Cellular network flow time sequence prediction method and system based on time-space-frequency fusion

Abstract

The invention discloses a cellular network flow time sequence prediction method and system based on time-space-frequency fusion, and relates to the technical field of communication network intelligent analysis, and the method comprises the steps: carrying out the adaptive learning of an implicit association relationship between different cells based on a graph structure; performing frequency decomposition on a historical flow sequence to obtain dominant periodic components and high-frequency disturbance characteristics, performing joint modeling on spatial characteristics and frequency domain representation to generate a flow prediction value of a future time slice, and adopting an improved loss function in a training process to obtain a flow prediction value of the future time slice; therefore, the fluctuation characteristics of the cellular flow can be better adapted and the training stability of the model can be improved. According to the method, the spatial dependency relationship and the frequency domain structure of the cellular traffic can be described simultaneously, the accuracy and stability of traffic prediction are improved, and the method is suitable for complex network scenes such as multi-base-station collaborative prediction.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis technology for communication networks, specifically to a spatiotemporal frequency joint traffic prediction method and system for cellular networks, belonging to network traffic prediction, mobile communication network management, and artificial intelligence modeling methods. Background Technology

[0002] As mobile communication networks grow in scale and service complexity, cellular traffic exhibits significant diversity and dynamic changes. Implicit spatial dependencies arise between cells due to geographical proximity, user mobility, and service associations, leading to spatial correlations in multi-cell traffic. Simultaneously, cellular traffic displays clear daily patterns over time, exhibiting discernible pattern differences across different date types within a week, accompanied by high-frequency fluctuations and sudden changes. In practical network management, accurate prediction of future traffic is crucial for resource scheduling, load balancing, capacity planning, and quality of service assurance. In real-world scenarios, sudden events, increased user mobility, and dynamic changes in network structure further exacerbate traffic uncertainty, making traditional single-domain modeling methods insufficient for real-time prediction requirements.

[0003] Existing traffic prediction methods primarily focus on time series modeling or graph structure modeling based on fixed topologies. However, they often struggle to simultaneously characterize complex spatial dependencies and periodic features, resulting in limited accuracy in dynamically changing scenarios. In the frequency domain, some methods fail to effectively utilize the periodic structure of traffic sequences, leading to insufficient representation of periodic features. In the spatial dimension, some models rely on pre-defined adjacency relationships, making it difficult to adaptively capture potential spatial dependencies in real networks. Furthermore, the significant variations in the fluctuation range of cellular traffic across different order of magnitudes result in a markedly unbalanced distribution of prediction errors, placing higher demands on the stability of the loss function.

[0004] Therefore, there is an urgent need to construct a prediction model that can jointly utilize spatial dependence and frequency domain structure while taking into account dynamic characteristics, in order to improve the overall accuracy of cellular traffic prediction. Summary of the Invention

[0005] This invention addresses the problem that cellular network traffic exhibits both implicit spatial correlation and temporal structure, while existing prediction methods struggle to capture both spatial dependence and frequency domain periodic information in a single model. It proposes a spatiotemporal joint modeling method and system for cellular traffic prediction.

[0006] In a first aspect, embodiments of the present invention provide a method for predicting time series of cellular network traffic based on spatiotemporal-frequency fusion, comprising:

[0007] Each cell in the cellular network is treated as a node. The historical network traffic data sequence of the node is obtained, and a unified representation is obtained through the feature embedding layer. The embedded features are then correlated and modeled through cross attention to obtain the cross attention output features.

[0008] Implicit spatial features between cells are adaptively extracted based on a self-attention mechanism using cross-attention output features;

[0009] Wavelet transform is performed on the cross-attention output features to obtain different frequency components, and periodic and fluctuation information is highlighted through the frequency attention mechanism to form time-frequency features;

[0010] After implicit spatial features and time-frequency features are adaptively integrated through fusion and weighting mechanisms, the results are output through the output layer to generate prediction results for future time steps.

[0011] In some optional implementations, the step of treating each cell in the cellular network as a node and obtaining the historical network traffic data sequence of the node includes:

[0012] Suppose a cellular network consists of N cell nodes. At time step t, the traffic of the i-th node is denoted as . The state of the entire cellular network at time step t is represented as a vector. The network states of the most recent L historical time steps are stacked in chronological order to form a historical network traffic data sequence: .

[0013] In some optional implementations, the adaptive extraction of implicit spatial features between cells based on a self-attention mechanism from the cross-attention output features includes:

[0014] The query vector is obtained by linearly projecting the cross-attention output features. Key vector Value vector ;

[0015] Depend on Obtain attention weights ,in, Indicates the dimension of the key vector;

[0016] Depend on The implicit space features are obtained by weighting the value vectors according to the attention weights. .

[0017] In some optional implementations, performing wavelet transform on the cross-attention output features to obtain different frequency components, and highlighting periodic and fluctuation information through a frequency attention mechanism to form time-frequency features, includes:

[0018] Wavelet transform is applied to the cross-attention output features to achieve multi-scale frequency domain decomposition, obtaining frequency domain features corresponding to different frequency bands. Where W(⋅) is the wavelet transform operator, Represents the coefficients for the corresponding frequency band;

[0019] Depend on Features of different frequency bands are scored using a learnable mapping function, and frequency weights are obtained after normalization. It is a learnable mapping function used to calculate the importance of frequency bands. This represents the normalized importance weight corresponding to each frequency band, where f represents the frequency band.

[0020] Depend on Based on the learned frequency weights Weighted fusion of multi-band features yields frequency-enhanced time-frequency features. .

[0021] In some alternative implementations, the adaptive integration of implicit spatial features and time-frequency features through fusion and weighting mechanisms includes:

[0022] Let the implicit space feature be The time-frequency characteristics are The implicit spatial features and time-frequency features are concatenated to obtain a joint representation as follows: ,in, Represents the implicit spatial feature dimension. Represents the time-frequency feature dimension;

[0023] The joint features are processed node by node, and by Global statistics of the i-th node are extracted using average pooling operations along the feature dimension. ,in, This represents the average pooling operation across the feature dimension;

[0024] The node features after average pooling are input into a weight generation structure consisting of two fully connected layers. First, the features are mapped by the first fully connected layer and transformed using ReLU non-linear activation. Then, the features are mapped by the second fully connected layer and combined with the Sigmoid function to output the attention weights corresponding to the nodes. ,in, , This reflects the importance of the i-th node to the overall prediction. and These are the learnable weight matrices for the first and second fully connected layers, respectively. The Sigmoid activation function is used to apply weights... It is limited to the range [0,1].

[0025] Depend on Apply node-level weights to the joint features of the corresponding nodes. .

[0026] In some optional implementations, the step of generating prediction results for future time steps through the output layer includes:

[0027] Depend on Obtain the traffic prediction result for the i-th cell in the next time step. , The prediction function is implemented through a single linear mapping.

[0028] In some alternative implementations, the training loss function is: ,in, This is an adaptive scaling term used to dynamically adjust the residual contribution of samples of different magnitudes. These are scale control parameters. As a stable term, and Let represent the actual value and the predicted value of the i-th cell, respectively.

[0029] Secondly, embodiments of the present invention provide a cellular network traffic time series prediction system based on spatiotemporal-frequency fusion, comprising:

[0030] The data preprocessing module treats each cell in the cellular network as a node, obtains the historical network traffic data sequence of the node, obtains a unified representation through the feature embedding layer, and performs correlation modeling on the embedded features through cross attention to obtain the cross attention output features.

[0031] The spatiotemporal feature module is used to adaptively extract implicit spatial features between cells based on the self-attention mechanism from the cross-attention output features;

[0032] The time-frequency feature module is used to perform wavelet transform on the cross-attention output features to obtain different frequency components, and to highlight periodic and fluctuation information through the frequency attention mechanism to form time-frequency features.

[0033] The dynamic weighted attention module is used to adaptively integrate implicit spatial features and time-frequency features through fusion and weighting mechanisms;

[0034] The prediction output module is used to generate prediction results for future time steps through the output layer.

[0035] Thirdly, the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the methods described above.

[0036] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the methods described above.

[0037] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0038] 1. Existing traffic prediction methods mostly rely on single time series modeling or graph structure modeling based on fixed adjacency matrices, which makes it difficult to simultaneously capture the complex spatial dependencies and periodic variation features in cellular networks, resulting in insufficient prediction accuracy in dynamic network environments. This invention introduces a spatial feature modeling method based on a self-attention mechanism, which can adaptively learn the implicit associations between cells without prior topology, more accurately reflecting the traffic interaction characteristics in real-world scenarios, and significantly improving the model's spatial expressive power.

[0039] 2. Traditional methods often struggle to balance the characteristics of different time scales when dealing with cellular traffic patterns that exhibit both periodicity and high-frequency disturbances, resulting in insufficient responsiveness to sudden fluctuations and changes in periodic patterns. This invention employs a wavelet transform-based frequency domain analysis method to decompose the traffic sequence into different frequency components and highlights the dominant periodic characteristics through a frequency attention mechanism. This enables the model to more comprehensively characterize daily cycles and local high-frequency variations, improving prediction stability under complex time series structures.

[0040] 3. In prediction tasks, the impact of different regional units on changes in traffic volume is not uniform, and traditional methods typically do not model the differences in contribution between different regional units. This invention introduces a weighted attention mechanism, which can adaptively assign higher weights to regional units with greater influence, thereby strengthening the feature representation capabilities of key regions and improving overall prediction performance.

[0041] 4. Traditional loss functions are easily affected by the amplitude differences in different flow ranges, which may cause the model to overemphasize high flow ranges and ignore low flow or subtle fluctuations during training. This invention designs an adaptive scaling LogCosh loss function, which adjusts the error contribution of samples of different magnitudes through a dynamic scaling factor, so that the model has a consistent optimization objective across the entire distribution range, thereby improving prediction accuracy.

[0042] It should be understood that the above-described beneficial effects are not a limiting description of the present invention, and other embodiments of the present invention may also produce similar technical effects. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a schematic diagram of a method provided in an embodiment of the present invention;

[0045] Figure 2 The following are the model prediction performance comparison results on the Milano dataset provided in the embodiments of the present invention, wherein (a) is the MAE comparison result and (b) is the RMSE comparison result;

[0046] Figure 3 The following are comparison results of model prediction performance on the Treno dataset provided in the embodiments of the present invention, wherein (a) is the MAE comparison result and (b) is the RMSE comparison result;

[0047] Figure 4 The following are the model prediction performance comparison results on the MobileNJ dataset provided in the embodiments of the present invention, wherein (a) is the MAE comparison result and (b) is the RMSE comparison result;

[0048] Figure 5 This is a schematic diagram of the system structure provided in an embodiment of the present invention;

[0049] Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0051] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0052] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0053] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use. These terms are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0054] Furthermore, terms such as "horizontal," "vertical," and "sag" do not imply that components must be absolutely horizontal or suspended, but rather that they can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal relative to "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0055] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0056] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0057] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0058] In this embodiment of the invention, a time series prediction method for cellular network traffic based on spatiotemporal-frequency fusion is provided, which aims to solve the problem of insufficient prediction accuracy of existing cellular network traffic prediction under complex spatial dependence, periodic structure and different traffic amplitude distribution. The method improves the accuracy of future traffic prediction by extracting spatiotemporal features, extracting time-frequency features, fusing spatiotemporal-frequency features and optimizing adaptive scaling loss function, so that the prediction model can more comprehensively characterize the spatiotemporal dynamic characteristics of cellular traffic.

[0059] like Figure 1 As shown, the model in this embodiment of the invention uses the historical network traffic data sequence of the cell as input. The input first passes through the feature embedding layer to obtain a unified representation, and then performs spatiotemporal modeling and time-frequency modeling in parallel. Spatiotemporal modeling is based on a self-attention mechanism to adaptively extract implicit spatial features between cells. Time-frequency modeling performs wavelet transform on the embedded features to obtain different frequency components, and highlights the periodic and fluctuation information that is more critical to the prediction task through a frequency attention mechanism to form time-frequency features. The features output by the two branches are adaptively integrated through a fusion and weighting mechanism and then input to the output layer to generate the prediction results of the target cell for the future time step.

[0060] In this embodiment of the invention, in spatiotemporal modeling, cells in a cellular network are used as the basic unit for spatial modeling. Each cell is considered a node, and its historical traffic sequence is used as the node feature input to the spatial modeling module. Without relying on predefined geographical adjacencies, an attention-based spatial relationship modeling method adaptively learns the implicit association strength between cell nodes, characterizing the spatial dependencies of different cells during traffic changes and obtaining spatially enhanced feature representations.

[0061] In this embodiment of the invention, in time-frequency modeling, the historical cellular traffic sequence is decomposed into frequency, and wavelet transform is used to extract the periodic dominant components, trend change features and high-frequency disturbance features to obtain a frequency domain representation.

[0062] In this embodiment of the invention, during attention-weighted fusion, spatial features and frequency domain features are concatenated along the feature dimension to construct a node-level joint feature representation. Based on this, according to the degree of influence of different cell nodes on overall traffic changes, corresponding weight coefficients are adaptively generated through a node-weighted attention mechanism, and these weights are applied to the joint features to achieve enhanced expression of key cell features.

[0063] In this embodiment of the invention, in traffic prediction, the attention-weighted joint feature representation is used as input, and the predicted cellular network traffic values ​​for multiple future time steps are output. During model training, an improved adaptive LogCosh loss function is used to calculate the prediction error, so as to improve the prediction accuracy and stability of the model at different traffic scales.

[0064] The implementation process of each step is explained in detail below:

[0065] Specifically, the cellular network is set up by Composed of individual cell nodes, in time step At that time, the first The traffic of each node is denoted as Therefore, at any given moment The overall network state can be represented as a vector. , will recently The network states at each historical time step are stacked in chronological order to form a historical input sequence. N is the number of nodes, and D is the feature dimension.

[0066] Cellular network traffic exhibits both implicit spatial correlation and multi-frequency temporal structure. Spatial correlation reflects the topological, service, or regional coupling relationships between different base stations, while temporal structure reflects the periodicity, trends, and short-term dynamic changes in traffic. Against this backdrop, the embodiments of the present invention are based on a given historical sequence. Under the given conditions, predict the flow sequence for the next time step, i.e., learn a parameterized mapping. .in This represents the prediction result for time t+1. These are learnable parameters.

[0067] Specifically, in cellular networks, there is usually a significant spatial correlation between different cells. Changes in cell traffic are not only related to their own service characteristics, but also affected by the load distribution and coverage relationships of surrounding cells. Traditional spatial modeling methods often rely on geographical distance or artificially constructed adjacency matrices to characterize the relationships between base stations, which is difficult to fully reflect the implicit correlations that exist in real networks.

[0068] To enhance the ability to characterize such implicit spatial dependencies, this embodiment of the invention employs a GraphTrans module based on a self-attention mechanism to learn implicit spatial dependencies between cells. The GraphTrans module uses cells in a cellular network as spatial modeling objects, characterizing the correlation between cells through attention weights among the traffic features of each cell, and using an attention weight matrix to replace the predefined adjacency matrix to express the spatial dependencies between cells. Simultaneously, a shared linear mapping is used to project the features of each cell, constructing query, key, and value representations to ensure the stability and consistency of high-dimensional traffic features in the spatial relationship modeling process. This design enables the GraphTrans module to adaptively learn implicit spatial dependencies between cells without requiring manual definition of the topology.

[0069] In the specific implementation, firstly, a historical input sequence of length L is used. As input. Indicates the historical sequence at time [time]. The corresponding input feature matrix is ​​fed into the feature embedding layer for linear mapping to obtain the embedded feature representation. Building upon this, a cross-attention module is further used to model the association between the embedded features, resulting in the cross-attention output features. and output features from the obtained cross-attention Performing a linear projection yields the query, key, and value vectors:

[0070]

[0071] in, These are trainable parameter matrices, used to apply cross-attention output features. Perform linear projections of the query, key, and value directions to obtain feature representations for attention computation.

[0072] Subsequently, the correlation weights between nodes are adaptively learned through a self-attention mechanism, thereby adaptively learning the correlation between each node:

[0073]

[0074] in, This represents the dimension of the key vector, used for normalizing and scaling the attention score.

[0075] The value vectors are weighted and combined according to the attention weights to obtain the spatially enhanced feature representation:

[0076]

[0077] In summary, the GraphTrans module generates spatially enhanced features containing inter-cell correlation information by weighting and combining base station features through a self-attention structure without relying on a predefined adjacency matrix. ,exist Figure 1 In the diagram, this spatial enhancement feature is represented as: It can learn the implicit dependency structure between cells without pre-setting the adjacency matrix, and can more realistically reflect the dynamic spatial relationships formed by factors such as user distribution, business interaction, and coverage overlap in the actual network.

[0078] Specifically, to characterize the periodic structure of cellular traffic, this embodiment of the invention introduces a frequency domain feature extraction module based on wavelet transform. Structurally, the frequency domain feature extraction module can be regarded as a frequency domain encoder, which mainly consists of two substructures: frequency domain decomposition and frequency attention modeling.

[0079] In the specific implementation, the frequency domain feature extraction module takes the input features as input features As input, firstly, the input features Applying wavelet transform to achieve multi-scale frequency domain decomposition yields frequency domain feature representations corresponding to different frequency bands:

[0080]

[0081] in, The wavelet transform operator can decompose low-frequency trends and high-frequency local disturbances. This represents the coefficients for the corresponding frequency band. This operation realizes the mapping from time domain representation to multi-scale frequency domain representation, providing a foundation for subsequent frequency selection and fusion.

[0082] Subsequently, a frequency attention mechanism is introduced to adaptively model the importance of each frequency band, and weights are assigned to different frequency bands through this mechanism, enabling the more critical frequency components to be highlighted adaptively according to task requirements. A learnable mapping function is used to score the features of different frequency bands, and the frequency weights are obtained after normalization.

[0083]

[0084] in, It is a learnable mapping function used to calculate the importance of frequency bands. This represents the normalized importance weight for each frequency band. This mechanism can adaptively highlight frequency components that are more critical to the prediction task based on the data distribution characteristics. Specifically, As a neural network, its parameters are optimized through backpropagation based on error feedback during training, thereby gradually adjusting the weighting of frequency bands to ensure that the model can automatically highlight the frequency components that are key to the prediction task.

[0085] Finally, the multi-band features are weighted and fused according to the learned frequency weights to obtain a frequency-domain enhanced feature representation:

[0086]

[0087] This frequency domain enhancement indicates That is Figure 1 Output characteristics of the mid-time-frequency modeling branch .

[0088] This step can effectively capture the dominant periodic components, local peak disturbances, and variation patterns at different times, thereby making up for the problem of insufficient utilization of periodic patterns when modeling only in the time domain.

[0089] This step uses a frequency attention mechanism to weight and combine features from each frequency band, forming a frequency-domain enhanced representation that includes temporal information. This results in frequency-domain features that can simultaneously characterize periodic structures, low-frequency trends, and high-frequency disturbances. Compared to modeling methods that rely solely on time-domain sequences, this feature can supplement the temporal details that cannot be captured in the spatial modeling stage, enabling the model to have a more comprehensive temporal structure representation capability in subsequent predictions and improving prediction accuracy.

[0090] Specifically, considering that different cells have varying influences on overall traffic changes, relying solely on mean aggregation may weaken key regional features. This embodiment of the invention further designs a node-level weighted attention mechanism to achieve explicit modeling of node importance. Let the node features output by the spatial modeling module be... The output of the frequency domain modeling module after wavelet decomposition and frequency attention is: , Represents the implicit spatial feature dimension. This represents the time-frequency feature dimension.

[0091] First, the two types of features are concatenated along the feature dimension to obtain a joint representation. Subsequently, the joint features are processed node by node, and global statistics of the nodes are extracted through average pooling operations along the feature dimension:

[0092]

[0093] in, This represents an average pooling operation on the feature dimension, used to compress high-dimensional node features into compact descriptive vectors, thereby characterizing the overall activity level of nodes at the current time step. Based on this, the compressed node features are input into a weight generation structure consisting of two fully connected layers. First, the weights are mapped by the first fully connected layer and then... Nonlinear activation is used for feature transformation, followed by a second fully connected layer mapping and combination. The function outputs the attention weights corresponding to the nodes, and then calculates the importance weights of the nodes through a non-linear mapping:

[0094]

[0095] in, This reflects the importance of the i-th node to the overall prediction. and These are the learnable weight matrices for the first and second fully connected layers, respectively. The sigmoid activation function is used to restrict the weights to a certain value. Within the range. Finally, the node-level weights are applied to the spatial-frequency joint features of the corresponding nodes to achieve weighted enhancement of key nodes:

[0096]

[0097] Specifically, It is a matrix The Row eigenvectors, and It is a feature matrix containing all location features. Finally, the weighted features are input into the prediction function. In this function, the feature is mapped to the target flow value, yielding the flow prediction result for the next time step. This prediction function is achieved by performing a linear transformation on the fused features; that is, by weighting and summing the features according to their weights and adding a bias term, the corresponding flow prediction value for the next time step is output. Its calculation form is as follows:

[0098]

[0099] Unlike traditional attention mechanisms that rely solely on adjacency similarity, this invention directly models the global importance of cells at the cell level. This highlights key cells that significantly impact overall traffic changes while suppressing interference from noisy cells and low-activity areas, achieving global optimization of spatial-frequency domain fusion features and thus significantly improving prediction performance.

[0100] By emphasizing the importance of key nodes in the global traffic pattern, this mechanism can enhance the model's ability to focus on key areas, thereby effectively improving overall prediction performance.

[0101] Specifically, cellular network traffic exhibits significant magnitude differences across different regions and time periods. High-traffic areas experience large fluctuations, while low-traffic areas are more susceptible to noise interference. If a traditional loss function is directly applied, the high-traffic residual will dominate the optimization process, leading to training imbalance. To improve the model's adaptability to different traffic volumes, this embodiment of the invention introduces an adaptive-scale LogCosh loss function. During model training, this loss function serves as the overall optimization objective, used to measure the model's prediction results. Compared with actual traffic value The differences between them are analyzed, and the learnable parameters in the prediction model are updated end-to-end through the backpropagation algorithm, thereby guiding the model to obtain a more stable and balanced fitting effect in different flow ranges.

[0102]

[0103] in This is an adaptive scaling term used to dynamically adjust the residual contribution of samples of different magnitudes; These are scale control parameters; As a stable term, and They represent the first The true value and the predicted value of each sample.

[0104] Adaptive scaling item This is used to dynamically adjust the error scale based on the magnitude of the true sample value, so that traffic samples of different magnitudes are mapped to a unified relative scale space during loss calculation. This avoids the dominant effect of high-traffic samples on the loss function and improves the learning stability of the model in the low-traffic range. Furthermore, it introduces... The term is used to prevent gradient instability caused by a denominator of zero or a value that is too small. Its value is usually set to a constant that is much smaller than the average flow amplitude of the dataset, so as to ensure numerical stability without affecting the scaling effect.

[0105] Meanwhile, the gradient of the standard LogCosh loss function remains bounded in the large error region, preventing training instability caused by excessively large gradients when dealing with large residuals. Combined with the adaptive scaling term proposed in this embodiment, the loss function maintains a more balanced gradient update across samples of different magnitudes, thus making the training process more stable and the prediction results more consistent. Furthermore, this loss function can balance the error ratio between high and low flow ranges, giving the model a more consistent optimization objective across the entire range, thereby significantly improving training stability and prediction accuracy.

[0106] like Figures 2-4 The figures show the comparison results of the proposed method with various baseline models on three public datasets—Milano, Trento, and MobileNJ—in terms of MAE and RMSE metrics. As shown in the figures, the proposed method outperforms other baseline models in both MSE and RMSE metrics on the three different public datasets. These comparison results demonstrate that the proposed method maintains relatively stable and excellent performance across different datasets.

[0107] In another embodiment of the present invention, to facilitate better implementation of the method provided in the embodiments of the present invention, the present invention also provides a system based on the above method. The meanings of the terms are the same as in the above method, and specific implementation details can be found in the description of the method embodiments.

[0108] Please see Figure 5 , Figure 5 This is a schematic diagram of the system provided in an embodiment of the present invention. The system may include a data preprocessing module 501, a spatiotemporal feature module 502, a time-frequency feature module 503, a dynamic weighted attention module 504, and a prediction output module 505, wherein:

[0109] The data preprocessing module 501 is used to treat each cell in the cellular network as a node, obtain the historical network traffic data sequence of the node, obtain a unified representation through the feature embedding layer, and perform correlation modeling on the embedded features through cross attention to obtain the cross attention output features.

[0110] The spatiotemporal feature module 502 is used to adaptively extract implicit spatial features between cells based on the self-attention mechanism from the cross-attention output features;

[0111] The time-frequency feature module 503 is used to perform wavelet transform on the cross-attention output features to obtain different frequency components, and to highlight periodic and fluctuation information through the frequency attention mechanism to form time-frequency features.

[0112] The dynamic weighted attention module 504 is used to adaptively integrate implicit spatial features and time-frequency features through fusion and weighting mechanisms.

[0113] The prediction output module 505 is used to generate prediction results for future time steps through the output layer.

[0114] The specific implementation methods of each module can be referred to the description of the above method embodiments, and the embodiments of the present invention will not be repeated.

[0115] In another embodiment of the present invention, a computer device is also provided, such as... Figure 6 As shown, it illustrates a structural schematic diagram of a computer device involved in an embodiment of the present invention, specifically:

[0116] The computer device may include components such as a processor 601 with one or more processing cores, a memory 602 with one or more computer-readable storage media, a power supply 603, and an input unit 604. Those skilled in the art will understand that... Figure 6 The computer device structure shown does not constitute a limitation on the computer device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein:

[0117] The processor 601 is the control center of the computer device. It connects various parts of the computer device via various interfaces and lines, and performs various functions and processes data by running or executing software programs and / or modules stored in the memory 602, and by calling data stored in the memory 602, thereby providing overall monitoring of the computer device. Optionally, the processor 601 may include one or more processing cores; preferably, the processor 601 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operation of the storage medium, user interface, and application programs, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into the processor 601.

[0118] The memory 602 can be used to store software programs and modules. The processor 601 executes various functional applications and data processing by running the software programs and modules stored in the memory 602. The memory 602 may mainly include a program storage area and a data storage area. The program storage area may store application programs required for operating the storage medium and at least one function (such as sound playback function, image playback function, etc.); the data storage area may store data created according to the use of the computer device. In addition, the memory 602 may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 602 may also include a controller to provide the processor 601 with access to the memory 602.

[0119] The computer device also includes a power supply 603 that supplies power to the various components. Preferably, the power supply 603 can be logically connected to the processor 601 via a power management storage medium, thereby enabling functions such as charging, discharging, and power consumption management through the power management storage medium. The power supply 603 may also include one or more DC or AC power supplies, recharge storage media, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0120] The computer device may also include an input unit 604, which can be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0121] Although not shown, the computer device may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 601 in the computer device loads the executable files corresponding to the processes of one or more applications into the memory 602 according to the following instructions, and the processor 601 runs the applications stored in the memory 602, thereby implementing the steps in the above method embodiment.

[0122] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0123] Therefore, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, the computer program being loaded by a processor to execute the steps of any method provided in the embodiments of the present invention.

[0124] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0125] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0126] Since the computer program stored in the computer-readable storage medium can execute the steps of any of the methods provided in the embodiments of the present invention, the beneficial effects that any of the methods provided in the embodiments of the present invention can achieve can be realized, as detailed in the preceding embodiments, and will not be repeated here.

[0127] The foregoing has provided a detailed description of a method and system for predicting cellular network traffic time series based on spatiotemporal frequency fusion, as provided in the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A spatio-temporal-frequency fusion-based cellular network traffic time series prediction method, characterized in that, The method comprises the following steps: Each cell in the cellular network is taken as a node, and a historical network traffic data sequence of the node is obtained, a uniform representation is obtained through a feature embedding layer, and cross-attention is used to model the association of the embedded features to obtain cross-attention output features; Implicit spatial features between cells are adaptively extracted from the cross-attention output features based on a self-attention mechanism; Wavelet transform is performed on the cross-attention output features to obtain different frequency components, and frequency attention mechanism is used to highlight periodic and fluctuation information to form time-frequency features; After the implicit spatial features and the time-frequency features are adaptively integrated through a fusion and weighting mechanism, the prediction result of the future time step is output through an output layer.

2. The method of claim 1, wherein, The method further comprises the following steps: Let the cellular network be composed of N cell nodes, at time step t, the traffic of the i-th node is denoted as The state of the whole cellular network at time step t is represented as a vector Stacking the network states of the last L historical time steps in chronological order, a historical network traffic data sequence is constructed: .

3. The method of claim 2, wherein, The method further comprises the following steps: linearly projecting the cross-attention output features to obtain a query vector , key vector , value vector ; By obtaining attention weights wherein, denotes the dimension of the key vector; By The value vectors are combined with attention weights to obtain implicit spatial features .

4. The method of claim 3, wherein, The method further comprises the following steps: A wavelet transform is applied to the cross-attention output feature to realize multi-scale frequency domain decomposition, to obtain frequency domain features corresponding to different frequency bands wherein W(·) is a wavelet transform operator, denotes the coefficient of the corresponding frequency band. By The different frequency band features are scored by a learnable mapping function, and the frequency weight is obtained by normalization processing, wherein, The learnable mapping function is used to calculate the importance of the frequency band, The normalized importance weight corresponding to each frequency band is represented, and f represents the frequency band. By According to the learned frequency weight The multi-band features are weighted and fused to obtain the frequency domain enhanced time-frequency features .

5. The method of claim 4, wherein, The method further comprises the following steps: Let the implicit spatial feature be , the time-frequency feature be , and the joint representation obtained by splicing the implicit spatial feature and the time-frequency feature be , where represents the dimension of the implicit spatial feature, represents the dimension of the time-frequency feature. The joint features are processed by node and are represented by The global statistical information of the i-th node is extracted by the average pooling operation on the feature dimension wherein, denotes the average pooling operation on the feature dimension; The node features after the average pooling operation are input into a weight generation structure composed of two fully connected networks. First, the features are transformed through the first fully connected mapping and ReLU nonlinear activation, and then the second fully connected mapping is performed and the corresponding attention weight of the node is output in combination with the Sigmoid function wherein, , reflects the importance of the i-th node to the overall prediction, and are the learnable weight matrices of the first and second fully connected networks, respectively, is the Sigmoid activation function, which is used to limit the weight to the range [0, 1]; by the node-level weight is applied to the joint features of the corresponding node .

6. The method of claim 5, wherein, The method further comprises the following steps: From get the traffic prediction result of the i-th cell at the next time step , is a prediction function, which is realized by a linear mapping.

7. The method of claim 6, wherein, The training loss function is: wherein, is an adaptive scaling term for dynamically adjusting the residual contribution of different magnitude samples, is a scale control parameter, is a stabilizing term, and and represent the true and predicted values of the i-th cell, respectively.

8. A spatio-temporal-frequency fusion based cellular network traffic time series prediction system, characterized in that, The method further comprises the following steps: The method comprises the following steps: A data preprocessing module is configured to take each cell in the cellular network as a node, obtain a historical network traffic data sequence of the node, obtain a uniform representation through a feature embedding layer, and use cross-attention to model the association of the embedded features to obtain cross-attention output features; A space-time feature module is configured to adaptively extract implicit spatial features between cells from the cross-attention output features based on a self-attention mechanism; A time-frequency feature module is configured to perform wavelet transform on the cross-attention output features to obtain different frequency components, and use frequency attention mechanism to highlight periodic and fluctuation information to form time-frequency features; A dynamic weighted attention module is configured to adaptively integrate the implicit spatial features and the time-frequency features through a fusion and weighting mechanism; 9. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, A prediction output module is configured to output the prediction result of the future time step through an output layer. The processor executes the computer program to realize the steps of the method of any one of claims 1 to 7. 10.A computer readable storage medium having stored thereon a computer program, the computer program being executed by a processor to realize the steps of the method of any one of claims 1 to 7.

Images