Base station flow prediction method based on GRUPatch and graph neural network

Abstract

The invention discloses a base station flow prediction method based on GRUPatch and a graph neural network. The method comprises the following steps: dividing long-time sequence data into a plurality of short-time sequence segments (Patch), inputting the Patch into a GRU model to extract local time sequence features, capturing a relationship between the Patches through an Attention mechanism, and extracting global time sequence features; clustering base station nodes in combination with historical flow data and geographic position characteristics of the base station, and constructing a dynamic adjacency matrix; and based on the adjacency matrix, extracting shared information and personalized features among the base stations by using GNN, and fusing the features to realize high-precision traffic prediction. According to the method, the time dependence in the base station flow time sequence data and the dynamic association information between different base stations can be effectively captured, and the accuracy and adaptability of base station flow prediction in a complex network environment are remarkably improved through the combination of the time sequence characteristics and the space association information.

Description

Technical Field

[0001] The present invention belongs to the technical field of communication network traffic prediction, and specifically relates to a base station traffic prediction method based on GRUPatch and graph neural network. Background Art

[0002] With the rapid expansion of modern communication networks, especially in the fields of 5G, Internet of Things (IoT), and intelligent transportation, base station traffic prediction has become a key task to improve network performance, optimize resource allocation, and improve user experience. Accurate traffic prediction helps to balance the load of communication networks, avoid network overload, and ensure the service quality of the network. However, most of the current traffic prediction methods focus on analyzing the historical data of a single base station, and fail to fully consider the correlation between base stations, especially in the scenario of multi-base station collaborative prediction, which shows the lack of prediction accuracy.

[0003] In the prior art, time series modeling methods such as the autoregressive moving average model (ARIMA) and the long short-term memory network (LSTM) can effectively capture the timing dependency characteristics of a single base station, but face significant challenges when processing long time series data. First, directly inputting long time series into the model can easily lead to gradient vanishing or gradient exploding problems, especially in recurrent neural networks (RNNs), where the gradient gradually decays or expands over time, affecting the training stability of the model. Secondly, long sequence information is easily forgotten, especially the early information in the sequence is crucial for predicting traffic at subsequent moments, and directly inputting long sequences may cause the model to be unable to effectively utilize this information. In addition, the computational complexity of processing long time series is high, and the training time is long, which makes it difficult to meet the needs of real-time prediction in actual scenarios.

[0004] In addition to the difficulty in dealing with long time series, existing base station traffic prediction methods generally ignore the interrelationships between base stations. In actual networks, base stations do not exist in isolation, and traffic fluctuations at a certain base station often have an impact on its neighboring base stations. Especially in load balancing and traffic scheduling scenarios, the synergy between base stations is crucial. Existing methods only rely on the historical traffic data of a single base station and fail to model this correlation between base stations, resulting in unsatisfactory results for multi-base station collaborative prediction.

[0005] As a model that can effectively capture the complex relationships between nodes, graph neural networks have been widely used in many fields. However, the current base station traffic prediction model fails to fully utilize the advantages of the network topology and fails to effectively combine the timing characteristics of the base station with the correlation information of its network topology, resulting in the inadequacy of the existing model in handling complex multi-base station collaborative prediction tasks. The main problems of the existing technology in base station traffic prediction include: 1) Direct input of long time series data is prone to gradient disappearance or explosion, affecting the stability of model training; 2) The traffic correlation between base stations is ignored and the synergy between base stations cannot be captured; 3) The timing characteristics and network structure information of the base station are not effectively combined, resulting in insufficient prediction accuracy. Summary of the invention

[0006] The main purpose of the present invention is to provide a base station traffic prediction method based on GRUPatch and graph neural network. The method of the present invention can effectively capture the time dependency in the base station traffic time series data and the dynamic correlation information between different base stations by performing block processing on the time series data of the base station traffic and combining the structured correlation information between base stations. By combining the time series characteristics with the spatial correlation information, the accuracy and adaptability of the base station traffic prediction are significantly improved in a complex network environment. It is suitable for resource optimization and traffic management in communication networks such as 5G and the Internet of Things.

[0007] The technical solution of the present invention is specifically described as follows.

[0008] The present invention provides a base station traffic prediction method based on GRUPatch and graph neural network, comprising the following steps:

[0009] (1) The long time series data of base station traffic is divided into multiple short time series fragments Patches, and these short time series fragments Patches are input into the gated recurrent unit GRU to extract local time series features. The relationship between the short time series fragments Patches is captured through the Attention mechanism, thereby extracting global time series features;

[0010] (2) The base station nodes are clustered by combining historical traffic data and the geographical location characteristics of the base stations, so as to construct a dynamic adjacency matrix that can accurately represent the spatial and temporal correlations between base stations. Based on the adjacency matrix, a graph neural network is used to extract the shared information and individual features between base stations, and these features are fused to achieve more accurate traffic prediction.

[0011] In the present invention, in step (1), in step (1), the long time series data is standardized by using the RevIN module and then sent to the Patch partitioning and bidirectional GRU in the GRUPatch module for time series feature extraction.

[0012] In the present invention, in step (1), the entire time series data is divided into multiple smaller time blocks (Patches) according to the time dimension. Each Patch is input into the GRU (Gated Recurrent Unit) to extract local time series features. Through its gating mechanism, the GRU can capture the time dependencies within each Patch and extract the time series features of each Patch. This feature extraction process allows the model to focus on a shorter time window, fully capture the feature information within the local time window, and ensure that key information will not be lost in a long time series. This method effectively solves the problem of information forgetting in long sequence processing. The size of the Patch can be dynamically adjusted according to different application scenarios, so as to flexibly adapt to the needs of different tasks.

[0013] In the present invention, in step (1), the relationship between different patches is modeled through the Attention mechanism, the similarity between each short time series fragment patch and other short time series fragment patches is calculated, and the attention weight is generated, so as to capture the dependency between different time periods; the feature information of each short time series fragment patch is weighted and aggregated by the Attention mechanism, and finally the global time series feature is generated. This method ensures that the information in different time windows can be effectively integrated, providing a more comprehensive feature representation for the subsequent traffic prediction task.

[0014] In the present invention, in step (1), the method of capturing the relationship between short time series fragments Patch through the Attention mechanism to extract the global time series features is as follows:

[0015] First, the similarity between different short time series fragments Patches is calculated through the additive Attention mechanism, and the similarity score e ij The calculation formula is:

[0016]

[0017] Among them, h i and h j Respectively represent the feature vectors of the ith patch and the jth patch, [h i ;h j ] is their concatenation operation, W is the learnable weight matrix, b is the bias vector, It is the weight vector that maps the concatenated feature vector to the scalar space;

[0018] Next, the Softmax function is used to normalize the similarity scores between all short time series fragments Patches and calculate the attention weight α ij :

[0019]

[0020] The final output global feature h′ i It is expressed as:

[0021]

[0022] Among them, Patch_num represents the total number of patches after splitting.

[0023] In the present invention, in step (1), the output global time series features are transformed by multi-layer linear mapping, and Leaky ReLU is introduced as an activation function. After linear mapping and activation, RevIN is used to perform denormalization processing to restore the standardized features to the original scale.

[0024] In the present invention, in step (1), by combining Patch partitioning with the Attention mechanism, GRUPatch not only improves the ability to capture time series information, but also significantly reduces the problem of information loss in long sequence processing, ensuring that the model can complete the task of predicting base station traffic more accurately.

[0025] In the present invention, in step (2), the KMeans algorithm is used to construct an adjacency matrix A based on the similarity between base stations. The element Aij of the adjacency matrix represents the base station S i and S j The relationship between them is defined as follows:

[0026]

[0027] In the present invention, in step (2), based on the adjacency matrix, a graph neural network is used to extract shared information and individual features between base stations, and these features are integrated to achieve a method for achieving higher-precision traffic prediction as follows:

[0028] Step 1: Calculation of shared information

[0029] Based on the clustering of base station nodes, the time series characteristics of all base stations in the same cluster are analyzed. Perform weighted averaging to obtain shared features

[0030]

[0031] Among them, N i represents the number of base stations in cluster i, Cluster is the time series feature of each base station in the cluster. i is the i-th cluster;

[0032] Step 2: Attention mechanism extracts common features

[0033] The role of the Attention mechanism is to assign different weights to measure the importance of the features of the base station nodes according to the similarity between the features and the shared features; attention weight The calculation formula is as follows:

[0034]

[0035] in, is the Query vector, is the Key vector, d k is the feature dimension, W Q and W K is the weight matrix of the linear transformation; the final common feature Calculated as:

[0036]

[0037] in, is the Value vector, W V is the weight matrix of linear transformation;

[0038] Step 3: Integration of shared information and personality traits

[0039] The common characteristics Timing characteristics with base station nodes Perform difference calculation to obtain personality traits And extract common and difference features through linear mapping operations respectively;

[0040] Linear mapping of common features:

[0041]

[0042] Linear mapping of personality traits:

[0043]

[0044] Finally, the common features and individual features are added together to generate the final output, which is the traffic prediction result.

[0045]

[0046] Among them, W1 and W2 are learnable weight matrices introduced in the process of extracting common features and individual features, which respectively control the common features and individual features on the final output degree of impact.

[0047] In the present invention, in step (2), the base station nodes are clustered using the KMeans algorithm by combining historical traffic data and geographic location features, thereby constructing a dynamic adjacency matrix. The elements of the adjacency matrix A represent the connection strength between base station nodes, and can dynamically reflect the spatial and traffic associations between base stations. Based on this adjacency matrix, a graph neural network model is used to extract shared feature information between base stations, and is integrated with the timing characteristics of each base station to generate more accurate traffic prediction results; this method achieves deep mining of base station traffic timing data by introducing correlation information between base stations, significantly improving the prediction effect in complex network environments.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] The base station traffic prediction method based on GRUPatch and graph neural network proposed in the present invention has the advantages of strong stability and high prediction accuracy in processing long time series, can overcome the gradient vanishing and gradient explosion problems of traditional time series prediction methods in long sequence processing, and effectively capture the correlation information between base stations. The present invention tests the performance based on traffic prediction tasks in multiple time periods (5 minutes, 15 minutes, 30 minutes, 45 minutes and 60 minutes). The results show that the prediction performance of the GRUPATCH_Model proposed in the present invention in different time spans is significantly improved, especially in long time period prediction, which better solves the many limitations of traditional time series prediction models in processing long sequence data. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a specific flow chart of the base station traffic prediction method based on GRUPatch and graph neural network of the present invention.

[0051] Figure 2 Schematic diagram of the GRUPatch module.

[0052] Figure 3 It is a schematic diagram of the network module of the present invention.

[0053] Figure 4 Schematic diagram extracted for the graph neural shared information module.

[0054] Figure 5 The line chart shows the comparison of the mean absolute percentage error (MAPE) of each model at different prediction durations.

[0055] Figure 6 This is a comparison chart of the predictions and true values ​​of each model under a prediction time of 5 minutes. DETAILED DESCRIPTION

[0056] The present invention is further described below in conjunction with specific embodiments and drawings.

[0057] In traditional methods, processing long time series data will encounter problems such as information loss and difficulty in capturing features. The present invention can effectively extract global time series features by dividing long time series into short time segments and introducing the Attention mechanism. In addition, the present invention uses a graph neural network model to model the correlation between base stations, integrates the shared information of each base station with individual characteristics, and finally achieves accurate prediction of traffic. In addition, during the collection process, base station traffic data is often affected by irrelevant noise or external data fluctuations, and these noisy data will affect the accurate prediction of base station traffic. Therefore, how to reduce the interference of data noise and irrelevant fluctuations and extract key time series features has become a key issue to improve the accuracy of base station traffic prediction. The base station traffic prediction method based on GRUPatch and graph neural network proposed in the present invention can effectively handle this noise interference in the data preprocessing stage and feature extraction process.

[0058] This paper proposes a base station traffic prediction method based on GRUPatch and graph neural network, aiming to improve the accuracy of traffic prediction by combining time series feature extraction and relationship modeling between base stations. The entire model structure GRUPATCH_Model includes a GRUPatch module and a graph neural network shared information fusion module. The structure and parameter settings of each layer are as follows:

[0059] The first part of the network is the GRUPatch module, which is mainly used for base station time series feature extraction. The GRUPatch module includes the Patch part, bidirectional GRU, and Attention mechanism. Figure 2 The specific structure of the GRUPatch model is shown. First, in the Patch part, the long sequence is split into n short sequences to reduce the complexity of the model in processing long sequences and avoid the gradient vanishing or exploding problem caused by too many time steps. Each short sequence is processed by a bidirectional GRU network. GRU can effectively capture the temporal features within the sequence, learn the temporal dependencies in the short sequence, and obtain the feature sequence representation h1, h2...h n Subsequently, the processed short sequence representation is sent to the Attention mechanism. Attention captures the global dependencies between different short sequences by calculating the weighted relationship between short sequences, and obtains a sequence representation that takes into account global features. Overall, this structure combines the advantages of GRU in local time series modeling and Attention in global dependency modeling, significantly enhancing the performance of the model in complex time series tasks.

[0060] The input data is B×N×T×F, where B is the batch size, N is the number of base stations, T is the time step, and F is the feature dimension of each time step. By dividing the long time series into multiple patches, the length of each patch is patch_len, and the final input shape is B×N×Patch_num×Patch_len×F, where patch_num represents the total number of patches. Next, a bidirectional GRU is used to extract features from each patch. The hidden layer dimension of the GRU is set to D, and the output feature size is B×N×Patch_num×2D, where 2D is the output dimension of the bidirectional GRU (the concatenation of the forward and backward features). Then, the relationship between patches is captured through the Attention mechanism. The temporal features of each patch are concatenated to form a feature matrix of B×N×Patch_num×Patch_num×4D, which represents the interactive information between different time segments. Through the additive Attention mechanism, the model calculates the similarity between different patches and normalizes the results. The calculation formula for the attention weight is:

[0061]

[0062] Among them, h i and h j Respectively represent the feature vectors of the ith patch and the jth patch, [h i ;h j ] is their concatenation operation, W is the learnable weight matrix, b is the bias vector, is the weight vector that maps the concatenated feature vector to the scalar space. The similarity score is normalized by the Softmax function to obtain the attention weight:

[0063]

[0064] Finally, the model performs weighted summation of the features of each patch according to the attention weight to obtain the global temporal features:

[0065]

[0066] The second part is the graph neural network shared information fusion module. Figure 3 The system structure of the present invention including multiple modules is shown. Figure 4 The process of shared feature extraction is shown. Figure 3 Starting from the bottom, the first is BS FeaturesEncoder, which corresponds to the GRUPatch structure and converts the base station input time series features into feature vector representation X t,0, followed by the KMeans Concept Module, which combines the adjacency matrix obtained from it with the output data X from the BS Features Encoder t,0 Get the common characteristics of base stations Next is the Individual Information Module, which receives the individual characteristics of the base station And output Finally, the FC Layer combines the predicted values ​​of the common features. and the predicted value of personality And generate the final predicted value Y t . Figure 4 The left side shows the clustering of base station data using the KMeans algorithm, using the adjacency matrix combined with base station vector features Weighted acquisition of base station public representation Figure 4 The right side shows the use of the Attention mechanism, which enables the representation of base stations in the base station common features to be dynamically adjusted according to the attention weight, so as to more accurately capture the relationship between base stations and base station sharing. The left and right parts respectively reflect different data processing and relationship building methods.

[0067] First, the KMeans clustering algorithm is used to cluster the base stations according to the historical traffic data and the geographical location of the base stations, and an adjacency matrix A is obtained, which represents the correlation between the base stations. The value of Aij is 1, which means that the base station S i and S j Belong to the same cluster, otherwise it is 0. This adjacency matrix is ​​used in the message passing process of the graph neural network to further model the synergy between base stations. Then the time series features of each base station node are passed through the adjacency matrix to pass information with the features of its neighboring base stations to generate shared features. The shared features and personalized features are fused through the Attention mechanism, and finally the fused features are output for the final traffic prediction.

[0068] Specifically, the present invention adopts a multi-layer network structure, and the core structure includes a GRUPatch module and a graph neural network shared information fusion module;

[0069] (1) GRUPatch module

[0070] Time series data is processed by dividing a long time series into multiple shorter time periods (Patch). Each Patch is passed as an independent input to GRU (Gated Recurrent Unit) for local time series feature extraction. GRU can capture the time dependency within each Patch, avoiding the common problem of information forgetting in long time series processing.

[0071] After that, the model uses the Attention mechanism to model the relationship between different patches. The Attention mechanism calculates the similarity between each patch and other patches and generates attention weights. The time series features of each patch are weighted and aggregated through Attention to finally generate global time series features. This method ensures that the information in different time windows in a long time series can be effectively integrated, improving the accuracy of traffic prediction. The specific algorithm is as follows

[0072] Step 1: Data input and reorganization In the task of traffic prediction, the input data usually has multiple feature dimensions, and the traffic of different base stations changes over time. In order to process these time series data, the input data needs to be preprocessed and normalized first. The input data X∈R B×N×T×F , where B represents the batch size, which refers to the number of base stations processed simultaneously in one forward propagation; N represents the number of base stations, each of which has an independent data sequence; T represents the time step, that is, the time span of the input data; F represents the number of features in each time step. In order to ensure the uniformity of different feature scales, the input data will first be normalized by the RevIN (Reversible Normalization) module, which is specifically expressed as follows:

[0073]

[0074] Among them, μ and σ are the mean and standard deviation of the input data respectively. This step can effectively improve the stability of the model when processing data of different scales and avoid training difficulties caused by large differences in feature values.

[0075] Step 2: Patch division and bidirectional GRU temporal feature extraction

[0076] After data normalization, the time series data will be sent to the Patch partitioning and bidirectional GRU (Bidirectional GRU) in the GRUPatch module for time series feature extraction. In order to effectively process long time series data, the present invention divides the input time series into multiple shorter time segments (Patch). Assuming that the length of the input time series is T, multiple Patches are obtained by partitioning, and the length of each Patch is Patch_len, that is: X patch = {P1, P2, ..., P Patch_num},in Indicates the number of patches.

[0077] Each Patch contains multiple time steps, and the GRU model extracts time series features within the Patch. This division can avoid the attenuation of information when it is transmitted in a long sequence and solve the problems of gradient disappearance and gradient explosion.

[0078] In each patch, the time series features are extracted by the bidirectional GRU in the GRUPatch module. GRU (Gated Recurrent Unit) is a recurrent neural network structure that can capture long-term dependency information through its gate mechanism (reset gate and update gate), solving the gradient vanishing problem that traditional RNNs are prone to when processing long time series.

[0079] Bidirectional GRU processes time series from two directions simultaneously: forward (from the past to the future) and backward (from the future to the past). This enables the model to capture the forward and backward dependencies in the time series at the same time. The calculation formula of the forward GRU is: The calculation formula of backward GRU is: The dimension of the output result is 2D, where D is the dimension of the GRU hidden layer. The bidirectional GRU extracts the hidden state h in two directions. t , which can provide a more comprehensive understanding of the timing characteristics of base station traffic.

[0080] After GRU processing, the time series data will be dimensionally adjusted from B×N×T×2D, where 2D represents the sum of the hidden state output dimensions of the forward and backward GRU. The main function of this step is to extract the bidirectional time series features of each base station at each time step from the input time series data, ensuring that both historical information and future information can participate in the feature extraction process.

[0081] After the GRU module completes the extraction of temporal features, the model needs to further process these features to adapt to the subsequent Attention mechanism. Specifically, the shape of the GRU output is B×N×Patch_num×2D, that is, the bidirectional temporal features of each base station on multiple patches. In order to integrate these temporal features into a whole, the output is first dimensionally adjusted, and the relationship between different patches is modeled through the Attention mechanism.

[0082] The core idea of ​​the Attention mechanism is to capture the dependencies between different patches. To achieve this goal, the model expands the features of each patch. The Attention mechanism captures the dependencies between patches by calculating the similarity between each patch. By repeating the features of each patch and concatenating them with the features of other patches, the generated feature vector is in the shape of B×N×Patch_num×Patch_num×4D. This operation makes the features of each patch contain not only its own information, but also the interactive information with other patches.

[0083] Through the Attention mechanism, the model can generate a query and key for each patch. Different from the traditional dot product similarity, this paper adopts the additive Attention mechanism to capture the dependency between patches. Specifically, additive attention concatenates the features of each pair of patches and calculates their similarity scores through a learnable feedforward neural network. ij The calculation formula is:

[0084]

[0085] Among them, h i and h j Respectively represent the feature vectors of the ith patch and the jth patch, [h i ;h j ] is their concatenation operation, W is the learnable weight matrix, b is the bias vector, It is the weight vector that maps the concatenated feature vector to the scalar space.

[0086] Next, the model uses the Softmax function to normalize the similarity scores between all patches and calculate the attention weight α ij :

[0087]

[0088] Through the attention weight, the model performs weighted summation on the features of all patches to obtain the global temporal feature representation of each base station:

[0089]

[0090] Finally, the model selects the features of the last patch as the global time series representation of the base station, and the output dimension is B × N × 2D. This process ensures that the model can capture interdependent global information between multiple time segments, thereby effectively improving the accuracy of traffic prediction.

[0091] Step 3: Linear mapping and anti-normalization

[0092] After the Attention mechanism is completed, the time series features will be transformed through multi-layer linear mapping to further extract higher-level features. The present invention introduces Leaky ReLU as the activation function. Leaky ReLU has better gradient flow characteristics than traditional ReLU and can avoid the "neuron death" problem in the negative value area. Its activation function expression is:

[0093]

[0094] Subsequently, in order to prevent the model from overfitting, Dropout technology was applied to the training process of the neural network. The use of Dropout technology further enhances the generalization ability of the model. By randomly "discarding" some neurons during the training process, Dropout can effectively reduce the risk of overfitting of the model.

[0095] After linear mapping and activation, the model uses RevIN to perform denormalization, restoring the standardized features back to the original scale, thereby ensuring that the prediction results have reasonable physical meaning at the original scale. The denormalization process is as follows:

[0096] X denorm =X norm ×σ+μ

[0097] Among them, σ and μ are the standard deviation and mean calculated during the previous normalization.

[0098] The graph neural network (such as Figure 3 The module clusters base stations using the KMeans algorithm, builds an adjacency matrix between base stations, and uses the model to model and fuse the correlation information between base stations to improve the accuracy and effect of base station traffic prediction. This module is particularly suitable for processing complex shared information and correlation between base stations, providing a more comprehensive feature representation for the prediction model.

[0099] (2) Graph Neural Network Shared Information Fusion Module

[0100] In order to further improve the accuracy of the model, the present invention introduces a graph neural network to model the correlation between base stations. First, the base stations are clustered using the KMeans clustering algorithm to generate an adjacency matrix A, which represents the connection strength between base station nodes. Then, the graph neural network fuses the timing features of adjacent base stations through a message passing mechanism to generate shared features. During the fusion process, the model weightedly fuses the shared features of the base station with its own personalized timing features to form the final traffic prediction result. This feature fusion method based on graph neural networks ensures information coordination between base stations and improves the accuracy and stability of traffic prediction. The specific algorithm is as follows:

[0101] Step 1: KMeans clustering and adjacency matrix construction

[0102] After extracting the global time series features, the present invention further clusters the base stations using the KMeans clustering algorithm. The historical traffic data and geographic location of the base stations are used to measure the similarity between the base stations, and the base stations are divided into different classes according to the similarity. The results of the base station clustering are used to construct the adjacency matrix A between the base stations. The element Aij of the adjacency matrix represents the distance between the base stations S i and S jThe relationship between them is defined as follows:

[0103]

[0104] This adjacency matrix is ​​used in the message passing process in the graph neural network to capture the synergy between base stations.

[0105] Step 2: Message passing and shared feature fusion of graph neural networks

[0106] In the present invention, graph neural network is used to further extract and fuse shared features between base stations (such as Figure 4 The time series features of each base station node are transmitted through the adjacency matrix with the features of its neighboring base stations to generate shared features Specifically, the shared features of a base station are obtained by weighted aggregation of the features of its neighboring base stations. The formula is as follows:

[0107]

[0108] Next, the Attention mechanism is used to further integrate the timing characteristics of the base station node and shared features Perform weighted fusion to generate common features Attention weight The calculation formula is as follows:

[0109]

[0110] Among them, Q i and K j They represent the query and key vectors respectively, and are used to measure the similarity between base station features and shared features. Finally, the shared features and personalized features of each base station are fused through the weighted operation of the Attention mechanism.

[0111] Step 3: Feature fusion and traffic prediction: In the feature fusion stage, the common features of base stations and personalized time series features It is further fused through linear mapping to generate the final traffic prediction result. The formula is as follows:

[0112]

[0113] Through this fusion method, the model can simultaneously utilize the personalized timing characteristics of base stations and shared global characteristics to generate more accurate traffic prediction values.

[0114] Experimental verification: We conducted experimental verification on the above embodiments and compared them through different models.

[0115] Experimental setup: In the experimental setup of the present invention, based on a real base station traffic data set, we divide the data set into a training set, a validation set, and a test set in a ratio of 6:2:2 to ensure the generalization ability of the model on different data sets. Before each training, the input data is first normalized using the Reversible Normalization (RevIN) method so that different base stations and different features are processed at a unified scale, thereby improving the stability of the model training. During the training process, we selected L1 Loss (mean absolute error) as the loss function, and its calculation formula is: Among them, y i and Represent the true value and the predicted value respectively. L1 Loss has good robustness in processing abnormal data and can effectively reduce the interference of outliers on model prediction. Adam was selected as the optimizer, with an initial learning rate of 0.001 and weight decay (weight_decay) set to le-4. In order to prevent gradient explosion, we used gradient clipping technology during training with a clipping threshold of 15 to ensure that the norm of the gradient is not too large to avoid instability in model training. In addition, in order to prevent overfitting of the model, the Early Stopping strategy was introduced on the validation set, that is, when the loss on the validation set has not been improved within 10 consecutive epochs, the training is stopped in advance. The model was trained for a total of 250 epochs, and the batch size (batchsize) was set to 64. All experiments were run on a machine equipped with an NVIDIA GeForce RTX 4060Ti graphics card to ensure the efficiency of the training process and the repeatability of the experimental results.

[0116] Experimental results:

[0117] Table 1 shows the performance comparison of the GRUPatch model of the present invention and other base station traffic prediction algorithms in different time windows (5 minutes, 15 minutes, 30 minutes, 45 minutes and 60 minutes). It can be seen from the experiment that the GRUPatch model has shown significant advantages in the traffic prediction tasks in each time window.

[0118] 5-minute prediction task: The MAPE of the GRUPatch model is 0.0963±0.0009, which is better than GRUModel (0.1142±0.0011) and MGAT_PGAT_Model (0.1064±0.0010). This result shows that in short-term prediction tasks, GRUPatch can better capture the dynamic changes of base station traffic.

[0119] 15-minute prediction task: In the 15-minute prediction task, the MAPE of the GRUPatch model is 0.1449±0.0013, which is better than iTransformerA (0.1496±0.0012) and PatchTSTA (0.1587±0.0015). This shows that the model has a significant advantage in capturing short- and medium-term time series features and can effectively improve prediction accuracy.

[0120] 30-minute prediction task: The GRUPatch model achieved a MAPE of 0.1808±0.0017 in the 30-minute prediction task, which is better than TCNA (0.1844±0.0017) and iTransformerA (0.1873±0.0016), further verifying its advantages in longer time series tasks.

[0121] 45-minute prediction task: In the 45-minute prediction task, GRUPatch achieved a MAPE of 0.1745±0.0022, which significantly reduced the error compared to other algorithms, such as PatchTSTA (0.2038±0.0026). This shows that the model can still maintain high prediction accuracy and robustness under longer time series.

[0122] 60-minute prediction task: In the 60-minute long-term window prediction, the MAPE of the GRUPatch model is 0.2076±0.0031, which is significantly lower than GRUModel (0.2561±0.0035) and PatchTSTA (0.2640±0.0038). This shows that GRUPatch not only performs well in short-term prediction tasks, but also has high accuracy in long-term prediction tasks.

[0123] Experimental results show that the GRUPatch model effectively improves the prediction accuracy of base station traffic by using patch partitioning, bidirectional GRU to extract time series features, Attention mechanism to capture dependencies between patches, and graph neural network modeling of associations between base stations. Especially when processing long time series data, the method of the present invention can avoid the problems of gradient disappearance and gradient explosion, and successfully improve the ability to capture time series features in traffic prediction tasks.

[0124] Through experimental verification, the GRUPatch model proposed in this paper shows significant advantages in processing long-term base station traffic prediction tasks, especially in capturing complex time series features and modeling the correlation between base stations. Experimental results show that the GRUPatch model surpasses many existing advanced algorithms in short-term, medium-term and long-term prediction tasks, and the attention mechanism and graph neural network introduced by it play a key role in improving prediction accuracy. In the future, the GRUPatch model can be widely used in base station traffic prediction and other tasks involving time series data analysis.

[0125] Table 1: Comparison of average MAPE performance on outdoor cellular traffic dataset

[0126]

[0127] Above, the MAPE (mean absolute percentage error) of the GRUPATCH_Model proposed in the present invention in multiple time periods ( Figure 5 As shown in the figure, the MAPE of GRUPATCH_Model is significantly lower than that of traditional models. For example, in the 5-minute traffic prediction task, GRUPATCH_Model achieved a MAPE of 0.0963±0.0009, which is about 9.5% higher than the traditional MGAT_PGAT_Model. In the 60-minute long-term traffic prediction task, the MAPE of GRUPATCH_Model is 0.2076±0.0031, which is about 16.7% higher than that of iTransformerA’s 0.2492±0.0035, significantly reducing the prediction error in long time series.

[0128] In addition, GRUPATCH_Model also performs well in medium-term prediction tasks (30 minutes and 45 minutes). For example, in the 30-minute task, the MAPE of GRUPATCH_Model is 0.1808±0.0017, which is about 12.6% higher than the 0.2069±0.0022 of the PatchTSTA model; in the 45-minute task, the MAPE of GRUPATCH_Model is 0.1745±0.0022, which is 0.1738±0.0018 of the TCNA model, showing higher stability. Figure 6 It also shows the comparison between the prediction results of each model and the actual value under the 5-minute prediction time. These quantifiable results show the significant performance improvement of the GRUPATCH_Model proposed in this invention in different time spans, especially in long-term prediction, which better solves the many limitations of traditional time series prediction models in processing long sequence data.

Claims

1. A base station traffic prediction method based on GRUPatch and graph neural network, characterized in that: The following steps are involved: (1) The long time series data of base station traffic is divided into multiple short time series fragments Patches, and these short time series fragments Patches are input into the gated recurrent unit GRU to extract local time series features. The relationship between the short time series fragments Patches is captured through the Attention mechanism, thereby extracting global time series features; (2) Clustering base station nodes based on historical traffic data and the geographical location characteristics of base stations, thereby constructing a dynamic adjacency matrix that accurately represents the spatial and temporal correlations between base stations; Based on the adjacency matrix, a graph neural network is used to extract shared information and individual features between base stations, and these features are fused to achieve more accurate traffic prediction.

2. The base station traffic prediction method according to claim 1, characterized in that: In step (1), the long time series data is standardized using the RevIN module and then sent to the Patch partitioning and bidirectional GRU in the GRUPatch module for time series feature extraction.

3. The base station traffic prediction method according to claim 1, characterized in that: In step (1), the similarity between each short time series fragment Patch and other short time series fragments Patch is calculated through the Attention mechanism to generate attention weights, thereby capturing the dependencies between different time periods; The feature information of each short time series fragment Patch is weighted and aggregated through the Attention mechanism to finally generate the global time series feature.

4. The base station traffic prediction method according to claim 1, characterized in that: In step (1), the method of capturing the relationship between short time series fragments Patches through the Attention mechanism to extract global time series features is as follows: First, the similarity between different short time series fragments Patches is calculated through the additive Attention mechanism, and the similarity score e ij The calculation formula is: Among them, h i and h j Respectively represent the feature vectors of the ith patch and the jth patch, [h i ;h j ] is their concatenation operation, W is the learnable weight matrix, b is the bias vector, It is the weight vector that maps the concatenated feature vector to the scalar space; Next, the Softmax function is used to normalize the similarity scores between all short time series fragments Patches and calculate the attention weight α ij : The final output global feature h i ′ It is expressed as: Among them, Patch_num represents the total number of patches after splitting.

5. The base station traffic prediction method according to claim 1, characterized in that: In step (1), the output global temporal features are transformed through multi-layer linear mapping, and Leaky ReLU is introduced as the activation function. After linear mapping and activation, the RevIN module is used to perform denormalization processing to restore the standardized features to the original scale.

6. The base station traffic prediction method according to claim 1, characterized in that: In step (2), the KMeans algorithm is used to construct an adjacency matrix A based on the similarity between base stations. The element Aij of the adjacency matrix represents the base station S i and S j The relationship between them is defined as follows:

7. The base station traffic prediction method according to claim 1, characterized in that: In step (2), based on the adjacency matrix, a graph neural network is used to extract shared information and individual features between base stations, and these features are fused to achieve a more accurate traffic prediction method as follows: Step 1: Calculation of shared information Based on the clustering of base station nodes, the time series characteristics of all base stations in the same cluster are analyzed. Perform weighted averaging to obtain shared features Among them, N i represents the number of base stations in cluster i, Cluster is the time series feature of each base station in the cluster. i is the i-th cluster; Step 2: Attention mechanism extracts common features The role of the Attention mechanism is to assign different weights to measure the importance of the features of the base station nodes according to the similarity between the features and the shared features; attention weight The calculation formula is as follows: in, is the Query vector, is the Key vector, d k is the feature dimension, W Q and W K is the weight matrix of the linear transformation; the final common feature Calculated as: in, is the Value vector, W V is the weight matrix of linear transformation; Step 3: Integration of shared information and personality traits The common characteristics Timing characteristics with base station nodes Perform difference calculation to obtain personality traits And extract common and difference features through linear mapping operations respectively; Linear mapping of common features: Linear mapping of personality traits: Finally, the common features and individual features are added together to generate the final output, which is the traffic prediction result. Among them: W1 and W2 are learnable weight matrices introduced in the process of extracting common features and individual features, which respectively control the common features and individual features on the final output degree of impact.

Images