Traffic prediction method, device, electronic device and storage medium in wireless communication

Abstract

This application proposes a traffic prediction method, apparatus, electronic device, and storage medium for wireless communication. The method includes: acquiring historical spatiotemporal traffic data of multiple spatial nodes in a target area; extracting multi-scale temporal features of each spatial node based on the historical spatiotemporal traffic data of each spatial node; performing spatial dependency modeling on the multi-scale temporal features to obtain a node embedding vector corresponding to each spatial node; determining a target expert model for the corresponding spatial node from multiple preset candidate expert models based on the node embedding vector corresponding to each spatial node; and predicting the multi-scale temporal features using the target expert model corresponding to each spatial node to obtain the target communication traffic prediction value for multiple spatial nodes. This application can effectively learn multi-scale temporal features in long-term sequences and accurately model the complex heterogeneous interaction relationship of multi-scale temporal features between different spatial nodes, thereby improving the prediction accuracy of spatiotemporal communication traffic.

Description

Technical Field

[0001] This application relates to the field of wireless communication, and more particularly to a traffic prediction method, apparatus, electronic device, and storage medium in wireless communication. Background Technology

[0002] Spatiotemporal prediction of communication traffic is a key task in intelligent network management and resource scheduling. Its goal is to accurately predict future changes in communication traffic across different geographical locations based on historical data. With the development of 5G and future networks, network traffic exhibits significant periodicity and burstiness in the time dimension, and strong regional coupling and topological dependence in the spatial dimension. Therefore, effectively modeling the temporal dependencies and spatial correlations in communication traffic to achieve high-precision spatiotemporal traffic prediction has become an important issue for improving network operating efficiency and user service quality.

[0003] In related technologies, spatial modeling widely employs graph neural networks, such as spatiotemporal graph convolutional networks utilizing predefined graphs or adaptive graph convolutional recurrent networks capable of adaptively learning graph structures; in temporal modeling, the Mamba architecture based on state-space models has attracted considerable attention due to its efficiency. However, these existing technologies suffer from two key drawbacks when processing long-term spatiotemporal data: first, the multi-scale information naturally contained in long-term sequences is difficult to learn effectively and simultaneously; second, the complex heterogeneous interactions between these multi-scale temporal patterns at different spatial nodes cannot be fully modeled. Summary of the Invention

[0004] This application provides a traffic prediction method, apparatus, electronic device, and storage medium for wireless communication, which can effectively learn multi-scale temporal features in long-term sequences and accurately model the complex heterogeneous interaction relationships of multi-scale temporal features between different spatial nodes, thereby improving the prediction accuracy of spatiotemporal communication traffic.

[0005] To achieve the above objectives, a first aspect of this application proposes a traffic prediction method in wireless communication, the method comprising:

[0006] Acquire historical spatiotemporal flow data of multiple spatial nodes in the target area;

[0007] Based on the historical spatiotemporal flow data of each spatial node, extract the multi-scale temporal features of the spatial node;

[0008] Spatial dependency modeling is performed on the multi-scale temporal features to obtain the node embedding vector corresponding to each spatial node.

[0009] Based on the node embedding vector corresponding to each spatial node, a target expert model is determined for the corresponding spatial node from a plurality of preset candidate expert models;

[0010] The multi-scale time features are predicted by the target expert model corresponding to each spatial node, thereby obtaining the target communication traffic prediction values ​​for multiple spatial nodes.

[0011] In some embodiments, extracting multi-scale temporal features of the spatial nodes based on the historical spatiotemporal flow data of each spatial node includes:

[0012] The historical spatiotemporal flow data of each spatial node is projected into a latent space feature tensor using a multilayer perceptron; wherein the latent space feature tensor includes a preset latent space feature dimension.

[0013] Each latent space feature tensor is convolved by multiple one-dimensional convolutional layers with different preset kernel sizes to obtain temporal feature representations at multiple time scales.

[0014] For each spatial node, the temporal feature representations at multiple time scales are stacked to obtain the multi-scale temporal features of the spatial node.

[0015] In some embodiments, acquiring historical spatiotemporal traffic data of multiple spatial nodes in the target area includes:

[0016] Acquire device measurement data reported by all terminal devices within the target area; wherein, the device measurement data includes geographical location data and traffic data;

[0017] The target area is divided into multiple preset geographic grids, and each geographic grid is defined as a spatial node;

[0018] Based on the geographic location data and the traffic data, the historical traffic data of each spatial node in each preset time segment is determined, and the historical spatiotemporal traffic data of each spatial node is obtained.

[0019] In some embodiments, the step of performing spatial dependency modeling on the multi-scale temporal features to obtain the node embedding vector corresponding to each spatial node includes:

[0020] Based on the preset initial node embedding vector of each of the spatial nodes, an adaptive graph adjacency matrix is ​​generated to represent the spatial dependencies between the multiple spatial nodes.

[0021] Based on the adaptive graph adjacency matrix, graph convolution processing is performed on the multi-scale temporal features to aggregate the spatial neighborhood information of the multiple spatial nodes and obtain multi-scale temporal aggregated features.

[0022] The multi-scale temporal aggregation features are fused across scales to obtain the node embedding vector corresponding to each spatial node.

[0023] In some embodiments, the step of performing cross-scale fusion on the multi-scale temporal aggregated features to obtain the node embedding vector corresponding to each spatial node includes:

[0024] The multi-scale temporal aggregation features are projected into a query matrix, a key matrix, and a value matrix, respectively.

[0025] Matrix multiplication is performed on the query matrix and the key matrix, and the multiplication result is scaled and masked to obtain an attention score; wherein, the masking process uses a preset lower triangular mask;

[0026] The attention score is normalized and then multiplied with the value matrix to obtain the attention output features;

[0027] The attention output features and the multi-scale temporal aggregation features are subjected to residual connection and layer normalization to generate fused spatiotemporal features.

[0028] Based on the fused spatiotemporal features, the node embedding vector corresponding to each of the spatial nodes is determined.

[0029] In some embodiments, the plurality of preset candidate expert models are configured to include a shared expert model and a plurality of dedicated expert models. The step of determining a target expert model for a corresponding spatial node from the plurality of preset candidate expert models based on the node embedding vector corresponding to each spatial node includes:

[0030] Based on the node embedding vector, calculate the routing score corresponding to each of the dedicated expert models;

[0031] The dedicated expert model with the highest routing score and the shared expert model are jointly determined as the target expert model.

[0032] In some embodiments, the step of predicting the multi-scale temporal features using the target expert model corresponding to each of the spatial nodes to obtain target communication traffic prediction values ​​for multiple spatial nodes includes:

[0033] The multi-scale temporal features are extracted using the selected dedicated expert model and the shared expert model, respectively, to obtain dedicated features and shared features; wherein the feature extraction process includes:

[0034] The multi-scale temporal features are extended by performing temporal scale expansion processing on the multi-scale temporal features using multiple one-dimensional convolutional layers with different kernel sizes to obtain the extended multi-scale temporal features.

[0035] Multi-scale feature extraction processing is performed on the extended multi-scale temporal features using a preset learnable multi-scale bias.

[0036] The specific features and the shared features are fused to obtain the target fused features;

[0037] The target communication traffic prediction value is obtained by performing a prediction spatial mapping on the target fusion features using a multilayer perceptron.

[0038] In some embodiments, after obtaining the target communication traffic prediction values ​​for the plurality of space nodes, the method further includes:

[0039] Based on a preset joint loss function, the loss value between the predicted target communication traffic value and the corresponding actual traffic value is calculated; wherein, the joint loss function includes mean absolute error loss and causal comparison loss, the mean absolute error loss is used to measure the prediction error, and the causal comparison loss is used to promote pattern decoupling among the multiple candidate expert models;

[0040] Based on the loss value, the multiple candidate expert models are updated.

[0041] To achieve the above objectives, a second aspect of this application provides a traffic prediction device for wireless communication, comprising:

[0042] The acquisition module is used to acquire historical spatiotemporal flow data of multiple spatial nodes in the target area;

[0043] The extraction module is used to extract multi-scale temporal features of each spatial node based on the historical spatiotemporal flow data of each spatial node.

[0044] The modeling module is used to perform spatial dependency modeling on the multi-scale temporal features to obtain the node embedding vector corresponding to each spatial node.

[0045] The determination module is used to determine a target expert model for the corresponding spatial node from a plurality of preset candidate expert models based on the node embedding vector corresponding to each spatial node.

[0046] The prediction module is used to predict the multi-scale time features by using the target expert model corresponding to each of the spatial nodes, so as to obtain the target communication traffic prediction values ​​of multiple spatial nodes.

[0047] To achieve the above objectives, a third aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the traffic prediction method in wireless communication as described in the first aspect.

[0048] To achieve the above objectives, a fourth aspect of the present application provides a storage medium, which is a computer-readable storage medium storing a computer program that, when executed by a processor, implements the traffic prediction method in wireless communication as described in the first aspect.

[0049] This application proposes a traffic prediction method, apparatus, electronic device, and storage medium for wireless communication. First, historical spatiotemporal traffic data of multiple spatial nodes in a target area are acquired. Then, based on the historical spatiotemporal traffic data of each spatial node, multi-scale temporal features of the spatial nodes are extracted. Next, spatial dependency modeling is performed on the multi-scale temporal features to obtain a node embedding vector corresponding to each spatial node. Then, based on the node embedding vector corresponding to each spatial node, a target expert model is determined from multiple preset candidate expert models for the corresponding spatial node. Finally, the multi-scale temporal features are predicted using the target expert model corresponding to each spatial node to obtain the target communication traffic prediction values ​​for multiple spatial nodes.

[0050] In contrast to the problems in the background art, the traffic prediction method in wireless communication provided in this application first extracts multi-scale temporal features from historical spatiotemporal traffic data, providing a feature foundation that simultaneously includes multiple time-scale patterns for subsequent processing. Then, by performing spatial dependency modeling on the multi-scale temporal features, a node embedding vector representing the inherent attributes of each spatial node can be learned and generated. Next, a target expert model is determined for each spatial node based on the node embedding vector. This approach assigns nodes with different traffic pattern characteristics to different expert models for processing according to the inherent attributes of the spatial nodes, thereby achieving adaptive modeling of the heterogeneity of spatial nodes. Finally, the multi-scale temporal features are predicted using the target expert model determined by its corresponding node embedding vector, enabling the prediction task to be completed using the expert model best suited for processing a specific pattern. Through the above methods, this solution can effectively learn multi-scale temporal features in long-term sequences and accurately model the complex heterogeneous interaction relationships of multi-scale temporal features among different spatial nodes, thereby improving the prediction accuracy of spatiotemporal communication traffic.

[0051] Other features and advantages of this application will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description

[0052] Figure 1 A flowchart of a traffic prediction method in wireless communication provided in an embodiment of this application;

[0053] Figure 2 A flowchart of a traffic prediction method in wireless communication provided in another embodiment of this application;

[0054] Figure 3 A flowchart of a traffic prediction method in wireless communication provided in another embodiment of this application;

[0055] Figure 4 A flowchart of a traffic prediction method in wireless communication provided in another embodiment of this application;

[0056] Figure 5 A flowchart of a traffic prediction method in wireless communication provided in another embodiment of this application;

[0057] Figure 6 A flowchart of a traffic prediction method in wireless communication provided in another embodiment of this application;

[0058] Figure 7 A flowchart of a traffic prediction method in wireless communication provided in another embodiment of this application;

[0059] Figure 8 A schematic diagram of a traffic prediction device in wireless communication provided in an embodiment of this application;

[0060] Figure 9 This is a schematic diagram of the hardware structure of an electronic device provided in another embodiment of this application. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0062] It should be noted that although functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart.

[0063] Unless otherwise defined, 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 application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0064] Spatiotemporal prediction of communication traffic is a key task in intelligent network management and resource scheduling. Its goal is to accurately predict future changes in communication traffic across different geographical locations based on historical data. With the development of 5G and future networks, network traffic exhibits significant periodicity and burstiness in the time dimension, and strong regional coupling and topological dependence in the spatial dimension. Therefore, effectively modeling the temporal dependencies and spatial correlations in communication traffic to achieve high-precision spatiotemporal traffic prediction has become an important issue for improving network operating efficiency and user service quality.

[0065] In related technologies, spatial modeling widely employs graph neural networks, such as spatiotemporal graph convolutional networks utilizing predefined graphs or adaptive graph convolutional recurrent networks capable of adaptively learning graph structures; in temporal modeling, the Mamba architecture based on state-space models has attracted considerable attention due to its efficiency. However, these existing technologies suffer from two key drawbacks when processing long-term spatiotemporal data: first, the multi-scale information naturally contained in long-term sequences is difficult to learn effectively and simultaneously; second, the complex heterogeneous interactions between these multi-scale temporal patterns at different spatial nodes cannot be fully modeled.

[0066] To address the aforementioned issues, the traffic prediction method for wireless communication provided in this application first extracts multi-scale temporal features from historical spatiotemporal traffic data, providing a feature foundation that simultaneously encompasses multiple time-scale patterns for subsequent processing. Then, by performing spatial dependency modeling on the multi-scale temporal features, a node embedding vector representing the inherent attributes of each spatial node can be learned and generated. Next, a target expert model is determined for each spatial node based on its node embedding vector. This approach assigns nodes with different traffic pattern characteristics to different expert models for processing according to the inherent attributes of the spatial nodes, thereby achieving adaptive modeling of spatial node heterogeneity. Finally, the multi-scale temporal features are predicted using the target expert model determined by its corresponding node embedding vector, enabling the prediction task to be completed using the expert model best suited for processing a specific pattern. Through this method, this solution can effectively learn multi-scale temporal features in long-term sequences and accurately model the complex heterogeneous interaction relationships of multi-scale temporal features among different spatial nodes, thereby improving the prediction accuracy of spatiotemporal communication traffic.

[0067] The following will further describe a traffic prediction method, apparatus, electronic device, and storage medium in wireless communication provided by embodiments of this application. First, the traffic prediction method in wireless communication in the embodiments of this application will be described in detail.

[0068] Reference Figure 1 This is an optional flowchart of a traffic prediction method in wireless communication provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps 101 to 105. It is also understood that this embodiment... Figure 1 The order of steps 101 to 105 is not specifically limited. The order of steps can be adjusted or some steps can be reduced or added according to actual needs.

[0069] Step 101: Obtain historical spatiotemporal flow data of multiple spatial nodes in the target area.

[0070] Step 102: Extract multi-scale temporal features of each spatial node based on its historical spatiotemporal flow data.

[0071] Step 103: Perform spatial dependency modeling for multi-scale temporal features to obtain the node embedding vector corresponding to each spatial node.

[0072] Step 104: Based on the node embedding vector corresponding to each spatial node, determine the target expert model for the corresponding spatial node from multiple preset candidate expert models.

[0073] Step 105: Predict the multi-scale time features using the target expert model corresponding to each spatial node to obtain the target communication traffic prediction values ​​for multiple spatial nodes.

[0074] In step 101 of some embodiments, a "spatial node" can be a basic spatial unit after the target area has been divided, such as a geographic grid or a base station coverage area. "Historical spatiotemporal traffic data" refers to communication traffic information recorded on these spatial nodes over a continuous period of time in the past. This data provides the necessary input for the subsequent model to learn the spatiotemporal evolution of traffic.

[0075] Please see Figure 2 In some embodiments, step 101 may include, but is not limited to, steps 201 to 203.

[0076] Step 201: Obtain device measurement data reported by all terminal devices within the target area.

[0077] Step 202: Divide the target area into multiple preset geographic grids, and define each geographic grid as a spatial node.

[0078] Step 203: Based on the geographic location data and traffic data, determine the historical traffic data of each spatial node in each preset time segment, and obtain the historical spatiotemporal traffic data of each spatial node.

[0079] In step 201 of some embodiments, device measurement data reported by all terminal devices within the target area covered by the edge node (such as a base station) of the communication network is first obtained. Here, "terminal device" typically refers to User Equipment (UE), and the "device measurement data" reported by it technically corresponds to Measurement Reports (MR). This raw data is high-frequency and massive, containing crucial information necessary for subsequent processing. Specifically, it includes geographic location data that identifies the device's location, such as latitude and longitude coordinates, and traffic data that reflects network usage, such as downlink throughput. This step is the data source for the entire prediction process.

[0080] In step 202 of some embodiments, in order to convert the continuous geographic space into a discrete structure that can be processed by the model, the target area needs to be spatially divided. Specifically, this step divides the map data of the entire target area into multiple pre-defined geographic rasters of uniform size and without overlap. For example, the area can be divided into a 10-meter × 10-meter grid. After rasterization, this embodiment logically defines each independent geographic raster as a "spatial node," providing a foundation for subsequent spatial dependency modeling using methods such as graph neural networks. This allows complex geographic spatial relationships to be abstracted into a graph structure composed of nodes and edges.

[0081] In step 203 of some embodiments, after discretizing the spatial dimension, the temporal dimension also needs to be discretized, that is, continuous time is divided into multiple preset time segments, for example, with 5 minutes as a time granularity. Then, based on the geographic location data and traffic data obtained in step 201, each device measurement data is assigned to a specific geographic raster that it falls into within a specific time segment. By fusing the traffic data of all terminal devices falling into the same geographic raster and the same time segment, the historical traffic data of that spatial node within that time segment can be determined. By traversing all spatial nodes and all historical time segments and repeating this process, the final data obtained is the well-formatted historical spatiotemporal traffic data of each spatial node that can be directly used by subsequent models.

[0082] Through steps 201 to 203 above, this embodiment can transform massive, unordered raw measurement data into a structured, discrete spatiotemporal data format. Specifically, spatial rasterization and temporal slicing provide a unified data view for subsequent spatiotemporal modeling. Geographic rasters are defined as spatial nodes, allowing continuous geographic space to be represented by a discrete graph structure, which is the foundation for spatial modeling operations such as graph convolution. Finally, by performing traffic statistics and aggregation on each spatiotemporal unit, data redundancy and noise are effectively reduced, enabling a macroscopic reflection of key information on traffic dynamics in a specific region within a specific time period. This provides high-quality, uniformly formatted data input for subsequent models to perform high-precision spatiotemporal traffic prediction.

[0083] In step 102 of some embodiments, since communication traffic exhibits both long-term periodicity (such as daily or weekly variations) and short-term burstiness (such as minute-level fluctuations) over time, this step aims to process the raw historical data to explicitly represent these time patterns at different scales at the feature level. The resulting "multi-scale time features" are a richer and more comprehensive feature representation that can simultaneously characterize both long-term trends and short-term changes in traffic.

[0084] Please see Figure 3 In some embodiments, step 102 may include, but is not limited to, steps 301 to 303.

[0085] Step 301: Project the historical spatiotemporal flow data of each spatial node into a latent space feature tensor using a multilayer perceptron; wherein the latent space feature tensor includes a preset latent space feature dimension.

[0086] Step 302: Perform convolution processing on each latent space feature tensor through multiple one-dimensional convolutional layers with different preset kernel sizes to obtain temporal feature representations at multiple time scales.

[0087] Step 303: For each spatial node, stack the temporal feature representations at multiple time scales to obtain the multi-scale temporal features of the spatial node.

[0088] In step 301 of some embodiments, after obtaining the representation as an input tensor After processing the historical spatiotemporal traffic data, it is first processed through an input mapping layer composed of a multi-layer perceptron (MLP). The multi-layer perceptron here is a type of neural network that projects the input tensor through linear layers. Transform into a high-dimensional latent space feature tensor , where T is the historical time step, N is the number of spatial nodes, C is the input feature dimension, and d is the preset latent space feature dimension, which is usually higher in dimension.

[0089] In step 302 of some embodiments, in order to decompose patterns of different scales in the time dimension, this embodiment employs multiple (e.g., q) preset one-dimensional convolutional layers with different kernel sizes. We then perform parallel convolution processing on the latent space feature tensor H obtained in the previous step. For each scale q, the corresponding convolution operation can be represented by the following formula:

[0090]

[0091] in, This represents the kernel size used in the q-th convolutional layer. Different kernel sizes have different temporal receptive fields; for example, larger kernels... Capable of capturing long-term trends, while smaller This approach focuses more on localized, sudden fluctuations. By executing the convolution operation represented by this formula q times in parallel, q sets of temporal feature representations representing different time scales can be obtained. .

[0092] In step 303 of some embodiments, after obtaining multiple sets of time feature representations corresponding to different time scales, this step integrates these feature representations. Specifically, this step integrates the q sets of time feature representations obtained in step 302. Stacking along a new dimension forms a higher-dimensional, structurally unified tensor. This final tensor represents the "multi-scale temporal features" of spatial nodes, and its dimension can be represented as... This operation preserves and integrates all the decomposed and extracted time information at different scales into a unified data structure, so that subsequent model layers can simultaneously acquire and utilize time information at all scales.

[0093] Through steps 301 to 303 described above, this embodiment effectively transforms the original spatiotemporal flow data into a deeper feature representation that is more conducive to model learning and includes multi-scale time dimensions. Specifically, the multilayer perceptron projection in step 301 elevates the data to a high-dimensional latent space, enhancing the expressive power of the features. Step 302 achieves effective decomposition of the time series by using multiple one-dimensional convolutional kernels of different sizes, which can simultaneously capture and separate various patterns in the data, such as long-term regularities and short-term sudden fluctuations. Finally, step 303, through a stacking operation, completely preserves and integrates the separated information at different scales into a unified data structure. This series of operations provides high-quality feature inputs that have decoupled information at different time scales for subsequent spatial modeling and prediction steps, thereby solving the problem of difficulty in simultaneously learning multi-scale features in the background technology.

[0094] In step 103 of some embodiments, the purpose of this step is to learn and quantify the interrelationships and inherent properties between different spatial nodes. Through spatial dependency modeling, complex network relationships can be analyzed and captured as to how traffic changes at one node are affected by other nodes. The resulting "node embedding vector" is a numerical vector that represents the unique static characteristics of each spatial node, including its geographical or functional attributes.

[0095] Please see Figure 4 In some embodiments, step 103 may include, but is not limited to, steps 401 to 403.

[0096] Step 401: Based on the preset initial node embedding vector of each spatial node, generate an adaptive graph adjacency matrix to represent the spatial dependencies between multiple spatial nodes.

[0097] Step 402: Based on the adaptive graph adjacency matrix, perform graph convolution processing on the multi-scale temporal features to aggregate the spatial neighborhood information of multiple spatial nodes and obtain multi-scale temporal aggregated features.

[0098] Step 403: Perform cross-scale fusion on the multi-scale temporal aggregated features to obtain the node embedding vector corresponding to each spatial node.

[0099] In step 401 of some embodiments, to model the interdependencies between spatial nodes, this method first generates an adaptive graph adjacency matrix. The core of this step lies in allowing the model to learn the strength of associations between nodes from the data, without relying on any predefined, fixed geographic topology information. Specifically, this step pre-sets a learnable "initial node embedding vector" for each spatial node, which is represented as a parameter matrix in the model. By calculating the pairwise similarity between these embedding vectors, an adaptive graph adjacency matrix A that represents the functional associations between nodes can be generated. In practical applications, its normalized form is usually used, which can be obtained through the following formula:

[0100]

[0101] Where D is the degree matrix. This process enables the model to capture spatial dependencies that are not intuitive but truly exist.

[0102] In step 402 of some embodiments, after constructing the adaptive graph adjacency matrix, graph convolution processing can be used to aggregate spatial information. This step applies the spatial relationships represented by the adaptive graph adjacency matrix generated in the previous step to "multi-scale temporal features". For each time scale, the graph convolution operation enables each spatial node to aggregate the feature information of its neighboring nodes in the adaptive graph, thereby allowing the node's representation to be integrated into the context of its spatial environment. A specific implementation of this operation can be approximated by the following formula:

[0103]

[0104] in, It is a specific scale feature of the input. It is the output after spatial information aggregation, namely "multi-scale temporal aggregation feature". It is a low-rank parameter used to reduce computational complexity.

[0105] In step 403 of some embodiments, after spatial information aggregation is completed for temporal features at different scales, cross-scale fusion is performed on the multi-scale temporal aggregated features. The purpose of this step is to effectively integrate feature information that is separated at different scales and already contains spatial context. A fusion rule can be designed to ensure that long-period, coarse-grained features can effectively guide short-period, fine-grained features, thereby avoiding mutual interference between information at different scales. After this cross-scale fusion step, the model finally learns and determines a "node embedding vector" that can accurately represent each spatial node in a multi-dimensional spatiotemporal context.

[0106] Please see Figure 5 In some embodiments, step 403 may include, but is not limited to, steps 501 to 505.

[0107] Step 501: Project the multi-scale temporal aggregation features into a query matrix, a key matrix, and a value matrix, respectively.

[0108] Step 502: Perform matrix multiplication on the query matrix and the key matrix, and perform scaling and masking operations on the multiplication result to obtain attention scores.

[0109] Step 503: Normalize the attention scores and perform matrix multiplication with the value matrix to obtain attention output features.

[0110] Step 504: Perform residual connection and layer normalization on the attention output features and the multi-scale temporal aggregation features to generate fused spatio-temporal features.

[0111] Step 505: Based on the fused spatio-temporal features, determine the node embedding vectors corresponding to each spatial node.

[0112] In step 501 of some embodiments, for attention calculation, first, the input multi-scale temporal aggregation features ( ) need to be linearly transformed. Specifically, through three linear projection layers, the multi-scale temporal aggregation features are respectively transformed into three different matrices, namely the query matrix (Query, Q), the key matrix (Key, K), and the value matrix (Value, V). This transformation process can be expressed as:

[0113]

[0114] This operation maps the original features to different subspaces, laying a foundation for subsequent calculation of feature correlations (through Q and K) and extraction of feature content (through V).

[0115] In step 502 of some embodiments, after obtaining the query, key, and value matrices, this step calculates the relationship between them to obtain attention scores. This process first performs matrix multiplication on the query matrix Q and the key matrix K to calculate the original correlation scores between different scale features. Then, for the stability of the training process, usually, this result is scaled, for example, divided by a scaling factor . The most crucial operation in this step is the masking operation, that is, adding a preset "lower triangular mask" (M) to the above result. The specific implementation of this mask matrix M can be that when the scale index i < j, the corresponding value is negative infinity, and the other positions are 0. This mathematically makes the features of any scale i unable to focus on the features of a finer scale j than it, thus forcibly implementing the unidirectional information flow from coarse-grained features to fine-grained features. The process formula can be expressed as:

[0116] In step 503 of some embodiments, in order to transform the attention scores obtained in the previous step into usable weights and extract information accordingly, this step first applies a normalization function, typically the Softmax function, to the masked attention scores. This operation transforms the scores into a probability distribution that sums to 1 along a specific dimension, i.e., attention weights, which quantitatively represent the scales of features that the model should pay attention to and to what extent during the fusion process. Subsequently, this normalized attention weight is multiplied by the value matrix V. This operation is functionally equivalent to a weighted summation, which selectively extracts and aggregates information from the value matrix V according to the attention weights to obtain the "attention output features" (…). The above calculation formula can be expressed as:

[0117]

[0118] In step 504 of some embodiments, to ensure that this deep attention module can be trained stably and to avoid performance degradation during information transmission, this embodiment performs a residual connection between the "attention output features" obtained in step 503 and the original input "multi-scale temporal aggregation features" of this fusion process. This allows information to propagate directly across layers, effectively alleviating the gradient vanishing problem in deep networks. Next, the summed result undergoes another layer normalization process to standardize the distribution of features at each layer, further stabilizing the training process and accelerating model convergence. The final output of this step is the "fused spatiotemporal features" (…). The formula can be expressed as:

[0119]

[0120] In step 505 of some embodiments, based on the "fused spatiotemporal features" obtained in the previous step, which have deeply fused multi-scale spatiotemporal information, this embodiment finally determines the "node embedding vector corresponding to each spatial node". The node embedding vector output in this step can be regarded as the final form of the initial node embedding vector after sufficient learning and optimization through the entire spatial dependency modeling and cross-scale fusion process.

[0121] By executing steps 501 to 505 above, this embodiment, through QKV projection and attention computation, enables the model to adaptively learn the mutual importance between features at different time scales. The core lies in the application of the lower triangular mask in step 502, which forces a unidirectional information flow from coarse to fine, ensuring that long-period patterns can serve as effective prior knowledge to guide the understanding of short-term patterns, while avoiding information leakage and interference, thus solving the problem of disordered interaction of multi-scale information. Furthermore, the introduction of residual connections and layer normalization in step 504 achieves multi-feature fusion. The final node embedding vector is a node representation that has undergone multi-scale spatiotemporal context association, greatly improving the discriminative power and effectiveness of the features.

[0122] In step 104 of some embodiments, this embodiment employs a hybrid expert approach, namely, pre-setting a set of "candidate expert models" with different processing capabilities. This step is an intelligent routing or selection process that uses the "node embedding vector" obtained in the previous step, which characterizes the node's features, as a criterion to assign one or more "target expert models" best suited to process its data pattern for each spatial node. This step enables the method to adaptively and differentiate the processing of spatial nodes with different characteristics.

[0123] Please see Figure 6 In some embodiments, multiple preset candidate expert models are set to include a shared expert model and multiple dedicated expert models. Step 104 may include, but is not limited to, steps 601 to 602.

[0124] Step 601: Calculate the routing score for each dedicated expert model based on the node embedding vector.

[0125] Step 602: Select the dedicated expert model and the shared expert model with the highest routing score as the target expert model.

[0126] In step 601 of some embodiments, in order to match the most suitable dedicated expert model to each spatial node, this embodiment first needs to calculate the routing score of each dedicated expert model for that node. The basis for calculating the routing score is not the dynamically changing time series itself, but the "node embedding vector" (which can be represented as...) obtained in the previous step (such as step 103) that can characterize the static inherent attributes of the spatial node. Specifically, a gating network will control the node's... As input, and for each candidate specialized expert model (e.g., K in total), a scalar value is output, which is the "routing score" corresponding to that specialized expert. The routing score for the k-th specialized expert can be calculated as follows: ,in This is the corresponding gating function. In some specific embodiments, to enhance the robustness of the routing process, a small random noise rk can be added to the calculated score.

[0127] In step 602 of some embodiments, after calculating the routing scores of all dedicated expert models for the current spatial node, this step will select the target expert model for the node based on the scores. The specific selection strategy is to select the dedicated expert model with the highest routing score, and this selection process can be expressed by the following formula:

[0128]

[0129] Here, g is the index of the selected dedicated expert model. The final "target expert model" determined in this step is a combination consisting of two parts: the first part is the dedicated expert model selected by the routing score that is best suited to handle the node-specific patterns; the second part is the "shared expert model" used by all spatial nodes. Subsequent prediction steps will utilize both expert models simultaneously.

[0130] By executing steps 601 to 602 above, this embodiment elaborates in detail an adaptive routing based on a Mixture-of-Experts (MoE) architecture. Instead of making routing decisions based on dynamic, time-varying time-series data itself, it utilizes the node embedding vectors input in step 601, which characterize the static inherent attributes of spatial nodes, as the routing basis. This makes the routing decision smoother and more stable, avoiding the problem of frequent jumps in expert selection due to minor fluctuations in the input sequence. The selection mechanism in step 602 ensures that the data of each spatial node is refined by a dedicated expert best suited to handle its specific pattern. Simultaneously, the existence of shared experts guarantees that all nodes can obtain general, common pattern information. This combination of shared and dedicated experts enhances the model's expressive and generalization capabilities, effectively solving the problem in the background art of modeling complex heterogeneous interactions between different spatial nodes.

[0131] In step 105 of some embodiments, the multi-scale temporal features extracted in the preceding steps are fed into the "target expert model" determined for the node in step 104. Because these expert models are tailored to the node's characteristics, they are able to perform more targeted, in-depth temporal pattern analysis and inference of the input features. Finally, the model outputs a prediction of communication traffic for one or more future time steps, i.e., the "target communication traffic prediction value."

[0132] In some embodiments, step 105 may include, but is not limited to:

[0133] Multi-scale temporal features are extracted using the selected dedicated expert model and shared expert model respectively, resulting in dedicated features and shared features. The feature extraction process includes: performing temporal scale expansion processing on the multi-scale temporal features through multiple one-dimensional convolutional layers with different kernel sizes to obtain expanded multi-scale temporal features.

[0134] Multi-scale feature extraction is performed on the extended multi-scale temporal features using a preset learnable multi-scale bias.

[0135] The target fused feature is obtained by fusing specific features and shared features.

[0136] By performing a prediction spatial mapping on the target fusion features using a multilayer perceptron, the predicted value of the target communication traffic is obtained.

[0137] First, this method uses a selected dedicated expert model and a shared expert model to extract features from the "multi-scale temporal features" obtained in the previous steps in parallel. The input feature data stream is copied here and simultaneously fed into two expert models with identical structures but independent parameters. After a series of internal processing by each expert model, they output two different types of features: "dedicated features" output by the dedicated expert model, which can reflect the unique patterns of the spatial node, and "shared features" output by the shared expert model, which can reflect the common patterns of all spatial nodes.

[0138] In the feature extraction process described above, each expert model internally performs the same set of processing procedures. This process first includes a temporal scale expansion process. This process uses a set of one-dimensional convolutional layers with different kernel sizes to further process the input multi-scale temporal features, thereby expanding the model's receptive field and enabling it to capture longer-term dependencies. Subsequently, multi-scale feature extraction is performed on the expanded multi-scale temporal features using a sequence processing method based on a state-space model (SSM). This is achieved by inputting parameters (specifically, selectivity parameters) into the method's internal parameters. In this paper, a "preset learnable multiscale bias" (Ω) is introduced, which can be influenced by the following formula:

[0139]

[0140] This bias enables a single processing model to effectively distinguish and identify spliced ​​features of different scales and to extract patterns from them independently and in a targeted manner, thus solving the problem that traditional sequence models have difficulty in processing multi-scale information.

[0141] After obtaining the "specific features" and "shared features" respectively, this method fuses these two types of features to obtain a "target fused feature" that contains both common and specific information. In some specific embodiments, this feature fusion can be achieved through a simple element-wise addition operation, the process of which can be represented as follows:

[0142]

[0143] in, The output represents the output of a dedicated expert model. The output of the shared expert model is represented by z, which is the "target fusion feature".

[0144] Finally, after obtaining the "target fusion features," this method maps them to a prediction space using an output mapping layer composed of a multi-layer perceptron (MLP). This step transforms the high-dimensional, abstract fusion features into a final output with explicit physical meaning. This output mapping layer maps the dimensions of the fusion features to dimensions consistent with the prediction target, ultimately obtaining the "target communication traffic prediction value" for one or more future time steps.

[0145] By performing the above steps, this embodiment, through the combination of dedicated and shared experts, can simultaneously capture the unique patterns of specific spatial nodes and the common general rules shared by all nodes, greatly enhancing the modeling capability for spatially heterogeneous data. Its core advantage lies in the feature extraction process within the expert model: this process, through time scale expansion and the introduction of learnable multi-scale bias, effectively solves the problem that traditional sequence models struggle to handle multi-scale information simultaneously. This multi-scale bias enables the model to independently and purposefully analyze and extract features at different time scales within a unified framework. Finally, through feature fusion and output mapping, these deeply and refined features are transformed into high-precision prediction results.

[0146] Please see Figure 7 In some embodiments, after step 105, steps 701 to 702 may also be included, but are not limited to.

[0147] Step 701: Based on the preset joint loss function, calculate the loss value between the predicted target communication traffic value and the corresponding actual traffic value.

[0148] Step 702: Update the multiple candidate expert models based on the loss values.

[0149] In step 701 of some embodiments, after the model generates the "target communication traffic prediction value," in order to quantitatively evaluate the model's performance and guide its parameter optimization, this step calculates the loss value between the predicted value and the corresponding actual traffic value based on a preset joint loss function. This joint loss function is an important component of the present invention; it consists of two key parts and can be expressed by the following total loss function formula:

[0150]

[0151] Where L is the total loss, and λ is a hyperparameter balancing the importance of the two loss components. The two components of this joint loss function are explained in detail below. The first part is the mean absolute error loss, expressed as... Its main function is to directly measure the numerical difference between the predicted value and the actual value, serving as a core indicator of the model's prediction accuracy. The formula is as follows:

[0152]

[0153] The second part is the causal comparison loss, expressed as... Its design aims to promote pattern decoupling among multiple candidate expert models, incentivizing different specialized experts to learn differentiated patterns. To achieve this, the loss first defines a special "asymmetric similarity measure":

[0154]

[0155] Here, γ1 and γ2 are two adjustable hyperparameters used to control the decay rate of similarity scores as scale differences change. Based on this similarity metric, the causal contrastive loss can be calculated using the following formula, with the goal of bringing positive sample pairs closer together and pushing negative sample pairs further apart:

[0156]

[0157] in, This represents the set of data samples that are routed to the k-th expert in a training batch. This represents "anchor samples", which are from the set. The feature representation of the i-th sample at the p-th scale. This represents the set of "positive samples" corresponding to the aforementioned "anchor samples." In contrastive learning, positive samples typically refer to samples that should be considered semantically or pattern-wise similar to the anchor samples. This represents the set of "negative samples" that correspond to the "anchor sample," i.e., samples that should be considered dissimilar in pattern. It is a positive sample taken from the set of positive samples. It is a negative sample taken from the set of negative samples, and θ is the temperature hyperparameter.

[0158] In step 702 of some embodiments, after calculating the total loss value L, this step adjusts and optimizes all learnable parameters in the model based on this loss value. The ultimate goal of this step is to minimize the loss value. In deep learning practice, this process is usually achieved by combining gradient optimization algorithms (such as Adam, SGD, etc.) with a backpropagation mechanism. Starting with the loss value, its gradient is backpropagated to every learnable parameter of the model, including the weights of all convolutional and linear layers, the learnable biases in multi-scale Mamba, and the initial node embedding vectors in adaptive graph networks. By iteratively performing this step, all parameters of the model are fine-tuned in the direction that continuously reduces the total loss, thereby enabling the model to reach an optimal state in terms of both prediction accuracy and internal expert function differentiation.

[0159] By executing steps 701 to 702 above, this embodiment ensures that the core objective of model training—continuously improving prediction accuracy and bringing predicted values ​​closer to true values—through the mean absolute error loss in step 701. On the other hand, by introducing an innovative causal contrast loss, this method guides different specialized experts to learn independent and differentiated data patterns through a training mechanism. This effectively avoids the problem of expert redundancy, ensuring that each expert can become a model for processing specific patterns, thereby maximizing the advantages of the hybrid expert architecture in dealing with complex and heterogeneous data. Finally, step 702 optimizes the model overall based on this joint loss value, improving overall prediction accuracy.

[0160] Please see Figure 8 This application also provides a traffic prediction device for wireless communication, which can implement the above-mentioned traffic prediction method for wireless communication, including:

[0161] The acquisition module is used to acquire historical spatiotemporal flow data of multiple spatial nodes in the target area;

[0162] The extraction module is used to extract multi-scale temporal features of spatial nodes based on the historical spatiotemporal flow data of each spatial node.

[0163] The modeling module is used to perform spatial dependency modeling for multi-scale temporal features and obtain the node embedding vector corresponding to each spatial node.

[0164] The determination module is used to determine the target expert model for the corresponding spatial node from multiple preset candidate expert models based on the node embedding vector corresponding to each spatial node.

[0165] The prediction module is used to predict multi-scale temporal features using the target expert model corresponding to each spatial node, and obtain the target communication traffic prediction values ​​for multiple spatial nodes.

[0166] This application proposes a traffic prediction method, apparatus, electronic device, and storage medium for wireless communication. First, historical spatiotemporal traffic data of multiple spatial nodes in a target area are acquired. Then, based on the historical spatiotemporal traffic data of each spatial node, multi-scale temporal features of the spatial nodes are extracted. Next, spatial dependency modeling is performed on the multi-scale temporal features to obtain a node embedding vector corresponding to each spatial node. Then, based on the node embedding vector corresponding to each spatial node, a target expert model is determined from multiple preset candidate expert models for the corresponding spatial node. Finally, the multi-scale temporal features are predicted using the target expert model corresponding to each spatial node to obtain the target communication traffic prediction values ​​for multiple spatial nodes.

[0167] In contrast to the problems in the background art, the traffic prediction method in wireless communication provided in this application first extracts multi-scale temporal features from historical spatiotemporal traffic data, providing a feature foundation that simultaneously includes multiple time-scale patterns for subsequent processing. Then, by performing spatial dependency modeling on the multi-scale temporal features, a node embedding vector representing the inherent attributes of each spatial node can be learned and generated. Next, a target expert model is determined for each spatial node based on the node embedding vector. This approach assigns nodes with different traffic pattern characteristics to different expert models for processing according to the inherent attributes of the spatial nodes, thereby achieving adaptive modeling of the heterogeneity of spatial nodes. Finally, the multi-scale temporal features are predicted using the target expert model determined by its corresponding node embedding vector, enabling the prediction task to be completed using the expert model best suited for processing a specific pattern. Through the above methods, this solution can effectively learn multi-scale temporal features in long-term sequences and accurately model the complex heterogeneous interaction relationships of multi-scale temporal features among different spatial nodes, thereby improving the prediction accuracy of spatiotemporal communication traffic.

[0168] This application also provides an electronic device, including:

[0169] At least one memory;

[0170] At least one processor;

[0171] At least one program;

[0172] The program is stored in a memory, and the processor executes the at least one program to implement the traffic prediction method in wireless communication described above. The electronic device can be any smart terminal, including mobile phones, tablets, personal digital assistants (PDAs), and in-vehicle computers.

[0173] Please see Figure 9 , Figure 9 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes:

[0174] The processor 901 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0175] The memory 902 can be implemented in the form of ROM (Read-Only Memory), static storage device, dynamic storage device, or RAM (Random Access Memory). The memory 902 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 902 and called by the processor 901 to execute the traffic prediction method in the wireless communication embodiments of this application.

[0176] The input / output interface 903 is used to implement information input and output;

[0177] The communication interface 904 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0178] Bus 905 transmits information between various components of the device (e.g., processor 901, memory 902, input / output interface 903, and communication interface 904);

[0179] The processor 901, memory 902, input / output interface 903, and communication interface 904 are connected to each other within the device via bus 905.

[0180] This application embodiment also provides a storage medium, which is a computer-readable storage medium, storing a computer program that, when executed by a processor, implements the traffic prediction method in the above-described wireless communication.

[0181] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0182] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of this application, and do not constitute a limitation on the technical solutions provided in this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in this application are also applicable to similar technical problems.

[0183] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0184] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0185] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0186] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0187] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c. One can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be a single or multiple.

[0188] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. The coupling or direct coupling or communication connection between the shown or discussed units may be through some interfaces, or the indirect coupling or communication connection between the apparatus or units may be electrical, mechanical, or other forms.

[0189] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0190] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0191] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0192] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A traffic prediction method in wireless communication, characterized in that, The method includes: Acquire historical spatiotemporal flow data of multiple spatial nodes in the target area; Based on the historical spatiotemporal flow data of each spatial node, extract the multi-scale temporal features of the spatial node; Spatial dependency modeling is performed on the multi-scale temporal features to obtain the node embedding vector corresponding to each spatial node. This includes: generating an adaptive graph adjacency matrix representing the spatial dependencies between multiple spatial nodes based on a preset initial node embedding vector for each spatial node; performing graph convolution processing on the multi-scale temporal features based on the adaptive graph adjacency matrix to aggregate the spatial neighborhood information of the multiple spatial nodes, obtaining multi-scale temporal aggregated features; projecting the multi-scale temporal aggregated features into a query matrix, a key matrix, and a value matrix; performing matrix multiplication on the query matrix and the key matrix, and scaling and masking the multiplication result to obtain an attention score; wherein the masking process uses a preset lower triangular mask; normalizing the attention score and performing matrix multiplication with the value matrix to obtain an attention output feature; performing residual connection and layer normalization processing on the attention output feature and the multi-scale temporal aggregated features to generate fused spatiotemporal features; and determining the node embedding vector corresponding to each spatial node based on the fused spatiotemporal features. Based on the node embedding vector corresponding to each spatial node, a target expert model is determined for the corresponding spatial node from a plurality of preset candidate expert models; The multi-scale time features are predicted by the target expert model corresponding to each spatial node, thereby obtaining the target communication traffic prediction values ​​for multiple spatial nodes.

2. The method according to claim 1, characterized in that, The extraction of multi-scale temporal features of each spatial node based on its historical spatiotemporal flow data includes: The historical spatiotemporal flow data of each spatial node is projected into a latent space feature tensor using a multilayer perceptron; wherein the latent space feature tensor includes a preset latent space feature dimension. Each latent space feature tensor is convolved by multiple one-dimensional convolutional layers with different preset kernel sizes to obtain temporal feature representations at multiple time scales. For each spatial node, the temporal feature representations at multiple time scales are stacked to obtain the multi-scale temporal features of the spatial node.

3. The method according to claim 1, characterized in that, The acquisition of historical spatiotemporal flow data of multiple spatial nodes in the target area includes: Acquire device measurement data reported by all terminal devices within the target area; wherein, the device measurement data includes geographical location data and traffic data; The target area is divided into multiple preset geographic grids, and each geographic grid is defined as a spatial node; Based on the geographic location data and the traffic data, the historical traffic data of each spatial node in each preset time segment is determined, and the historical spatiotemporal traffic data of each spatial node is obtained.

4. The method according to claim 1, characterized in that, The plurality of preset candidate expert models are configured to include one shared expert model and multiple dedicated expert models. The step of determining the target expert model for the corresponding spatial node from the plurality of preset candidate expert models based on the node embedding vector corresponding to each spatial node includes: Based on the node embedding vector, calculate the routing score corresponding to each of the dedicated expert models; The dedicated expert model with the highest routing score and the shared expert model are jointly determined as the target expert model.

5. The method according to claim 4, characterized in that, The step of predicting the multi-scale temporal features using the target expert model corresponding to each of the spatial nodes to obtain target communication traffic prediction values ​​for multiple spatial nodes includes: The multi-scale temporal features are extracted using the selected dedicated expert model and the shared expert model, respectively, to obtain dedicated features and shared features; wherein the feature extraction process includes: The multi-scale temporal features are extended by performing temporal scale expansion processing on the multi-scale temporal features using multiple one-dimensional convolutional layers with different kernel sizes to obtain the extended multi-scale temporal features. Multi-scale feature extraction processing is performed on the extended multi-scale temporal features using a preset learnable multi-scale bias. The specific features and the shared features are fused to obtain the target fused features; The target communication traffic prediction value is obtained by performing a prediction spatial mapping on the target fusion features using a multilayer perceptron.

6. The method according to claim 1, characterized in that, After obtaining the target communication traffic prediction values ​​for multiple space nodes, the method further includes: Based on a preset joint loss function, the loss value between the predicted target communication traffic value and the corresponding actual traffic value is calculated; wherein, the joint loss function includes mean absolute error loss and causal comparison loss, the mean absolute error loss is used to measure the prediction error, and the causal comparison loss is used to promote pattern decoupling among multiple candidate expert models; Based on the loss value, the multiple candidate expert models are updated.

7. A traffic prediction device for wireless communication, characterized in that, include: The acquisition module is used to acquire historical spatiotemporal flow data of multiple spatial nodes in the target area; The extraction module is used to extract multi-scale temporal features of each spatial node based on the historical spatiotemporal flow data of each spatial node. A modeling module is used to perform spatial dependency modeling on the multi-scale temporal features to obtain a node embedding vector corresponding to each spatial node. This includes: generating an adaptive graph adjacency matrix representing the spatial dependencies between multiple spatial nodes based on a preset initial node embedding vector for each spatial node; performing graph convolution processing on the multi-scale temporal features based on the adaptive graph adjacency matrix to aggregate the spatial neighborhood information of the multiple spatial nodes, obtaining multi-scale temporal aggregated features; projecting the multi-scale temporal aggregated features into a query matrix, a key matrix, and a value matrix; performing matrix multiplication on the query matrix and the key matrix, and scaling and masking the multiplication results to obtain an attention score; wherein the masking process uses a preset lower triangular mask; normalizing the attention score and performing matrix multiplication with the value matrix to obtain an attention output feature; performing residual connection and layer normalization processing on the attention output feature and the multi-scale temporal aggregated features to generate fused spatiotemporal features; and determining the node embedding vector corresponding to each spatial node based on the fused spatiotemporal features. The determination module is used to determine a target expert model for the corresponding spatial node from a plurality of preset candidate expert models based on the node embedding vector corresponding to each spatial node. The prediction module is used to predict the multi-scale time features by using the target expert model corresponding to each of the spatial nodes, so as to obtain the target communication traffic prediction values ​​of multiple spatial nodes.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the traffic prediction method in wireless communication as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the traffic prediction method in wireless communication as described in any one of claims 1 to 7.

Images