convolutional neural network and a Mamba neural network to obtain multi-scale time features;

[0011] High-dimensional spatiotemporal sequence features are input into the spatial modeling branch. Through a densely connected convolutional neural network and a multi-head attention mechanism, local and global spatial features are extracted and fused to obtain multi-scale spatial features.

[0012] S4. Fuse multi-scale temporal features with multi-scale spatial features and generate unified spatiotemporal coding features through a fully connected layer;

[0013] S5. Adaptive fusion of spatiotemporal coding features, high-dimensional cross-domain features and high-dimensional metadata features using a gated fusion mechanism to obtain the final fused features;

[0014] S6. The final fused features are linearly transformed by the output layer to generate the final cellular network traffic prediction result.

[0015] Furthermore, the specific process of S1 includes:

[0016] S1.1 Obtain historical cellular network traffic data, cross-domain external data, and metadata from a specific scenario. The historical cellular network traffic data includes time-series data of three service types—SMS, Call, and Internet—on an H×W spatial grid. The cross-domain external data includes base station distribution information. Point of interest distribution information and social activity information All of these are processed into H×W matrices aligned with the traffic data space; the metadata includes time context features, specifically hour, day of the week, and whether it is a holiday, forming a metadata vector. ;

[0017] S1.2 Clean and aggregate historical cellular network traffic data, process missing and outlier values, and aggregate them to a preset time granularity.

[0018] S1.3. Normalize the aggregated historical traffic data and cross-domain external data, and construct them into a tensor format with a unified spatial dimension that the model can input; at the same time, standardize the metadata and construct it into a vector format that the model can input; for a specific service type s∈{SMS, Call, Internet}, cellular traffic is represented as a spatiotemporal data point sequence. ={ |t=1,2,…,T}, where T is the total number of time intervals. It is the flow matrix in the H×W geographic region at time t; each sample contains historical flow data for P consecutive time steps. , ..., }, corresponding cross-domain data for each time step and metadata vector for each time step Used to predict the flow at the next time step. .

[0019] Furthermore, the specific data processing procedure of the S2 data embedding layer includes:

[0020] The preprocessed historical traffic sequences are standardized, and for service type s, the sequence is constructed as a three-dimensional tensor. Then, it is mapped to high-dimensional spatiotemporal sequence features through a linear projection layer. ;

[0021] For cross-domain external data, including base station distribution information Point of interest distribution information and social activity information Processed into tensors through join operations High-dimensional cross-domain features are obtained by encoding through a two-layer convolutional neural network. ;

[0022] Metadata vector The data is standardized and then embedded and encoded using a two-layer fully connected neural network to obtain high-dimensional metadata features. .

[0023] Furthermore, the specific process S3 includes:

[0024] High-dimensional spatiotemporal sequence features output from the data embedding layer Input a one-dimensional causal convolutional neural network to capture short-term temporal features in traffic sequences. ;

[0025] High-dimensional spatiotemporal sequence features The input is fed into a Mamba neural network to model long-term temporal dependencies in the sequence and obtain the output features. ;

[0026] Will and By fusing the data, multi-scale temporal features can be generated. ;

[0027] High-dimensional spatiotemporal sequence features output from the data embedding layer Input a densely connected convolutional neural network and extract local features of network traffic in the spatial dimension. ;

[0028] Will The input is fed into a multi-head attention mechanism to calculate the association weights between all spatial locations and capture global spatial dependency features. ;

[0029] Will and By fusing the data, multi-scale spatial features can be generated. .

[0030] Furthermore, the specific process of S4 includes:

[0031] Multiscale temporal features output by the temporal modeling branch Spatial features output by the spatial modeling branch Perform dimensional alignment to ensure consistency between the two in terms of time, space, and batch size, resulting in aligned features. and ;

[0032] Aligned multi-scale temporal features Spatial features The features are then concatenated to generate a comprehensive spatiotemporal coding feature. .

[0033] Furthermore, the specific process of S5 includes:

[0034] The spatiotemporal features output by the spatiotemporal coding layer Cross-domain features independently encoded via convolutional neural networks Metadata features encoded via fully connected neural network embedding Spatiotemporal features projected onto the same dimension are obtained by performing projection transformations using different learnable matrices. Cross-domain characteristics and metadata features ;

[0035] Based on the projected spatiotemporal features Cross-domain characteristics and metadata features Generate adaptive gating weights respectively The projected spatiotemporal features, cross-domain features, and metadata features are weighted and fused using the gate weights to output the final fused features. .

[0036] Furthermore, the specific process of S6 includes:

[0037] The fused features output by the feature fusion layer The input and output layers are mapped to a low-dimensional tensor with the same dimension as the predicted target space through a linear projection operation. ;

[0038] For low-dimensional tensors After performing inverse normalization to restore the data to the original range of traffic data, the final cellular network traffic prediction result is obtained. .

[0039] Furthermore, multi-scale temporal features The extraction is accomplished collaboratively by a one-dimensional causal convolutional neural network and a Mamba neural network: the one-dimensional causal convolutional neural network focuses on capturing local mutations and short-term patterns between adjacent time points through limited receptive fields and causal constraints; the Mamba neural network first suppresses feature distribution shifts through layer normalization, then performs two convolutions to extract local temporal information and introduces residual connections to alleviate gradient vanishing, and then performs layer normalization again to stabilize intermediate features; in addition, internally, it uses linear layer dimension mapping, convolutional feature transformation, nonlinear enhancement of activation functions, combined with the dynamic state update mechanism of the selective state space model (SSM) to collaboratively model the long-term macroeconomic trends and periodic patterns of the sequence;

[0040] Multi-scale spatial features The extraction is accomplished collaboratively by a densely connected convolutional neural network and a multi-head attention mechanism: the densely connected convolutional neural network consists of multiple cascaded "batch normalization + ReLU activation + convolution" modules, with feature reuse achieved between modules through dense connections, while residual connections are introduced to ensure effective feature transfer; the multi-head attention module uses multiple parallel attention heads, after querying ( ),key( ),value( Projection, scaling dot product calculation, layer normalization, and feedforward neural network processing are used to capture global spatial dependencies.

[0041] Finally, the temporal and spatial features are adaptively weighted by the gated fusion module. The two types of features are used as input 1 and input 2, respectively. Each input is first activated by tanh to generate a feature representation. Input 1 is first activated by the Sigmoid activation function to generate an adaptive weight, which is then processed by "1-" to obtain another weight. Subsequently, the two weights are weighted by input 1 and input 2, respectively. Finally, the weighted results are summed and output to achieve a unified representation and modeling of the dynamic temporal and spatial characteristics of network traffic.

[0042] The second aspect of the present invention provides a spatiotemporal prediction system for cellular network traffic based on a hybrid neural network, which is applied to the spatiotemporal prediction method for cellular network traffic as described in the first aspect; it includes a data embedding layer, a spatiotemporal coding layer, a feature fusion layer, and an output layer;

[0043] The data embedding layer is used to embed features of historical traffic sequences, cross-domain external data and metadata. It generates high-dimensional spatiotemporal sequence features through a linear projection layer, encodes high-dimensional cross-domain features through a convolutional neural network, and embeds high-dimensional metadata features through a fully connected neural network.

[0044] The spatiotemporal coding layer includes a parallel temporal modeling branch and a spatial modeling branch. The temporal modeling branch uses one-dimensional causal convolution and Mamba neural network to extract multi-scale temporal features, while the spatial modeling branch uses densely connected convolutional network and multi-head attention mechanism to extract multi-scale spatial features. The multi-scale temporal features and multi-scale spatial features are integrated into a unified spatiotemporal coding feature.

[0045] The feature fusion layer uses a gated fusion mechanism to adaptively weight and fuse spatiotemporal coding features, high-dimensional cross-domain features and high-dimensional metadata features to obtain the final fused features.

[0046] The output layer performs linear transformation and inverse normalization on the final fused features to generate cellular network traffic prediction results.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. This invention effectively overcomes the computational bottleneck of traditional Transformer models when processing long sequences by constructing a time modeling branch that works in collaboration with a one-dimensional causal convolutional network and a Mamba neural network. It significantly improves the efficiency of long sequence processing while ensuring prediction accuracy, and solves the technical problem of insufficient accuracy of existing methods in long-term prediction.

[0049] 2. This invention achieves unified extraction of local spatial features and global spatial correlations through a spatial modeling branch that combines densely connected convolutional networks with a multi-head attention mechanism;

[0050] 3. This invention comprehensively considers external influencing factors such as base station distribution, point of interest distribution, and social activity levels, as well as metadata information such as hours, days of the week, and whether it is a holiday. Through an adaptive fusion mechanism based on a gated neural network, it achieves deep interaction between spatiotemporal features, cross-domain external data, and metadata, thereby comprehensively improving the ability to analyze the patterns of network traffic changes. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the method of the present invention.

[0052] Figure 2 This is a schematic diagram of the network architecture of the system of the present invention.

[0053] Figure 3 A schematic diagram of the structure of the long-short time fusion timing modeling module.

[0054] Figure 4 This is a schematic diagram of the dual-path spatial feature extraction module.

[0055] Figure 5 This is a schematic diagram of the gating fusion module structure.

[0056] Figure 6 A comparison chart of prediction results for SMS services in a certain region.

[0057] Figure 7 A comparison chart of prediction results for a certain area of ​​call services.

[0058] Figure 8 A comparison chart of prediction results for a certain region of network services. Detailed Implementation

[0059] like Figure 1 As shown, the spatiotemporal prediction method for cellular network traffic based on hybrid neural networks proposed in this invention includes the following steps:

[0060] S1. Obtain historical cellular network traffic data, cross-domain external data, and metadata, and preprocess them to obtain historical traffic sequences, cross-domain data matrices, and metadata vectors;

[0061] S2. Input the processed historical traffic sequence, cross-domain data matrix and metadata vector into the data embedding layer to obtain high-dimensional spatiotemporal sequence features, high-dimensional cross-domain features and high-dimensional metadata features respectively.

[0062] S3. Input the high-dimensional spatiotemporal sequence features into the time modeling branch, and extract and fuse short-term and long-term time features through a one-dimensional causal convolutional neural network and a Mamba neural network to obtain multi-scale time features;

[0063] High-dimensional spatiotemporal sequence features are input into the spatial modeling branch. Through a densely connected convolutional neural network and a multi-head attention mechanism, local and global spatial features are extracted and fused to obtain multi-scale spatial features.

[0064] S4. Fuse multi-scale temporal features with multi-scale spatial features and generate unified spatiotemporal coding features through a fully connected layer;

[0065] S5. The spatiotemporal coding features, high-dimensional cross-domain features and high-dimensional metadata features are adaptively fused using a gating fusion mechanism to obtain the final fused features.

[0066] S6. The final fused features are linearly transformed by the output layer to generate the final cellular network traffic prediction result.

[0067] like Figure 2 As shown, the spatiotemporal prediction system for cellular network traffic based on a hybrid neural network model includes a data embedding layer, a spatiotemporal coding layer, a feature fusion layer, and an output layer.

[0068] The data embedding layer is used to embed features of historical traffic sequences, cross-domain external data and metadata. It generates high-dimensional spatiotemporal sequence features through a linear projection layer, encodes high-dimensional cross-domain features through a convolutional neural network, and embeds high-dimensional metadata features through a fully connected neural network.

[0069] The spatiotemporal coding layer includes a parallel temporal modeling branch and a spatial modeling branch. The temporal modeling branch uses one-dimensional causal convolution and Mamba neural network to extract multi-scale temporal features, while the spatial modeling branch uses densely connected convolutional network and multi-head attention mechanism to extract multi-scale spatial features. The multi-scale temporal features and multi-scale spatial features are integrated into a unified spatiotemporal coding feature.

[0070] The feature fusion layer uses a gated fusion mechanism to adaptively weight and fuse spatiotemporal coding features, high-dimensional cross-domain features and high-dimensional metadata features to obtain the final fused features.

[0071] The output layer performs linear transformation and inverse normalization on the final fused features to generate cellular network traffic prediction results.

[0072] The present invention will be further described below with reference to embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0073] S1. The specific method for obtaining historical cellular network traffic data, cross-domain external data, and metadata, and preprocessing them to obtain historical traffic sequences, cross-domain data matrices, and metadata vectors includes the following sub-steps:

[0074] S1.1 Obtain historical cellular network traffic data, cross-domain external data, and metadata from a specific scenario. The cellular network traffic data includes time-series data of three service types—SMS, Call, and Internet—on an H×W spatial grid. The cross-domain external data includes base station distribution information. Point of interest distribution information and social activity information All of these are processed into H×W matrices aligned with the traffic data space; the metadata includes time context features, specifically hour, day of the week, and whether it is a holiday, forming a metadata vector. ;

[0075] S1.2 Perform data cleaning and aggregation on cellular network traffic data, process missing and outlier values, and aggregate them to a preset time granularity.

[0076] S1.3. Normalize the aggregated traffic data and cross-domain external data, and construct them into a tensor format with a unified spatial dimension that the model can input; at the same time, standardize the metadata and construct it into a vector format that the model can input; for a specific service type s∈{SMS, Call, Internet}, cellular traffic can be represented as a spatiotemporal data point sequence. ={ |t=1, 2, ..., T}, where T is the total number of time intervals. For any grid location (i, j), where i∈[1, H], j∈[1, W], the flow matrix in the H×W geographic region at time t is defined as: This is used to measure cellular traffic in the cell at coordinates (i, j); each training sample contains historical traffic data for P consecutive time steps. , ..., }, corresponding cross-domain data for each time step and metadata vector for each time step Used to predict the flow at the next time step. .

[0077] The cross-domain external data is also represented in the same spatial grid as the traffic data, as specifically defined below:

[0078] For any grid location (i, j), where i ∈ [1, H], j ∈ [1, W], the base station distribution matrix in the H×W geographical region is defined as follows: , is used to represent the number of base stations in the cell with coordinates (i, j);

[0079] For any grid location (i, j), where i ∈ [1, H], j ∈ [1, W], the point of interest distribution matrix in the H×W geographic region is defined as follows: , is used to represent the number of interest points in the cell with coordinates (i, j);

[0080] For any grid location (i, j), where i ∈ [1, H], j ∈ [1, W], the social activity distribution matrix in the H×W geographic region is defined as follows: , used to represent the number of social activities in the unit with coordinates (i, j);

[0081] Finally, a spatiotemporal sequence dataset was constructed for model training and validation, where each training sample contains historical traffic data for P consecutive time steps. , ..., }, corresponding cross-domain data for each time step and metadata vector for each time step As input, to predict the flow at the next time step. .

[0082] S2. The specific method for inputting the processed historical traffic sequence, cross-domain data matrix, and metadata vector into the data embedding layer to obtain high-dimensional spatiotemporal sequence features, high-dimensional cross-domain features, and high-dimensional metadata features includes the following sub-steps:

[0083] S2.1. Standardize the preprocessed historical cellular network traffic sequences. For service type s, its sequence... ={ , , ..., It can be constructed as a three-dimensional tensor. Then, through a linear projection, it is mapped to a high-dimensional flow feature tensor. :

[0084] ;

[0085] in, (·) represents the reshaping operation, used to adjust the tensor structure. (·) represents the linear projection layer, responsible for feature dimension mapping; through (·)and The (·) operation ultimately outputs a unified three-dimensional tensor. , It can be directly used as input for each modeling branch in time and space;

[0086] S2.2, Cross-domain datasets , and Processed into tensors through join operations Cross-domain features are obtained by encoding through a two-layer convolutional neural network. , is represented as: ;

[0087] ;

[0088] in It's a splicing operation. , The convolution operation is represented by Conv1, which is used to align the input dimensions, and Conv2, which is used to extract features and encode external knowledge into feature representations.

[0089] S2.3, Metadata Vector The data is standardized by embedding and encoding using a two-layer fully connected neural network, followed by a reshaping operation to obtain high-dimensional metadata features. ;;

[0090] ;

[0091] in and These are the learnable parameters of the Lth layer, where L∈{1,2}. This represents the Sigmoid activation function.

[0092] S3. Input the high-dimensional spatiotemporal sequence features into the time modeling branch, and extract and fuse short-term and long-term time features to obtain multi-scale time features through a one-dimensional causal convolutional neural network and a Mamba neural network; input the high-dimensional spatiotemporal sequence features into the spatial modeling branch, and extract and fuse local and global spatial features to obtain multi-scale spatial features through a densely connected convolutional neural network and a multi-head attention mechanism. The specific method includes the following sub-steps:

[0093] S3.1, High-dimensional traffic features output from the data embedding layer Inputting a one-dimensional causal convolutional neural network to capture short-term temporal features in a flow sequence, its output features It can be represented as: ;

[0094] S3.2, will The input is fed into a Mamba neural network. Mamba, as a selective state-space model, has parameters that change according to the current input. The dynamic changes effectively capture long-term time dependencies in the sequence; the output sequence of this module is then subjected to a subsequent linear transformation to obtain the output features. :

[0095] ;

[0096] ;

[0097] ;

[0098] in, and These represent the hidden states at the current and previous time steps, respectively. This is the discretized state transition matrix. The input weight matrix is ​​the discretized result. To output the weight matrix, Represents the Hadama product. Given the current input vector, Output for the current moment, Y={ , …, } represents the sequence of outputs at all times;

[0099] S3.3, will and Perform dual-input gating fusion, for example based on An adaptive gating weight Z is generated, and short-term and long-term time features of the gating weight are weighted and fused to generate multi-scale time features. :

[0100] ;

[0101] ;

[0102] in, It is the Sigmoid activation function. and Let represent the learnable weights and biases of the gated network, and ⊙ represent the Hadamard product;

[0103] S3.4, High-dimensional traffic features output from the data embedding layer Input a densely connected convolutional neural network and extract local features of network traffic in the spatial dimension. :

[0104] ;

[0105] Wherein, DenseConvBlock represents a convolutional neural network module with a densely connected structure;

[0106] S3.5, will The input is fed into a multi-head attention mechanism and transformed into a sequence. This is used to calculate the association weights between all spatial locations and capture global spatial dependency features. :

[0107] ;

[0108] ;

[0109] in, , , and The projected weight matrix is ​​a learnable matrix. Where is the dimension of the key vector, and Concat is the concatenation operation;

[0110] S3.6, will and Perform dual-input gating fusion to generate multi-scale spatial features. .

[0111] In this embodiment, the extraction of temporal features is accomplished collaboratively by a one-dimensional causal convolutional neural network and a Mamba neural network, as follows: Figure 3 As shown: One-dimensional causal convolutional neural networks, with their limited receptive field and causal constraints, focus on capturing local abrupt changes and short-term patterns between adjacent time points; the Mamba neural network first suppresses feature distribution shift through layer normalization, then performs two convolutions to extract local temporal information and introduces residual connections to alleviate gradient vanishing, followed by another layer normalization to stabilize intermediate features; in addition, internally, it uses linear layer dimension mapping, convolutional feature transformation, nonlinear enhancement of activation functions, combined with the dynamic state update mechanism of the Selective State Space Model (SSM), to collaboratively model the long-term macro-trends and periodic patterns of the sequence; the extraction of spatial features is completed collaboratively by densely connected convolutional neural networks and multi-head attention mechanisms, as shown in the process. Figure 4 As shown: The densely connected convolutional neural network consists of multiple cascaded "batch normalization + ReLU activation + convolution" modules. Features are reused between modules through dense connections, while residual connections are introduced to ensure effective feature transfer. The multi-head self-attention module uses multiple parallel attention heads, after querying ( ),key( ),value( Projection, scaling dot product calculation, layer normalization, and feedforward neural network processing are used to capture global spatial dependencies; finally, the temporal and spatial features are processed separately... Figure 5The dual-input gating fusion module shown in part (a) adaptively weights the two types of features to be fused as input 1 and input 2, respectively. Each input is first activated by tanh to generate a feature representation. Input 1 is first activated by the Sigmoid activation function to generate an adaptive weight, which is then processed by "1-" to obtain another weight. Subsequently, the two weights are weighted by input 1 and input 2, respectively. Finally, the weighted results are summed and output to achieve a unified representation and accurate modeling of the dynamic temporal and spatial characteristics of network traffic.

[0112] S4. The specific method for fusing multi-scale temporal features with multi-scale spatial features and generating unified spatiotemporal coding features through a fully connected layer includes the following sub-steps:

[0113] S4.1 Multi-scale temporal features output from the temporal modeling branch Spatial features output by the spatial modeling branch Perform dimensional alignment to ensure consistency between the two in terms of time, space, and batch size, resulting in aligned features. and ;

[0114] S4.2, Align the multi-scale temporal features Spatial features The features are then concatenated to generate a comprehensive spatiotemporal coding feature. .

[0115] S5. The specific method for adaptively fusing spatiotemporal encoded features, high-dimensional cross-domain features, and high-dimensional metadata features using a gated fusion mechanism to obtain the final fused feature includes the following sub-steps:

[0116] S5.1 Spatiotemporal features output by the spatiotemporal coding layer Cross-domain features independently encoded via convolutional neural networks Metadata features encoded via fully connected neural network embedding Spatiotemporal features projected onto the same dimension are obtained by performing projection transformations using different learnable matrices. Cross-domain characteristics and metadata features ;

[0117] S5.2, Based on the spatiotemporal features after projection Cross-domain characteristics and metadata features Generate adaptive gating weights respectively The projected spatiotemporal features, cross-domain features, and metadata features are weighted and fused using the gate weights to output the final fused features. :

[0118] ;

[0119] ;

[0120] ;

[0121] ;

[0122] in, It is the Sigmoid activation function. and (i∈{1,2,3}) represents the learnable weights and biases of the gated network, and ⊙ represents the Hadamard product.

[0123] In this embodiment, the workflow of the three-input gating fusion mechanism is as follows: Figure 5 As shown in part (b): the three features to be fused are used as input 1, input 2 and input 3 respectively. Each input is first activated by tanh to generate a feature representation, and then activated by sigmoid to generate its own adaptive gating weights. Subsequently, each gating weight is weighted by its corresponding feature representation. Through this process, the module can dynamically weigh the importance of different features, so that the model can flexibly adjust its dependence on different features according to the specific context, thereby effectively improving prediction accuracy and robustness.

[0124] S6. The specific method for generating the final cellular network traffic prediction result by linearly transforming the final fused features through the output layer includes the following sub-steps:

[0125] S6.1, The fused features output by the feature fusion layer The input and output layers are mapped to a low-dimensional tensor with the same dimension as the predicted target space through a linear projection operation. :

[0126] ;

[0127] in, and For learnable weights and biases;

[0128] S6.2, For low-dimensional tensors After performing inverse normalization to restore the data to the original range of traffic data, the final cellular network traffic prediction result is obtained. :

[0129] ;

[0130] in, and These are the normalized boundary values ​​obtained from the original data during the training phase.

[0131] This embodiment describes a spatiotemporal prediction method for cellular network traffic based on a hybrid neural network model. The implementation environment is based on Ubuntu 20.04 operating system, with Python 3.8 as the programming environment, PyTorch 2.1.1 as the deep learning framework, and CUDA 11.8. The model underwent 300 training rounds on a system equipped with an NVIDIA GeForce RTX 4090 GPU (24GB VRAM). Experiments compared the fitting effects of this method on the actual and predicted values ​​for three different services. Figure 6 , Figure 7 and Figure 8 As shown, the spatiotemporal dynamics are accurately captured, and the predicted curve closely matches the actual value. The results also compare the RMSE, MAE, and other performance metrics of three different methods. RMSE and MAE represent root mean square error and mean absolute error, respectively; the smaller the values ​​of both, the better. The larger the value, the better, representing the proportion of real data variation that the model can explain. Specific comparative experimental results are shown in Table 1:

[0132] Table 1. RMSE, MAE, and... (The rest of the text appears to be a table or table, possibly related to different methods and business metrics.) contrast

[0133]

[0134] In summary, this invention addresses the challenges of balancing efficiency and performance in long-sequence time-series modeling, the lack of spatiotemporal feature collaboration mechanisms, and insufficient cross-domain and metadata information fusion capabilities in existing methods. While maintaining linear complexity, it achieves efficient capture of long and short-period time dependencies and deep fusion of local and global spatial features. Based on a gating mechanism, it completes adaptive weighted fusion of spatiotemporal features, cross-domain external data, and metadata, thereby improving the accuracy and efficiency of cellular network traffic prediction.

[0135] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0136] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A spatiotemporal prediction method for cellular network traffic based on hybrid neural networks, characterized in that, The process includes the following: S1. Obtain historical cellular network traffic data, cross-domain external data, and metadata, and preprocess them to obtain historical traffic sequences, cross-domain data matrices, and metadata vectors; S2. Input the processed historical traffic sequence, cross-domain data matrix and metadata vector into the data embedding layer to obtain high-dimensional spatiotemporal sequence features, high-dimensional cross-domain features and high-dimensional metadata features respectively. S3. Input the high-dimensional spatiotemporal sequence features into the time modeling branch, and extract and fuse short-term and long-term time features through a one-dimensional causal convolutional neural network and a Mamba neural network to obtain multi-scale time features; High-dimensional spatiotemporal sequence features are input into the spatial modeling branch. Through a densely connected convolutional neural network and a multi-head attention mechanism, local and global spatial features are extracted and fused to obtain multi-scale spatial features. Specifically, this includes: high-dimensional spatiotemporal sequence features output from the data embedding layer. Input a one-dimensional causal convolutional neural network to capture short-term temporal features in traffic sequences. ; High-dimensional spatiotemporal sequence features The input is fed into a Mamba neural network to model long-term temporal dependencies in the sequence and obtain the output features. ; Will and By fusing the data, multi-scale temporal features can be generated. ; High-dimensional spatiotemporal sequence features output from the data embedding layer Input a densely connected convolutional neural network and extract local features of network traffic in the spatial dimension. ; Will The input is fed into a multi-head attention mechanism to calculate the association weights between all spatial locations and capture global spatial dependency features. ; Will and By fusing the data, multi-scale spatial features can be generated. ; Among them, multi-scale time features The extraction is accomplished collaboratively by a one-dimensional causal convolutional neural network and a Mamba neural network: the one-dimensional causal convolutional neural network, through its limited receptive field and causal constraints, focuses on capturing local mutations and short-term patterns between adjacent time points; the Mamba neural network first suppresses feature distribution shift through layer normalization, then performs two convolutions to extract local temporal information and introduces residual connections to alleviate gradient vanishing, and then performs layer normalization again to stabilize intermediate features; in addition, the Mamba neural network internally models the long-term macroeconomic trends and periodic patterns of sequences through linear layer dimension mapping, convolutional feature transformation, nonlinear enhancement of activation functions, and the dynamic state update mechanism of the selective state space model (SSM). Multi-scale spatial features The extraction is accomplished collaboratively by a densely connected convolutional neural network and a multi-head attention mechanism. The densely connected convolutional neural network consists of multiple cascaded "batch normalization + ReLU activation + convolution" modules. Features are reused between modules through dense connections, while residual connections are introduced to ensure effective feature transfer. The multi-head attention module uses multiple parallel attention heads, after querying... ,key ,value Projection, scaling dot product calculation, layer normalization, and feedforward neural network processing are used to capture global spatial dependencies. Finally, the temporal and spatial features are adaptively weighted by the gated fusion module. The two types of features are used as input 1 and input 2, respectively. Each input is first activated by tanh to generate a feature representation. Input 1 is first activated by the Sigmoid activation function to generate an adaptive weight, which is then processed by "1-" to obtain another weight. Subsequently, the two weights are weighted by input 1 and input 2, respectively. Finally, the weighted results are summed and output to achieve a unified representation and modeling of the dynamic temporal and spatial characteristics of network traffic. S4. Fuse multi-scale temporal features with multi-scale spatial features and generate unified spatiotemporal coding features through a fully connected layer; S5. Adaptive fusion of spatiotemporal coding features, high-dimensional cross-domain features and high-dimensional metadata features using a gated fusion mechanism to obtain the final fused features; S6. The final fused features are linearly transformed by the output layer to generate the final cellular network traffic prediction result.

2. The spatiotemporal prediction method for cellular network traffic based on a hybrid neural network as described in claim 1, characterized in that: The specific process of S1 includes: S1.1 Obtain historical cellular network traffic data, cross-domain external data, and metadata from a specific scenario. The historical cellular network traffic data includes time-series data of three service types: SMS, Call, and Internet, on an H×W spatial grid. The cross-domain external data includes base station distribution information. Point of interest distribution information and social activity information All of these are processed into H×W matrices aligned with the traffic data space; the metadata includes time context features, specifically hour, day of the week, and whether it is a holiday, forming a metadata vector. ; S1.2 Clean and aggregate historical cellular network traffic data, process missing and outlier values, and aggregate them to a preset time granularity. S1.

3. Normalize the aggregated historical traffic data and cross-domain external data, and construct them into a tensor format with a unified spatial dimension that the model can input; at the same time, standardize the metadata and construct it into a vector format that the model can input; for a specific service type s∈{SMS, Call, Internet}, cellular traffic is represented as a spatiotemporal data point sequence. ={ |t=1,2,…,T}, where T is the total number of time intervals. It is the flow matrix in the H×W geographic region at time t; each sample contains historical flow data for P consecutive time steps. Cross-domain data corresponding to each time step and metadata vector for each time step Used to predict the flow at the next time step. .

3. The spatiotemporal prediction method for cellular network traffic based on a hybrid neural network as described in claim 1, characterized in that: The specific data processing procedure of the S2 data embedding layer includes: The preprocessed historical traffic sequences are standardized, and for service type s, the sequence is constructed as a three-dimensional tensor. Then, it is mapped to high-dimensional spatiotemporal sequence features through a linear projection layer. ; For cross-domain external data, including base station distribution information Point of interest distribution information and social activity information Processed into tensors through join operations High-dimensional cross-domain features are obtained by encoding through a two-layer convolutional neural network. ; Metadata vector The data is standardized and then embedded and encoded using a two-layer fully connected neural network to obtain high-dimensional metadata features. .

4. The spatiotemporal prediction method for cellular network traffic based on a hybrid neural network as described in claim 1, characterized in that: The specific process of S4 includes: Multiscale temporal features output by the temporal modeling branch Spatial features output by the spatial modeling branch Perform dimensional alignment to ensure consistency between the two in terms of time, space, and batch size, resulting in aligned features. and ; Aligned multi-scale temporal features Spatial features The features are then concatenated to generate a comprehensive spatiotemporal coding feature. .

5. The spatiotemporal prediction method for cellular network traffic based on a hybrid neural network as described in claim 1, characterized in that: The specific process of S5 includes: The spatiotemporal features output by the spatiotemporal coding layer Cross-domain features independently encoded via convolutional neural networks Metadata features encoded via fully connected neural network embedding Spatiotemporal features projected onto the same dimension are obtained by performing projection transformations using different learnable matrices. Cross-domain characteristics and metadata features ; Based on the projected spatiotemporal features Cross-domain characteristics and metadata features Generate adaptive gating weights respectively The projected spatiotemporal features, cross-domain features, and metadata features are weighted and fused using the gate weights to output the final fused features. .

6. The spatiotemporal prediction method for cellular network traffic based on a hybrid neural network as described in claim 1, characterized in that: The specific process of S6 includes: The fused features output by the feature fusion layer The input and output layers are mapped to a low-dimensional tensor with the same dimension as the predicted target space through a linear projection operation. ; For low-dimensional tensors After performing inverse normalization to restore the data to the original range of traffic data, the final cellular network traffic prediction result is obtained. .

7. A spatiotemporal prediction system for cellular network traffic based on a hybrid neural network, characterized in that: Applied to the spatiotemporal prediction method for cellular network traffic as described in any one of claims 1 to 6; It includes a data embedding layer, a spatiotemporal coding layer, a feature fusion layer, and an output layer; The data embedding layer is used to embed features of historical traffic sequences, cross-domain external data and metadata. It generates high-dimensional spatiotemporal sequence features through a linear projection layer, encodes high-dimensional cross-domain features through a convolutional neural network, and embeds high-dimensional metadata features through a fully connected neural network. The spatiotemporal coding layer includes a parallel temporal modeling branch and a spatial modeling branch. The temporal modeling branch uses one-dimensional causal convolution and Mamba neural network to extract multi-scale temporal features, while the spatial modeling branch uses densely connected convolutional network and multi-head attention mechanism to extract multi-scale spatial features. The multi-scale temporal features and multi-scale spatial features are integrated into a unified spatiotemporal coding feature. The feature fusion layer uses a gated fusion mechanism to adaptively weight and fuse spatiotemporal coding features, high-dimensional cross-domain features and high-dimensional metadata features to obtain the final fused features. The output layer performs linear transformation and inverse normalization on the final fused features to generate cellular network traffic prediction results.

Images