Mobile traffic prediction method and device based on diffusion model

Abstract

The invention provides a mobile traffic prediction method and device based on a diffusion model, and relates to the technical field of mobile traffic prediction.The method comprises the steps that on the basis of obtained historical mobile traffic, noise mobile traffic and dynamic characteristics of the historical mobile traffic are determined, and the noise mobile traffic is the mobile traffic added with noise; the noise priori estimation unit based on the diffusion model is used for carrying out noise estimation on the noise movement flow and the dynamic characteristics to obtain noise priori, and carrying out noise prediction on the noise movement flow based on a denoising network of the diffusion model to obtain residual noise; and performing flow prediction according to the noise prior, the residual noise and the noise moving flow to obtain future moving flow. According to the invention, the accuracy of mobile traffic prediction can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of mobile traffic prediction, and in particular, to a mobile traffic prediction method and device based on a diffusion model. Background Art

[0002] With the popularization of mobile traffic, timely and accurate traffic prediction plays a key role in aspects such as resource allocation, base station energy saving, network planning and optimization, and is also crucial for service providers aiming to align network capacity with user needs and the development of smart city infrastructure.

[0003] To better process mobile traffic data, existing diffusion model-based solutions mainly focus on designing innovative denoising networks or conditional mechanisms to integrate various features. Although providing conditional features to the denoising network can provide certain background information for prediction, the noise distribution gradually learned during the denoising process actually plays a more crucial role. This noise distribution directly shapes the model's generation process and has a significant impact on the quality and accuracy of the prediction.

[0004] However, the above methods mainly improve the prediction effect by innovating the network architecture, while ignoring the reconstruction noise, resulting in poor traffic prediction results. Therefore, an effective solution is urgently needed to solve the above problems. Summary of the Invention

[0005] To solve the above problems, the present invention provides a mobile traffic prediction method and device based on a diffusion model.

[0006] The present invention provides a mobile traffic prediction method based on a diffusion model, including: Based on the obtained historical mobile traffic, determine the noise mobile traffic and the dynamic characteristics of the historical mobile traffic, where the noise mobile traffic is the mobile traffic after adding noise; Based on the noise prior estimation unit of the diffusion model, perform noise estimation on the noise mobile traffic and the dynamic characteristics to obtain a noise prior, and perform noise prediction on the noise mobile traffic based on the denoising network of the diffusion model to obtain residual noise; According to the noise prior, the residual noise, and the noise mobile traffic, perform traffic prediction to obtain future mobile traffic.

[0007] According to the mobile traffic prediction method based on a diffusion model provided by the present invention, the step of performing traffic prediction according to the noise prior, the residual noise, and the noise mobile traffic to obtain future mobile traffic includes: Perform weighted fusion on the noise prior and the residual noise to obtain reconstructed noise; Based on the reconstructed noise, denoise the noise mobile traffic multiple times to obtain the future mobile traffic.

[0008] According to a mobile traffic prediction method based on a diffusion model provided by the present invention, the step of denoising the noise mobile traffic multiple times based on the reconstructed noise to obtain the future mobile traffic includes: Denoise the noise mobile traffic based on the reconstructed noise to obtain the denoised mobile traffic; When the number of denoising times is less than the diffusion step size, use the denoised mobile traffic as the noise mobile traffic, and re - execute the noise prior estimation unit based on the diffusion model to perform noise estimation on the noise mobile traffic and the dynamic characteristics and subsequent steps. The diffusion step size is the total number of steps in the diffusion process of the diffusion model; When the number of denoising times is equal to the diffusion step size, use the denoised mobile traffic as the future mobile traffic.

[0009] According to a mobile traffic prediction method based on a diffusion model provided by the present invention, the step of determining the noise mobile traffic and the dynamic characteristics of the historical mobile traffic based on the obtained historical mobile traffic includes: Based on the forward noise - adding process and the diffusion step size of the diffusion model, add noise to the obtained historical mobile traffic multiple times to obtain the noise mobile traffic. The diffusion step size is the total number of steps in the diffusion process of the diffusion model; According to the prediction step size, extract the dynamic characteristics of the historical mobile traffic to obtain the dynamic characteristics of the historical mobile traffic. The prediction step size represents the prediction length between the future mobile traffic and the historical mobile traffic.

[0010] According to a mobile traffic prediction method based on a diffusion model provided by the present invention, the step of extracting the dynamic characteristics of the historical mobile traffic according to the prediction step size to obtain the dynamic characteristics of the historical mobile traffic includes: When the prediction step size is greater than the set value, use a periodic dynamic extraction method to extract the dynamic characteristics of the historical mobile traffic to obtain the dynamic characteristics of the historical mobile traffic; When the prediction step size is less than or equal to the set value, use a proximity - based dynamic extraction method to extract the dynamic characteristics of the historical mobile traffic to obtain the dynamic characteristics of the historical mobile traffic.

[0011] According to a mobile traffic prediction method based on a diffusion model provided by the present invention, the step of performing noise estimation on the noise mobile traffic and the dynamic characteristics by the noise prior estimation unit based on the diffusion model to obtain the noise prior includes: Calculate the noise prior using the following formula: Among them, is the noise prior, is the noise moving traffic, is the cumulative parameter of the signal retention degree in the th step of the diffusion process, is the dynamic characteristic.

[0012] According to a mobile traffic prediction method based on a diffusion model provided by the present invention, the training process of the diffusion model includes: Obtain training mobile traffic, where the training mobile traffic includes sample mobile traffic and target mobile traffic; Sample random white noise from a standard Gaussian distribution; Based on the random white noise, the noise schedule of the diffusion model, and the diffusion step size, add noise to the target mobile traffic multiple times to obtain the fake target mobile traffic after adding noise, where the diffusion step size is the total number of steps in the diffusion process of the diffusion model; Extract the dynamic characteristics of the sample mobile traffic to obtain the dynamic characteristics of the sample mobile traffic; Based on the noise prior estimation unit of the diffusion model, perform noise estimation on the noisy target mobile traffic and the dynamic characteristics of the sample mobile traffic to obtain a predicted noise prior, and perform noise prediction on the noisy target mobile traffic based on the denoising network of the diffusion model to obtain a predicted residual noise; Calculate a loss value based on the predicted noise prior, the predicted residual noise, and the random white noise; Based on the loss value, perform gradient update on the diffusion model, continue to train the diffusion model until the diffusion model converges, and obtain a trained diffusion model.

[0013] The present invention also provides a mobile traffic prediction device based on a diffusion model, including: A determination module, configured to determine the noise mobile traffic and the dynamic characteristics of the historical mobile traffic based on the obtained historical mobile traffic, where the noise mobile traffic is the mobile traffic after adding noise; A noise processing module, configured to perform noise estimation on the noise mobile traffic and the dynamic characteristics based on the noise prior estimation unit of the diffusion model to obtain a noise prior, and perform noise prediction on the noise mobile traffic based on the denoising network of the diffusion model to obtain a residual noise; A prediction module, configured to perform traffic prediction based on the noise prior, the residual noise, and the noise mobile traffic to obtain future mobile traffic.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the method for predicting mobile traffic based on a diffusion model as described in any one of the above is implemented.

[0015] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the method for predicting mobile traffic based on a diffusion model as described in any one of the above is implemented.

[0016] The present invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, the method for predicting mobile traffic based on a diffusion model as described in any one of the above is implemented.

[0017] The method and device for predicting mobile traffic based on a diffusion model provided by the present invention determine the dynamic characteristics of the noise mobile traffic and the historical mobile traffic based on the obtained historical mobile traffic, where the noise mobile traffic is the mobile traffic after adding noise; the noise prior estimation unit based on the diffusion model performs noise estimation on the noise mobile traffic and the dynamic characteristics to obtain a noise prior, and performs noise prediction on the noise mobile traffic based on the denoising network of the diffusion model to obtain residual noise; according to the noise prior, the residual noise, and the noise mobile traffic, traffic prediction is performed to obtain future mobile traffic. By capturing the intrinsic dynamic characteristics of mobile traffic, a reference benchmark is provided for noise estimation, ensuring the accuracy of the noise prior. By obtaining the noise residual to capture the additional changes that the noise prior fails to fully capture, the diffusion model is given the ability to handle data complexity and unpredictability, and the noise is reconstructed based on the noise prior and the noise residual, thereby improving the accuracy of mobile traffic prediction. Description of the Drawings

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0019] Figure 1 It is a flowchart showing the method for predicting mobile traffic based on a diffusion model provided by the present invention.

[0020] Figure 2 It is a flowchart showing the denoising process of the diffusion model based on the noise prior provided by the present invention.

[0021] Figure 3 It is a schematic diagram showing the extraction of dynamic characteristics provided by the present invention.

[0022] Figure 4 It is a schematic diagram of the diffusion process in the diffusion model provided by the present invention.

[0023] Figure 5 It is a schematic structural diagram of a mobile traffic prediction device based on the diffusion model provided by the present invention.

[0024] Figure 6 It is a schematic structural diagram of an electronic device provided by the present invention. Detailed implementation manners

[0025] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without making creative efforts based on the embodiments in the present invention belong to the scope of protection of the present invention.

[0026] The following will be combined with Figures 1-6 Describe the mobile traffic prediction method and device based on the diffusion model of the present invention.

[0027] First, a brief description of the related content involved in the present invention will be given.

[0028] Mobile traffic prediction refers to using historical and real-time mobile traffic data to predict the change trend of mobile traffic in a specific future time period, which is of great significance for optimizing network performance and improving user experience. In the context of the popularization of the fifth-generation mobile communication technology (5G) network and mobile devices, mobile network operators are facing the explosive growth of data traffic and complex dynamic usage patterns, which pose challenges to network operation. Therefore, timely and accurate traffic prediction has become one of the research hotspots.

[0029] In the past ten-odd years, traditional deep learning methods, such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Transformers, Graph Convolutional Networks (GCNs), etc., directly map the historical traffic input to the output predicted traffic through a deterministic process and have achieved good results in prediction.

[0030] The recent emergence of diffusion models has provided a more promising alternative for this task. Different from traditional models, it provides a probability framework that captures the inherently complex distribution of time data and can effectively handle and model the uncertainty and complexity of mobile traffic data. Most current diffusion model-based solutions simply adopt the noise modeling method in the image domain and design a dedicated denoising network for it. In actual scenarios, the dynamic characteristics of mobile traffic data are very different from those in the image domain. It stems from human activities in daily life and has extremely high dynamics and periodicity.

[0031] However, so far, the method of reconstructing noise to retain the intrinsic characteristics of mobile traffic data remains largely an under-explored area.

[0032] Implementing the above idea is not an easy task. First, due to the driving of multiple factors such as human activity patterns and environmental changes, mobile traffic data exhibits high non-stationarity, which requires the model to be sensitive to sudden changes while capturing the regularity and predictable patterns of the data. Existing traditional deep learning-based and diffusion model-based methods often have difficulty effectively meeting these requirements and finding it hard to balance the capture of regular patterns and the sensitivity to sudden changes. Second, in the process of reconstructing noise, the Gaussian distribution characteristic of the predicted noise must be maintained, which is a key factor to ensure the stability and convergence of the diffusion model. Therefore, for existing diffusion model-based traffic prediction methods, how to effectively address these challenges to achieve noise reconstruction remains an urgent problem to be solved.

[0033] Figure 1 is a schematic flow diagram of the mobile traffic prediction method based on the diffusion model provided by the present invention. As Figure 1 shown, the method includes steps 101 to 103.

[0034] Step 101: Based on the obtained historical mobile traffic, determine the noise mobile traffic and the dynamic characteristics of the historical mobile traffic, where the noise mobile traffic is the mobile traffic after adding noise.

[0035] Specifically, the format of mobile traffic (data) is usually represented as a three-dimensional tensor , where represents the set of mobile traffic (data), represents the time length of the time series, represents the number of different base stations or regions, represents the number of traffic characteristics, is a real number. For mobile traffic prediction, this task is mainly to predict the future under the condition of given historical mobile traffic (data) with a length of ​ The traffic of the step, i.e., the future mobile traffic , where is context the abbreviation of is the time of historical mobile traffic is target the abbreviation of

[0036] In practical applications, the historical mobile traffic for prediction can be obtained first, and then the historical mobile traffic is preprocessed to obtain the noise mobile traffic corresponding to the historical mobile traffic and the dynamic characteristics of the historical mobile traffic , where represents the th step in the diffusion process of the diffusion model

[0037] Exemplarily, through the forward process of the diffusion model, the historical mobile traffic is denoised to obtain the noise mobile traffic; and based on a preset dynamic characteristic extraction strategy, the dynamic characteristics of the historical mobile traffic are extracted to obtain the dynamic characteristics of the historical mobile traffic

[0038] It should be noted that represents the set of mobile traffic (data), represents a mobile traffic sequence or data in , where is the total number of steps in the diffusion process of the diffusion model, i.e., the diffusion step length

[0039] Step 102: Based on the noise prior estimation unit of the diffusion model, perform noise estimation on the noise mobile traffic and the dynamic characteristics to obtain a noise prior, and perform noise prediction on the noise mobile traffic based on the denoising network of the diffusion model to obtain residual noise

[0040] In practical applications, referring to Figure 2 , Figure 2 is the flow chart of denoising based on the diffusion model with noise prior provided by the present invention: The noise can be calculated through two independent paths. The first path relies on the data dynamic background information, i.e., the dynamic characteristics to calculate the noise prior used as the benchmark for the denoising process, that is, the noise mobile traffic and the dynamic characteristics (data dynamics) are input into the noise prior estimation unit (noise prior estimation) of the diffusion model to obtain the noise prior The second path predicts the residual noise through a denoising network. , that is, moving the noise traffic Input it into the denoising network for noise estimation to obtain the residual noise .

[0041] Step 103: According to the noise prior, the residual noise, and the noise moving traffic, perform traffic prediction to obtain future moving traffic.

[0042] In practical applications, refer to Figure 2 , after obtaining the noise prior and the residual noise , based on the noise prior and the residual noise perform cyclic denoising on the noise moving traffic to obtain the final future moving traffic . In addition, the input of the denoising network also includes the current step (Step) of the diffusion process .

[0043] The mobile traffic prediction method based on the diffusion model provided by the present invention determines the noise moving traffic and the dynamic characteristics of the historical mobile traffic based on the obtained historical mobile traffic, where the noise moving traffic is the mobile traffic after adding noise; based on the noise prior estimation unit of the diffusion model, perform noise estimation on the noise moving traffic and the dynamic characteristics to obtain a noise prior, and based on the denoising network of the diffusion model, perform noise prediction on the noise moving traffic to obtain a residual noise; according to the noise prior, the residual noise, and the noise moving traffic, perform traffic prediction to obtain future moving traffic. By capturing the intrinsic dynamic characteristics of the mobile traffic, it provides a reference benchmark for noise estimation, ensures the accuracy of the noise prior, obtains additional changes that the noise prior fails to fully capture through the noise residual, endows the diffusion model with the ability to handle data complexity and unpredictability, and realizes the reconstruction of the noise based on the noise prior and the noise residual, thereby improving the accuracy of mobile traffic prediction.

[0044] In one or more alternative embodiments of the present invention, the performing traffic prediction according to the noise prior, the residual noise, and the noise moving traffic to obtain future moving traffic includes: Perform weighted fusion on the noise prior and the residual noise to obtain a reconstructed noise; Based on the reconstructed noise, perform denoising on the noise moving traffic multiple times to obtain future moving traffic.

[0045] Specifically, the reconstructed noise refers to the noise obtained after reconstructing the noise again.

[0046] In practical applications, refer toFigure 2 , the noise prior and the residual noise can be weighted and fused, so as to obtain an accurate noise result, that is, the reconstructed noise . Therefore, the reconstructed noise is defined here as the weighted combination of the noise prior and the residual noise predicted by the model. An adjustable weighting coefficient can be set to balance the contribution degrees of the noise prior and the residual noise , as shown in the specific formula (1): (1) By combining the noise prior with the residual noise output by the diffusion model , the diffusion model can integrate the inherent dynamic characteristics of the data by directly manipulating the noise at any diffusion step during the diffusion process. Further, based on the reconstructed noise , cyclic denoising is performed on the noise mobile traffic, that is, multiple denoising operations are performed, and the future mobile traffic can be obtained, that is, first obtain , and finally obtain . In this way, the accuracy of the future mobile traffic can be further improved.

[0047] In one or more alternative embodiments of the present invention, the step of performing multiple denoising operations on the noise mobile traffic based on the reconstructed noise to obtain the future mobile traffic includes: Performing denoising on the noise mobile traffic based on the reconstructed noise to obtain the denoised mobile traffic; When the number of denoising operations is less than the diffusion step length, using the denoised mobile traffic as the noise mobile traffic, and re-executing the noise prior estimation unit based on the diffusion model to perform noise estimation on the noise mobile traffic and the dynamic characteristics and subsequent steps, where the diffusion step length is the total number of steps in the diffusion process of the diffusion model; When the number of denoising operations is equal to the diffusion step length, using the denoised mobile traffic as the future mobile traffic.

[0048] In practical applications, after obtaining the reconstructed noise , the reconstructed noise can be used to gradually denoise the noise mobile traffic corresponding to the historical mobile traffic according to formula (2) until the final result, that is, the future mobile traffic is obtained: (2) where, Characterize the standard Gaussian distribution, Characterize the signal retention degree in the cumulative parameter, The noise schedule for the step of the diffusion process.

[0049] Exemplarily, a diffusion model sampling algorithm with the following noise prior can be used to predict future mobile traffic: Input: Historical mobile traffic Dynamic characteristics The denoising network in the trained diffusion model Noisy mobile traffic Output: Target prediction data (future mobile traffic) ; for n = N to 1 do Calculate the noise prior Predict the residual noise Obtain the denoised mobile traffic by reverse denoising through formula (2) end for Return: Future mobile traffic 。

[0050] It should be noted that the noise schedule for the entire diffusion process can be represented by , which is a key parameter increasing sequence that controls the noise addition process, controls the noise intensity at each step, and defines the size and evolution manner of the noise.

[0051] In one or more alternative embodiments of the present invention, determining the noisy mobile traffic and the dynamic characteristics of the historical mobile traffic based on the obtained historical mobile traffic includes: Based on the forward noise addition process and the diffusion step size of the diffusion model, adding noise to the obtained historical mobile traffic multiple times to obtain the noisy mobile traffic, where the diffusion step size is the total number of steps in the diffusion process of the diffusion model; According to the prediction step size, extract the dynamic characteristics of the historical mobile traffic to obtain the dynamic characteristics of the historical mobile traffic, where the prediction step size characterizes the prediction length between the future mobile traffic and the historical mobile traffic.

[0052] In practical applications, random white noise can be collected from a standard Gaussian distribution through the forward noise addition process of the denoising network and the diffusion steps of the diffusion model, and the historical mobile traffic can be noise-added multiple times to obtain the noisy mobile traffic.

[0053] Exemplarily, first, the diffusion model samples random white noise from a standard Gaussian distribution . Then, in the forward noise addition process (forward process) of the diffusion model, the random white noise will be gradually added to the historical mobile traffic under the constraint of the noise schedule to obtain the noise-added mobile traffic , until the forward noise addition process is completed steps to obtain the final noisy mobile traffic . Among them, the target data damaged at any diffusion step, that is, the noise-added mobile traffic , can be directly calculated by the following formula (3): (3) where , and its physical meaning is mainly reflected in the trade-off between the data signal fidelity and the noise ratio.

[0054] In addition, for different prediction steps, different methods can be used to extract dynamic characteristics to obtain the dynamic characteristics of the historical mobile traffic, so as to improve the accuracy of the dynamic characteristics.

[0055] In one or more alternative embodiments of the present invention, the extracting the dynamic characteristics of the historical mobile traffic according to the prediction step to obtain the dynamic characteristics of the historical mobile traffic includes: When the prediction step is greater than a set value, a periodic dynamic extraction method is used to extract the dynamic characteristics of the historical mobile traffic to obtain the dynamic characteristics of the historical mobile traffic.

[0056] Specifically, the set value can be 1.

[0057] In practical applications, when the prediction step is greater than 1, that is, in the case of multi-step prediction, a periodic dynamic extraction method can be used to extract dynamic characteristics. The periodic dynamic extraction method is a pattern that captures regular and predictable patterns.

[0058] See Figure 3 , Figure 3Schematic diagram for extracting dynamic characteristics provided by the present invention: The periodic dynamic extraction method utilizes the Fast Fourier Transform (FFT) to capture periodic dynamic characteristics and is a technique widely used in time series analysis. Specifically, the periodic dynamic extraction method obtains several key components through FFT and converts them back to the time domain. This process can be expressed by formulas (4)-(6): where, is the amplitude of the th frequency component obtained by applying FFT to the data (here referring to historical mobile traffic), represents the phase of the th frequency component. represents the length of the input sequence (historical mobile traffic). is the maximum selected through amplitudes, used for the selection of key components. is the number of selected frequency components used to calculate the periodic dynamics. refers to the frequency corresponding to the th component, while represents the corresponding conjugate complex number. is the value after mapping the th point in the input sequence to the time domain.

[0059] Then, a specified period length is specified, and the time domain signals at the same time point for different periods are averaged to extract the final periodic dynamic characteristics , that is, the dynamic characteristic , as shown in formula (7): where, represents the total number of complete periods, while represents the sampling value at the time point within each period .

[0060] In the embodiment of the present invention, the periodic dynamic characteristic is extended to the entire data set (historical mobile traffic) within the period to obtain the dynamic characteristic corresponding to each time point. In most historical mobile traffic, the period is defined as one week. If the historical mobile traffic is large enough, a longer period corresponding to the behavior pattern of human activities, such as one month or one year, can be selected.

[0061] In one or more alternative embodiments of the present invention, the dynamic characteristics of the historical mobile traffic are extracted according to the prediction step length to obtain the dynamic characteristics of the historical mobile traffic, including: When the prediction step length is less than or equal to the set value, an adjacency dynamic extraction method is used to extract the dynamic characteristics of the historical mobile traffic to obtain the dynamic characteristics of the historical mobile traffic.

[0062] Specifically, the set value can be 1.

[0063] In practical applications, when the prediction step length is equal to 1, that is, in the case of single-step prediction, an adjacency dynamic extraction method can be used to extract dynamic characteristics. The adjacency dynamic extraction method can reflect subtle but sudden changes in mobile traffic.

[0064] See Figure 3 , the periodic dynamic characteristics reflect the periodic pattern, but it may not be suitable for capturing short-term changes. Mobile traffic (data) usually depends on continuous time steps, and the current value has a strong correlation with the most recent previous step. Therefore, the embodiments of the present invention also take adjacency dynamics into consideration. Specifically, the embodiments of the present invention use the first-order lag value to describe the adjacency dynamic characteristics , that is, the change of the dynamic characteristics . Specifically, the adjacency dynamic characteristics can be calculated by formula (8) : Where is the data for calculating the dynamic characteristics, which refers to the historical mobile traffic here.

[0065] In one or more alternative embodiments of the present invention, the noise prior estimation unit based on the diffusion model estimates the noise of the noise mobile traffic and the dynamic characteristics to obtain the noise prior, including: The noise prior is calculated using the following formula (9): Where is the noise prior, is the noise mobile traffic, is the cumulative parameter of the signal retention degree in the th step of the diffusion process, is the dynamic characteristic.

[0066] Specifically, the dynamic characteristics are extracted (Periodic dynamic characteristics and / or proximity dynamic characteristics ), first assume that the dynamic characteristics are closely related to the corresponding true values (future mobile traffic). Based on this assumption, the future mobile traffic can be reconstructed as shown in Equation (10): where represents a small residual term that quantifies the difference between the target value and the extracted dynamics. Then, after transforming Equation (3), Equation (11) is obtained: where the random white noise is the noise to be predicted. Subsequently, substituting Equation (10) into Equation (11), Equation (12) can be obtained: (12) Furthermore, the noise prior can be as shown in Equation (13): In one or more alternative embodiments of the present invention, the training process of the diffusion model includes: Obtaining training mobile traffic, where the training mobile traffic includes sample mobile traffic and target mobile traffic; Sampling random white noise from a standard Gaussian distribution; Based on the random white noise, the noise schedule of the diffusion model, and the diffusion step size, adding noise to the target mobile traffic multiple times to obtain the fake target mobile traffic after adding noise, where the diffusion step size is the total number of steps in the diffusion process of the diffusion model; Extracting the dynamic characteristics of the sample mobile traffic to obtain the dynamic characteristics of the sample mobile traffic; Based on the noise prior estimation unit of the diffusion model, estimating the noise of the fake target mobile traffic after adding noise and the dynamic characteristics of the sample mobile traffic to obtain the predicted noise prior, and predicting the residual noise of the fake target mobile traffic after adding noise based on the denoising network of the diffusion model; Calculating the loss value based on the predicted noise prior, the predicted residual noise, and the random white noise; Updating the gradient of the diffusion model based on the loss value, and continuing to train the diffusion model until the diffusion model converges to obtain the trained diffusion model.

[0067] In practical applications, before using the diffusion model, it needs to be trained. During the training process, the diffusion model first samples random white noise from the standard Gaussian distribution , and according to formula (3), is converted to , and the noise prior is calculated using the extracted dynamic characteristics . The parameters of the diffusion model are optimized by minimizing the loss function , and minimizing the loss function is shown in formula (14): Exemplarily, the following diffusion model training algorithm with noise prior can be used to train the diffusion model: Input: Training mobile traffic Noise schedule Data dynamic characteristics Output: The trained diffusion model and the denoising network in the diffusion model Repeat: Divide the input data into sample mobile traffic (historical background data) and target mobile traffic Sample and Calculate the noise sample (the target mobile traffic after adding noise) Calculate the predicted noise prior Perform gradient update: Until: Convergence.

[0068] The mobile traffic prediction method based on the diffusion model provided by the present invention will be further described below.

[0069] 1. Overall framework.

[0070] Referring to Figure 2 , in the embodiment of the present invention, the dynamic characteristics of mobile traffic are incorporated into the diffusion process as noise prior, and the final noise is calculated through two independent paths. The first path relies on the data dynamic background information to calculate the noise prior as the benchmark for the denoising process. The second path predicts the residual noise through the denoising network. By cleverly weighting and fusing these two parts, the diffusion model can obtain accurate noise results.

[0071] Regarding the acquisition of background information on data dynamics (dynamic characteristics), such as Figure 3 As shown, two core dynamics are accurately extracted from mobile traffic data: periodic dynamics and proximity dynamics. Among them, the periodic dynamics are obtained through the FFT technique, while the proximity dynamics are represented by the first-order lag value, which not only enhances the model's ability to capture data dynamics but also improves the accuracy of denoising and prediction.

[0072] 2. Mobile traffic prediction based on the diffusion model.

[0073] The format of mobile traffic data is usually represented as a three-dimensional tensor . Specifically, this task mainly involves learning a diffusion model that predicts the future traffic for steps given historical data of length .

[0074] See Figure 4 , Figure 4 which is a schematic diagram of the diffusion process in the diffusion model provided by the present invention: In the forward process (forward noise addition process) of the diffusion model, noise is gradually added to the target data under the constraint of the noise schedule . The target data damaged at any diffusion step can be directly calculated by the following formula: In the reverse process (reverse denoising process), the diffusion model first samples random white noise from the standard Gaussian distribution and is gradually denoised by the denoising network under the condition of historical data. This process can be defined as a Markov process, as shown in the following formula: where represents the conditional probability distribution of the model in the reverse generation process, is the mean of this distribution, which needs to be fitted by the model during the diffusion process, is the variance of this conditional distribution, which is fixed as a constant during the diffusion process, is the noise target sample distribution sampled at the initial stage of prediction. Finally, the model is trained to estimate the noise vector added to the data . The parameters of the model will be updated by solving the following optimization problem: 3. Extraction of the dynamics of mobile traffic data By analyzing the dynamic characteristics of mobile traffic data, two representative dynamics are selected to calculate the noise prior, which are respectively defined as the periodic dynamic characteristic and the proximity dynamic characteristic. The periodic dynamic characteristic captures regular and predictable patterns, while the proximity dynamic characteristic reflects subtle but sudden changes.

[0075] Periodic dynamic characteristic: The fast Fourier transform is used to capture the periodic dynamic characteristic, which is a widely used technique in time series analysis. Specifically, several key components obtained by FFT are selected in the present invention and converted back to the time domain. This process can be expressed as follows: Then, the period length is specified , and the time domain signals at the same time points of different periods are averaged to extract the final periodic dynamic characteristic , that is, the dynamic characteristic , as shown in the following formula: Proximity dynamic characteristic: It reflects the periodic pattern, which may not be suitable for capturing short-term changes. Mobile traffic data usually depends on consecutive time steps, and there is a strong correlation between the current value and the nearest previous step. Therefore, the proximity dynamic can be taken into consideration. Specifically, the first-order lag value can be used to describe the proximity change, as shown in the following formula: 3. Derivation and fusion of the noise prior.

[0076] Derivation of the noise prior: After extracting the periodic dynamic characteristic and the proximity dynamic characteristic, assuming that they are closely related to the corresponding true values, the mobile traffic data can be reconstructed, as shown in the following formula: where represents a small residual term, which quantifies the difference between and the extracted dynamics. Then is transformed to obtain the following formula: Here is the noise that the model needs to predict during the training process, the true value is random white noise, and the actually predicted one is the reconstructed noise .

[0077] Subsequently, formula is substituted into formula , the following formula can be obtained: Therefore, the noise prior can be defined as: The target noise that the diffusion model needs to predict can be rewritten as: where is the residual noise that the denoising network needs to estimate.

[0078] Noise prior fusion: According to the above content, when the provided dynamic prior is very close to the target value, that is, is very small, the calculated noise prior can naturally serve as the reference noise for the target noise . Considering that the dynamic prior extracted from the data represents the regular pattern of the data and may not always be accurate enough in the face of irregular changes, the denoising network is still needed to estimate the residual noise. Therefore, the reconstructed noise is defined as the weighted combination of the prior noise and the residual noise predicted by the model. An adjustable weighting coefficient can be set to balance the contribution degrees of the noise prior and the residual noise as shown in the following formula: During the diffusion process, the present invention can calculate the corresponding noise prior at any diffusion step according to the formula for defining the noise prior. And through the reconstructed noise, the noise prior can be combined with the output of the model, enabling the model to integrate the inherent dynamic characteristics of the data by directly manipulating the noise at any diffusion step in the diffusion process.

[0079] 4. Training and Inference Training: During the training process, the diffusion model first samples random white noise from the standard Gaussian distribution, and converts the target data into a noise version , and calculates the noise prior using the extracted dynamic prior. The parameters of the model are optimized by minimizing the following loss function: Inference: First, sample the noise data at the denoising start stage from the standard Gaussian distribution. Then estimate the noise prior , and denoise step by step according to the following formula until the final result is obtained : 5. Structure of the denoising network.

[0080] Specifically, the denoising network can be any kind of denoising network. Preferably, the denoising network can be a representative model from the spatio-temporal domain, such as CSDI (Conditional Score-based Diffusion Models for Probabilistic Time Series Imputation), ConvLSTM (Convolutional LSTM Network), and STID (Spatial-Temporal Identity). CSDI Transformers is used as the denoising network, ConvLSTM is based on Convolutional LSTM, and STID uses a multi-layer perceptron (MLP) as its backbone network. These models cover the main methods in deep learning, such as CNN, LSTM, multi-layer perceptron (MLP), and Transformers.

[0081] 6. Specific implementation and testing.

[0082] In the specific verification, the following four real-world mobile traffic datasets are selected to verify the effectiveness of the present invention: MobileA, MobileB, MobileC1, and MobileC2. These datasets come from three major cities: A, B, and C. The detailed descriptions of these datasets are shown in the following table.

[0083] Thirteen existing models are used in the verification process to compare with the present invention. They can be roughly divided into four categories: Classical methods: HA, ARIMA.

[0084] Time series models: PatchTST, iTransformer, Time-LLM.

[0085] City spatio-temporal models: STResNet, ATFM, STNorm, STGSP, TAU, PromptST.

[0086] Video prediction models: MAU, MIM.

[0087] The final results are evaluated using two classic metrics in the spatio-temporal domain, namely MAE (Mean Absolute Error) and RMSE (Root Mean Square Error). The calculation methods are as follows: Finally, the present invention achieved excellent results on four datasets. The best results of various baseline models and the results of the present invention (using CSDI as the denoising network) are recorded in the following table.

[0088] In the embodiments of the present invention, the data dynamic pattern is used as prior information to reconstruct the noise, thereby improving the prediction accuracy. The noise is divided into two key parts: Firstly, the noise prior, which captures the inherent dynamic characteristics of mobile traffic data, including periodic patterns and local variations, provides a reference benchmark for noise estimation at each step; Secondly, the noise residual, which explains the additional variations that the noise prior fails to fully capture, endows the model with the ability to handle data complexity and unpredictability. The present invention can utilize the noise prior on the premise of conforming to the noise distribution, reconstruct the noise predicted by the model during the diffusion process to enhance the prediction performance. As a general solution, the present invention can be seamlessly integrated into existing diffusion models, enhancing their prediction performance and enabling them to better adapt to the uniqueness of mobile traffic data.

[0089] The mobile traffic prediction device based on the diffusion model provided by the present invention is described below. The mobile traffic prediction device based on the diffusion model described below can be mutually referred to the mobile traffic prediction method based on the diffusion model described above.

[0090] Figure 5 is a schematic structural diagram of the mobile traffic prediction device based on the diffusion model provided by the present invention, as Figure 5 shown. The mobile traffic prediction device based on the diffusion model includes: A determination module 501, configured to determine the noise mobile traffic and the dynamic characteristics of the historical mobile traffic based on the acquired historical mobile traffic, where the noise mobile traffic is the mobile traffic after adding noise; A noise processing module 502, configured to perform noise estimation on the noise mobile traffic and the dynamic characteristics based on the noise prior estimation unit of the diffusion model to obtain a noise prior, and perform noise prediction on the noise mobile traffic based on the denoising network of the diffusion model to obtain residual noise; A prediction module 503, configured to perform traffic prediction based on the noise prior, the residual noise, and the noise mobile traffic to obtain future mobile traffic.

[0091] Optionally, the prediction module 503 is specifically configured to: Perform weighted fusion on the noise prior and the residual noise to obtain a reconstructed noise; Based on the reconstructed noise, denoise the noise mobile traffic multiple times to obtain future mobile traffic.

[0092] Optionally, the prediction module 503 is specifically configured to: Based on the reconstructed noise, denoise the noise mobile traffic to obtain the denoised mobile traffic; When the number of denoising times is less than the diffusion step size, use the denoised mobile traffic as the noise mobile traffic, and re - execute the noise prior estimation unit based on the diffusion model to perform noise estimation on the noise mobile traffic and the dynamic characteristics and subsequent steps. The diffusion step size is the total number of steps in the diffusion process of the diffusion model; When the number of denoising times is equal to the diffusion step size, use the denoised mobile traffic as the future mobile traffic.

[0093] Optionally, the determination module 501 is specifically configured to: Based on the forward noise - adding process and the diffusion step size of the diffusion model, add noise to the obtained historical mobile traffic multiple times to obtain noise mobile traffic. The diffusion step size is the total number of steps in the diffusion process of the diffusion model; According to the prediction step size, extract the dynamic characteristics of the historical mobile traffic to obtain the dynamic characteristics of the historical mobile traffic. The prediction step size represents the prediction length between the future mobile traffic and the historical mobile traffic.

[0094] Optionally, the determination module 501 is specifically configured to: When the prediction step size is greater than a set value, use a periodic dynamic extraction method to extract the dynamic characteristics of the historical mobile traffic to obtain the dynamic characteristics of the historical mobile traffic; When the prediction step size is less than or equal to the set value, use a proximity - based dynamic extraction method to extract the dynamic characteristics of the historical mobile traffic to obtain the dynamic characteristics of the historical mobile traffic.

[0095] Optionally, the noise processing module 502 is specifically configured to: Calculate the noise prior using the following formula: Where, is the noise prior, is the noise mobile traffic, is the The cumulative parameter of signal retention degree in the step, is the dynamic characteristic.

[0096] Optionally, the mobile traffic prediction device based on the diffusion model further includes a training module configured to: Obtain training mobile traffic, where the training mobile traffic includes sample mobile traffic and target mobile traffic; Sample random white noise from the standard Gaussian distribution; Based on the random white noise, the noise schedule of the diffusion model, and the diffusion step size, add noise to the target mobile traffic multiple times to obtain the faked target mobile traffic after adding noise, where the diffusion step size is the total number of steps in the diffusion process of the diffusion model; Extract the dynamic characteristics of the sample mobile traffic to obtain the dynamic characteristics of the sample mobile traffic; Based on the noise prior estimation unit of the diffusion model, perform noise estimation on the faked target mobile traffic after adding noise and the dynamic characteristics of the sample mobile traffic to obtain a predicted noise prior, and perform noise prediction on the faked target mobile traffic after adding noise based on the denoising network of the diffusion model to obtain a predicted residual noise; Calculate a loss value based on the predicted noise prior, the predicted residual noise, and the random white noise; Update the gradient of the diffusion model based on the loss value, continue to train the diffusion model until the diffusion model converges, and obtain a trained diffusion model.

[0097] Figure 6 Illustrates a schematic diagram of the physical structure of an electronic device, as Figure 6 shown. The electronic device may include: a processor 610, a communication interface 620, a memory 630, and a communication bus 640. Among them, the processor 610, the communication interface 620, and the memory 630 complete mutual communication through the communication bus 640. The processor 610 can call the logical instructions in the memory 630 to execute the mobile traffic prediction method based on the diffusion model. The method includes: based on the obtained historical mobile traffic, determining the noisy mobile traffic and the dynamic characteristics of the historical mobile traffic, where the noisy mobile traffic is the mobile traffic after adding noise; based on the noise prior estimation unit of the diffusion model, performing noise estimation on the noisy mobile traffic and the dynamic characteristics to obtain a noise prior, and performing noise prediction on the noisy mobile traffic based on the denoising network of the diffusion model to obtain a residual noise; and performing traffic prediction based on the noise prior, the residual noise, and the noisy mobile traffic to obtain future mobile traffic.

[0098] In addition, when the logical instructions in the above-mentioned memory 630 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0099] On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the mobile traffic prediction method based on the diffusion model provided by the above-mentioned various methods. The method includes: based on the acquired historical mobile traffic, determining the dynamic characteristics of the noise mobile traffic and the historical mobile traffic, where the noise mobile traffic is the mobile traffic after adding noise; based on the noise prior estimation unit of the diffusion model, performing noise estimation on the noise mobile traffic and the dynamic characteristics to obtain a noise prior, and performing noise prediction on the noise mobile traffic based on the denoising network of the diffusion model to obtain residual noise; according to the noise prior, the residual noise, and the noise mobile traffic, performing traffic prediction to obtain future mobile traffic.

[0100] On another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is implemented to execute the mobile traffic prediction method based on the diffusion model provided by the above-mentioned various methods. The method includes: based on the acquired historical mobile traffic, determining the dynamic characteristics of the noise mobile traffic and the historical mobile traffic, where the noise mobile traffic is the mobile traffic after adding noise; based on the noise prior estimation unit of the diffusion model, performing noise estimation on the noise mobile traffic and the dynamic characteristics to obtain a noise prior, and performing noise prediction on the noise mobile traffic based on the denoising network of the diffusion model to obtain residual noise; according to the noise prior, the residual noise, and the noise mobile traffic, performing traffic prediction to obtain future mobile traffic.

[0101] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative effort.

[0102] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A mobile traffic prediction method based on a diffusion model, characterized in that, Including: Based on the obtained historical mobile traffic, determine the noise mobile traffic and the dynamic characteristics of the historical mobile traffic, where the noise mobile traffic is the mobile traffic after adding noise; Based on the noise prior estimation unit of the diffusion model, perform noise estimation on the noise mobile traffic and the dynamic characteristics to obtain a noise prior, and perform noise prediction on the noise mobile traffic based on the denoising network of the diffusion model to obtain residual noise; According to the noise prior, the residual noise, and the noise mobile traffic, perform traffic prediction to obtain future mobile traffic.

2. The mobile traffic prediction method based on the diffusion model according to claim 1, wherein The step of performing traffic prediction according to the noise prior, the residual noise, and the noise mobile traffic to obtain future mobile traffic includes: Perform weighted fusion on the noise prior and the residual noise to obtain reconstructed noise; Based on the reconstructed noise, perform denoising on the noise mobile traffic multiple times to obtain future mobile traffic.

3. The mobile traffic prediction method based on the diffusion model according to claim 2, characterized in that The step of performing denoising on the noise mobile traffic multiple times based on the reconstructed noise to obtain future mobile traffic includes: Based on the reconstructed noise, perform denoising on the noise mobile traffic to obtain denoised mobile traffic; When the number of denoising times is less than the diffusion step size, use the denoised mobile traffic as the noise mobile traffic, and re-execute the noise prior estimation unit based on the diffusion model to perform noise estimation on the noise mobile traffic and the dynamic characteristics and subsequent steps. The diffusion step size is the total number of steps in the diffusion process of the diffusion model; When the number of denoising times is equal to the diffusion step size, use the denoised mobile traffic as future mobile traffic.

4. The mobile traffic prediction method based on a diffusion model according to claim 1, wherein The step of determining the noise mobile traffic and the dynamic characteristics of the historical mobile traffic based on the obtained historical mobile traffic includes: Based on the forward noise addition process and the diffusion step size of the diffusion model, perform noise addition on the obtained historical mobile traffic multiple times to obtain noise mobile traffic. The diffusion step size is the total number of steps in the diffusion process of the diffusion model; According to the prediction step size, perform dynamic characteristic extraction on the historical mobile traffic to obtain the dynamic characteristics of the historical mobile traffic. The prediction step size represents the prediction length between the future mobile traffic and the historical mobile traffic.

5. The method for predicting mobile traffic based on a diffusion model according to claim 4, wherein, The step of performing dynamic characteristic extraction on the historical mobile traffic according to the prediction step size to obtain the dynamic characteristics of the historical mobile traffic includes: When the prediction step size is greater than a set value, use a periodic dynamic extraction method to perform dynamic characteristic extraction on the historical mobile traffic to obtain the dynamic characteristics of the historical mobile traffic; When the prediction step size is less than or equal to the set value, use a proximity dynamic extraction method to perform dynamic characteristic extraction on the historical mobile traffic to obtain the dynamic characteristics of the historical mobile traffic.

6. The method for predicting mobile traffic based on a diffusion model according to claim 1, characterized in that The step of the noise prior estimation unit based on the diffusion model performing noise estimation on the noise mobile traffic and the dynamic characteristics to obtain a noise prior includes: Calculate the noise prior using the following formula: Among them, is the noise prior,[ is the noise moving flow,[ is the cumulative parameter of the signal retention degree in the th step of the diffusion process,[ is the dynamic characteristic.[ 7. The mobile traffic prediction method based on the diffusion model according to claim 1, wherein The training process of the diffusion model includes: Obtain training mobile traffic, where the training mobile traffic includes sample mobile traffic and target mobile traffic; Sample random white noise from a standard Gaussian distribution; Based on the random white noise, the noise schedule, and the diffusion step size of the diffusion model, add noise to the target mobile traffic multiple times to obtain the fake target mobile traffic after noise addition, where the diffusion step size is the total number of steps in the diffusion process of the diffusion model; Extract the dynamic characteristics of the sample mobile traffic to obtain the dynamic characteristics of the sample mobile traffic; Based on the noise prior estimation unit of the diffusion model, estimate the noise for the noise-added target mobile traffic and the dynamic characteristics of the sample mobile traffic to obtain a predicted noise prior, and based on the denoising network of the diffusion model, predict the noise for the noise-added target mobile traffic to obtain a predicted residual noise; Calculate a loss value based on the predicted noise prior, the predicted residual noise, and the random white noise; Based on the loss value, perform gradient update on the diffusion model, and continue to train the diffusion model until the diffusion model converges to obtain a trained diffusion model.

8. A mobile traffic prediction device based on a diffusion model, characterized in that, Including: A determination module configured to determine, based on the acquired historical mobile traffic, a noise mobile traffic and the dynamic characteristics of the historical mobile traffic, where the noise mobile traffic is the mobile traffic after adding noise; A noise processing module configured to estimate the noise for the noise mobile traffic and the dynamic characteristics based on the noise prior estimation unit of the diffusion model to obtain a noise prior, and predict the noise for the noise mobile traffic based on the denoising network of the diffusion model to obtain a residual noise; A prediction module configured to perform traffic prediction based on the noise prior, the residual noise, and the noise mobile traffic to obtain future mobile traffic.

9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the mobile traffic prediction method based on the diffusion model according to any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the mobile traffic prediction method based on the diffusion model according to any one of claims 1 to 7.

Images