Construction method and system of space-time two-dimensional service flow distribution generation model
By constructing a spatio-temporal two-dimensional business traffic distribution generation model and using the Transformer model and hierarchical clustering algorithm, the problem of difficulty in accurately analyzing communication network business traffic distribution in the existing technology is solved, and accurate spatio-temporal distribution characteristic analysis of communication business traffic is achieved. and model parameter adjustability.
Patent Information
- Application Number
- CN202411476830.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-22
- Publication Date
- 2025-01-24
AI Technical Summary
The prior art is difficult to provide a model that can accurately adjust function parameters and generate spatiotemporal distribution function expressions, which is used to analyze the service traffic distribution of communication networks and meet the network planning and optimization needs of different scenarios.
A method of building a model for the distribution of time and space business flow distribution models, by determining the time generation model and space generation model, and combining the Transformer model and hierarchical cluster algorithm to generate the model when generating the flow of flow. This model can decompose time sequences, fit trends, seasons and noise items, and analyze space characteristics through layers.
It achieves accurate space-time distribution characteristic analysis of communication business traffic, has good parameter adjustability and interpretability, is suitable for data sets with different computational complexity, and improves the similarity between the model and real data.
Smart Images
Figure CN119364381A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of mobile communications, and in particular, to a method and system for constructing a spatiotemporal two-dimensional business traffic distribution generation model. Background Art
[0002] With the popularity of smart phones and the rapid development of mobile Internet services, the traffic in mobile communication networks is growing exponentially, and the types of services are becoming more and more diverse. Analyzing the traffic distribution of mobile communication networks is of great significance for network planning and optimization, network resource allocation, and business development and promotion based on traffic prediction. The business distribution generation model is mainly used for traffic prediction. By predicting network traffic, network resources such as bandwidth and server resources can be allocated more effectively to meet network needs in different time periods and regions, avoid resource waste or shortage, and improve network stability and reliability. With the development of machine learning and deep learning technologies, deep neural networks have become a powerful tool for traffic prediction in communication networks. Deep neural networks have strong nonlinear fitting capabilities and can automatically learn and extract features during training, which reduces the need for manual feature engineering. As the size of the data set grows, neural networks can adapt to larger data by increasing the depth and width of the network. Traditional statistical models may increase computational complexity due to too many parameters, and even have overfitting problems. Neural networks can learn multiple related tasks simultaneously by sharing hidden layers, which improves the efficiency of the model and may improve the accuracy of predictions. When processing high-dimensional data, neural networks can identify complex patterns that are difficult for humans to understand. Neural networks can improve the accuracy of predictions by integrating multiple models, such as by stacking different neural networks or using ensemble techniques in deep learning frameworks. However, deep learning prediction models that introduce neural networks are often regarded as "black boxes" because their decision-making process is opaque, there is no explicit functional form, and no explicit mathematical expression of the prediction results is provided. This means that it is impossible to directly obtain a clear relationship between input and output from the model, which may be insufficient for theoretical research and certain engineering applications. The methods for modeling business distribution generation models through traditional general statistical theory mainly include: Bayesian modeling method, autoregressive integrated moving average model (ARIMA), etc. Traditional modeling methods such as ARIMA, exponential smoothing method, seasonal model, etc. are widely used because they usually have good fitting ability for historical data, strong interpretability of model parameters, and high model calculation efficiency. They are still widely used in actual scenarios such as network traffic, traffic flow, and power load. ARIMA is mainly used for data analysis and prediction with time dependence. However, for nonlinear patterns or more complex data structures, ARIMA is not as effective as some advanced machine learning models. In practical applications, ARIMA may need to be combined with other methods or models to improve the accuracy of predictions. The key assumption of naive Bayes modeling is the conditional independence between features, that is, each feature has an independent impact on the classification result. This assumption simplifies the probability model and makes the calculation feasible, but it may also ignore the potential correlation between features. In the communication network service request data, there is temporal correlation and spatial correlation, and time and space are also correlated.
[0003] Therefore, how to provide a model construction method that can adjust function parameters for different scenarios to give an accurate spatiotemporal distribution function expression and provide theoretical support for subsequent network planning and optimization research based on spatiotemporal correlation and statistical characteristics analysis has become an urgent problem to be solved in this field. Summary of the invention
[0004] The present application proposes a method for constructing a two-dimensional space-time business traffic distribution generation model, including the following steps: determining a time generation model; determining a space generation model; and obtaining a traffic space-time generation model based on the space generation model and the time generation model.
[0005] As above, the determination of the time generation model includes the following sub-steps: obtaining the total flow calculation request time series of the entire region; inputting the total flow calculation request time series into the Transformer model, and the Transformer model performs time series decomposition on the total flow calculation request time series to obtain trend terms, seasonal terms, and noise terms; fitting the trend terms, seasonal terms, and noise terms respectively, and summing the fitted trend terms, seasonal terms, and noise terms to obtain the calculation request time generation model of the entire region.
[0006] As above, before using the Fourier series function to fit the seasonal term, it also includes determining the data period T and determining the fundamental frequency f0 according to the data period T.
[0007] As above, the seasonal term after fitting is expressed as:
[0008]
[0009] a0 represents the DC component, a n and b n Both represent the amplitude of harmonics.
[0010] As above, the total traffic calculation request time series expression f(t) is decomposed as follows:
[0011] f(t)=f trend (t)+f season (t)+f residual (t)
[0012] f trend (t) is the expression of the trend term, f season (t) is the expression of seasonal term, f residual (t) is the expression of the noise term.
[0013] A system for constructing a spatiotemporal two-dimensional business traffic distribution generation model, specifically comprising: a time generation model determination unit, a space generation model determination unit and a space-time generation model determination unit; the time generation model determination unit is used to determine the time generation model; the space generation model determination unit is used to determine the space generation model; the space-time generation model determination unit is used to obtain a traffic space-time generation model based on the space generation model and the time generation model.
[0014] As above, the time generation model determination unit determines the time generation model, including the following sub-steps: obtaining the total flow calculation request time series of the entire region; inputting the total flow calculation request time series into the Transformer model, and the Transformer model performs time series decomposition on the total flow calculation request time series to obtain trend terms, seasonal terms, and noise terms; fitting the trend terms, seasonal terms, and noise terms respectively, and summing the fitted trend terms, seasonal terms, and noise terms to obtain the calculation request time generation model for the entire region.
[0015] As above, before the time generation model determination unit uses the Fourier series function to fit the seasonal term, it also includes determining the data period T and determining the fundamental frequency f0 according to the data period T.
[0016] As above, the seasonal term after the temporal generation model determines the unit fitting is expressed as:
[0017]
[0018] a0 represents the DC component, a n and b n Both represent the amplitude of harmonics.
[0019] As above, in the time generation model determination unit, the total traffic calculation request time series expression f(t) is decomposed as follows:
[0020] f(t)=f trend (t)+f season (t)+f residual (t)
[0021] f trend (t) is the expression of the trend term, f season (t) is the expression of seasonal term, f residual (t) is the expression of the noise term.
[0022] This application has the following beneficial effects:
[0023] Aiming at the spatial-temporal distribution characteristics of communication business traffic, this application proposes a Transformer-based and hierarchical clustering spatial-temporal generation model suitable for communication network business traffic. The proposed spatial-temporal generation model has good parameter adjustability, is suitable for data sets of different computational complexity, and can explain the data characteristics of different data sets. The similarity between the generation model of this application and the real data is improved, and the established spatial-temporal two-dimensional generation model is more in line with the real data. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.
[0025] Figure 1 It is a flowchart of a method for constructing a spatiotemporal two-dimensional business traffic distribution generation model provided in an embodiment of the present application;
[0026] Figure 2 It is a design diagram for a time generation model provided according to an embodiment of the present application;
[0027] Figure 3 It is a design diagram for a space generation model provided according to an embodiment of the present application;
[0028] Figure 4 It is a schematic diagram of the internal structure of a system for constructing a spatiotemporal two-dimensional business traffic distribution generation model provided in an embodiment of the present application. DETAILED DESCRIPTION
[0029] The following is a clear and complete description of the technical solutions in the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.
[0030] In the time dimension, this application uses Transformer to decompose the data to obtain data sequences representing trends, seasons and noise characteristics, and then uses different function expressions to fit different parts of the data sequence; in the spatial dimension, a hierarchical clustering method is used, and then the two are multiplied to achieve a two-dimensional space-time union. The proposed model construction method makes the model function have good parameter adjustability, is suitable for data sets of different computational complexity, and can explain the data characteristics of different data sets.
[0031] Embodiment 1
[0032] like Figure 1 As shown, this embodiment provides a method for constructing a spatiotemporal two-dimensional service traffic distribution generation model, which specifically includes the following steps:
[0033] Step S110: Determine the time generation model.
[0034] like Figure 2 As shown, this is a design diagram of the time generation model. In the design of the time generation model, this embodiment introduces the Transformer model to decompose the time series to obtain the trend term, the seasonal term and the noise term, and then fits the three data items respectively to complete the design of the time generation model. The introduction of the Transformer model can obtain a more accurate time generation model, and can be applied to data sets of different complexities. Compared with traditional modeling to predict traffic, it is more versatile.
[0035] Based on the above, step S110 includes the following sub-steps:
[0036] Step S1101: Obtain the total traffic calculation request time series of the entire area.
[0037] The traffic calculation requests in different regions at the same time are summed up to obtain the total traffic calculation requests for the entire region, which is the calculation request time series for the entire region.
[0038] Step S1102: input the total flow calculation request time series into the Transformer model, and the Transformer model decomposes the total flow calculation request time series to obtain trend terms, seasonal terms, and noise terms.
[0039] Get the traffic calculation request for the entire region, and then divide the entire region into N regions based on geographic location or business relevance. Sum the traffic calculation requests for the N regions at the same time to get the overall traffic time series of the region.
[0040] The regional overall traffic time series is input into the Transformer model, and the temporal features in the time series are usually embedded into the sequence through position encoding.
[0041] After inputting into the Transformer model, the Transformer model can capture the inherent temporal characteristics in the time series. Secondly, the core of the Transformer model is the self-attention mechanism, which can capture the dependencies between any time points in the time series, not just local adjacent time points. Through this mechanism, the Transformer model can learn the global dependencies within the entire sequence, thereby helping to identify trends and periodicities in the sequence. The self-attention mechanism can capture long-term dependencies, thereby learning the long-term trend of changes in the time series and decomposing them into trend terms. The same self-attention mechanism can capture periodic patterns, such as annual and quarterly changes, which helps to identify the periodic components in the time series and decompose them into seasonal terms. Random errors or noise in the data are noise terms. The transformer model is used to decompose the total traffic calculation request time series to obtain trend terms, seasonal terms, and residual terms. According to the different data characteristics of these three parts, different function forms are selected to fit the data of each part.
[0042] The total traffic calculation request time series expression can be decomposed as follows:
[0043] f(t)=f trend (t)+f season (t)+f residual (t)
[0044] Among them, f(t) is the expression of the total traffic calculation request time series, f trend (t is the expression of the trend term, f season (t) is the expression of seasonal term, f residual (t) is the expression of the noise term.
[0045] Step S1103: Fit the trend term, seasonal term and noise term respectively, sum the fitted trend term, seasonal term and noise term, and obtain the calculation request time generation model of the entire region.
[0046] For the trend item, since the trend item gradually rises and finally flattens out, choose to use a polynomial function to fit:
[0047]
[0048] Among them, a1...a n+3 are all independent parameters, t is time, and each parameter can be obtained through the traditional polynomial fitting algorithm. According to the different characteristics of each data set, different functions and parameters can be added for different trend items to improve the accuracy of fitting.
[0049] For seasonal terms, trend terms have strong regularity, so the Fourier series function fitting method is selected. First, determine the period T of the data, and then calculate the fundamental frequency
[0050]
[0051] Among them, a0 is the DC component (the average value of the signal), and a n and b n is the amplitude of the harmonic, n represents a natural number, t represents time, and f(t) represents a function.
[0052] Since infinite summation is not possible in practical applications, it is necessary to truncate the Fourier series to N terms, where N is a sufficiently large integer to ensure the accuracy of the approximation. Depending on the fitting effect, it may be necessary to adjust the truncation term N or recalculate the Fourier coefficients to optimize the fit.
[0053] For the noise term, since this part conforms to the Gaussian random distribution, it can be regarded as a variance of σ 2 , a Gaussian random variable with mean μ, i.e. f residual (t).
[0054] Then sum the above three items to get the complete fitting time series formula for the calculation request data, which is specifically expressed as:
[0055] f(t)=f trend (t)+f season (t)+f residual (t)
[0056] That is, the time generation model design of traffic calculation request data is completed.
[0057] Step S120: Determine the space generation model.
[0058] like Figure 3 As shown, this is a design diagram of the space generation model. In the design of the space generation model, this embodiment introduces a hierarchical clustering algorithm based on the correlation and dependency of each area, groups areas with similar spatial features into a cluster, and then models different clusters to realize the design of the space generation model.
[0059] Based on the above, step S120 includes the following sub-steps:
[0060] Step S1201: Obtain the proportion of traffic calculation request data in each area, and obtain N groups of proportion curve data based on the proportion of traffic calculation request data in all areas.
[0061] The calculation request data of different regions at the same time are summed up, and the proportion of calculation request data in each region is obtained. The same operation is performed at each time to obtain the proportion curve of each region changing with time. Since there are N regions, there are N groups of data in total, and the N groups of data correspond to the N regions one by one.
[0062] Step S1202: clustering N groups of proportion curve data.
[0063] The N groups of data are clustered using a hierarchical clustering method, specifically using the Ward method in hierarchical clustering.
[0064] The Ward method merges clusters by minimizing variance, which can maintain the homogeneity within the cluster, that is, the points within the cluster are as similar as possible. The calculation process of the Ward method is as follows:
[0065] Step 1: Initialization. Each data point in each set of data is initially considered as a separate cluster.
[0066] Step 2: Calculate the distance between all possible cluster pairs. In Ward's method, the distance is determined by calculating the increase in variance after merging two clusters.
[0067] Step 3: Select the two clusters with the smallest variance increase after merging. This is the core of Ward's method, which is to minimize the variance increase.
[0068] The fourth step is to update the clustering, merge the two selected clusters into a new cluster, and update the cluster list.
[0069] Step 5: Repeat steps 2 to 4 until all data points are merged into one cluster.
[0070] Finally, a dendrogram is generated. Through this process, a dendrogram is generated, showing the order and distance of cluster merging.
[0071] Cluster N groups of data and determine the difference, which is the distance between clusters calculated by the Ward method, to indicate the similarity between each group of data. Then you can get M clusters of data, and the data in each cluster is similar. The difference is an indicator used to judge the similarity between data in hierarchical clustering. The smaller the difference, the more detailed the clustering of the overall data, and the more similar the data between the clustered groups. If the data group has other data such as user preferences, it can be added to the evaluation indicators of the clustering algorithm, and more accurate similar groups can be obtained based on different indicators.
[0072] Then, each cluster of data is fitted separately. Therefore, the fitting basis function of each cluster is consistent, and the corresponding fitting function can be obtained by adjusting the parameters. According to the characteristics of each group of data, different fitting functions and parameters can be adjusted to improve accuracy.
[0073] F1(t)=a1sin(b1t+c1)+d1cos(g1t+h1)+l1
[0074] F2(t)=a2sin(b2t+c2)+d2cos(g2t+h2)+l2+i2t ...
[0076] F s (t)=a2sin(b2t+c2)+d2cos(g2t+h2)+l2+i s t
[0077]
[0078] a1 to a N ,b1 to b N ,c1 to c N ,d1 to d N ,g2 to g N ,h1 to h N ,l1 to l N ,k N , i2 to i s , k N
[0079] , m are parameters, The e in is the exponent.
[0080] F1(t) to F N (t) is the proportion fitting function of N groups of data, thereby completing the design of the spatial generation model.
[0081] Step S130: Obtain a traffic space-time generation model according to the space generation model and the time generation model.
[0082] The time generation model and the space generation model are multiplied to obtain the traffic space-time generation model. The expression of the traffic space-time generation model H(t,s) is:
[0083] H(t,s)=f(t)*F s (t)
[0084] t is time, and s is the mapping number of a region among N regions.
[0085] The algorithm of this embodiment is implemented based on the general open source Transformer model. The model dimension is set to 64, the number of attention heads is set to 4, the number of encoder layers is set to 2, the number of decoder layers is set to 2, the feedforward network dimension is 2048, the drop rate is 0.1, and a training cycle is set to 10 training rounds (Epoch). The data set is put into time series decomposition to obtain trend terms, seasonal terms, and noise terms. The trend terms are fitted with polynomial functions, the polynomial series is modified multiple times, and additional functions such as exponential functions and logarithmic functions are added or subtracted according to the features to improve the accuracy of the fitting; the seasonal terms are fitted with Fourier series, and N is set to 1000. The ward method is selected for hierarchical clustering to calculate the distance between clusters.
[0086] In the training environment setting of this embodiment, all algorithms are implemented in Python language, and are trained and tested on Ubuntu 20.04 system, with CUDA version 11.4. The compilation platform used is Pycharm, and the deep learning library version used is Pytorch 1.10.1.
[0087] In the evaluation index setting of this embodiment, the root mean square error (RMSE) and prediction accuracy are used to judge the prediction effect during the test, and R 2 Also called the coefficient of determination, it measures how well the model fits the data. 2 The value of is between 0 and 1 and can be interpreted as the percentage of the variance explained by the model to the total variance. The prediction accuracy is defined as follows:
[0088]
[0089] Embodiment 2
[0090] like Figure 4 As shown, this embodiment provides a system for constructing a spatiotemporal two-dimensional business traffic distribution generation model, specifically including: a time generation model determination unit 410, a space generation model determination unit 420, and a space-time generation model determination unit 430.
[0091] The time generation model determination unit 410 is used to determine the time generation model.
[0092] The time generation model determination unit includes the following submodules: a time series determination module, an input decomposition module, and a time generation model determination module.
[0093] The time series determination module is used to obtain the total flow calculation request time series of the entire region.
[0094] The traffic calculation requests in different areas at the same time are summed up to obtain the total traffic calculation requests for the entire area, which is the calculation request time series of the entire area.
[0095] The input decomposition module is used to input the total traffic calculation request time series into the Transformer model. The Transformer model decomposes the total traffic calculation request time series to obtain trend items, seasonal items, and noise items. The traffic calculation request for the entire region is obtained, and then the entire region under consideration is divided into N regions based on geographic location or business relevance. The traffic calculation requests of N regions at the same time are summed to obtain the overall regional traffic time series.
[0096] In the input decomposition module, the regional overall traffic time series is input into the Transformer model, and the time features in the time series are usually embedded into the sequence through position encoding. After input into the Transformer model, the Transformer model can capture the inherent time characteristics in the time series. Secondly, the core of the Transformer model is the self-attention mechanism, which can capture the dependencies between any time points in the time series, not just local adjacent time points. Through this mechanism, the Transformer model can learn the global dependencies within the entire sequence, thereby helping to identify trends and periodicities in the sequence. The self-attention mechanism can capture long-term dependencies, thereby learning the long-term trend of changes in the time series and decomposing them into trend terms. The same self-attention mechanism can capture periodic patterns, such as annual and quarterly changes, which helps to identify the periodic components in the time series and decompose them into seasonal terms. Random errors or noise in the data are noise terms. The total traffic calculation request time series is decomposed using the transformer model to obtain trend terms, seasonal terms, and residual terms. According to the different data characteristics of these three parts, different function forms are selected to fit the data of each part.
[0097] The total traffic calculation request time series expression can be decomposed as follows:
[0098] f(t)=f trend (t)+f season (t)+f residual (t)
[0099] Among them, f(t) is the expression of the total traffic calculation request time series, f trend (t is the expression of the trend term, f season (t) is the expression of seasonal term, f residual (t) is the expression of the noise term.
[0100] The time generation model determination module is used to fit the trend term, seasonal term and noise term respectively, and sum the fitted trend term, seasonal term and noise term to obtain the calculation request time generation model of the entire region. For the trend term, since the trend term gradually rises and finally flattens, a polynomial function is selected for fitting:
[0101]
[0102] Among them, a1...a n+3 are all independent parameters, t is time, and each parameter can be obtained through the traditional polynomial fitting algorithm. According to the different characteristics of each data set, different functions and parameters can be added for different trend items to improve the accuracy of fitting.
[0103] For seasonal terms, trend terms have strong regularity, so the Fourier series function fitting method is selected. First, determine the period T of the data, and then calculate the fundamental frequency
[0104]
[0105] Among them, a0 is the DC component (the average value of the signal), and a n and b n is the amplitude of the harmonic. Since infinite summation is not possible in practical applications, it is necessary to truncate the Fourier series to N terms, where N is a sufficiently large integer to ensure the accuracy of the approximation. Depending on the fitting effect, it may be necessary to adjust the truncation term N or recalculate the Fourier coefficients to optimize the fit.
[0106] For the noise term, since this part conforms to the Gaussian random distribution, it can be regarded as a variance of σ 2 , a Gaussian random variable with mean μ, i.e. f residual (t).
[0107] Then sum the above three items to get the complete fitting time series formula for the calculation request data,
[0108] f(t)=f trend (t)+f season (t)+f residual (t)
[0109] That is, the time generation model design of traffic calculation request data is completed.
[0110] The space generation model determination unit is used to determine the space generation model.
[0111] The spatial generation model determination unit includes the following modules: a data determination module and a clustering module.
[0112] The data determination module is used to obtain the proportion of traffic calculation request data in each area, and obtain N groups of proportion curve data according to the proportion of traffic calculation request data in all areas. The calculation request data of different areas at the same time are summed, and the proportion of calculation request data in each area is obtained. The same operation is performed at each time, and the proportion curve of each area changing with time can be obtained. Since there are N areas, there are N groups of data in total, and the N groups of data correspond to the N areas one by one.
[0113] The clustering module is used to cluster N groups of proportion curve data. The N groups of data are clustered using the hierarchical clustering method, specifically the Ward method in hierarchical clustering. The Ward method merges clusters by minimizing the variance, which can maintain the homogeneity within the cluster, that is, the points within the cluster are as similar as possible. The calculation process of the Ward method is as follows:
[0114] Step 1: Initialization. Each data point in each set of data is initially considered as a separate cluster.
[0115] Step 2: Calculate the distance between all possible cluster pairs. In Ward's method, the distance is determined by calculating the increase in variance after merging two clusters.
[0116] Step 3: Select the two clusters with the smallest variance increase after merging. This is the core of Ward's method, which is to minimize the variance increase.
[0117] The fourth step is to update the clustering, merge the two selected clusters into a new cluster, and update the cluster list.
[0118] Step 5: Repeat steps 2 to 4 until all data points are merged into one cluster.
[0119] Finally, a dendrogram is generated. Through this process, a dendrogram is generated, showing the order and distance of cluster merging.
[0120] Cluster N groups of data and determine the difference, which is the distance between clusters calculated by the Ward method, to indicate the similarity between each group of data. Then you can get M clusters of data, and the data in each cluster is similar. The difference is an indicator used to judge the similarity between data in hierarchical clustering. The smaller the difference, the more detailed the clustering of the overall data, and the more similar the data between the clustered groups. If the data group has other data such as user preferences, it can be added to the evaluation indicators of the clustering algorithm, and more accurate similar groups can be obtained based on different indicators.
[0121] Then, each cluster of data is fitted separately. Therefore, the fitting basis function of each cluster is consistent, and the corresponding fitting function can be obtained by adjusting the parameters. According to the characteristics of each group of data, different fitting functions and parameters can be adjusted to improve accuracy.
[0122] F1(t)=a1sin(b1t+c1)+d1cos(g1t+h1)+l1
[0123] F2(t)=a2sin(b2t+c2)+d2cos(g2t+h2)+l2+i2t ...
[0125] F s (t)=a2sin(b2t+c2)+d2cos(g2t+h2)+l2+i s t
[0126]
[0127] a1 to a N ,b1 to b N ,c1 to c N ,d1 to d N ,g2 to g N ,h1 to h N ,l1 to l N ,k N , i2 to i s , k N
[0128] , m are parameters, The e in is the exponent.
[0129] F1(t) to F N (t) is the proportion fitting function of N groups of data, thereby completing the design of the spatial generation model.
[0130] The space-time generation model determination unit 430 is used to obtain the traffic space-time generation model according to the space generation model and the time generation model.
[0131] The time generation model and the space generation model are multiplied to obtain the traffic space-time generation model. The expression of the traffic space-time generation model H(t,s) is: H(t,s)=f(t)*F s (t)
[0132] t is time, and s is the mapping number of a region among N regions.
[0133] This application has the following beneficial effects:
[0134] Aiming at the spatial-temporal distribution characteristics of communication business traffic, this application proposes a Transformer-based and hierarchical clustering spatial-temporal generation model suitable for communication network business traffic. The proposed spatial-temporal generation model has good parameter adjustability, is suitable for data sets of different computational complexity, and can explain the data characteristics of different data sets. The similarity between the generation model of this application and the real data is improved, and the established spatial-temporal two-dimensional generation model is more in line with the real data.
[0135] Although the present application is described with reference to examples, this is for illustrative purposes only and is not intended to limit the present application, and changes, additions and / or deletions to the embodiments may be made without departing from the scope of the present application.
[0136] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.
Claims
1. A method for constructing a spatiotemporal two-dimensional business traffic distribution generation model, characterized in that: The following steps are involved: Determine the time generation model; Determine the spatial generation model; According to the space generation model and the time generation model, the traffic space-time generation model is obtained.
2. The method for constructing a spatiotemporal two-dimensional business traffic distribution generation model according to claim 1, characterized in that: Determining the temporal generation model includes the following sub-steps: Get the total traffic calculation request time series for the entire region; The total flow calculation request time series is input into the Transformer model, and the Transformer model decomposes the total flow calculation request time series to obtain trend terms, seasonal terms, and noise terms; The trend term, seasonal term and noise term are fitted respectively, and the fitted trend term, seasonal term and noise term are summed to obtain the calculation request time generation model of the entire region.
3. The method for constructing a spatiotemporal two-dimensional business traffic distribution generation model according to claim 2, characterized in that: Before using the Fourier series function to fit the seasonal term, the process also includes determining the data period T and determining the fundamental frequency f0 based on the data period T.
4. The method for constructing a spatiotemporal two-dimensional business traffic distribution generation model according to claim 3, characterized in that: The fitted seasonal term is expressed as: a0 represents the DC component, a n and b n Both represent the amplitude of harmonics.
5. The method for constructing a spatiotemporal two-dimensional business traffic distribution generation model according to claim 4, characterized in that: The total traffic calculation request time series expression f(t) is decomposed as follows: f(t)=f trend (t)+f season (t)+f residual (t) f trend (t) is the expression of the trend term, f season (t) is the expression of seasonal term, f residual (t) is the expression of the noise term.
6. A system for constructing a spatiotemporal two-dimensional business traffic distribution generation model, characterized in that: Specifically including: a time generation model determination unit, a space generation model determination unit and a space-time generation model determination unit; A time generation model determination unit, used for determining the time generation model; A space generation model determination unit is used to determine the space generation model; The space-time generation model determination unit is used to obtain the traffic space-time generation model according to the space generation model and the time generation model.
7. The system for constructing a spatiotemporal two-dimensional business traffic distribution generation model according to claim 6, characterized in that: The time generation model determination unit determines the time generation model including the following sub-steps: Get the total traffic calculation request time series for the entire region; The total flow calculation request time series is input into the Transformer model, and the Transformer model decomposes the total flow calculation request time series to obtain trend terms, seasonal terms, and noise terms; The trend term, seasonal term and noise term are fitted respectively, and the fitted trend term, seasonal term and noise term are summed to obtain the calculation request time generation model of the entire region.
8. The system for constructing a spatiotemporal two-dimensional business traffic distribution generation model according to claim 7, characterized in that: Before the time generation model determination unit uses the Fourier series function to fit the seasonal term, the method also includes determining the data period T and determining the fundamental frequency f0 according to the data period T.
9. The system for constructing a spatiotemporal two-dimensional business traffic distribution generation model according to claim 8, characterized in that: The seasonal term after fitting the temporal generation model is expressed as: a0 represents the DC component, a n and b n Both represent the amplitude of harmonics.
10. The system for constructing a spatiotemporal two-dimensional business traffic distribution generation model according to claim 9, characterized in that: In the time generation model determination unit, the total traffic calculation request time series expression f(t) is decomposed as follows: f(t)=f trend (t)+f season (t)+f residual (t) f trend (t) is the expression of the trend term, f season (t) is the expression of seasonal term, f residual (t) is the expression of the noise term.