A multi-node base station traffic prediction method based on a self-organizing graph neural network

Abstract

The application discloses a kind of multi-node base station traffic prediction methods based on self-organizing graph neural network, belong to space-time traffic prediction technical field, including base station node data processing and influence evaluation: node set subgraph is formed, and the size of influence of adjacent matrix and evaluation node is generated;Base station node reorganization: through mapping matrix, linear combination is carried out to time sequence node, generates linearly independent vector base, and according to influence guide mapping matrix convergence direction;Using part attention aggregation spatial feature: node set is sequentially carried out static graph convolution and dynamic graph convolution and captures spatial connection;In convolution, long short-term memory network is inserted: capture time connection and carry out time series prediction;Finally, standardization is carried out and is generated prediction result by multilayer perception machine.The application uses the above-mentioned one kind of multi-node base station traffic prediction methods based on self-organizing graph neural network, proposes the spatial feature extraction mechanism of multi-node base station traffic, provides effective prediction method for the change trend of base station traffic.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of space-time flow prediction, and particularly relates to a multi-node base station flow prediction method based on a self-organizing graph neural network. BACKGROUND

[0002] With the development of equipment and the emergence of new user demands such as metadata and large language models, cellular communication traffic is growing exponentially in a diverse and heterogeneous manner. The huge cellular traffic poses many challenges to the next generation of mobile communication systems, such as cellular traffic management, base station energy control, network signal delay, and concurrent network requests. For 5G base stations, the energy consumption problem is particularly serious. According to statistics, the energy consumption of a 5G base station is about three to four times that of a 4G base station. Therefore, energy-saving strategies are particularly important in mobile communication systems. As one of the efficient energy-saving strategies, cellular traffic resource optimization relying on large-scale urban high-precision cellular traffic prediction has attracted much attention.

[0003] In the field of deep learning, people initially applied recurrent neural networks (RNN) to this task. The recurrent neural network is a model architecture commonly used in time series prediction tasks composed of gating units. This architecture captures time series features by forgetting, adding, and updating memory units. The process is to input the data at each time of the time series into the same model one by one and output the prediction of the future time series at the last time. However, it was later found that cellular network traffic has strong spatial correlation. Considering only the time characteristics of cellular network traffic cannot well predict the trend of cellular network traffic.

[0004] 5G traffic prediction problem can be modeled as a spatio-temporal sequence prediction problem. Graph convolutional network (GCN) is a neural network for processing abstract graph relationships such as social networks, knowledge graphs, complex file systems, etc. Combining graph convolutional network with RNN can better complete this task. However, GCN can only capture static and fixed node correlation, which is disastrous for cellular traffic with frequently changing correlation. In view of the shortcomings of the previous graph convolution or its approximate method, some researchers try to combine GCN with attention to improve the accuracy of GCN in the task of cellular traffic prediction. However, no one has carefully studied how GCN should be combined with attention to achieve the best results. The common practice is to multiply the adjacency matrix of GCN with the dynamic graph formed by attention element by element. The cellular traffic prediction method based on adaptive hybrid spatio-temporal graph neural network (AHSTGNN) proposed by Wang et al. in 2023 adopts the method of adding a gate to add the static graph and the dynamic graph by weight. However, these combination methods still need to be improved, and they cannot take advantage of the attention mechanism to capture the complex spatio-temporal dependence of base station traffic.

[0005] Therefore, it is necessary to design a better algorithm that can coordinate the capture of static features and the capture of dynamic features to achieve better prediction performance. SUMMARY

[0006] The purpose of the present application is to provide a multi-node base station traffic prediction method based on self-organizing graph neural network, and to provide an effective prediction method for the change trend of base station traffic by proposing a spatial feature extraction mechanism for multi-node base station traffic.

[0007] To achieve the above-mentioned purpose, the present application provides a multi-node base station traffic prediction method based on self-organizing graph neural network, comprising the following steps:

[0008] S1, divide a region into a plurality of sub-regions, divide the base stations in each sub-region into N nodes, collect node data, generate an adjacency matrix and evaluate influence;

[0009] S2, construct a neural network model;

[0010] S3, define a loss function and an optimizer, take Smooth_l1_loss as the loss function, and use the Adam optimizer to update the model parameters;

[0011] S4, bring the prediction result and the true value into the loss function to calculate the loss value, use the gradient descent algorithm for back propagation, and update the model parameters;

[0012] S5, repeating S4 until the neural network model converges, inputting the test set into the neural network model to evaluate the prediction performance of the neural network model.

[0013] Preferably, the neural network model in S2 comprises the following component modules:

[0014] Node fusion module: including a randomly initialized mapping matrix, which realizes self-organizing recombination of the node set and maps the node set into a vector base, and guides the convergence of the mapping matrix according to the node influence;

[0015] Node spatial feature convolution module: containing static convolution and dynamic convolution; first, the attention score matrix is calculated in time sequence for the node set and the vector base, the node set is mapped into the vector base again by using the mapping matrix after static graph convolution, and dynamic convolution is performed by using the attention score matrix;

[0016] Node time sequence feature extraction module: long short-term memory network is inserted between multiple layers of the node spatial feature convolution module to learn time sequence features, and the outputs of each layer are combined together and aggregated by a fully connected layer;

[0017] Output module: after the outputs of each layer are aggregated by a fully connected layer, standardization is performed according to layer standardization, and then a prediction result is output by a multilayer perceptron.

[0018] Preferably, the mapping matrix is generated for base station node recombination, and the convergence direction of the mapping matrix is guided by the node influence.

[0019] Preferably, the specific process of the base station node recombination comprises: in S1, performing row normalization on the adjacency matrix, and then calculating the column sum to represent the influence size of each base station node; and in the node fusion module of S2, the convergence direction of the mapping matrix is guided by converting the base station node influence into a diagonal matrix and multiplying it with the mapping matrix; then the node set is mapped into K linearly independent vector bases by the mapping matrix with a shape of (S*N, K) to participate in the calculation of the attention score matrix in S3.

[0020] Wherein, S is the number of sub-regions, N is the number of nodes in a node set, K is the number of vector bases generated by condensing the node set, and is between 1 and S*N.

[0021] Preferably, a partial attention mechanism is proposed by using the mapping matrix to improve the calculation strategy of attention, which comprises: in the node spatial feature convolution module of S2, the attention score matrix is calculated between the node set and the vector base, and after the graph convolution is performed on the node set, the transpose of the mapping matrix, the transpose of the attention score matrix are multiplied in turn to obtain the node set of the next layer.

[0022] Therefore, the application adopts the above-mentioned multi-node base station traffic prediction method based on a self-organizing graph neural network, and proposes a spatial feature extraction mechanism for multi-node base station traffic, thereby providing an effective prediction method for the change trend of base station traffic.

[0023] The technical solutions of the application will be further described in detail below with reference to the drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 is a whole flow chart of a multi-node base station traffic prediction method based on a self-organizing graph neural network of the application;

[0025] Figure 2 is a self-organizing graph neural network structure diagram of an embodiment of the multi-node base station traffic prediction method based on a self-organizing graph neural network of the application;

[0026] Figure 3 is a schematic diagram of a node dynamic relationship generation process of an embodiment of the multi-node base station traffic prediction method based on a self-organizing graph neural network of the application;

[0027] Figure 4 is a prediction effect diagram of a final model on cellular traffic of an embodiment of the multi-node base station traffic prediction method based on a self-organizing graph neural network of the application. DETAILED DESCRIPTION

[0028] The technical solutions of the application will be further described in detail below with reference to the drawings and embodiments.

[0029] Unless otherwise defined, the technical terms or scientific terms used in the application shall have the usual meanings understood by those skilled in the art to which the application belongs.

[0030] Embodiment one

[0031] As shown in Figure 1 , the application provides a multi-node base station traffic prediction method based on a self-organizing graph neural network, comprising the following steps:

[0032] S1, divide a region into multiple sub-regions, divide the base stations in each sub-region into N nodes, collect node data, generate an adjacency matrix and evaluate influence;

[0033] S2, construct a neural network model, comprising the following component modules:

[0034] The node fusion module includes a random initialization mapping matrix, which will realize self-organizing reorganization of the node set, map the node set to a vector base, and guide the mapping matrix to converge according to the node influence;

[0035] Node spatial feature convolution module: contains static convolution and dynamic convolution; first, the node set and vector base are calculated in time sequence to obtain the attention score matrix, and the node set is mapped into the vector base again after static graph convolution, and dynamic convolution is performed using the attention score matrix;

[0036] Node temporal feature extraction module: long short-term memory network is used to learn temporal features between multiple layers of the node spatial feature convolution module, and the outputs of each layer are combined together and aggregated by a fully connected layer;

[0037] Output module: after aggregating the outputs of each layer by a fully connected layer, standardization is performed according to layer standardization, and then a multi-layer perceptron is used to output the prediction result.

[0038] S3, define a loss function and an optimizer, use Smooth_l1_loss as the loss function, and use Adam optimizer to update the model parameters;

[0039] S4, input the prediction result and the true value into the loss function to calculate the loss value, use the gradient descent algorithm for back propagation, and update the model parameters;

[0040] S5, repeat S4 until the neural network model converges, and input the test set into the neural network model to evaluate the prediction performance of the neural network model.

[0041] The mapping matrix is generated for base station node reorganization, and the node influence is used to guide the convergence direction of the mapping matrix. The base station node reorganization includes row normalization of the adjacency matrix in S1, and then the column sum is calculated to represent the influence size of each base station node; and in the node fusion module of S2, the base station node influence is converted into a diagonal matrix, and multiplied by the mapping matrix to guide the convergence direction of the mapping matrix; then the node set is mapped into K linearly independent vector bases by the mapping matrix with a shape of (S*N, K) to participate in the calculation of the attention score matrix in S3;

[0042] Wherein, S is the number of sub-regions, N is the number of nodes in a node set, K is the number of vector bases generated by condensing the node set, and is between 1 and S*N.

[0043] The mapping matrix is used to improve the calculation strategy of attention, and a partial attention mechanism is proposed, including S2, the node spatial feature convolution module, the attention score matrix is calculated between the node set and the vector base, and after the graph convolution is performed on the node set, the transpose of the mapping matrix, the transpose of the attention score matrix are multiplied in turn to obtain the node set of the next layer.

[0044] Embodiment two

[0045] A prediction method based on a self-organizing graph neural network, comprising the following steps:

[0046] S1: Collect datasets. Directly collect existing public datasets, such as Milan and Trentino datasets, or regional traffic datasets uploaded by telecommunications operators such as China Telecom and China Mobile to public data platforms. This embodiment uses the Milan public dataset for experimentation.

[0047] S2: Dataset Preprocessing. The dataset is processed according to the dimensions of time axis, spatial length, spatial width, and feature types to obtain a standardized dataset. Then, subgraphing is performed along the spatial length and spatial width dimensions, and multiple samples are divided according to the time axis dimension. Finally, the sample set is divided into two parts: a training set and a test set. Optionally, the training set accounts for 80% and the test set accounts for 20%.

[0048] S3: Generate training aids. For each subset of the normalized dataset after subgraphization, calculate the cosine similarity between any two nodes, using the following formula:

[0049]

[0050] in, It is the flow value of the i-th node in the s-th subgraph at time t. It is the flow value of the j-th node in the s-th subgraph at time t, a s,i,j Let be the cosine similarity between the i-th node and the j-th node in the s-th subgraph. Use the cosine similarity between nodes to construct an adjacency matrix.

[0051] S4: As Figure 2 As shown, the adjacency matrices and node sets of multiple subgraphs are input into the model. The model mainly consists of a fusion module (FM), a static convolution module (SCM), a dynamic convolution module (DCM), and a recurrent neural network gated structure (RNN Gated Structure).

[0052] The fusion module includes row diagonalization of the adjacency matrix, two linear layers, and a diagonalization operation. The specific operation can be represented by the following formula:

[0053]

[0054]

[0055]

[0056] E = Diag({e s，j} s＝1，2，...S，j＝1,2，...N );

[0057] M = E × W r ;

[0058] Among them, A sis the adjacency matrix of subgraph s, a s，i，j represents A s , the element in the i-th row and j-th column of A represents the row sum of the elements in A s . Diag(*) represents the operation of converting a vector into a diagonal matrix, D s is the degree matrix of A s , A′ s is the result of row normalization of A s using the degree matrix. a′ s，i，j represents the element in the i-th row and j-th column of A′ s . represents the column sum of the elements in A′ s . e s，j is the influence of the j-th node in the s-th subgraph. w e and b e are learnable parameters. E is the node influence matrix, W r is the linear layer weight for generating the mapping matrix, which is also a learnable parameter. M is the mapping matrix, which will finally be passed to the dynamic convolution layer.

[0059] The static convolution module includes graph static convolution and position static convolution. The graph static convolution is to perform graph neural network convolution on each subgraph according to the respective adjacency matrix, and the position static convolution is to reassemble all subgraphs according to the original positional relationship and perform convolution on adjacent nodes.

[0060] The dynamic convolution module includes calculating an attention score matrix and attention convolution. Calculating the attention score matrix refers to calculating the relevance between the node set before static convolution as the query side and the vector base (also known as the node representative) obtained by converting the node set before static convolution by the mapping matrix as the queried side. The attention convolution is to perform convolution on the node set after static convolution using the attention score matrix. The node connection change diagram in this process is shown in FIG. 3. At the same time, this process can be represented by the following formula: Figure 3

[0061]

[0062]

[0063]

[0064]

[0065]

[0066] attn=softmax(attn1||attn2);

[0067] ​

[0068] where W e is the weight matrix used to map vectors to embedding dimension, W q is the weight matrix used to map vectors to query dimension, W k is the weight matrix used to map vectors to key dimension, all of which are learnable parameters. Q allnodes denotes the query matrix formed by mapping the node set to the query dimension, K allnodes denotes the key matrix formed by mapping the node set to the key dimension, K keynodes denotes the key matrix formed by mapping the node representatives to the key dimension. attn'1 is the attention matrix of the node set to the node representatives, attn'2 is the vector formed by each node in the node set to itself attention, ∑ col (Q allnodes ·K allnodes ) denotes the dot product of Q allnodes matrix and K allnodes matrix and then sum up in the encoding dimension, and the result will be a vector. d model is the dimension number of the embedding layer. attn is the attention score matrix of each node in the node set to the node representatives and itself, the biggest difference with attn'1 is to consider the weight of itself by using attn'2, to avoid being completely affected by the node representatives and ignoring the original features of itself. softmax(attn'1||attn'2) denotes that attn'1 matrix and attn'2 vector are combined together, and then the attention weight distribution is converted to score distribution by using softmax function, and the total score of any node to other nodes is 1. And attn1 and attn2 are the reverse operation of combining attn'1 matrix and attn'2 vector together, to re-split attn and act on node representatives and node itself for convolution respectively. X is the node set before the static convolution module, X' is the node set after the static convolution module, and X" is the output of the dynamic convolution module.

[0069] The recurrent neural network gating unit module uses long short-term memory network (LSTM) to extract time sequence features.

[0070] Finally, a multi-layer stacking method is needed to repeatedly extract spatial features and time features, and finally use a fully connected layer with an input dimension equal to the number of stacked layers to integrate the outputs of each layer, then use layer normalization for standardization, and output the prediction result through a three-layer perceptron with relu as the activation function.

[0071] S5: the loss function is selected as smooth_l1_loss, and the optimizer is selected as Adam optimizer. By calculating the loss value between the predicted value and the true value, the gradient of each model parameter is obtained, and the gradient descent method is used for back propagation, and the model parameters are updated by using the Adam optimizer.

[0072] S6: repeat S4 and S5 until the model converges. Test the model with the test set to ensure that the model prediction is effective, and output the final model.

[0073] As shown in Figure 4 , the prediction effect of the present example experiment is shown. The true value of the predicted data comes from the traffic variation of a certain central area node in the Milan data set over a period of time.

[0074] Therefore, the present application adopts the above-mentioned multi-node base station traffic prediction method based on self-organizing graph neural network, and proposes a spatial feature extraction mechanism for multi-node base station traffic, providing an effective prediction method for the change trend of base station traffic.

[0075] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application and not to limit them, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: it can still modify or equivalently replace the technical solutions of the present application, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A method for predicting traffic to a multi-node base station based on a self-organizing graph neural network, characterized in that: Includes the following steps: S1. Divide a region into multiple sub-regions. Each sub-region has N base stations. Collect node data, generate an adjacency matrix, and evaluate the influence. S2. Construct a neural network model; S3. Define the loss function and optimizer, using Smooth_l1_loss as the loss function and Adam optimizer to update the model parameters; S4. Substitute the predicted results and the true values ​​into the loss function to calculate the loss value, and use the gradient descent algorithm for backpropagation to update the model parameters; S5. Repeat S4 until the neural network model converges, and input the test set into the neural network model to evaluate the prediction performance of the neural network model. The neural network model in S2 includes the following modules: a node fusion module, which includes a randomly initialized mapping matrix that enables the self-organization and recombination of the node set, mapping the node set to a vector basis, and guiding the convergence of the mapping matrix based on the node influence; a node spatial feature convolution module, which includes static convolution and dynamic convolution; firstly, the attention score matrix is ​​calculated temporally on the node set and the vector basis, and after static graph convolution, the node set is mapped to a vector basis again using the mapping matrix, and then dynamic convolution is performed using the attention score matrix; and a node temporal feature extraction module, which interweaves a long short-term memory network to learn temporal features between the multiple layers of the node spatial feature convolution module, and merges the outputs of each layer together and aggregates them using a fully connected layer. Output module: The outputs of each layer are aggregated by a fully connected layer, standardized according to layer standardization, and then output as prediction results through a multilayer perceptron; The specific process of base station node reorganization includes: in S1, performing row normalization on the adjacency matrix and then calculating the column sum to represent the influence of each base station node; and in the node fusion module of S2, transforming the influence of base station nodes into a diagonal matrix and multiplying it with the mapping matrix to guide the convergence direction of the mapping matrix; then, mapping the node set into K linearly independent vector bases using a mapping matrix of shape (S*N,K) to participate in the attention score matrix calculation in S3; where S is the number of sub-regions, N is the number of nodes in a node set, and K refers to the number of vector bases generated by condensing the node set, which is between 1 and S*N.

2. The method for predicting traffic of a multi-node base station based on a self-organizing graph neural network according to claim 1, characterized in that: The mapping matrix is ​​generated for base station node reorganization, and the node influence is used to guide the convergence direction of the mapping matrix.

3. The method for predicting traffic of a multi-node base station based on a self-organizing graph neural network according to claim 2, characterized in that: A partial attention mechanism is proposed by using a mapping matrix to improve the attention calculation strategy. This mechanism includes the node spatial feature convolution module in S2, which calculates the attention between the node set and the vector basis to generate an attention score matrix. After performing graph convolution on the node set, it is multiplied by the transpose of the mapping matrix and the transpose of the attention score matrix in sequence to obtain the node set of the next layer.

Images