Urban network traffic prediction method and system based on covariate interaction and graph enhancement

Abstract

The application provides a city network flow prediction method and system based on covariate interaction and graph enhancement, and belongs to the field of mobile communication network flow prediction. The method comprises the following steps: acquiring city cellular network flow data and multi-source covariate data, and constructing a multi-source data set; obtaining a unified feature set through a linear mapping layer; constructing a bidirectional interaction path based on the feature set, and obtaining bidirectional interaction features through a cross attention mechanism; weighting and fusing various covariate features through an attention mechanism to obtain covariate fusion features; dynamically constructing an adjacency matrix according to a flow state, extracting spatial dependence features and long-term time sequence features, and obtaining final space-time features; and inputting the final space-time features into a multi-layer perception machine to output network flow prediction results. The application can accurately depict time-varying spatial dependence between city regions, and improves the precision and stability of city cellular network flow prediction.

Description

Technical Field

[0001] This invention belongs to the field of mobile communication network traffic prediction technology, and particularly relates to a method and system for predicting urban network traffic based on covariate interaction and graph augmentation. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the rapid development of 5G / 6G mobile communication, mobile internet, and smart city construction, and the continuous popularization of new services such as live video streaming, vehicle-to-everything (V2X) communication, and the Internet of Things (IoT), urban cellular network traffic exhibits complex dynamic characteristics characterized by high spatiotemporal coupling, drastic fluctuations, and strong suddenness. Traffic in different regions and at different times is affected by multiple factors such as user commuting, commercial activities, weather changes, and holidays, showing significant periodic and sudden fluctuations. Accurate prediction of urban cellular network traffic plays a crucial supporting role in dynamic scheduling of network resources, load balancing, intelligent operation and maintenance, and infrastructure planning.

[0004] Existing traffic flow prediction methods are mainly divided into three categories, all of which have significant limitations. Traditional statistical methods such as ARIMA and SVR rely on linear assumptions and struggle to fit the strong nonlinearity and complex spatiotemporal dependencies of traffic flow data. Deep learning methods such as LSTM and GRU can capture temporal features, but they primarily focus on modeling the time dimension and are insufficient in characterizing the spatial relationships between urban areas. Graph Neural Networks (GNNs), while capable of jointly modeling spatiotemporal relationships, still suffer from three major drawbacks: First, the multi-source covariates are not fully utilized. Existing methods only perform simple feature splicing on external information such as weather, time, holidays, and POIs, without modeling the bidirectional dynamic dependence between flow and covariates, ignoring the real law that flow status inversely affects the importance of covariates, resulting in low utilization of external information.

[0005] Secondly, spatial relationship modeling is static and fixed. Most models use a fixed adjacency matrix to describe regional relationships, which cannot reflect the dynamic changes in urban traffic spatial relationships over time, activities, and events, and is difficult to adapt to the spatial structure migration in scenarios such as morning rush hour and holidays.

[0006] Third, the model lacks robustness. In real-world scenarios, issues such as missing covariates, sensor noise, and data anomalies are common. Existing methods lack targeted compensation mechanisms, and errors tend to accumulate in the network, resulting in poor prediction stability and weak generalization ability in complex environments. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, this invention provides a method and system for predicting urban network traffic based on covariate interaction and graph augmentation. By constructing a bidirectional covariate interaction and dynamic graph augmentation network traffic prediction model (CDBG-Net), it solves the technical problems of insufficient utilization of multi-source covariates, static and fixed spatial correlation modeling, and insufficient model robustness in the prior art, thereby improving the accuracy and stability of urban network traffic prediction.

[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a method for predicting urban network traffic based on covariate interaction and graph augmentation; Urban network traffic prediction methods based on covariate interaction and graph augmentation include: Acquire urban cellular network traffic data and multi-source covariate data to construct a multi-source dataset; A linear mapping layer maps data from multiple source datasets to a unified feature space, resulting in a unified feature set. Based on the unified feature set, a bidirectional interaction path between network traffic data and covariate data is constructed. The bidirectional dependency between target traffic and covariates is modeled through a cross-attention mechanism to obtain bidirectional interaction features. The bidirectional interactive features are input into the covariate attention fusion module to dynamically calculate the contribution weights of various covariates to the current prediction task, and the multi-source covariate features are weighted and fused to obtain the covariate fusion features. Based on the covariate fusion features and the current traffic status information of each region, an inter-regional adjacency matrix is ​​dynamically constructed. The dynamic adjacency matrix and the covariate fusion features are then input into a graph convolutional network to extract spatial dependency features. Finally, a temporal convolutional network is used to extract long-term temporal features to obtain the final spatiotemporal features. The final spatiotemporal features are input into a multilayer perceptron, which outputs the predicted urban cellular network traffic for future moments.

[0009] As a further technical solution, the process of acquiring urban cellular network traffic data and multi-source covariate data also includes: Perform quality checks on the acquired data to determine whether there are any missing or abnormal fluctuations in each data point; For outlier data points, the historical statistical information of their location at the current time is used to make preliminary corrections to obtain time-compensated data. Based on time dimension compensation, the neighborhood aggregation features of the spatial location of the data point are obtained, and the time-compensated data is further corrected using the neighborhood aggregation features to obtain the final compensation result. Replace the original abnormal data with the final compensation result.

[0010] As a further technical solution, the multi-source covariate data includes meteorological data, time information, holiday markers, and city point of interest (POI) data.

[0011] As a further technical solution, the process of constructing a bidirectional interaction path between network traffic data and covariate data based on the unified feature set to obtain bidirectional interaction features includes: Construct the first interaction path, using historical traffic features as query vectors and covariate features as key and value vectors. Calculate the degree of attention traffic pays to covariates through cross-attention to obtain the first interaction result from traffic to covariates. A second interaction path is constructed, using covariate features as query vectors and historical traffic features as key and value vectors. The degree of inverse influence of covariates on traffic is calculated through cross-attention, resulting in the second interaction between covariates and traffic. The first interaction result and the second interaction result are fused to obtain a bidirectional fusion feature.

[0012] As a further technical solution, the bidirectional interactive features are input into the covariate attention fusion module to dynamically calculate the contribution weights of various covariates to the current prediction task, and the multi-source covariate features are weighted and fused to obtain covariate fusion features, including: Calculate the attention scores of various covariates under the current prediction task using learnable attention parameters; The attention scores of covariates across all categories are normalized to obtain the contribution weights of the covariates. The weighted summation of the features of various covariates is performed using the aforementioned contribution weights to obtain the weighted fusion features.

[0013] As a further technical solution, the process of extracting spatially dependent features includes: Obtain the historical traffic coding features of each region at the current moment, and use the feature vector of each region as the state representation of that region at this moment; For any two regions, calculate the cosine similarity between their feature vectors, and use the calculated similarity value as the dynamic connection strength between the two regions at the current time. The connection strength between all pairs of regions together constitutes the initial dynamic adjacency matrix at the current time. Obtain the historical graph structure matrix of the previous moment, and perform a weighted fusion of the initial dynamic adjacency matrix of the current moment with the historical graph structure matrix. Use a preset update coefficient to control the fusion ratio of current information and historical information to obtain the dynamic graph structure matrix of the current moment. The dynamic graph structure matrix at the current moment and the fused features are input into the graph convolutional network. The spatial information of each region and its neighboring regions is aggregated through graph convolution operations, and the spatial features containing dynamic spatial dependencies are output.

[0014] As a further technical solution, long-term temporal features are extracted through a temporal convolutional network to obtain the final spatiotemporal features, including: The spatial features output by the graph convolutional network are organized according to the time dimension to form a spatial feature sequence containing multiple historical time steps; The spatial feature sequence is input into a temporal convolutional network. Through the causal convolution operation in the temporal convolutional network, the output of the current time step depends only on the input of the current and previous historical time steps. Expanding the temporal receptive field through dilated convolution operations in temporal convolutional networks allows for the capture of long-range temporal dependencies spanning multiple time steps. The temporal features output by the temporal convolutional network and the spatial features output by the graph convolutional network are added and fused element by element to obtain the final spatiotemporal features that simultaneously contain dynamic spatial dependencies and long-term temporal dependencies.

[0015] A second aspect of the present invention provides an urban network traffic prediction system based on covariate interaction and graph enhancement.

[0016] A city network traffic prediction system based on covariate interaction and graph augmentation includes: The data acquisition module is configured to: acquire urban cellular network traffic data and multi-source covariate data, and construct a multi-source dataset; The multi-source feature encoding module is configured to map data from a multi-source dataset to a unified feature space through a linear mapping layer, thereby obtaining a unified feature set. The bidirectional covariate interaction module is configured to: construct a bidirectional interaction path between network traffic data and covariate data based on the unified feature set, model the bidirectional dependency between target traffic and covariates through a cross-attention mechanism, and obtain bidirectional interaction features; The covariate attention fusion module is configured to: input the bidirectional interactive features into the covariate attention fusion module, dynamically calculate the contribution weights of various covariates to the current prediction task, and perform weighted fusion of multi-source covariate features to obtain covariate fusion features; The dynamic graph spatiotemporal modeling module is configured to: dynamically construct an inter-regional adjacency matrix based on the covariate fusion features and the current regional traffic status information, and input the dynamic adjacency matrix and covariate fusion features into a graph convolutional network to extract spatial dependency features, and then extract long-term temporal features through a temporal convolutional network to obtain the final spatiotemporal features; The prediction output module is configured to input the final spatiotemporal features into the multilayer perceptron and output the predicted urban cellular network traffic at future moments.

[0017] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of the urban network traffic prediction method based on covariate interaction and graph augmentation as described in the first aspect of the present invention.

[0018] A fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the urban network traffic prediction method based on covariate interaction and graph augmentation as described in the first aspect of the present invention.

[0019] The above one or more technical solutions have the following beneficial effects: (1) This invention overcomes the technical deficiency of traditional methods that simply splice covariates, leading to insufficient information utilization, by constructing a bidirectional covariate interaction mechanism. Specifically, this invention constructs two cross-attention paths: "flow to covariate" and "covariate to flow," enabling the model to actively select key external factors based on the current flow status and to perceive the driving influence of the external environment on flow, thus achieving deep bidirectional dependency modeling between the two. On this basis, this invention further introduces a covariate attention fusion strategy to dynamically allocate the contribution weights of different covariates in various prediction scenarios.

[0020] (2) This invention addresses the problem that existing graph neural networks, which use fixed adjacency matrices, struggle to describe time-varying spatial relationships. It proposes a dynamic graph spatiotemporal modeling method. Specifically, this invention dynamically calculates the similarity between regions based on their real-time traffic status, constructs an initial dynamic adjacency matrix for the current moment, and weights it with the historical graph structure to achieve smooth updates of spatial dependencies. This mechanism effectively captures the dynamic changes in regional relationships during peak hours, holidays, and emergencies. Furthermore, this invention combines a temporal convolutional network with causal convolution and dilated convolution operations to simultaneously model short-term fluctuations and long-term periodic patterns.

[0021] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0023] Figure 1This is a flowchart of the method in the first embodiment.

[0024] Figure 2 This is the overall architecture diagram of the CDBG-Net model in the first embodiment.

[0025] Figure 3 This is a schematic diagram comparing the cumulative distribution function of absolute error of different models in the first embodiment on the network traffic prediction task.

[0026] Figure 4 This is a schematic diagram illustrating the impact of the bidirectional covariate interaction module on the prediction results of a typical region in the ablation experiment of the first embodiment.

[0027] Figure 5 This is a visualization diagram of the spatial distribution prediction results of urban cellular network traffic in the first embodiment. Detailed Implementation

[0028] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0029] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0030] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0031] The overall approach of this invention is as follows: This invention proposes a city cellular network traffic prediction model (CDBG-Net) based on bidirectional covariate interaction and dynamic graph enhancement. First, it introduces multi-source covariates such as meteorological data, time information, holiday markers, and city points of interest, and reduces representational differences between heterogeneous data through unified feature encoding. Based on this, a bidirectional covariate interaction mechanism is constructed, utilizing bidirectional attention to model the bidirectional dependency between target traffic and external covariates; simultaneously, a covariate attention fusion strategy is combined to dynamically characterize the importance of different covariates to the prediction task. Considering the dynamic changes in spatial relationships between urban regions over time, a dynamic graph spatiotemporal modeling module is further designed, using a dynamic adjacency matrix to characterize the time-varying spatial dependencies between regions, and combining a temporal convolutional network to mine long-term temporal features. This invention offers higher prediction accuracy and stronger robustness, effectively adapting to complex dynamic traffic scenarios in city cellular networks, and providing reliable technical support for network resource scheduling, intelligent operation and maintenance, and optimized deployment.

[0032] Example 1 This embodiment discloses a method for predicting urban network traffic based on covariate interaction and graph augmentation. By uniformly encoding multi-source data, using bidirectional attention to model the dynamic relationship between traffic and covariates, and combining dynamic graph learning of time-varying spatial dependence, it achieves high-precision and high-robust traffic prediction.

[0033] like Figure 1 As shown, the urban network traffic prediction method based on covariate interaction and graph augmentation includes: Step S1: Obtain urban cellular network traffic data and multi-source covariate data to construct a multi-source dataset.

[0034] Step S1.1: Urban cellular network traffic exhibits significant temporal dependence and spatial correlation. In actual prediction, it is necessary not only to consider the changing patterns of historical traffic sequences but also to comprehensively analyze the influence of various external factors such as meteorological conditions, time information, holiday arrangements, and the distribution of urban functions. Therefore, urban network traffic prediction is essentially a typical multivariate spatiotemporal series prediction problem. To facilitate subsequent model construction, the structure, input variables, and prediction objectives of the acquired urban cellular network traffic data and multi-source covariate data are first formally defined.

[0035] In this embodiment, the urban network traffic is assumed to form a spatiotemporal sequence in the time dimension. .in This indicates the number of urban spatial grid divisions in both directions, and C represents the number of channels for traffic data. The network traffic sequence on each grid cell... This indicates the network traffic status of each grid area in the city at time step t.

[0036] In real-world communication environments, urban network traffic is not only influenced by historical traffic patterns but also closely related to various external factors. To more accurately characterize network traffic variation patterns, multi-source covariates from different domains are introduced to enhance the model's ability to depict complex changes in communication demands. In this embodiment, the multi-source covariate data includes meteorological data, time information, holiday markers, and city point of interest (POI) data.

[0037] (1) Meteorological factors can affect residents' travel behavior and communication needs. For example, rainy weather may lead users to use mobile internet services more. By selecting meteorological factors such as temperature, humidity, wind speed, and precipitation as input features, they are defined as:

[0038] Where M represents the meteorological feature dimension; This represents the meteorological data characteristics at time step t.

[0039] (2) Urban communication behavior typically exhibits clear periodic characteristics. For example, there are differences between weekdays and weekends, and communication demands also vary regularly throughout the day. These can be represented by time characteristics such as hours, days of the week, and months:

[0040] in, Representing the time feature dimension, It represents time information characteristics.

[0041] (3) Holidays typically alter users' travel patterns and communication behaviors, thus affecting network traffic distribution. Special date information can be represented using holiday-marked variables:

[0042] Where L represents the number of holiday categories; This indicates the characteristics of holiday data.

[0043] (4) POI information can reflect the functional attributes of urban areas, such as commercial areas, educational areas, transportation hubs, and residential areas. Differences in the functional attributes of different areas will affect network communication needs; therefore, POI information is defined as:

[0044] Where P represents the feature dimension of POI; This represents the characteristics of points of interest (POIs). Unlike other time-varying covariates, POI information typically remains relatively stable over a short period of time, and therefore can be considered a static spatial feature.

[0045] Step S1.2: In real-world communication environments, issues such as missing meteorological data, delayed POI updates, and sensor noise are common. If abnormal data is used directly for prediction, errors may accumulate layer by layer in the model. Therefore, to effectively address these issues and improve robustness, this embodiment employs a joint compensation mechanism of time and spatial dimensions to correct abnormal data.

[0046] First, a comprehensive quality inspection was conducted on the collected urban cellular network traffic, weather, time, holiday, and POI data to determine whether there were any missing data, outliers, or noise interference.

[0047] For the identified outlier and missing data points, a time-dimensional compensation correction is first performed. Based on the historical time-series statistical mean of the feature corresponding to that location, the original outlier and the historical mean are weighted and fused to complete the initial correction, resulting in the time-compensated data, as shown below:

[0048] in, This is the data after time compensation; This is the original input; This is the historical average. For compensation weights.

[0049] Based on time compensation, the flow and covariate features of the spatial grid where the data point is located and its surrounding neighboring grids are further extracted. Neighborhood aggregation features are obtained through spatial weighted aggregation, and these features are used to perform a second correction on the time-compensated data. By combining time-dimensional smoothing with spatial neighborhood constraints, a stable and reliable final compensation result is output, as shown below:

[0050] in, Neighborhood features; This is the spatial compensation coefficient; This is the final revised result.

[0051] Finally, the final compensation result is used to replace the original outlier data points, completing the overall robust data correction process. Through a joint mechanism of historical compensation and neighborhood compensation, the problems of missing data and noise interference can be effectively mitigated. This provides high-quality input data for subsequent model training and traffic prediction, effectively improving the model's prediction stability and generalization ability in scenarios with missing data and strong noise.

[0052] Step S1.3, in order to uniformly describe the model input, the preprocessed historical data of urban network traffic and multi-source covariates are represented together as follows:

[0053] in, It is a collection of data from multiple sources.

[0054] Step S2: The data in the multi-source dataset is mapped to a unified feature space through a linear mapping layer to obtain a unified feature set.

[0055] Data from different sources exhibit significant differences in organization and statistical attributes, posing challenges to multi-source information fusion. Network traffic data is a typical spatiotemporal series data, containing both temporal evolution patterns and reflecting spatial distribution characteristics across regions; however, time information, holiday information, meteorological information, and POI information differ in temporal attributes, spatial attributes, and expression dimensions. Directly splicing features without unified processing can easily lead to inconsistencies in representation space and insufficient correlation modeling, thus affecting the model's efficiency in utilizing multi-source heterogeneous information. Therefore, it is necessary to reduce representational differences between heterogeneous data through unified encoding and, based on this, further conduct interactive modeling to fully explore the impact of external covariates on changes in urban network traffic.

[0056] Specifically, the multi-source dataset is input into a trained bidirectional covariate interaction and dynamic graph augmentation network traffic prediction model (CDBG-Net), such as... Figure 2 As shown, the linear mapping layer of the multi-source feature encoding module projects data from different sources into the same latent space. For historical traffic data, a linear transformation is used to extract the basic traffic representation:

[0057] in, The hidden feature representation after encoding historical traffic; Encode the weight matrix for traffic; This is the bias term. This process can preserve the basic spatiotemporal information in the original flow sequence.

[0058] For meteorological data, since it has a certain nonlinear relationship with flow rate changes, a mapping layer is also used for feature transformation:

[0059] in, This is the meteorological coding weight matrix; For bias terms; The results are the results of the meteorological feature coding.

[0060] Time information and holiday information mainly describe the periodic patterns and special fluctuation patterns in traffic flow, therefore they are coded independently:

[0061] in, The weight matrix is ​​encoded for time. For bias terms; This represents the temporal characteristics.

[0062] Time features are primarily used to characterize daily, weekly, and monthly cycles:

[0063] in, This is a holiday coding matrix; For bias terms; This represents the characteristics of holidays.

[0064] POI information reflects the functional attributes of a region, and its spatial characteristics are relatively stable.

[0065] in, Encode the weight matrix for POI; For bias terms; Output POI encoded.

[0066] Ultimately, a unified set of features is formed:

[0067] Where H is a multi-source unified feature set.

[0068] This module essentially unifies heterogeneous feature representations, providing an input foundation for subsequent deep interactions.

[0069] Step S3: Construct a bidirectional interaction path between network traffic data and covariate data based on the unified feature set, and model the bidirectional dependency between target traffic and covariates through a cross-attention mechanism to obtain bidirectional interaction features.

[0070] Traditional multi-source flow forecasting methods typically employ a simple data concatenation approach:

[0071] in, Historical network traffic characteristics; To represent the characteristics of multi-source covariates, we include meteorological characteristics, temporal characteristics, holiday characteristics, and city point of interest (POI) characteristics.

[0072] This approach assumes that covariates only have a one-way impact on the target flow, making it difficult to characterize the complex dynamic coupling between the two. In reality, external environmental factors influence flow changes, and the flow state itself determines the model's focus on covariates. For example, during peak commuting hours, time characteristics are usually more important; under extreme weather conditions, meteorological factors may become key influencing variables. Therefore, there is essentially a two-way dependency between the target flow and covariates.

[0073] To address the aforementioned issues, a bidirectional interaction mechanism is constructed in this embodiment. First, important covariate information is actively selected based on traffic characteristics. By constructing a first interaction path, historical traffic characteristics are used as the query vector, and covariate characteristics are used as the key and value vectors. Cross-attention is used to calculate the degree of attention traffic pays to the covariate, thus obtaining the first interaction result from traffic to the covariate.

[0074] Target traffic actively selects key covariate information:

[0075] in, This is a query vector generated from historical traffic characteristics; Key vectors generated for covariate features; The value vector generated for the covariate features; d is the feature dimension; This represents the interaction result from traffic to covariates. This process indicates which external factors the current traffic state is concerned with.

[0076] Secondly, by constructing a second interaction path, covariate features are used as query vectors, and historical traffic features are used as key and value vectors. The degree of inverse influence of covariates on traffic is calculated through cross-attention, thus obtaining the second interaction result from covariates to traffic.

[0077] Covariate inverse filtering of important traffic information:

[0078] in, Query vectors generated for covariates; Key vectors generated for historical traffic features; A value vector generated for historical traffic characteristics; This represents the interaction between covariates and traffic patterns. The process indicates how the external environment influences traffic patterns.

[0079] Finally, the first interaction result and the second interaction result are fused to obtain the bidirectional interaction features:

[0080] in, It features two-way interaction.

[0081] Through a two-way interaction mechanism, the model can more fully explore the deep dependencies between target traffic and external covariates.

[0082] Step S4: Input the bidirectional interactive features into the covariate attention fusion module, dynamically calculate the contribution weights of various covariates to the current prediction task, and perform weighted fusion of multi-source covariate features to obtain covariate fusion features.

[0083] Although the bidirectional interaction mechanism constructed in step S3 can enhance the modeling ability of the association between target traffic and covariates and obtain bidirectional interaction features, the contribution of different covariates to the network traffic prediction task still varies significantly under different scenarios. For example, during weekday commuting peak hours, time features usually play a more important role; during holidays, holiday features are more important; and under extreme weather conditions, meteorological features may become the dominant factor. If the bidirectional fusion result output from step S3 is directly input into the subsequent spatiotemporal modeling module, redundant covariate information can easily be introduced, reducing the model's ability to express key features.

[0084] Therefore, in this embodiment, the bidirectional interaction features obtained in step S3 are used as input for covariate importance assessment. Learnable attention parameters are used to further calculate the attention scores of different covariates under the current prediction task, thereby measuring the contribution of each type of covariate to the target traffic prediction. For the i-th type of covariate, its attention score is calculated as follows:

[0085] in, This represents the attention score for the i-th type of covariate; Indicates the learnable weight parameters; This indicates the bias term.

[0086] Furthermore, the attention scores of all covariates are normalized to obtain the contribution weights of each covariate in the current prediction task:

[0087] in, represents the contribution weight of the i-th type of covariate; K represents the total number of covariate categories; This represents the attention score of the j-th type of covariate.

[0088] Subsequently, based on the calculated contribution weights, the bidirectional interaction features output in step S3 are weighted and fused to obtain the final fused features:

[0089] in, This indicates the characteristics of covariate fusion.

[0090] Through the aforementioned dynamic weight allocation mechanism, the model can further highlight the key covariate information in the current prediction task and suppress irrelevant noise features, providing more effective input for subsequent dynamic graph spatiotemporal modeling.

[0091] Step S5: Based on the covariate fusion features and the current traffic status information of each region, dynamically construct the inter-regional adjacency matrix, and input the dynamic adjacency matrix and covariate fusion features into a graph convolutional network to extract spatial dependency features, and then extract long-term temporal features through a temporal convolutional network to obtain the final spatiotemporal features.

[0092] Urban network traffic exhibits significant spatial correlation, with traffic changes between different areas not being independent. However, traditional graphical models typically use fixed adjacency matrices to describe regional relationships, failing to reflect the dynamic changes in urban spatial dependencies. For example, during the morning rush hour, the connection between residential and commercial areas strengthens, while the link between nighttime entertainment areas and transportation hubs becomes stronger.

[0093] Therefore, after completing the covariate attention fusion, the dynamic graph spatiotemporal modeling process begins.

[0094] Step S5.1: Calculate the inter-regional correlation strength based on historical traffic coding characteristics. Using the coding feature vectors of each region as state representations, calculate the dynamic connection strength between any region i and region j at time t using cosine similarity. The formula is:

[0095] in, This is the initial dynamic adjacency matrix at the current time.

[0096] Then, the historical graph structure matrix from the previous time step is read, and the current adjacency matrix and the historical graph structure are weighted and fused according to preset update coefficients to obtain the stable dynamic graph structure matrix at the current time step. The calculation formula is as follows:

[0097] in, The current dynamic graph structure matrix; λ is the update coefficient; This is the structure matrix of the previous timeframe.

[0098] The features output from fusing the dynamic graph structure matrix and covariate attention are jointly input into the graph convolutional layer, where graph convolution operations are performed to aggregate spatial dependency information. The computation method is as follows:

[0099] in, For graph convolution weights; The output spatial features. Features are output by covariate attention fusion.

[0100] Step S5.2: After completing the spatial information aggregation based on the dynamic graph structure matrix and fusion features, a temporal convolutional network is further introduced to extract long-term temporal features.

[0101] Specifically, the spatial features output by the graph convolutional network are first... Organize the sequence according to the time dimension to construct a spatial feature sequence containing multiple historical time steps:

[0102] in, It is a spatial feature sequence organized according to the time dimension; k is the length of the historical time step.

[0103] The spatial feature sequence is then input into a temporal convolutional network, where causal convolutional layers are used to model the sequence, ensuring the output at the current time step t. It relies solely on inputs from the current and previous historical time steps to avoid future information leaks.

[0104] Meanwhile, to expand the network's temporal receptive field to capture long-range temporal dependencies, dilated convolution operations are used to enhance causal convolutions. The sampling interval of the convolution kernel is controlled by setting the dilation coefficient d, and the calculation formula is as follows:

[0105] in, These are the convolution kernel parameters; Temporal features output by temporal convolution; This is a one-dimensional temporal convolution operation; For sequence padding operations, ensure consistent timing lengths; This is an expansion coefficient used to broaden the temporal receptive field and capture long-period dependencies.

[0106] The temporal features extracted by the temporal convolutional network and the spatial features output by the graph convolutional network are fused element-wise to obtain the final spatiotemporal joint features, namely:

[0107] in, This is the final spatiotemporal feature representation. This feature accurately depicts the dynamic spatial dependence and long-term temporal fluctuation patterns of urban cellular network traffic, providing complete feature input for subsequent high-precision traffic prediction.

[0108] Step S6: Input the final spatiotemporal features into the multilayer perceptron and output the urban cellular network traffic prediction results for future moments.

[0109] After multi-source feature encoding, bidirectional covariate interaction, covariate attention fusion, and dynamic graph spatiotemporal modeling, the CDBG-Net model has obtained a high-level feature representation that includes historical traffic information, multi-source external covariate information, and dynamic spatiotemporal dependencies. This feature not only integrates information from different data sources but also fully characterizes the spatiotemporal evolution of urban network traffic.

[0110] To map high-dimensional spatiotemporal features to the final traffic prediction space, a multilayer perceptron (MLP) is used as the prediction output module. This module performs a nonlinear mapping on the final feature representation, outputting the urban network traffic prediction results for future timeframes. The expression is as follows:

[0111] in, This is the result of urban cellular network traffic prediction.

[0112] Furthermore, to measure the error between the model's predictions and the actual values, the mean absolute error (MAE) is used as the loss function, and its expression is:

[0113] in, This represents the actual traffic volume. This represents the model's predicted value; N represents the number of samples. The goal of model training is to minimize the loss function by continuously optimizing the parameters.

[0114] Through the prediction output module, the model can further convert the high-dimensional spatiotemporal features extracted by the aforementioned modules into the final prediction results, and achieve overall performance optimization through end-to-end training, thereby completing the task of predicting urban cellular network traffic.

[0115] To evaluate the performance of the method of this invention in urban network traffic prediction tasks, it was experimentally analyzed on a real urban cellular communication network dataset, along with several existing methods. The experimental data used was the widely used public dataset Milan Telecom Datase, which contains multiple data sources such as telecommunications, weather, news, social activities, and electricity in a city. The richness of the data sources provides conditions for studying the impact of external factors on urban network traffic.

[0116] In the comparative experiments, three commonly used evaluation metrics—root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²)—were selected for comparison. Existing methods included STCNet, STACN, TWACNet, MST-DenseNet, GLSTTN, and CFFTNet, encompassing both models utilizing cross-domain auxiliary information and those without, thus enabling a more comprehensive analysis of model performance differences. Table 1 shows the prediction results of each model for SMS, Call, and Internet traffic on the urban network traffic dataset.

[0117] Table 1. Comparison of prediction performance of different models on all network traffic datasets.

[0118] Table 1 shows that while models like GLSTTN, which rely solely on historical traffic data, can capture spatiotemporal dependencies well, their predictive performance remains limited in complex scenarios due to the lack of modeling external environmental factors. In contrast, models such as STCNet, STACN, TWACNet, and MST-DenseNet, which incorporate external auxiliary information, demonstrate superior prediction accuracy, indicating that meteorological, temporal, holiday, and regional functional information can effectively supplement network traffic prediction. Furthermore, CFFTNet further enhances predictive performance by employing cross-domain spatiotemporal alignment and multi-source covariate fusion mechanisms, demonstrating that unified modeling of multi-source heterogeneous data can effectively enhance the model's ability to express complex spatiotemporal relationships.

[0119] The CDBG-Net model proposed in this invention achieved optimal prediction results. This is mainly due to the bidirectional covariate interaction mechanism, which can more fully explore the dynamic dependency between target traffic and external covariates; the covariate attention fusion module, which can improve the efficiency of multi-source information utilization; the dynamic graph modeling module, which can more accurately characterize the time-varying spatial relationships between urban areas; and the robust compensation mechanism, which further improves the stability of the model in complex environments.

[0120] To further verify the effectiveness of the key modules proposed in this invention, an analysis was conducted from two perspectives: overall error distribution and prediction performance in typical regions. Specifically, the overall prediction error distribution of the model was first analyzed using the cumulative distribution function of absolute error (CDF). Subsequently, visualization experiments were carried out in typical regions for bidirectional covariate interaction (BCI), dynamic graph spatiotemporal modeling (DGM), and robust compensation (RC) to further analyze the specific role of each module.

[0121] To compare the overall error distribution of different models, the prediction errors of each model across all city grids were further statistically analyzed, and the cumulative distribution function (CDF) curves of absolute errors were plotted, as shown below. Figure 3 As shown. Among them, Figure 3 Figure (a) shows the distribution of the cumulative absolute error of the method of the present invention in the SMS traffic prediction task; Figure 3 Figure (b) shows the distribution of the cumulative absolute error of the method of the present invention in the Call traffic prediction task; Figure 3 Figure (c) shows the cumulative distribution of absolute error of the method of the present invention in Internet traffic prediction tasks. In the figure, the horizontal axis represents the absolute prediction error, and the vertical axis represents the cumulative distribution function (CDF) value. CDBG-Net represents the complete prediction model proposed in this invention; w / o BCI represents the prediction result after removing the bidirectional covariate interaction module; w / o DGM represents the prediction result after removing the dynamic graph modeling module; w / o RC represents the prediction result after removing the robust compensation module.

[0122] Depend on Figure 3 As can be seen, the CDF curve corresponding to the complete model of this invention is closer to the upper left, indicating that under the same error threshold, the method of this invention can keep more samples with smaller prediction errors and the prediction performance is better than the comparison model after removing any module. This further verifies the effectiveness of the bidirectional covariate interaction module, dynamic graph modeling module and robust compensation module in improving the prediction accuracy of urban cellular network traffic.

[0123] To further verify the effectiveness of the bidirectional covariate interaction module, ablation experiments were conducted in three typical regions: Bocconi (university area), Navigli (entertainment area), and Duomo (commercial center area). The experimental results are as follows: Figure 4 As shown.

[0124] Figure 4 The attached figure (a) shows a comparison of the prediction results for the three regions. Figure 4 Appendix (b) shows a comparison of the relative errors in the three regions. Figure 4 Figure (c) shows a comparison of the cumulative distribution function (CDF) of absolute error in the three regions. Figure 4 It can be seen that the prediction results of the method of this invention are closer to the actual traffic flow trend, especially maintaining high prediction accuracy during traffic peaks and sudden fluctuations. In contrast, after removing the bidirectional covariate interaction module, the model is more prone to underprediction or overprediction during peak periods, and the relative error fluctuates more. At the same time, the absolute error CDF curve shifts to the right overall, indicating an increase in the proportion of samples with larger errors.

[0125] The results show that the bidirectional covariate interaction module can effectively enhance the correlation modeling ability between target traffic data and external covariates, thereby improving the accuracy and stability of urban cellular network traffic prediction.

[0126] To more intuitively observe the model's predictive performance in the spatial dimension, urban cellular network traffic data at a typical moment was selected for visualization analysis. The experimental results are as follows: Figure 5 As shown.

[0127] Figure 5 The attached figure (a) shows the spatial distribution of real network traffic in the urban grid. Figure 5 Figure (b) shows the prediction results for the CDBG-Net model. Figure 5 The attached figure (c) shows the distribution of the absolute error between the predicted result and the actual value.

[0128] From an overall spatial distribution perspective, the prediction results of the model in this invention show a high degree of consistency with the actual traffic flow in terms of spatial structure. The model can effectively characterize the traffic flow differences between different areas of the city, achieving relatively accurate predictions for both high-traffic hotspots and low-traffic ordinary areas.

[0129] The distribution of hotspot areas shows that commercial centers, transportation hubs, and densely populated areas typically exhibit higher network traffic demands. The model in this invention can accurately identify the location and traffic intensity of these high-traffic areas, demonstrating that the model can effectively learn the spatial distribution patterns of urban network traffic.

[0130] As can be seen from the error distribution map, the prediction error is relatively small in most areas, with some deviation only appearing in a few areas with more drastic traffic fluctuations. This is mainly because hotspot areas are usually affected by factors such as sudden events, user gathering behavior, and short-term changes in communication demand, which makes traffic changes more complex and increases the difficulty of prediction.

[0131] In summary, Figure 5 The visualization results further validate the effectiveness of the method of the present invention in spatial prediction tasks, indicating that CDBG-Net can more accurately characterize the spatial dynamic distribution characteristics of urban cellular network traffic.

[0132] Example 2 This embodiment discloses an urban network traffic prediction system based on covariate interaction and graph augmentation; A city network traffic prediction system based on covariate interaction and graph augmentation includes: The data acquisition module is configured to: acquire urban cellular network traffic data and multi-source covariate data, and construct a multi-source dataset; The multi-source feature encoding module is configured to map data from a multi-source dataset to a unified feature space through a linear mapping layer, thereby obtaining a unified feature set. The bidirectional covariate interaction module is configured to: construct a bidirectional interaction path between network traffic data and covariate data based on the unified feature set, model the bidirectional dependency between target traffic and covariates through a cross-attention mechanism, and obtain bidirectional interaction features; The covariate attention fusion module is configured to: input the bidirectional interactive features into the covariate attention fusion module, dynamically calculate the contribution weights of various covariates to the current prediction task, and perform weighted fusion of multi-source covariate features to obtain covariate fusion features; The dynamic graph spatiotemporal modeling module is configured to: dynamically construct an inter-regional adjacency matrix based on the covariate fusion features and the current regional traffic status information, and input the dynamic adjacency matrix and covariate fusion features into a graph convolutional network to extract spatial dependency features, and then extract long-term temporal features through a temporal convolutional network to obtain the final spatiotemporal features; The prediction output module is configured to input the final spatiotemporal features into the multilayer perceptron and output the predicted urban cellular network traffic at future moments.

[0133] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0134] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the urban network traffic prediction method based on covariate interaction and graph augmentation as described in Example 1.

[0135] Example 4 The purpose of this embodiment is to provide an electronic device.

[0136] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the urban network traffic prediction method based on covariate interaction and graph augmentation as described in Embodiment 1.

[0137] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0138] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0139] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is 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 method for predicting urban network traffic based on covariate interaction and graph augmentation, characterized in that, include: Acquire urban cellular network traffic data and multi-source covariate data to construct a multi-source dataset; A linear mapping layer maps data from multiple source datasets to a unified feature space, resulting in a unified feature set. Based on the unified feature set, a bidirectional interaction path between network traffic data and covariate data is constructed. The bidirectional dependency between target traffic and covariates is modeled through a cross-attention mechanism to obtain bidirectional interaction features. The bidirectional interactive features are input into the covariate attention fusion module to dynamically calculate the contribution weights of various covariates to the current prediction task, and the multi-source covariate features are weighted and fused to obtain the covariate fusion features. Based on the covariate fusion features and the current traffic status information of each region, an inter-regional adjacency matrix is ​​dynamically constructed. The dynamic adjacency matrix and the covariate fusion features are then input into a graph convolutional network to extract spatial dependency features. Finally, a temporal convolutional network is used to extract long-term temporal features to obtain the final spatiotemporal features. The final spatiotemporal features are input into a multilayer perceptron, which outputs the predicted urban cellular network traffic for future moments.

2. The urban network traffic prediction method based on covariate interaction and graph augmentation as described in claim 1, characterized in that, The process of acquiring urban cellular network traffic data and multi-source covariate data also includes: Perform quality checks on the acquired data to determine whether there are any missing or abnormal fluctuations in each data point; For outlier data points, the historical statistical information of their location at the current time is used to make preliminary corrections to obtain time-compensated data. Based on time dimension compensation, the neighborhood aggregation features of the spatial location of the data point are obtained, and the time-compensated data is further corrected using the neighborhood aggregation features to obtain the final compensation result. Replace the original abnormal data with the final compensation result.

3. The urban network traffic prediction method based on covariate interaction and graph augmentation as described in claim 1, characterized in that, The multi-source covariate data includes meteorological data, time information, holiday markers, and city point of interest (POI) data.

4. The urban network traffic prediction method based on covariate interaction and graph augmentation as described in claim 1, characterized in that, The process of constructing a bidirectional interaction path between network traffic data and covariate data based on the unified feature set, and obtaining bidirectional interaction features, includes: Construct the first interaction path, using historical traffic features as query vectors and covariate features as key and value vectors. Calculate the degree of attention traffic pays to covariates through cross-attention to obtain the first interaction result from traffic to covariates. A second interaction path is constructed, using covariate features as query vectors and historical traffic features as key and value vectors. The degree of inverse influence of covariates on traffic is calculated through cross-attention, resulting in the second interaction between covariates and traffic. The first interaction result and the second interaction result are fused to obtain a bidirectional fusion feature.

5. The urban network traffic prediction method based on covariate interaction and graph augmentation as described in claim 1, characterized in that, The bidirectional interactive features are input into the covariate attention fusion module to dynamically calculate the contribution weights of various covariates to the current prediction task, and the multi-source covariate features are weighted and fused to obtain the covariate fusion features, including: Calculate the attention scores of various covariates under the current prediction task using learnable attention parameters; The attention scores of covariates across all categories are normalized to obtain the contribution weights of the covariates. The weighted summation of the features of various covariates is performed using the aforementioned contribution weights to obtain the weighted fusion features.

6. The urban network traffic prediction method based on covariate interaction and graph augmentation as described in claim 1, characterized in that, The process of extracting spatially dependent features includes: Obtain the historical traffic coding features of each region at the current moment, and use the feature vector of each region as the state representation of that region at this moment; For any two regions, calculate the cosine similarity between their feature vectors, and use the calculated similarity value as the dynamic connection strength between the two regions at the current time. The connection strength between all pairs of regions together constitutes the initial dynamic adjacency matrix at the current time. Obtain the historical graph structure matrix of the previous moment, and perform a weighted fusion of the initial dynamic adjacency matrix of the current moment with the historical graph structure matrix. Use a preset update coefficient to control the fusion ratio of current information and historical information to obtain the dynamic graph structure matrix of the current moment. The dynamic graph structure matrix at the current moment and the fused features are input into the graph convolutional network. The spatial information of each region and its neighboring regions is aggregated through graph convolution operations, and the spatial features containing dynamic spatial dependencies are output.

7. The urban network traffic prediction method based on covariate interaction and graph augmentation as described in claim 1, characterized in that, Long-term temporal features are extracted using a temporal convolutional network to obtain the final spatiotemporal features, including: The spatial features output by the graph convolutional network are organized according to the time dimension to form a spatial feature sequence containing multiple historical time steps; The spatial feature sequence is input into a temporal convolutional network. Through the causal convolution operation in the temporal convolutional network, the output of the current time step depends only on the input of the current and previous historical time steps. Expanding the temporal receptive field through dilated convolution operations in temporal convolutional networks allows for the capture of long-range temporal dependencies spanning multiple time steps. The temporal features output by the temporal convolutional network and the spatial features output by the graph convolutional network are added and fused element by element to obtain the final spatiotemporal features that simultaneously contain dynamic spatial dependencies and long-term temporal dependencies.

8. A city network traffic prediction system based on covariate interaction and graph augmentation, characterized in that, include: The data acquisition module is configured to: acquire urban cellular network traffic data and multi-source covariate data, and construct a multi-source dataset; The multi-source feature encoding module is configured to map data from a multi-source dataset to a unified feature space through a linear mapping layer, thereby obtaining a unified feature set. The bidirectional covariate interaction module is configured to: construct a bidirectional interaction path between network traffic data and covariate data based on the unified feature set, model the bidirectional dependency between target traffic and covariates through a cross-attention mechanism, and obtain bidirectional interaction features; The covariate attention fusion module is configured to: input the bidirectional interactive features into the covariate attention fusion module, dynamically calculate the contribution weights of various covariates to the current prediction task, and perform weighted fusion of multi-source covariate features to obtain covariate fusion features; The dynamic graph spatiotemporal modeling module is configured to: dynamically construct an inter-regional adjacency matrix based on the covariate fusion features and the current regional traffic status information, and input the dynamic adjacency matrix and covariate fusion features into a graph convolutional network to extract spatial dependency features, and then extract long-term temporal features through a temporal convolutional network to obtain the final spatiotemporal features; The prediction output module is configured to input the final spatiotemporal features into the multilayer perceptron and output the predicted urban cellular network traffic at future moments.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the urban network traffic prediction method based on covariate interaction and graph augmentation as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the urban network traffic prediction method based on covariate interaction and graph augmentation as described in any one of claims 1-7.

Images