An asymmetric spatiotemporal data generation method and system based on a skew diffusion model
By introducing a channel-independent extended skewed distribution into traditional DDPM, the problem of modeling asymmetric spatiotemporal data in traditional DDPM is solved, and more efficient asymmetric spatiotemporal data generation and prediction are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGZHOU UNIVERSITY
- Filing Date
- 2026-03-04
- Publication Date
- 2026-06-09
AI Technical Summary
Traditional Denoising Diffusion Model (DDPM) is difficult to effectively fit the asymmetric spatiotemporal data distribution that is widespread in the real world, especially the obvious asymmetric features in the spatiotemporal dimension, which increases the difficulty of modeling.
A noise reduction diffusion model based on extended skewed distribution (ESN-DDPM) is adopted to model the spatiotemporal data as a dynamic graph structure. It is defined as a channel-independent univariate extended skewed distribution at each node and at each time point, which explicitly captures asymmetry while maintaining computational efficiency.
It significantly improves the model's ability to fit complex asymmetric spatiotemporal data, achieving a balance between quality and efficiency in data generation. The generated data is closer to the true statistical characteristics, resulting in lower prediction errors.
Smart Images
Figure CN122173922A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of spatiotemporal data generation technology, and in particular to an asymmetric spatiotemporal data generation method and system based on a skewed diffusion model. Background Technology
[0002] The Denoising Diffusion Probabilistic Model (DDPM), as an advanced generative model, has been widely used in various data generation tasks, including images, videos, and audio. In recent years, this model has also been extended to the field of spatiotemporal data generation, such as traffic flow prediction. DDPM mainly consists of two processes: forward denoising and backward denoising. The forward process gradually adds Gaussian noise to perturb the original data distribution to a standard normal distribution; the backward process learns a denoising network to gradually reconstruct the true data distribution from the noise. Traditional DDPM typically assumes that the denoised distribution is a learnable normal distribution.
[0003] However, the normal distribution, due to its symmetry, struggles to effectively fit the asymmetric data distributions that are prevalent in the real world. This is especially true for spatiotemporal data, which often exhibit significant asymmetry in their spatiotemporal dimensions, further complicating the modeling difficulty of DDPM on such data. The Extended Skew-Normal Distribution (ESN), as a probability distribution capable of flexibly characterizing asymmetry, theoretically offers the possibility of improving the diffusion model's ability to model asymmetric distributions.
[0004] To enhance the expressive power of DDPM, existing research has attempted to extend its denoising distribution. For example, some works have proposed a mechanism of learning variance through the mean vector to enhance parameter flexibility; others have generalized the denoising distribution to a continuous mixed normal distribution to improve the fitting ability to multimodal distributions. However, these methods still do not explicitly model the asymmetry of the data distribution, nor do they fully incorporate the structural characteristics of spatiotemporal data, thus remaining limited in generating spatiotemporal data with complex asymmetric distributions.
[0005] To address this, this invention proposes a denoising diffusion model based on extended skewed distribution, termed ESN-DDPM. This method models spatiotemporal data as a dynamic graph structure and defines the denoising distribution as a channel-independent univariate extended skewed distribution at each node and at each time step. This design not only explicitly captures asymmetry through the skewness parameter but also maintains computational efficiency comparable to the classic DDPM. Using a multivariate skewed distribution for overall modeling introduces cubic time complexity, while this invention, through the channel independence assumption, improves expressive power while avoiding a significant increase in computational complexity, thus achieving a balance between generation quality and efficiency. Summary of the Invention
[0006] The purpose of this invention is to provide an asymmetric spatiotemporal data generation method and system based on a skewed diffusion model, which extends the denoising distribution of the traditional denoising diffusion model to a channel-independent extended skewed distribution, making the diffusion model more effectively approximate complex asymmetric distributions.
[0007] To achieve the above objectives, this invention provides an asymmetric spatiotemporal data generation method based on a skewed diffusion model, comprising the following steps: S1. Obtain the context information of the target spatiotemporal data, and calculate the hyperparameter set of the denoising diffusion model in multiple diffusion steps based on the preset noise scheduling strategy; S2. Based on the noisy samples and context information of each diffusion step, determine the corresponding extended skewness distribution parameter set through a spatiotemporal denoising network. The parameter set includes location parameters, scale parameters, skewness parameters, and auxiliary parameters. S3. Construct a loss function and train the spatiotemporal denoising network based on the difference between the parameter set of the extended skewed distribution and the target denoising distribution calculated based on real samples. S4. Using the trained spatiotemporal denoising network, starting from the initial noise samples, the target spatiotemporal data is generated through an iterative denoising process with multiple diffusion steps.
[0008] Preferably, the extended skewed distribution is a channel-independent extended skewed distribution.
[0009] Preferably, the preset noise scheduling strategy is a linear noise scheduling strategy.
[0010] Preferably, step S1 specifically includes: S11, Given the number of diffusion steps The first step is to calculate the linear noise scheduling strategy. noise variance parameter of the step The expression is: ; in, The preset initial noise variance, This represents the final noise variance. S12, Order , , Calculate the variance of the noisy distribution. ; S13. Calculate the scale parameters used for subsequent calculations of the denoised distribution. .
[0011] Preferably, step S2 specifically includes: S21. For real samples From the standard normal distribution Medium sampling noise Calculate the first Noisy samples of the step The expression is: ; S22, The first time interval of all nodes at all times. Step band noise samples With context information Input to the spatiotemporal denoising network, encoded as the first... Hidden layer representation of the step In this context, the hidden layer representation of each node at each time step is a 4-dimensional vector: ; S23, Node At any moment Hidden layer representation Analysis as an extended skewed distribution The four parameters include the position parameter. Scale parameters skewness parameter and auxiliary parameters The specific formula is as follows: ; ; ; in, It is a four-dimensional vector. , , , Let represent the 0th to 3rd elements respectively. Substitute the four parameters into the probability density function of the extended skewed distribution: ; in, For the first Noisy samples of the step, The mean is variance is The probability density function of the normal distribution. This is the cumulative distribution function of the standard normal distribution.
[0012] Preferably, step S3 specifically includes: S31, based on the first node of each node Noisy samples of the step and real samples Calculate the target denoising distribution The expression is: ; Among them, the mean ; S32. Calculate the single-step denoising loss. The single-step denoising loss is the target denoising distribution. With the extended skewed distribution learned by the model The Kullback-Leibler divergence (KL divergence) between them is expressed as: ; The Kullback-Leibler divergence between two normal distributions. Expected item Estimated using the Monte Carlo method, The preset positive integer number of samples, yes Each sample was obtained from independent sampling. Sampled from normal distribution ; S33. By weighted summing of the single-step denoising losses of all denoising steps, the final loss function is obtained: ; The spatiotemporal denoising network is then trained using the final loss function.
[0013] Preferably, step S4 specifically includes: S41, from the standard normal distribution Initial noise samples were obtained by mid-sampling. ; S42, For each noise reduction step , Obtain the corresponding extended skewed distribution parameters. ; S43, Based on the skewness parameter Calculate the normalized skewness parameter ; S44, From the two-dimensional standard normal distribution Mid-sampling and calculate : ; S45. Through reparameterization techniques, Scaling and translation yield a distribution that follows an extended skewed distribution. samples ; S46. Repeat steps S42 to S45 until... The final generated sample is obtained. .
[0014] This invention also provides an asymmetric spatiotemporal data generation system based on a skewed diffusion model, comprising: The data preprocessing module is used to obtain contextual information of the target spatiotemporal data; The parameter calculation module is used to calculate the hyperparameter set based on a preset noise scheduling strategy. The spatiotemporal denoising network module is used to determine the parameter set of the extended skewed distribution based on noisy samples and contextual information; The training module is used to construct the loss function and train the spatiotemporal denoising network; The generation module is used to perform an iterative denoising process using the trained spatiotemporal denoising network module to generate target spatiotemporal data.
[0015] Therefore, this invention employs the aforementioned asymmetric spatiotemporal data generation method and system based on a skewed diffusion model. By innovatively replacing the symmetric normal distribution denoising process of the traditional diffusion model with an asymmetric extended skewed distribution (ESN), the model's ability to fit complex asymmetric spatiotemporal data (such as traffic flow and base station communication) distributions is significantly improved. Simultaneously, the model maintains computational efficiency comparable to benchmark methods, achieving a balance between generation quality and efficiency, and providing a superior solution for spatiotemporal data generation and prediction tasks.
[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0017] Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2 This is a diagram illustrating the training process of the skewed diffusion model according to an embodiment of the present invention. Figure 3 This is a sampling process diagram of the skewed diffusion model in an embodiment of the present invention; Figure 4 This is a framework diagram of the skewed diffusion model in an embodiment of the present invention. Detailed Implementation
[0018] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0019] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.
[0020] Example This embodiment provides an asymmetric spatiotemporal data generation method based on a skewed diffusion model, such as... Figure 1 As shown, it includes the following steps: S1. Obtain the context information of the target spatiotemporal data, and calculate the hyperparameter set of the denoising diffusion model in multiple diffusion steps based on the linear noise scheduling strategy. In this invention, the denoising distribution is selected as the channel-independent extended skewed distribution, named ESN-DDPM. S11, Given the number of diffusion steps The first step is to calculate the linear noise scheduling strategy. noise variance parameter of the step The expression is: ; S12, Order , , Calculate the variance of the noise distribution of ESN-DDPM. ; S13. Calculate the scale parameters used for subsequent calculations of the denoised distribution. .
[0021] S2. Based on the noisy samples and context information of each diffusion step, determine the corresponding extended skewness distribution parameter set through a spatiotemporal denoising network. The parameter set includes location parameters, scale parameters, skewness parameters, and auxiliary parameters. S21. For real samples From the standard normal distribution Medium sampling noise Calculate the first Noisy samples of the step The expression is: ; S22, The first time interval of all nodes at all times. Step band noise samples With context information Input to the spatiotemporal denoising network, encoded as the first... Hidden layer representation of the step In this context, the hidden layer representation of each node at each time step is a 4-dimensional vector: The spatiotemporal denoising network of this invention maintains the same network structure as the spatiotemporal denoising network of the Denoising Diffusion Probabilistic Model (DDPM), only changing the output dimension of the network from 1 to 4. S23, Node At any moment Hidden layer representation The parameters are parsed into a denoised distribution. Since this invention processes the hidden layer representation of each node at each time step independently for each channel, the labels are omitted to simplify the notation. and And the hidden layer representation is abbreviated as . Analysis as an extended skewed distribution The four parameters include the position parameter. Scale parameters skewness parameter and auxiliary parameters The specific formula is as follows: ; ; ; in, It is a four-dimensional vector. , , , Let represent the 0th to 3rd elements respectively. Substitute the four parameters into the probability density function of the extended skewed distribution: ; in, For the first Noisy samples of the step, The mean is variance is The probability density function of the normal distribution. This is the cumulative distribution function of the standard normal distribution.
[0022] S3. Construct a loss function based on the difference between the parameter set of the extended skewed distribution and the target denoising distribution calculated from real samples, and train the spatiotemporal denoising network accordingly. Figure 2 As shown, the learning rate is set to The number of training rounds is set to 300. S31, based on the first node of each node Noisy samples of the step and real samples Calculate the target denoising distribution The expression is: ; Among them, the mean ; S32. Calculate the single-step denoising loss. The single-step denoising loss is the target denoising distribution. With the extended skewed distribution learned by the model The Kullback-Leibler divergence (KL divergence) between two normal distributions, where the first term is the KL divergence between the two normal distributions, is expressed as follows: ; The second term is the expected term, which has no closing expression. This invention utilizes sampling... Sample The mean is calculated and estimated; its expression is as follows: ; Each of them Sampled from a normal distribution ; S33. By weighted summing of the single-step denoising losses of all denoising steps, the final loss function is obtained: ; The spatiotemporal denoising network is trained using the final loss function. This design aims to balance the denoising loss at different time steps, maintaining consistency with the original DDPM, to enhance the model's ability to reconstruct noise.
[0023] S4. Using the trained spatiotemporal denoising network, starting from the initial noisy samples, it generates the target spatiotemporal data through an iterative denoising process with multiple diffusion steps, such as... Figure 3 As shown.
[0024] S41, from the standard normal distribution Initial noise samples were obtained by mid-sampling. ; S42, For each noise reduction step , Obtain the corresponding extended skewed distribution parameters. ; S43, Based on the skewness parameter Calculate the normalized skewness parameter ; S44, From the two-dimensional standard normal distribution Mid-sampling and calculate : ; S45. Through reparameterization techniques, Scaling and translation yield a distribution that follows an extended skewed distribution. samples ; S46. Repeat steps S42 to S45 until... The final generated sample is obtained. ,like Figure 4 .
[0025] An asymmetric spatiotemporal data generation system based on a skewed diffusion model is also provided for performing the above-described method, including: The data preprocessing module is used to obtain contextual information of the target spatiotemporal data; The parameter calculation module is used to calculate the hyperparameter set based on a preset noise scheduling strategy. The spatiotemporal denoising network module is used to determine the parameter set of the extended skewed distribution based on noisy samples and contextual information; The training module is used to construct the loss function and train the spatiotemporal denoising network; The generation module is used to perform an iterative denoising process using the trained spatiotemporal denoising network module to generate target spatiotemporal data.
[0026] Example 1 To verify the effectiveness of the present invention, the effect of ESN-DDPM was verified on two spatiotemporal datasets, including the Beijing mobile traffic generation dataset and the PEMS08 traffic flow prediction dataset.
[0027] The Beijing mobile traffic generation dataset contains hourly mobile traffic data from 960 base stations over a week. This invention selects STK-Diff, the best-performing method on this dataset, as a comparison method. Furthermore, this invention compares two variants of ESN-DDPM to illustrate the effectiveness of the extended skewed distribution as a denoising distribution: 1. N-DDPM: The denoising distribution of this variant is a channel-independent normal distribution, and its mean and variance can be learned.
[0028] 2. SN-DDPM: The denoising distribution of this variant is a channel-independent skewed distribution, and its mean, scale parameter and skewness parameter can be learned.
[0029] This invention uses Jensen–Shannon Divergence (JSD) to measure the difference between the generated data distribution and the real data distribution from different aspects, specifically including the following four indicators: 1. T-JSD: This metric measures the difference in data distribution across the time dimension for each node.
[0030] 2. S-JSD: This metric measures the spatial distribution differences of all nodes.
[0031] 3. T-JSD-FO: This metric measures the distributional differences of the first-order differences of each node's data over time.
[0032] 4. S-JSD-FO: This metric measures the spatial distribution differences of the first-order differences of all node data.
[0033] The experimental results on the Beijing mobile traffic generation dataset are shown in Table 1: Table 1 Comparison of experimental results for the Beijing mobile traffic generation dataset
[0034] Experimental results show that ESN-DDPM achieves optimal values on all metrics (e.g., T-JSD is 0.3181, S-JSD-FO is 0.2256), significantly outperforming the comparison methods, demonstrating the superiority of extended skewed distribution in modeling asymmetric distributions.
[0035] The PEMS08 traffic flow prediction dataset records traffic flow data from the California highway system, including data from 170 sensors at 17,856 time points, with a 5-minute time interval. This invention selects advanced diffusion model methods on this dataset for comparison, including the TimeGrad diffusion model, the Conditional Diffusion Inclusion (CSDI) model, and the Spatiotemporal Graph Diffusion (DiffSTG) model. The average of multiple generation results from the diffusion model is used as the prediction result. Three metrics are used to measure the model's prediction performance: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Continuous Ranked Probability Score (CRPS). CRPS measures the uncertainty of the predicted value relative to reality; a smaller CRPS indicates a more accurate prediction. Experimental results on the PEMS08 traffic flow prediction dataset are shown in Table 2. Table 2 Comparison of experimental results for the PEMS08 traffic flow prediction dataset
[0036] On the PEMS08 traffic flow prediction dataset, the results show that ESN-DDPM achieved the lowest prediction error (MAE of 16.89 and RMSE of 26.31), and its CRPS was on par with the best baseline (0.06), demonstrating its high accuracy and strong uncertainty modeling ability in spatiotemporal data prediction tasks.
[0037] In summary, the results in Table 1 show that ESN-DDPM can reduce the difference between the generated data distribution and the real data distribution compared to existing methods; the results in Table 2 show that ESN-DDPM can reduce the prediction error of the diffusion model. Furthermore, ESN-DDPM outperforms SN-DDPM and N-DDPM, demonstrating the superiority of using the extended skewed distribution as the denoising distribution over the skewed and normal distributions.
[0038] Therefore, this invention employs an asymmetric spatiotemporal data generation method and system based on a skewed diffusion model. By innovatively replacing the symmetric normal distribution denoising process of the traditional diffusion model with an asymmetric extended skewed distribution (ESN), it significantly improves the model's ability to fit complex asymmetric spatiotemporal data distributions (such as traffic flow and base station communication). Experiments show that on standard datasets, this method outperforms existing state-of-the-art methods in both distribution similarity index (JSD) and prediction accuracy index (MAE / RMSE), generating data that more closely reflects true statistical characteristics and exhibiting lower prediction errors. Simultaneously, the model maintains computational efficiency comparable to benchmark methods, achieving a balance between generation quality and efficiency, and providing a superior solution for spatiotemporal data generation and prediction tasks.
[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A method for generating asymmetric spatiotemporal data based on a skewed diffusion model, characterized in that, Includes the following steps: S1. Obtain the context information of the target spatiotemporal data, and calculate the hyperparameter set of the denoising diffusion model in multiple diffusion steps based on the preset noise scheduling strategy; S2. Based on the noisy samples and context information of each diffusion step, determine the corresponding extended skewness distribution parameter set through a spatiotemporal denoising network. The parameter set includes location parameters, scale parameters, skewness parameters, and auxiliary parameters. S3. Construct a loss function and train the spatiotemporal denoising network based on the difference between the parameter set of the extended skewed distribution and the target denoising distribution calculated based on real samples. S4. Using the trained spatiotemporal denoising network, starting from the initial noise samples, the target spatiotemporal data is generated through an iterative denoising process with multiple diffusion steps.
2. The asymmetric spatiotemporal data generation method based on a skewed diffusion model according to claim 1, characterized in that, The extended skewed distribution is a channel-independent extended skewed distribution.
3. The asymmetric spatiotemporal data generation method based on a skewed diffusion model according to claim 1, characterized in that, The preset noise scheduling strategy is a linear noise scheduling strategy.
4. The asymmetric spatiotemporal data generation method based on a skewed diffusion model according to claim 3, characterized in that, Step S1 specifically includes: S11, Given the number of diffusion steps The first step is to calculate the linear noise scheduling strategy. noise variance parameter of the step The expression is: ； in, The preset initial noise variance, This represents the final noise variance. S12, Order , , Calculate the variance of the noisy distribution. ; S13. Calculate the scale parameters used for subsequent calculations of the denoised distribution. .
5. The asymmetric spatiotemporal data generation method based on a skewed diffusion model according to claim 2, characterized in that, Step S2 specifically includes: S21. For real samples From the standard normal distribution Medium sampling noise Calculate the first Noisy samples of the step The expression is: ； S22, The first time interval of all nodes at all times. Step band noise samples With context information Input to the spatiotemporal denoising network, encoded as the first... Hidden layer representation of the step In this context, the hidden layer representation of each node at each time step is a 4-dimensional vector: ; S23, Node At any moment Hidden layer representation Analysis as an extended skewed distribution The four parameters include the position parameter. Scale parameters skewness parameter and auxiliary parameters The specific formula is as follows: ； ； ； in, It is a four-dimensional vector. , , , Let represent the 0th to 3rd elements respectively. Substitute the four parameters into the probability density function of the extended skewed distribution: ； in For the first Noisy samples of the step, The mean is variance is The probability density function of the normal distribution. This is the cumulative distribution function of the standard normal distribution.
6. The asymmetric spatiotemporal data generation method based on a skewed diffusion model according to claim 5, characterized in that, Step S3 specifically includes: S31, based on the first node of each node Noisy samples of the step and real samples Calculate the target denoising distribution The expression is: ； Among them, the mean ; S32. Calculate the single-step denoising loss. The single-step denoising loss is the target denoising distribution. With the extended skewed distribution learned by the model The Kullback-Leibler divergence between them is expressed as: ； The Kullback-Leibler divergence between two normal distributions. Expected item Estimated using the Monte Carlo method, The preset positive integer number of samples, yes Each sample was obtained from independent sampling. Sampled from normal distribution ; S33. By weighted summing of the single-step denoising losses of all denoising steps, the final loss function is obtained: ； The spatiotemporal denoising network is then trained using the final loss function.
7. The asymmetric spatiotemporal data generation method based on a skewed diffusion model according to claim 6, characterized in that, Step S4 specifically includes: S41, From the standard normal distribution Initial noise samples were obtained by mid-sampling. ; S42, For each noise reduction step , Obtain the corresponding extended skewed distribution parameters. ; S43, Based on the skewness parameter Calculate the normalized skewness parameter ; S44, From the two-dimensional standard normal distribution Mid-sampling and calculate : ； S45. Through reparameterization techniques, Scaling and translation yield a distribution that follows an extended skewed distribution. samples ; S46. Repeat steps S42 to S45 until... The final generated sample is obtained. .
8. An asymmetric spatiotemporal data generation system based on a skewed diffusion model, used to execute the asymmetric spatiotemporal data generation method based on a skewed diffusion model as described in any one of claims 1-7, characterized in that, include: The data preprocessing module is used to obtain contextual information of the target spatiotemporal data; The parameter calculation module is used to calculate the hyperparameter set based on a preset noise scheduling strategy. The spatiotemporal denoising network module is used to determine the parameter set of the extended skewed distribution based on noisy samples and contextual information; The training module is used to construct the loss function and train the spatiotemporal denoising network; The generation module is used to perform an iterative denoising process using the trained spatiotemporal denoising network module to generate target spatiotemporal data.