A base station traffic prediction method based on joint representation of space-time features

Abstract

The present application relates to the technical field of electric digital data processing, and especially relates to a base station traffic prediction method based on space-time feature joint representation. A discrete Fourier transform identification operation is performed on historical traffic data of each base station, and a continuous traffic time sequence is split to obtain a traffic data block; a base station traffic space-time correlation graph is constructed with the traffic data block as a node; the structural entropy of the base station traffic space-time correlation graph is minimized to obtain an optimal coding tree; a multi-depth convolution operation is performed on the base station traffic space-time correlation graph in the frequency domain to obtain a space-time feature extraction result; a base station traffic prediction model is trained according to a composite loss function and the space-time feature extraction result to obtain a target prediction model; the base station traffic value of the current time step output by the target prediction model is taken as the input of the next time step for self-recurrence iterative prediction to obtain traffic prediction values of at least one future time point of each base station. The purpose of improving the accuracy and efficiency of base station traffic prediction is achieved.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and in particular to a base station traffic prediction method based on joint spatiotemporal feature characterization. Background Technology

[0002] With the continuous increase in energy consumption of communication networks, people have higher requirements for energy saving of base stations. Relying solely on directly reducing power consumption to achieve energy saving in base stations will obviously lead to insufficient signal coverage and a decline in user experience. Therefore, while achieving green energy saving in base stations, it is also necessary to ensure the quality of network services.

[0003] To predict base station traffic and help operators optimize resource allocation and ensure user speeds before peak traffic periods, while reducing unnecessary energy consumption during off-peak periods to balance energy efficiency and user experience, the method typically involves capturing the inherent dependencies between historical traffic patterns. For example, a multi-head self-attention (MLSA) method can be developed to enhance feature weights, capture long-short-term temporal dependencies using a Long Short-Term Memory (LSTM) network, and extract global multi-scale features through dilated causal convolutions and residual blocks in a Temporal Convolutional Network (TCN). This addresses the limitations of traditional models in characterizing long-short-term dependencies and limiting prediction accuracy. For example, after encoding the traffic sequence into a hidden state using a gated recurrent unit, TCN (temporal convolutional network) and LSTM (long short-term memory network) are used in parallel to extract local short-term features and global long-term features, respectively. Then, the two types of features are fused by bidirectional weighting and finally output the traffic values ​​for future multiple time periods through a fully connected layer. This solves the problem of base station traffic prediction methods that can only predict in the short term or have complex long-term models and poor generalization.

[0004] However, focusing solely on capturing the intrinsic dependencies between historical traffic flows while ignoring the impact of spatial correlations between base stations on traffic prediction limits the accuracy of base station traffic forecasting. In other words, base station traffic prediction methods that only consider the traffic changes of a single base station over time and do not account for the spatial correlations between base stations suffer from limited prediction accuracy.

[0005] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0006] The main objective of this invention is to provide a base station traffic prediction method based on joint spatiotemporal feature representation, which aims to solve the problem of limited prediction accuracy in base station traffic prediction methods that only focus on the traffic changes of a single base station in the time dimension.

[0007] To achieve the above objectives, the present invention provides a base station traffic prediction method based on joint spatiotemporal feature representation, the method comprising the following steps:

[0008] Step S100: Perform Discrete Fourier Transform identification on the historical traffic data of each base station, and segment the continuous traffic time series to obtain traffic data blocks;

[0009] Step S200: Based on the metric function, construct a spatiotemporal correlation graph of base station traffic with the traffic data block as the graph node, wherein the metric function is a metric function that integrates the spatiotemporal characteristics of traffic;

[0010] Step S300: According to structural information theory, minimize the structural entropy of the spatiotemporal correlation graph of the base station traffic to obtain the optimal coding tree;

[0011] Step S400: Perform multi-depth convolution operation on the spatiotemporal correlation graph of the base station traffic in the frequency domain to obtain spatiotemporal feature extraction results;

[0012] Step S500: Construct a contrastive loss based on the optimal coding tree partition, and construct a composite loss function based on the Huber loss and the contrastive loss;

[0013] Step S600: Train the base station traffic prediction model based on the composite loss function and the spatiotemporal feature extraction results to obtain the target prediction model;

[0014] Step S700: Use the base station traffic value at the current time step output by the target prediction model as the input for the next time step, perform autoregressive iterative prediction, and obtain the traffic prediction value for each base station at least one future time point.

[0015] Optionally, the step of constructing a spatiotemporal correlation graph of base station traffic with the traffic data blocks as graph nodes based on the metric function includes:

[0016] Step S210: Embed all the traffic data blocks, and n Each of the traffic data blocks is converted into a vector of a preset length. ;

[0017] Step S220: Based on the metric function, with nEach of the traffic data blocks is a graph node, constructing a spatiotemporal correlation graph of the base station traffic, wherein the metric function is a similarity metric function between graph nodes; and,

[0018] Step S230: Quantify the comprehensive similarity of the two graph nodes in terms of geographical location, temporal proximity, and traffic pattern according to the quantification formula, which is expressed as follows:

[0019] ,

[0020] In the formula, , and These are the weighting coefficients. and express and The time corresponding to the two nodes, and express and The base station locations corresponding to the two nodes Represents a node A vector of a preset length.

[0021] Optionally, the step of minimizing the structural entropy of the spatiotemporal correlation graph of the base station traffic to obtain the optimal coding tree according to structural information theory includes:

[0022] Step S311: Based on the root node of the base station traffic spatiotemporal correlation graph, obtain the structural entropy of the base station traffic spatiotemporal correlation graph under the coding tree partition;

[0023] Step S312: Minimize the structural entropy of the spatiotemporal correlation graph of base station traffic based on the preset coding tree height and the structural entropy of the coding tree-divided spatiotemporal correlation graph.

[0024] Optionally, according to structural information theory, the step of minimizing the structural entropy of the spatiotemporal correlation graph of the base station traffic to obtain the optimal coding tree further includes:

[0025] Step S321: Initialize the graph nodes of the spatiotemporal correlation graph of the base station traffic to obtain n The traffic data block of the first independent base station is divided;

[0026] Step S322: Traverse all the first base station traffic data block divisions and merge the first base station traffic data block divisions based on the preset structural entropy minimization condition to obtain the second base station traffic data block divisions;

[0027] Step S323: Combine the second base station traffic data block partitions with different temporal and spatial semantics to obtain the third base station traffic data block partitions;

[0028] Step S324: Repeat steps S322 to S323 until the first base station traffic data block partition cannot be merged and the second base station traffic data block partition cannot be combined, to obtain the target base station traffic data block partition.

[0029] Step S325: Construct the optimal coding tree based on the target base station traffic data block partitioning.

[0030] Optionally, the step of performing multi-depth convolution operations on the spatiotemporal correlation graph of the base station traffic in the frequency domain to obtain spatiotemporal feature extraction results includes:

[0031] Step S410: Based on a parameter matrix, map the connection relationship between nodes in the spatiotemporal correlation graph of the base station traffic to a Green kernel;

[0032] Step S420: Perform a discrete Fourier transform on the Green kernel to obtain a frequency domain convolution kernel;

[0033] Step S430: Perform multi-layer convolution on the spatiotemporal correlation graph of the base station traffic according to the frequency domain convolution kernel.

[0034] Optionally, the step of performing multi-depth convolution operations on the spatiotemporal correlation graph of the base station traffic in the frequency domain to obtain spatiotemporal feature extraction results further includes:

[0035] Step S440: Obtain the spatiotemporal correlation graph operator based on the convolution kernel;

[0036] Step S450: Pass the multi-depth convolution stacking formula of the spatiotemporal correlation graph operator to perform multi-depth convolution operation on the spatiotemporal correlation graph of base station traffic to capture the correlation of different distance levels in the spatiotemporal correlation graph of base station traffic; the multi-depth convolution stacking formula is expressed as follows:

[0037] ,

[0038] In the formula, P The convolution depth is . This is the result of superimposing base station traffic across multiple depthwise convolutions in the frequency domain. for The bias parameters corresponding to the convolutional layers. For activation function, The feature matrix represents the characteristics of base station traffic data blocks. Spatiotemporal correlation graph operator conduct Multiplication by multiplication. This represents the discrete Fourier transform of a matrix.

[0039] Optionally, the step of constructing a contrastive loss based on the optimal coding tree partition, and constructing a composite loss function based on the Huber loss and the contrastive loss, includes:

[0040] Step S510: Divide the positive and negative samples of the contrast loss according to the division of base station traffic data blocks in the optimal coding tree;

[0041] Step S520: Sum the Huber loss and the contrast loss to obtain the composite loss function;

[0042] The step of dividing the contrastive loss into positive and negative samples includes:

[0043] Base station traffic data blocks within the same partition are considered positive samples;

[0044] Base station traffic data blocks between different partitions are used as negative samples.

[0045] Optionally, the step of training the base station traffic prediction model based on the composite loss function and the spatiotemporal feature extraction results to obtain the target prediction model includes:

[0046] Step S610: Based on the spatiotemporal feature extraction results, obtain a temporal domain flow data block representation;

[0047] Step S620: Based on the traffic prediction weight matrix, map the time-domain traffic data block representation to the output of the prediction length to obtain the base station traffic prediction value;

[0048] Step S630: Train the base station traffic prediction model based on the stochastic gradient descent algorithm, the predicted base station traffic value, and the composite loss function to obtain the target prediction model.

[0049] In addition, to achieve the above objectives, the present invention also provides a base station traffic prediction device, which includes a memory, a processor, and a base station traffic prediction program based on spatiotemporal feature joint characterization stored in the memory and executable on the processor. When the base station traffic prediction program based on spatiotemporal feature joint characterization is executed by the processor, it implements the steps of the base station traffic prediction method based on spatiotemporal feature joint characterization as described above.

[0050] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a base station traffic prediction program based on spatiotemporal feature joint characterization, wherein when the base station traffic prediction program based on spatiotemporal feature joint characterization is executed by a processor, it implements the steps of the base station traffic prediction method based on spatiotemporal feature joint characterization as described above.

[0051] This invention provides a base station traffic prediction method based on joint spatiotemporal feature representation. It constructs a spatiotemporal correlation graph of base station traffic with traffic data blocks as nodes by employing a base station traffic data block segmentation strategy that ensures the semantic integrity of base station traffic. Based on structural information theory, it extracts spatiotemporal hierarchical features of base station traffic, using Fourier graph operators to achieve simultaneous extraction of spatiotemporal features, thus achieving efficient extraction and generating a base station traffic data block representation with spatiotemporal characteristics, i.e., the spatiotemporal feature extraction results. A contrastive learning mechanism is introduced to enhance the ability to identify complex spatiotemporal features between traffic data blocks, thereby improving the accuracy of base station traffic prediction. This improves the accuracy and efficiency of base station traffic prediction, providing a reference for base station energy efficiency optimization strategy formulation and intelligent network resource scheduling. Attached Figure Description

[0052] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present invention and, together with the description, serve to explain the principles of the invention. To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0053] Figure 1 This is a schematic diagram of the hardware operating environment of the base station traffic prediction device according to an embodiment of the present invention;

[0054] Figure 2 This is a flowchart illustrating the first embodiment of the base station traffic prediction method based on joint spatiotemporal feature representation of the present invention.

[0055] Figure 3 This is a schematic diagram of the architecture of the first embodiment of the base station traffic prediction method based on joint spatiotemporal feature representation of the present invention.

[0056] Figure 4 This is a flowchart illustrating the second embodiment of the base station traffic prediction method based on joint spatiotemporal feature representation of the present invention.

[0057] Figure 5 This is a schematic diagram of the construction of the spatiotemporal correlation graph of base station traffic in a second embodiment of the base station traffic prediction method based on joint spatiotemporal feature representation of the present invention.

[0058] Figure 6 This is a flowchart illustrating the third embodiment of the base station traffic prediction method based on joint spatiotemporal feature representation of the present invention.

[0059] Figure 7 This is a schematic diagram of the coding tree construction in the third embodiment of the base station traffic prediction method based on spatiotemporal feature joint representation of the present invention.

[0060] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0061] To better understand the above technical solutions, exemplary embodiments of this disclosure will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of this disclosure to those skilled in the art.

[0062] As one implementation scheme, Figure 1 This is a schematic diagram of the hardware operating environment of the base station traffic prediction device involved in the embodiments of the present invention.

[0063] like Figure 1 As shown, the base station traffic prediction device may include: a processor 101, such as a central processing unit (CPU), a memory 102, and a communication bus 103. The memory 102 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. Optionally, the memory 102 may also be a storage device independent of the aforementioned processor 101. The communication bus 103 is used to enable communication between these components.

[0064] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the base station traffic prediction device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0065] like Figure 1 As shown, the memory 102, which is a computer-readable storage medium, may include an operating system, a data storage module, a network communication module, a user interface module, and a base station traffic prediction method program based on spatiotemporal feature joint characterization.

[0066] exist Figure 1 In the base station traffic prediction device shown, the processor 101 and the memory 102 can be installed in the base station traffic prediction device. The base station traffic prediction device calls the base station traffic prediction program based on spatiotemporal feature joint representation stored in the memory 102 through the processor 101, and performs the following operations:

[0067] Perform Discrete Fourier Transform identification on the historical traffic data of each base station, and segment the continuous traffic time series to obtain traffic data blocks;

[0068] Based on the metric function, a spatiotemporal correlation graph of base station traffic is constructed with the traffic data blocks as graph nodes. The metric function is a metric function that integrates the spatiotemporal characteristics of traffic.

[0069] According to structural information theory, the optimal coding tree is obtained by minimizing the structural entropy of the spatiotemporal correlation graph of the base station traffic.

[0070] Multi-depth convolution operation is performed on the spatiotemporal correlation graph of the base station traffic in the frequency domain to obtain spatiotemporal feature extraction results;

[0071] Construct a contrastive loss based on the optimal coding tree partition, and construct a composite loss function based on the Huber loss and the contrastive loss;

[0072] The base station traffic prediction model is trained based on the composite loss function and the spatiotemporal feature extraction results to obtain the target prediction model;

[0073] The base station traffic value at the current time step output by the target prediction model is used as the input for the next time step to perform autoregressive iterative prediction, thereby obtaining the traffic prediction value for each base station at least one future time point.

[0074] In one embodiment, the processor 101 can be used to call a base station traffic prediction program based on spatiotemporal feature joint representation stored in the memory 102, and perform the following operations:

[0075] Embed all of the aforementioned traffic data blocks, n Each of the traffic data blocks is converted into a vector of a preset length. ;

[0076] Based on the aforementioned metric function, n Each of the traffic data blocks is a graph node, constructing a spatiotemporal correlation graph of the base station traffic, wherein the metric function is a similarity metric function between graph nodes; and,

[0077] The quantitative formula is used to quantify the comprehensive similarity between two graph nodes in terms of geographical location, temporal proximity, and traffic patterns.

[0078] In one embodiment, the processor 101 can be used to call a base station traffic prediction program based on spatiotemporal feature joint representation stored in the memory 102, and perform the following operations:

[0079] Based on the root node of the spatiotemporal correlation graph of base station traffic, obtain the structural entropy of the spatiotemporal correlation graph of base station traffic under the coding tree partition;

[0080] Based on the preset coding tree height and the structural entropy of the spatiotemporal correlation graph of base station traffic under the coding tree division, the structural entropy of the spatiotemporal correlation graph of base station traffic is minimized.

[0081] In one embodiment, the processor 101 can be used to call a base station traffic prediction program based on spatiotemporal feature joint representation stored in the memory 102, and perform the following operations:

[0082] Initialize the graph nodes of the spatiotemporal correlation graph of the base station traffic to obtain n The traffic data block of the first independent base station is divided;

[0083] Traverse all the first base station traffic data block partitions and merge the first base station traffic data block partitions based on the preset structural entropy minimization condition to obtain the second base station traffic data block partitions;

[0084] By combining the second base station traffic data block partitions with different temporal and spatial semantics, a third base station traffic data block partition is obtained;

[0085] Repeat the process until the first base station traffic data block partition cannot be merged and the second base station traffic data block partition cannot be combined, to obtain the target base station traffic data block partition;

[0086] The optimal coding tree is constructed based on the target base station traffic data block partitioning.

[0087] In one embodiment, the processor 101 can be used to call a base station traffic prediction program based on spatiotemporal feature joint representation stored in the memory 102, and perform the following operations:

[0088] Based on a parameter matrix, the connection relationship between nodes in the spatiotemporal correlation graph of base station traffic is mapped to a Green kernel;

[0089] Perform a discrete Fourier transform on the Green kernel to obtain a frequency domain convolution kernel;

[0090] Based on the frequency domain convolution kernel, multi-layer convolution is performed on the spatiotemporal correlation graph of the base station traffic.

[0091] In one embodiment, the processor 101 can be used to call a base station traffic prediction program based on spatiotemporal feature joint representation stored in the memory 102, and perform the following operations:

[0092] Based on the convolution kernel, a spatiotemporal correlation graph operator is obtained;

[0093] The multi-depth convolution stacking formula of the spatiotemporal correlation graph operator is passed to perform multi-depth convolution operation on the spatiotemporal correlation graph of base station traffic in order to capture the correlation of different distance levels in the spatiotemporal correlation graph of base station traffic.

[0094] In one embodiment, the processor 101 can be used to call a base station traffic prediction program based on spatiotemporal feature joint representation stored in the memory 102, and perform the following operations:

[0095] Based on the division of base station traffic data blocks in the optimal coding tree, the positive and negative samples of the contrast loss are divided.

[0096] The Huber loss and the contrast loss are summed to obtain the composite loss function.

[0097] In one embodiment, the processor 101 can be used to call a base station traffic prediction program based on spatiotemporal feature joint representation stored in the memory 102, and perform the following operations:

[0098] Based on the spatiotemporal feature extraction results, a temporal traffic data block representation is obtained;

[0099] Based on the traffic prediction weight matrix, the time-domain traffic data block representation is mapped to the output of the prediction length to obtain the base station traffic prediction value;

[0100] The base station traffic prediction model is trained based on the stochastic gradient descent algorithm, the predicted base station traffic value, and the composite loss function to obtain the target prediction model.

[0101] Based on the hardware architecture of the aforementioned base station traffic prediction device, an embodiment of the base station traffic prediction method based on spatiotemporal feature joint characterization of the present invention is proposed.

[0102] Reference Figure 2 , Figure 3 In the first embodiment, the base station traffic prediction method based on joint spatiotemporal feature representation includes the following steps:

[0103] Step S100: Perform Discrete Fourier Transform identification on the historical traffic data of each base station, and segment the continuous traffic time series to obtain traffic data blocks;

[0104] Step S200: Based on the metric function, construct a spatiotemporal correlation graph of base station traffic with the traffic data blocks as graph nodes.

[0105] In this embodiment, steps S100 and S200 are preprocessing steps for base station traffic spatiotemporal information. Specifically, to ensure the semantic integrity of the segmented traffic data, a Discrete Fourier Transform (DFT) is used to identify the main period of historical traffic data for each base station, and this period is used as the segmentation interval to divide the continuous traffic time series into multiple traffic data blocks. Then, a metric function that comprehensively considers the spatiotemporal characteristics of traffic is used to calculate the spatiotemporal correlation degree between any two traffic data blocks, and based on this, a base station traffic spatiotemporal correlation graph with traffic data blocks as nodes is constructed.

[0106] Step S300: According to structural information theory, minimize the structural entropy of the spatiotemporal correlation graph of the base station traffic to obtain the optimal coding tree;

[0107] Step S400: Perform multi-depth convolution operation on the spatiotemporal correlation graph of the base station traffic in the frequency domain to obtain spatiotemporal feature extraction results.

[0108] In this embodiment, steps S300 and S400 are steps for extracting spatiotemporal correlation features of base station traffic.

[0109] Structural entropy, an extension of Shannon entropy, measures the uncertainty of graph structure and is calculated by hierarchically partitioning nodes by degree. A coding tree, as the optimal hierarchical structure of a graph, reveals the relationships between different levels of data.

[0110] Based on structural information theory, the optimal coding tree for the spatiotemporal correlation graph of base station traffic is generated by minimizing the structural entropy of the graph. This optimal coding tree, which serves as a spatiotemporal coding tree for base station traffic, can be used as a spatiotemporal hierarchical representation of base station traffic, thereby enabling the simultaneous capture of the hierarchical correlation of traffic data in both time and space dimensions.

[0111] For the above-mentioned multi-depth convolution operation, the frequency domain multi-depth convolution of the base station traffic spatiotemporal correlation graph designed based on the Fourier graph operator is adopted. The graph convolution operation is transformed into the frequency domain. By performing multi-depth convolution on the base station traffic spatiotemporal correlation graph in the frequency domain, the feature information of different base station traffic data blocks is aggregated simultaneously to generate a base station traffic data block representation that integrates spatiotemporal features.

[0112] Optionally, step S300 includes obtaining the structural entropy of the base station traffic spatiotemporal correlation graph under the coding tree division based on the root node of the base station traffic spatiotemporal correlation graph; and then minimizing the structural entropy of the base station traffic spatiotemporal correlation graph according to the preset coding tree height and the structural entropy of the base station traffic spatiotemporal correlation graph under the coding tree division.

[0113] Understandably, spatially, the hierarchy of base stations is reflected in the coexistence and collaborative operation of base stations with different coverage areas and operating frequency bands. For example, macro base stations operating in low-frequency bands achieve wide signal coverage due to their strong penetration capabilities, while micro base stations and pico base stations exist under the coverage of macro base stations, playing roles such as load balancing and signal enhancement. Temporally, the degree of influence between traffic data blocks at different time spans varies; therefore, the degree of mutual influence between traffic data blocks at different time spans also exhibits a certain hierarchy.

[0114] In this embodiment, the optimal coding tree is generated by minimizing the structural entropy of the spatiotemporal correlation graph of base station traffic, and then the hierarchical features of base station traffic are extracted based on the coding tree. That is, hierarchical feature extraction of base station traffic based on structural entropy.

[0115] Specifically, by minimizing the spatiotemporal correlation graph of base station traffic. structural entropy Construct the optimal encoding tree for a spatiotemporal correlation graph of base station traffic. This not only reduces structural noise in the spatiotemporal correlation graph of base station traffic that interferes with base station traffic prediction, but also effectively represents the spatiotemporal hierarchical information of base station traffic. The expression for minimum structural entropy can be expressed as:

[0116] .

[0117] in, Represents the coding tree height, It is the preset coding tree height. Covering all heights not exceeding The coding tree, It is a coding tree Delineate the structural entropy of the spatiotemporal correlation graph of base station traffic. The calculation method is as follows:

[0118] ;

[0119] .

[0120] in, A partition is a set consisting of several graph nodes or subpartitions. ; for The total weight of the edges between it and its complement; for The parent node is used for partitioning; The total weight of the edges in the partition; for The root node.

[0121] Optionally, step S400 includes:

[0122] Step S410: Based on a parameter matrix, map the connection relationship between nodes in the spatiotemporal correlation graph of the base station traffic to a Green kernel;

[0123] Step S420: Perform a discrete Fourier transform on the Green kernel to obtain a frequency domain convolution kernel;

[0124] Step S430: Perform multi-layer convolution on the spatiotemporal correlation graph of the base station traffic according to the frequency domain convolution kernel.

[0125] Understandably, when the number of base stations and the scale of base station traffic data are too large, the spatiotemporal correlation graph of base station traffic constructed through steps S100 to S200 will show a significant expansion in scale, which in turn leads to a severe challenge to the feature extraction process.

[0126] In this embodiment, the Fourier transform mechanism is used to achieve joint modeling and efficient extraction of the structural features and dynamic traffic features of large-scale base station networks.

[0127] Specifically, the formula Give the process of one convolution in the temporal domain of graph convolution. Wherein, This is the initial node feature representation, i.e., the feature representation after embedding the traffic data blocks. , The adjacency matrix of the spatiotemporal correlation graph of the corresponding base station traffic. For activation function, For parameter matrices, For the embedded dimension.

[0128] The above convolution operation can be transformed into the frequency domain using the Discrete Fourier Transform, where the convolution operation is represented as a product: .in, This represents the discrete Fourier transform of a matrix. It is a frequency domain convolution kernel. It is a node feature representation.

[0129] Spatiotemporal correlation diagram of base station traffic It is an undirected graph, that is .here, It is a set containing n nodes. This is achieved through a parameter matrix. Will Mapped to a special Green kernel ,in .

[0130] Will By performing a Discrete Fourier Transform, the spatiotemporal correlation graph of base station traffic can be obtained. Frequency domain convolution kernel:

[0131] .

[0132] For the Fourier graph operator of the spatiotemporal correlation graph of base station traffic, during the convolution process, Each node shares the same frequency domain convolution kernel, therefore it can Simplifying one dimension yields the spatiotemporal correlation graph operator for base station traffic. .

[0133] The feature matrix representing the characteristics of base station traffic data blocks By performing a Fourier transform and then multiplying it with the base station traffic spatiotemporal correlation graph operator, multi-layer convolution of the base station traffic spatiotemporal correlation graph can be achieved.

[0134] .

[0135] in, and They represent and conduct Multiplication by multiplication. , express The parameter matrix corresponding to the convolutional layer. This represents the discrete inverse Fourier transform of a matrix.

[0136] Furthermore, step S400 also includes:

[0137] Step S440: Obtain the spatiotemporal correlation graph operator based on the convolution kernel;

[0138] Step S450: Pass the multi-depth convolution stacking formula of the spatiotemporal correlation graph operator to perform multi-depth convolution operation on the spatiotemporal correlation graph of base station traffic to capture the correlation of different distance levels in the spatiotemporal correlation graph of base station traffic; the multi-depth convolution stacking formula is expressed as follows:

[0139] ,

[0140] In the formula, P The convolution depth is . This is the result of superimposing base station traffic across multiple depthwise convolutions in the frequency domain. for The bias parameters corresponding to the convolutional layers. For activation function, The feature matrix represents the characteristics of base station traffic data blocks. Spatiotemporal correlation graph operator conduct Multiplication by multiplication. This represents the discrete Fourier transform of a matrix. Understandably, That is, the spatiotemporal feature extraction results of base station traffic data blocks.

[0141] It should be noted that steps S410 to S440 are steps for synchronously extracting the spatiotemporal features of base station traffic based on Fourier graph operators.

[0142] In this embodiment, the spatiotemporal correlation graph operator Depend on Obtained by simplifying one dimension.

[0143] Because the degree of mutual influence between traffic data blocks varies among base stations at different distances, in order to comprehensively capture the correlation at different distance levels in the spatiotemporal correlation graph of base station traffic, a multi-depth convolution stacking formula is used to improve convolution efficiency by passing convolution results from shallow convolution to deep convolution. In one embodiment, .

[0144] The purpose of this is to migrate traditional graph convolution operations from the time domain to the frequency domain based on Fourier graph operators in large-scale base station traffic data processing scenarios. This effectively alleviates the problem of the dramatic increase in convolution computation overhead caused by the expansion of base station traffic data scale. Furthermore, through a multi-scale deep frequency domain convolution mechanism, it is possible to fully capture the spatial dependence and traffic evolution patterns of base station traffic while maintaining computational convergence, thereby achieving synchronous extraction and modeling of base station spatiotemporal features and breaking through the limitations of traditional graph convolution in asynchronous feature processing and generalization performance.

[0145] Step S500: Construct a contrastive loss based on the optimal coding tree partition, and construct a composite loss function based on the Huber loss and the contrastive loss;

[0146] Step S600: Train the base station traffic prediction model based on the composite loss function and the spatiotemporal feature extraction results to obtain the target prediction model;

[0147] Step S700: Use the base station traffic value at the current time step output by the target prediction model as the input for the next time step, perform autoregressive iterative prediction, and obtain the traffic prediction value for each base station at least one future time point.

[0148] In this embodiment, steps S500, S600, and S700 are the training and prediction steps of the base station traffic prediction model.

[0149] First, a composite loss function consisting of contrastive loss and Huber loss is constructed. The contrastive loss uses the aforementioned spatiotemporal coding tree partitioning of base station traffic to define positive and negative samples, thereby enhancing the model's ability to identify and represent complex spatiotemporal patterns. Then, the frequency domain features represented by the base station traffic data blocks are transformed back to the time domain through inverse Fourier transform, and feature weighted fusion and dimension mapping are performed through a fully connected layer to obtain the traffic prediction values ​​for each base station at multiple future time points.

[0150] Optionally, step S500 includes:

[0151] Step S510: Divide the positive and negative samples of the contrast loss according to the division of base station traffic data blocks in the optimal coding tree;

[0152] Step S520: Sum the Huber loss and the contrast loss to obtain the composite loss function;

[0153] The step of dividing the contrastive loss into positive and negative samples includes:

[0154] Base station traffic data blocks within the same partition are considered positive samples;

[0155] Base station traffic data blocks between different partitions are used as negative samples.

[0156] It should be noted that steps S510 to S520 are the loss function construction steps.

[0157] In this embodiment, to enhance the model's ability to capture the complex spatiotemporal characteristics between traffic data blocks, a contrastive loss is constructed to strengthen the constraint on the relationships between nodes in the spatiotemporal correlation graph of base station traffic. By comparing the characteristics of traffic data block nodes in different spatiotemporal regions, the model can better learn the spatial mutual influence of traffic between base stations and the dynamic changes in time, thereby achieving full learning of features in the spatiotemporal synchronization modeling of base station traffic.

[0158] Combined with the optimal coding tree obtained in step S300 A contrastive loss based on coding tree partitioning is constructed to simultaneously ensure the effectiveness of base station traffic hierarchical feature extraction.

[0159] Specifically, the method of partitioning base station traffic data blocks in the coding tree is used to divide the positive and negative samples in the contrast loss. Base station traffic data blocks in the same partition are regarded as positive samples, and base station traffic data blocks between different partitions are regarded as negative samples, thereby constructing the following contrast loss:

[0160] .

[0161] in, Temperature coefficient; To characterize the base station traffic data blocks after spatiotemporal feature fusion, correspond Nodes in , will with Nodes located in the same minimal partition are represented as positive samples. Otherwise, it is treated as a negative sample. ; This is a similarity function used to calculate vectors. and Similarity between them:

[0162] .

[0163] To address issues such as outlier interference, gradient instability, and difficulty in balancing training efficiency and accuracy in base station traffic data, the following Huber loss function is introduced to improve the model's robustness to outlier data, accelerate convergence, and accurately fit the predicted and actual values.

[0164] .

[0165] in, This represents the actual value of base station traffic. For the predicted flow rate, The threshold used to distinguish error intervals.

[0166] By introducing contrast loss To enhance the model's learning of the traffic correlation structure between base stations, enabling it to capture more refined spatial dependencies and temporal evolution patterns; simultaneously, Huber loss is employed. To improve the model's robustness and convergence stability to outliers in the actual prediction stage, a composite loss function is obtained by combining the constraint capability of contrastive learning with the fitting advantage of regression tasks, as follows:

[0167] .

[0168] Optionally, step S600 includes:

[0169] Step S610: Based on the spatiotemporal feature extraction results, obtain a temporal domain flow data block representation;

[0170] Step S620: Based on the traffic prediction weight matrix, map the time-domain traffic data block representation to the output of the prediction length to obtain the base station traffic prediction value;

[0171] Step S630: Train the base station traffic prediction model based on the stochastic gradient descent algorithm, the predicted base station traffic value, and the composite loss function to obtain the target prediction model.

[0172] In this embodiment, in order to achieve the goal of predicting the traffic of each base station in the time domain, the spatiotemporal features of the base station traffic data blocks are first extracted. Convert back to the time domain, then use function from The base station to which each traffic data block belongs is decomposed from the base station traffic data block, and the dimensions of the number of base station blocks and the embedding of base station blocks are combined. By merging, a time-domain traffic data block representation is obtained. ,in This indicates the number of traffic data blocks for a single base station:

[0173] .

[0174] Next, a fully connected layer is constructed to represent the time-domain traffic data blocks. Mapped to prediction length of The output yields the base station traffic prediction value. This includes a traffic prediction weight matrix. and bias vector :

[0175] .

[0176] Finally, the composite loss function is calculated, and the weight parameters of each module are iteratively updated using the stochastic gradient descent algorithm to obtain the target prediction model.

[0177] After the model is trained, iterative prediction is performed using an autoregressive approach during the prediction phase. This means that the base station traffic value at the current time step is output by the model and used as the input for the next time step. This process is repeated to form a recursive input for the prediction results, thereby supporting continuous prediction of base station traffic for the next day to week.

[0178] In the technical solution provided in this embodiment, a base station traffic data block segmentation strategy that ensures the semantic integrity of base station traffic is adopted to construct a base station traffic spatiotemporal correlation graph with traffic data blocks as nodes. Based on structural information theory, spatiotemporal hierarchical feature extraction of base station traffic is performed, and Fourier graph operators are used to achieve synchronous extraction of spatiotemporal features of base station traffic, realizing efficient extraction of spatiotemporal features and generating a base station traffic data block representation with spatiotemporal characteristics, i.e., the spatiotemporal feature extraction result. A contrastive learning mechanism is introduced to enhance the ability to identify complex spatiotemporal features between traffic data blocks, thereby improving the accuracy of base station traffic prediction. This achieves improved accuracy and efficiency of base station traffic prediction, providing a reference for base station energy efficiency optimization strategy formulation and intelligent network resource scheduling.

[0179] Reference Figure 4 , Figure 5 Based on the above embodiments, in the second embodiment, step S200 includes the following steps:

[0180] S210: Embed all the aforementioned traffic data blocks, n Each of the traffic data blocks is converted into a vector of a preset length. ;

[0181] S220: Based on the aforementioned metric function, n Each traffic data block is used as a graph node to construct the spatiotemporal correlation graph of the base station traffic;

[0182] S230: Quantify the comprehensive similarity between two graph nodes in terms of geographical location, temporal proximity, and traffic pattern according to the quantification formula, which is expressed as follows:

[0183] ,

[0184] In the formula, , and These are the weighting coefficients. and express and The time corresponding to the two nodes, and express and The base station locations corresponding to the two nodes Represents a node A vector of a preset length.

[0185] It should be noted that base station traffic typically exhibits obvious periodic fluctuations in time, while in space it is affected by the service range of adjacent base stations, geographical location, and user behavior patterns, showing a strong spatiotemporal correlation.

[0186] In this embodiment, steps S210 to S230 are steps for constructing a spatiotemporal correlation graph of base station traffic. First, the traffic data blocks segmented from all base stations are embedded. Each traffic data block is embedded sequentially using an embedding formula, thereby embedding the traffic data blocks... Convert to length of vector The embedded formula is expressed as follows:

[0187] .

[0188] in, This is a dimension transformation matrix. This is the offset vector. It is obtained after embedding all traffic data blocks. .

[0189] Then, Using traffic data blocks as graph nodes, a spatiotemporal correlation graph of base station traffic is constructed by designing a similarity measurement function between graph nodes. ,in, For includes A set of nodes This indicates the connection relationships between the nodes in the graph.

[0190] By comprehensively considering the physical distance between base stations, the differences in corresponding traffic data blocks on the time axis, and the similarity of data block embedding representations in the feature space, a formula is given to quantify the comprehensive similarity between two nodes in terms of geographical location, temporal proximity, and traffic patterns. A similarity threshold is set. If node and Between ,but ,otherwise .

[0191] in, and For weighting coefficients, through and They represent and The time and space distance between them.

[0192] Further, step S100 is the base station traffic data segmentation step. This indicates that the base station's traffic data is represented in the time domain. Indicates the total number of base stations. Indicates the base station number. This indicates the length of the base station traffic data. Here, step S100 includes using the formula... This transforms the base station's traffic data representation in the time domain to the frequency domain. The frequency domain representation of base station traffic is a complex vector, consisting of a real part. and the virtual part composition; This is the discrete Fourier transform of a vector.

[0193] The intensity of a period in traffic data can be identified by the magnitude of the amplitude values ​​corresponding to different periods. This can be achieved using the formula... Solve for the magnitude of the flow data The larger the amplitude value, the more pronounced the period; therefore, the frequency corresponding to the maximum amplitude value is selected. and will As the dividing interval.

[0194] Base station traffic data Use size is Dividing the data at intervals, we can obtain a set of sizes. Traffic data blocks .Will N The traffic data of each base station is divided into blocks according to the above steps to obtain... ( ) One traffic data block.

[0195] Understandably, due to the periodicity of base station traffic changes, using Discrete Fourier Transform to identify the main period of traffic data before segmenting traffic data can ensure the temporal semantic integrity of the segmented traffic data.

[0196] The purpose of this is to effectively preserve the spatiotemporal semantics of traffic data by automatically detecting significant periods in base station traffic sequences. Compared with traditional fixed window segmentation methods, it has stronger adaptive capabilities and expressive effects in capturing periodic features.

[0197] In the technical solution provided in this embodiment, in order to achieve unified temporal and spatial modeling of base station traffic, after dividing the traffic data of all base stations into blocks, a metric function that comprehensively considers the temporal and spatial correlation of base station traffic data is designed to construct a spatiotemporal correlation graph of the entire base station network traffic data. By fusing a composite spatiotemporal similarity function that integrates geographical location, temporal proximity, and traffic pattern, a weighted undirected graph of base station traffic is constructed, achieving accurate modeling of the complex spatiotemporal correlations contained in base station traffic and overcoming the problem of independently modeling spatial and temporal information.

[0198] Reference Figure 6 , Figure 7 Based on the above embodiments, in the third embodiment, step S300 further includes the following steps:

[0199] Step S321: Initialize the graph nodes of the spatiotemporal correlation graph of the base station traffic to obtain n The traffic data block of the first independent base station is divided;

[0200] Step S322: Traverse all the first base station traffic data block divisions and merge the first base station traffic data block divisions based on the preset structural entropy minimization condition to obtain the second base station traffic data block divisions;

[0201] Step S323: Combine the second base station traffic data block partitions with different temporal and spatial semantics to obtain the third base station traffic data block partitions;

[0202] Step S324: Repeat steps S322 to S323 until the first base station traffic data block partition cannot be merged and the second base station traffic data block partition cannot be combined, to obtain the target base station traffic data block partition.

[0203] Step S325: Construct the optimal coding tree based on the target base station traffic data block partitioning.

[0204] In this embodiment, the spatiotemporal correlation graph of base station traffic is used. nodes Initialize to sequentially The first base station traffic data block is divided into independent blocks. ,get ,in express The sub-partitioning is then performed, and the optimal coding tree for the spatiotemporal correlation graph of base station traffic is constructed through the following steps:

[0205] The first step is to iterate through all the traffic data block partitions of the first base station. If the partitions are... and Satisfy the preset condition of minimizing structural entropy If the value is the largest and positive, then... and Merging, that is, combining The sub-partitions or nodes contained therein are merged to obtain Then from Delete This allows for the division of traffic data blocks at the second base station. Understandably, For the updated partitioning. To find a partitioning scheme that minimizes the structural entropy of the spatiotemporal correlation graph of base station traffic, it should be ensured that the merging operation can minimize the structural entropy between base station traffic data blocks, i.e., satisfy... The value of is the largest and positive, among which for and The merged encoding tree.

[0206] The second step involves iterating through all the partitions updated in the first step, i.e., the partitions of the second base station traffic data blocks mentioned above. If the second base station traffic data block partitions... and satisfy If the value is the largest and positive, then... and The third base station traffic data block is jointly formed, that is, a new partition is created. ,Will and Two partitions are placed ,make and become The sub-partitioning. Through joint operations, base station traffic data blocks with different temporal and spatial semantics are associated, reflecting the hierarchical structure of the base station traffic spatiotemporal correlation graph, where... for and The combined coding tree.

[0207] Repeat steps one and two above until no more mergeable or joint partitions exist, thus obtaining the final base station partition set for the region. The root node is at level 0 of the coding tree, which ultimately divides the base stations into sets. All partitions are treated as Level 1 nodes. If a partition contains sub-partitions, its sub-partitions are treated as child nodes and the process continues until base station traffic data blocks are treated as leaf nodes, thus obtaining the optimal coding tree for the spatiotemporal correlation graph of base station traffic. .

[0208] In the technical solution provided in this embodiment, based on structural information theory, the spatiotemporal hierarchical features of base station traffic are extracted for different spatiotemporal hierarchical relationships existing in the base station traffic data block, thereby capturing effective spatiotemporal information and effectively eliminating redundant structures that are irrelevant to traffic prediction in the spatiotemporal correlation graph of base station traffic.

[0209] For example, suppose the base station traffic data segments are as shown in Table 1, containing Historical traffic data for each base station. Traffic records for each base station were collected for 90 consecutive days, recorded every 5 minutes; therefore, the length of the traffic sequence for each base station is [length missing]. A specific point in time.

[0210] Table 1: Base Station Traffic Data Segments

[0211]

[0212] In the base station traffic data segmentation stage, the first step is periodic identification. Taking base station number 1 in Table 1 as an example, its complete traffic sequence is as follows: Input formula Performing Discrete Fourier Transform yields Specifically, =[ ],in , And so on. It is a vector represented by a complex number, and Through formula The amplitude corresponding to each frequency component was calculated. After analysis, the most significant period in the data was 288, i.e. This perfectly matches the daily traffic peak and trough pattern recorded every 5 minutes, ensuring that each data block can capture a complete daily variation pattern.

[0213] Then, based on the identified main cycles Perform data block segmentation. Divide the traffic sequence of each base station, with a length of 25920 bytes, sequentially along the time axis into... Each of the three independent traffic data blocks is 288 bytes long, thus obtaining... The traffic from other base stations can be segmented into traffic data blocks using the same steps. .

[0214] In the stage of constructing the spatiotemporal correlation graph of base station traffic, after obtaining traffic data blocks, a spatiotemporal correlation graph is constructed using these data blocks as basic units, such as... Figure 5 As shown. First, use the formula. ,right Each data block is embedded, mapping each 288-byte data block to a single data block. A low-dimensional dense vector. Multiply each flow block by a transformation matrix. , It is a learnable dimension transformation matrix with dimension . Then, a bias vector is added to complete the embedding. This step transforms the original traffic pattern into feature vectors that are easy for the model to process. After embedding all data blocks, a total feature matrix is ​​obtained. Its dimensions are These 90,000 vectors are used as a graph. node set and use the formula The similarity measurement function in the graph is used to calculate the spatiotemporal correlation between any two nodes, thereby defining the weight of the edges in the graph.

[0215] by and For example, the formula The first term is to calculate the cosine similarity, through... and The first term to be calculated is The second factor is temporal proximity. = for The corresponding start time point, for The temporal proximity of the two points can be obtained by taking the reciprocal of their time difference from their corresponding start times. The third term is spatial proximity, expressed as the reciprocal of the geographical distance to the base station. Location of corresponding base station 1 , Location of corresponding base station 2 The spatial proximity of the two is The weighting coefficient is set to , , Thus, it can be calculated The threshold is determined by the formula. Set to 0.3, because ,but Similar calculations and Similarity between ,because Therefore The above steps construct a weighted undirected graph that comprehensively reflects the spatiotemporal dynamic correlation of traffic data. .

[0216] In the stage of base station traffic hierarchical feature extraction based on structural entropy, the hierarchical structure representation of the base station traffic spatiotemporal correlation graph is generated as follows: Figure 7 As shown in the figure. Each of the 90,000 nodes is initialized as an independent partition, forming a set containing 90,000 initialized partitions. To clearly illustrate the subsequent construction process, a diagram is used. Taking four representative nodes as examples, their specific meanings are shown in Table 2.

[0217] Table 2: Figure Representative node description

[0218]

[0219] Initially, these four nodes were divided into independent partitions, i.e. , , , Set the maximum height of the coding tree. Following the above method for constructing the coding tree, traverse all possible partition pairs and use the formula... Calculate the reduction in structural entropy after merging them. Because and Highly correlated in both time and space (same day, adjacent base stations), the calculated reduction in structural entropy is the largest and positive among all combinations. Therefore, the choice is to... and Merge to form a new division The original After merging, the node is removed, thus updating the current partition set. Other nodes are merged in the same way.

[0220] Then, perform the joint operation according to the above coding tree construction steps, using the formula. The structural entropy reduction benefit of the joint operation is evaluated, and two partitions are selected to maximize the reduction in structural entropy after joint operation. Since the joint operation creates new nodes in the coding tree, which may increase the tree height, the height must be limited to a certain limit. The iteration stops when no more partitions satisfy the merging or joining conditions. A root node of the coding tree is then constructed, with sub-partitions or nodes from the final partition as children of the root node. These sub-partitions are then expanded layer by layer until the leaf nodes contain 90,000 traffic block nodes, ultimately constructing the optimal coding tree for the spatiotemporal correlation graph of base station traffic. .

[0221] In the stage of synchronous extraction of spatiotemporal features of base station traffic based on Fourier graph operators, the Green kernel of the spatiotemporal correlation graph of base station traffic is first constructed. This is achieved through the constructed adjacency matrix. This represents the spatiotemporal similarity between the nodes, as described above. and The similarity between them is Construct the Green's kernel by taking the values ​​from the first row of the adjacency matrix. The first line contains 90,000 numbers, which are sequentially connected to the parameter matrix. , obtain Green's kernel ={ },in , , .Will The Fourier plot operator is obtained after performing a matrix Fourier transform. Then, after simplifying it by one dimension, we obtain the base station traffic spatiotemporal graph operator. .

[0222] The initial input to the frequency domain convolution is the aforementioned spatiotemporal node feature matrix. The dimensions are 90000×64. This was obtained through the above steps. It is a global, 64×64-dimensional complex matrix. To comprehensively capture the correlations at different distances, the convolution depth is set. Activation function Select Gaussian error linear unit First, the spatiotemporal domain node feature matrix... Applying the Fast Fourier Transform, we transform it to the frequency domain to obtain the frequency domain characteristic matrix. Its dimensions remain 90000×64, but the elements in the matrix are all complex numbers. Next, the model will perform a process of depth... The process involves multi-level frequency domain convolution. This process is decomposed into four independent rounds of computation to aggregate neighborhood information at different distances from near to far:

[0223] Convolution depth is 1 ( ): Calculate frequency domain characteristics With Fourier plot operator The matrix multiplication itself is then added to the result by the first layer's dedicated learnable bias vector. Finally passed The activation function processes the data to obtain the first layer output. , It aggregates the 1-hop neighbor information of each node in the graph, that is, the directly connected nodes.

[0224] Convolution depth of 2 :calculate and power of two ,Right now The result of multiplying by itself, plus the bias vector of the second layer. Then, through the activation function, the output of the second layer is obtained. , It aggregates information from two-hop neighbors and captures spatiotemporal correlation features over a wider range.

[0225] Convolution depth is 3 Similarly, calculation and cube The product of, and the third layer bias Activated after addition, resulting in This is used to aggregate information about 3-hop neighbors.

[0226] Convolution depth of 5 Finally, calculate and fourth power The product of, and the fourth layer bias Activated after addition, resulting in This is used to aggregate information about 4-hop neighbors.

[0227] After completing all convolution calculations at 4 depths, according to the formula The result , , These four intermediate result matrices, each with a dimension of 90000×64, are summed element by element to complete the final feature aggregation, thus obtaining... .

[0228] The final output of this step is a complex matrix of 90000×64 dimensions. Each row incorporates spatiotemporal correlation features from 1-hop to 4-hop neighbors at different distance levels, which will serve as the core input for subsequent prediction tasks.

[0229] In the loss function construction phase, the model will utilize the deep spatiotemporal features extracted in the previous steps. The system is trained using a composite loss function and ultimately outputs the traffic prediction result. This composite loss function... It consists of two parts, which together serve as the optimization objective for model training. The first part is the contrastive loss. In contrastive loss, the definition of positive and negative samples is directly derived from the constructed optimal encoding tree. For nodes Because it is related to They are assigned to the same minimum partition in the coding tree. The corresponding feature is then considered a positive sample; while the feature corresponding to the positive sample is considered a positive sample. Nodes in different partitions The feature is then treated as a negative sample. The calculation in this part follows the formula... and set the temperature coefficient The second part of the loss function is the Huber loss used for the final regression task. Through formula Calculate the error between the predicted value and the actual value, and set a threshold. To enhance the model's robustness to data noise.

[0230] In the base station traffic prediction phase, during the prediction process, the model first uses the frequency domain feature matrix obtained in step 2.2. The 90000×64-dimensional matrix is ​​transformed back to the time domain using the inverse fast Fourier transform algorithm. Then, through a transformation operation, this 90000×64-dimensional matrix is ​​transformed into a 1000×(90×64)-dimensional base station aggregation feature matrix. Each row uniquely corresponds to a base station and aggregates the spatiotemporal characteristics of all traffic blocks from that base station over the past 90 days. Then, this... The matrix is ​​fed into a fully connected layer. The input dimension of this layer is 90 × 64 = 5760, and the output dimension is set to the time step to be predicted. This corresponds to 288 five-minute time points in the coming day. The output of this fully connected layer is the final output. dimensional prediction matrix This represents the traffic prediction results for 1000 base stations over the next 288 time steps. Some examples of the results are shown in Table 3.

[0231] Table 3: Traffic Forecast Results for Some Base Stations in the Next Day

[0232]

[0233] The present invention also provides a computer-readable storage medium storing a base station traffic prediction program based on spatiotemporal feature joint characterization. When the base station traffic prediction program based on spatiotemporal feature joint characterization is executed by a processor, it implements the various steps of the base station traffic prediction method based on spatiotemporal feature joint characterization as described in the above embodiments.

[0234] The computer-readable storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0235] It should be noted that, since the storage medium provided in the embodiments of this application is the storage medium used to implement the methods of the embodiments of this application, those skilled in the art can understand the specific structure and variations of the storage medium based on the methods described in the embodiments of this application, and therefore will not be repeated here. All storage media used in the methods of the embodiments of this application fall within the scope of protection of this application.

[0236] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0237] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0238] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0239] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0240] It should be noted that any reference signs placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, third, etc., does not indicate any order. These words can be interpreted as names.

[0241] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0242] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A base station traffic prediction method based on spatiotemporal feature joint representation, characterized in that, The base station traffic prediction method based on joint spatiotemporal feature representation includes the following steps: Step S100: Perform Discrete Fourier Transform identification on the historical traffic data of each base station, and segment the continuous traffic time series to obtain traffic data blocks; Step S200: Based on the metric function, construct a spatiotemporal correlation graph of base station traffic with the traffic data blocks as graph nodes. The metric function is a metric function that comprehensively considers the spatiotemporal characteristics of traffic, including: Step S210: Embed all the traffic data blocks, and n Each traffic data block is converted into a vector of a preset length; Step S220: Based on the metric function, with n Each of the traffic data blocks is a graph node, constructing a spatiotemporal correlation graph of the base station traffic, wherein the metric function is a similarity metric function between graph nodes; and, Step S230: Quantify the comprehensive similarity of the two graph nodes in terms of geographical location, temporal proximity, and traffic pattern according to the quantification formula, which is expressed as follows: ， In the formula, , and These are the weighting coefficients. and express and The time corresponding to the two nodes, and express and The base station locations corresponding to the two nodes, Represents a node A vector of a preset length; Step S300: According to structural information theory, minimize the structural entropy of the spatiotemporal correlation graph of the base station traffic to obtain the optimal coding tree, including: Step S321: initializing the graph nodes of the base station traffic spatio-temporal graph, obtaining n an independent first base station traffic data block division; Step S322: Traverse all the first base station traffic data block divisions and merge the first base station traffic data block divisions based on the preset structural entropy minimization condition to obtain the second base station traffic data block divisions; Step S323: Combine the second base station traffic data block partitions with different temporal and spatial semantics to obtain the third base station traffic data block partitions; Step S324: Repeat steps S322 to S323 until the first base station traffic data block partition cannot be merged and the second base station traffic data block partition cannot be combined, to obtain the target base station traffic data block partition. Step S325: Construct the optimal coding tree based on the target base station traffic data block partitioning; Step S400: Perform multi-depth convolution operation on the spatiotemporal correlation graph of the base station traffic in the frequency domain to obtain spatiotemporal feature extraction results; Step S500: Construct a contrastive loss based on the optimal coding tree partition, and construct a composite loss function based on the Huber loss and the contrastive loss; Step S600: Train the base station traffic prediction model based on the composite loss function and the spatiotemporal feature extraction results to obtain the target prediction model; Step S700: Use the base station traffic value at the current time step output by the target prediction model as the input for the next time step, perform autoregressive iterative prediction, and obtain the traffic prediction value for each base station at least one future time point. 2.The base station traffic prediction method based on joint representation of spatio-temporal features according to claim 1, wherein, The step of minimizing the structural entropy of the spatiotemporal correlation graph of base station traffic and obtaining the optimal coding tree according to structural information theory includes: Step S311: Based on the root node of the base station traffic spatiotemporal correlation graph, obtain the structural entropy of the base station traffic spatiotemporal correlation graph under the coding tree partition; Step S312: Minimize the structural entropy of the spatiotemporal correlation graph of base station traffic based on the preset coding tree height and the structural entropy of the coding tree-divided spatiotemporal correlation graph. 3.The base station traffic prediction method based on joint representation of spatio-temporal features according to claim 1, wherein, The step of performing multi-depth convolution operations on the spatiotemporal correlation graph of base station traffic in the frequency domain to obtain spatiotemporal feature extraction results includes: Step S410: Based on a parameter matrix, map the connection relationship between nodes in the spatiotemporal correlation graph of the base station traffic to a Green kernel; Step S420: Perform a discrete Fourier transform on the Green kernel to obtain a frequency domain convolution kernel; Step S430: Perform multi-layer convolution on the spatiotemporal correlation graph of the base station traffic according to the frequency domain convolution kernel. 4.The method of claim 3, wherein, The step of performing multi-depth convolution operations on the spatiotemporal correlation graph of base station traffic in the frequency domain to obtain spatiotemporal feature extraction results further includes: Step S440: Obtain the spatiotemporal correlation graph operator based on the convolution kernel; Step S450: Pass the multi-depth convolution stacking formula of the spatiotemporal correlation graph operator to perform multi-depth convolution operation on the spatiotemporal correlation graph of base station traffic to capture the correlation of different distance levels in the spatiotemporal correlation graph of base station traffic; the multi-depth convolution stacking formula is expressed as follows: ， In the formula, P The convolution depth is . This is the result of superimposing base station traffic across multiple depthwise convolutions in the frequency domain. for The bias parameters corresponding to the convolutional layers. For activation function, The feature matrix represents the characteristics of base station traffic data blocks. Spatiotemporal correlation graph operator conduct Multiplication by multiplication. This represents the discrete Fourier transform of a matrix. The number of base station traffic data blocks. The embedding dimension for the base station block. 5.The base station traffic prediction method based on joint representation of spatio-temporal features according to claim 1, wherein, The steps of constructing a contrastive loss based on the optimal coding tree partition, and constructing a composite loss function based on the Huber loss and the contrastive loss, include: Step S510: Divide the positive and negative samples of the contrast loss according to the division of base station traffic data blocks in the optimal coding tree; Step S520: Sum the Huber loss and the contrast loss to obtain the composite loss function; The step of dividing the contrastive loss into positive and negative samples includes: Base station traffic data blocks within the same partition are considered positive samples; Base station traffic data blocks between different partitions are used as negative samples. 6.The method of claim 1, wherein, The step of training the base station traffic prediction model based on the composite loss function and the spatiotemporal feature extraction results to obtain the target prediction model includes: Step S610: Based on the spatiotemporal feature extraction results, obtain a temporal domain flow data block representation; Step S620: Based on the traffic prediction weight matrix, map the time-domain traffic data block representation to the output of the prediction length to obtain the base station traffic prediction value; Step S630: Train the base station traffic prediction model based on the stochastic gradient descent algorithm, the predicted base station traffic value, and the composite loss function to obtain the target prediction model.

7. A base station traffic prediction device characterized by comprising: The base station traffic prediction device includes: a memory, a processor, and a base station traffic prediction program based on spatiotemporal feature joint characterization stored in the memory and executable on the processor. The base station traffic prediction program based on spatiotemporal feature joint characterization is configured to implement the steps of the base station traffic prediction method based on spatiotemporal feature joint characterization as described in any one of claims 1 to 6.

8. A readable storage medium, characterized by, The readable storage medium stores a base station traffic prediction program based on spatiotemporal feature joint characterization. When the base station traffic prediction program based on spatiotemporal feature joint characterization is executed by a processor, it implements the steps of the base station traffic prediction method based on spatiotemporal feature joint characterization as described in any one of claims 1 to 6.

Images