Infectious disease transmission prediction method based on SEIR parameter prediction and diffusion generation
By combining SEIR model parameter prediction with a diffusion generation model, the problems of nonlinear fluctuations and uncertainties in infectious disease prediction are solved, generating multi-scenario probabilistic prediction results that conform to the transmission law, thus improving the accuracy and interpretability of the prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-15
- Publication Date
- 2026-03-27
AI Technical Summary
Existing methods for predicting infectious diseases are ill-suited to adapting to nonlinear fluctuations in parameters in complex real-world scenarios. Mechanistic models lack epidemiological constraints, and data-driven models lack the ability to quantify uncertainty, resulting in insufficient accuracy and interpretability of prediction results.
A bidirectional long short-term memory network is used to predict SEIR model parameters. The basic propagation trajectory is generated by combining it with an ordinary differential equation solver. The residual distribution is learned through a U-Net diffusion generation model to generate multi-scenario propagation prediction trajectories and output probabilistic results.
It achieves that the prediction results of infectious disease transmission conform to the epidemiological laws, has the adaptability to complex real-world factors and the ability to quantify uncertainty, and provides scientific and reliable multi-scenario prediction support.
Smart Images
Figure CN121747993A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of infectious disease transmission prediction, and more specifically to an infectious disease transmission prediction method based on SEIR parameter prediction and diffusion generation. Background Technology
[0002] Infectious disease transmission prediction is a core issue in the field of public health. Its goal is to analyze historical data to predict future disease trends and provide key information for the allocation of prevention and control resources and the formulation of intervention policies.
[0003] Currently, a major type of prediction method is based on models of infectious disease dynamics, such as the classic SEIR model. These models describe the transitions between susceptible, latent, infected, and recovered states in a population through a set of ordinary differential equations. The parameters have clear epidemiological meanings, and the prediction results have good interpretability. However, these models usually assume that key parameters are constant or follow simple preset change patterns, making it difficult to effectively characterize the nonlinear fluctuations of parameters caused by factors such as changes in social behavior, population mobility, and seasonal changes in the real world. This limits their prediction accuracy in complex and variable scenarios.
[0004] Another mainstream approach is the fully data-driven deep learning model. This type of model can automatically learn complex nonlinear relationships directly from massive time series data without the need for an explicit hypothesis propagation mechanism, thus possessing strong fitting capabilities. However, their drawback is that the prediction process lacks epidemiological physical constraints, and the results are only statistically reasonable. When the development deviates from the historical data distribution, the predictive performance of this type of model is prone to significant decline, and its generalization ability is insufficient.
[0005] In addition, some studies have attempted to combine the two types of methods mentioned above, such as using neural networks to predict the time-varying parameters of the SEIR model or constructing a neural ordinary differential equation framework. These hybrid methods have improved the adaptability and accuracy of the model to some extent. However, most of them are still limited to the paradigm of deterministic prediction, that is, for a given input, the model only outputs a unique future propagation curve. This cannot characterize the inherent randomness caused by factors such as individual heterogeneity, contact network randomness and uncertainty of intervention effects in evolution. It lacks the ability to quantify uncertainty, making it difficult for the prediction results to support refined decision-making based on risk probability.
[0006] Therefore, how to design an infectious disease transmission prediction method based on SEIR parameter prediction and diffusion generation that can generate probabilistic prediction results that conform to the transmission law and reflect multiple possibilities in the future, and provide more comprehensive and reliable support for scientific decision-making, is an urgent problem to be solved by those skilled in the art. Summary of the Invention
[0007] In view of this, the present invention provides an infectious disease transmission prediction method based on SEIR parameter prediction and diffusion generation. It aims to solve the problems in existing infectious disease prediction technologies, such as the inability of mechanistic models to adapt to complex real-world changes and the lack of physical constraints and uncertainty quantification capabilities in purely data-driven models. The method aims to achieve probabilistic prediction results that not only conform to epidemiological laws but also reflect multiple future possibilities, thereby providing more scientific and richer quantitative support for public health decision-making.
[0008] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting the spread of infectious diseases based on SEIR parameter prediction and diffusion generation includes the following steps: S1. Acquire historical multi-source time series data and preprocess it to obtain standardized time series feature sequences; S2. Using the standardized temporal feature sequence, a bidirectional long short-term memory network is used to predict the SEIR model parameters that change over time; S3. Based on the SEIR model parameters, solve the SEIR dynamic equations using an ordinary differential equation solver to generate a basic transmission trajectory that conforms to the infectious disease transmission mechanism. S4. Using the basic propagation trajectory as input, a U-Net-based diffusion generation model is used to model the residual distribution between real disease propagation data and the basic propagation trajectory, generating multi-scenario propagation prediction trajectories. S5. Based on the multi-scenario propagation prediction trajectory, output the probabilistic prediction result.
[0009] Preferably, S1 includes: Collect multi-source time-series data including daily new confirmed cases, cumulative number of cured cases, number of deaths, population mobility intensity, policy intervention intensity, meteorological conditions and social behavior indicators; The multi-source time-series data are time-aligned and missing value imputed to eliminate data inconsistencies and missing information. The processed data is normalized to obtain a standardized time series feature sequence. L represents the length of the historical time window. Let be the multidimensional eigenvector at time t.
[0010] Preferably, in S2, the SEIR model parameters include propagation rate. Latency rate and recovery rate ; Bidirectional Long Short-Term Memory (LSTM) networks encode the input sequence in the forward and backward directions using forward LSTM layers, respectively, to obtain the forward hidden state sequence. With backward hidden state sequence ; The two are concatenated into a comprehensive timing representation H=[ ; The SEIR parameter sequence is output through a fully connected layer and the Softplus activation function:
[0011] Where W and b are the network weights and biases, and the Softplus activation function ensures that the parameters are non-negative.
[0012] Preferably, the SEIR parameter sequence further includes temporal difference regularization, which involves adding a regularization term to the model training loss. accomplish:
[0013] Where T is the prediction time step and λ is the regularization coefficient. Preferably, in S3, the SEIR kinetic equations are:
[0014] Where S, E, I, and R represent the number of susceptible individuals, latent individuals, infected individuals, and recovered individuals, respectively; SI represents the contact opportunities of susceptible individuals and infected individuals; and N is the total population. The ordinary differential equation solver employs a fourth-order Runge-Kutta numerical integration algorithm to integrate the SEIR model parameters. Substituting the equations into the system of equations and solving them, a continuous basic propagation trajectory is generated. .
[0015] Preferably, the update formula for the fourth-order Runge-Kutta numerical integration algorithm is:
[0016] in, Let i be the current state vector, and h be the iteration step index. , , , This is the intermediate gradient estimate.
[0017] Preferably, in S4, the diffusion generation model structure includes a symmetric encoder and decoder; The encoder employs multi-layer one-dimensional convolution and downsampling operations to extract multi-scale temporal features; The decoder uses upsampling operations to gradually restore the temporal resolution and fuses the features of the corresponding layers of the encoder through skip connections; At each layer of the encoder and decoder, SEIR model parameters and their time derivative features are injected as physical mechanism constraints through a conditional embedding mechanism.
[0018] Preferably, in step S4, the training process of the diffusion generation model includes: Forward diffusion process: data on the actual spread of the disease Gaussian noise is added step by step to obtain a noisy sequence. , … T is the number of diffusion steps; Inverse denoising process: Training the U-Net network To recover the original data from the noisy sequence, the loss function is:
[0019] in, This represents the noisy data at step t. This represents the added Gaussian noise, and c represents the conditional information.
[0020] Preferably, the method employs a two-stage training strategy: In the first stage, the bidirectional long short-term memory network and the ordinary differential equation solver are trained independently to minimize the observation error between the basic propagation trajectory and the real data. In the second stage, the model parameters from the first stage are fixed, and the diffusion generation model is trained to learn the residual distribution.
[0021] Preferably, in S5, the probabilistic prediction results include: the predicted mean and confidence interval of the number of infections at each future time point, and the probability distribution of key transmission indicators.
[0022] As can be seen from the above technical solution, compared with the prior art, the technical solution of the present invention has the following beneficial effects: 1. This method dynamically predicts the time-varying parameters of the SEIR model through a bidirectional long short-term memory network and generates the basic propagation trajectory using an ordinary differential equation solver, ensuring that the prediction results strictly follow the laws of infectious disease dynamics and have clear epidemiological interpretability; at the same time, it uses deep neural networks to automatically learn from multi-source time series data, overcoming the problem of insufficient adaptability caused by fixed parameters or simple assumptions.
[0023] 2. A diffusion generation model based on U-Net is introduced. Under the condition of the basic trajectory under the aforementioned mechanism constraints, the residual distribution between real data and mechanism prediction is learned. Through reverse denoising sampling, multiple future propagation trajectories that conform to physical laws and reflect the uncertainty of reality can be naturally generated, thereby realizing the probabilistic prediction of key indicators.
[0024] 3. By linking the core steps of parameter prediction, mechanism solving and generative modeling, and co-optimizing them through a two-stage training strategy, we not only ensure gradient propagation and overall performance optimization throughout the entire process from data to prediction results, but also make the model have good structural clarity and scalability, making it easy to integrate into other dynamic models or adapt to different infectious disease prediction scenarios. Attached Figure Description
[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0026] Figure 1 This is a flowchart of an infectious disease transmission prediction method based on SEIR parameter prediction and diffusion generation, provided as an embodiment of the present invention. Detailed Implementation
[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0028] like Figure 1 As shown, this embodiment provides a method for predicting the spread of infectious diseases based on SEIR parameter prediction and diffusion generation, including the following steps: S1. Acquire historical multi-source time series data and preprocess it to obtain standardized time series feature sequences; S2. Using the standardized temporal feature sequence, a bidirectional long short-term memory network is used to predict the SEIR model parameters that change over time; S3. Based on the SEIR model parameters, solve the SEIR dynamic equations using an ordinary differential equation solver to generate a basic transmission trajectory that conforms to the infectious disease transmission mechanism. S4. Using the basic propagation trajectory as input, a U-Net-based diffusion generation model is used to model the residual distribution between real disease propagation data and the basic propagation trajectory, generating multi-scenario propagation prediction trajectories. S5. Based on the multi-scenario propagation prediction trajectory, output the probabilistic prediction result.
[0029] This method integrates time-varying SEIR mechanism modeling with diffusion generation models, which improves adaptability and prediction accuracy to complex real-world factors while maintaining epidemiological interpretability. It also enables probabilistic multi-scenario extrapolation of future transmission trends, thus providing decision-making with prediction results that combine mechanistic reliability and uncertainty quantification.
[0030] The following provides a further detailed explanation of each step and related features in the above method; In this embodiment, S1 involves acquiring historical multi-source time-series data and preprocessing it to obtain a standardized time-series feature sequence; including: Collect multi-source time-series data including daily new confirmed cases, cumulative number of cured cases, number of deaths, population mobility intensity, policy intervention intensity, meteorological conditions and social behavior indicators; The multi-source time-series data are time-aligned and missing value imputed to eliminate data inconsistencies and missing information. The processed data is normalized to obtain a standardized time series feature sequence. L represents the length of the historical time window. Let be the multidimensional eigenvector at time t.
[0031] By systematically cleaning, aligning, interpolating, and normalizing multi-source heterogeneous time-series data, it obtained standardized feature sequences with uniform quality and scale, providing a high-quality, low-noise input foundation for subsequent parameter prediction and generative modeling based on deep learning.
[0032] In this embodiment S2, the SEIR model parameters that change over time are predicted using a bidirectional long short-term memory network based on the standardized temporal feature sequence; wherein, the SEIR model parameters include the propagation rate. Latency rate and recovery rate ; Furthermore, the bidirectional Long Short-Term Memory network encodes the input sequence in the forward and backward directions using forward LSTM layers and backward LSTM layers, respectively, to obtain the forward hidden state sequence. With backward hidden state sequence ; The two are concatenated into a comprehensive timing representation H=[ ; The SEIR parameter sequence is output through a fully connected layer and the Softplus activation function:
[0033] Where W and b are the network weights and biases, and the Softplus activation function ensures that the parameters are non-negative.
[0034] Furthermore, the SEIR parameter sequence also includes temporal difference regularization, which involves adding a regularization term to the model training loss. accomplish:
[0035] Where T is the prediction time step and λ is the regularization coefficient. This step utilizes a bidirectional long short-term memory network to automatically extract dynamic patterns from standardized temporal features, predicts the core parameters of the SEIR model that change over time, and ensures the physical rationality and temporal stability of the parameters through the Softplus function and temporal smoothing regularization, thereby encoding complex external influencing factors into interpretable dynamic parameters.
[0036] In this embodiment S3, based on the SEIR model parameters, the SEIR dynamic equations are solved using an ordinary differential equation solver to generate a basic transmission trajectory that conforms to the infectious disease transmission mechanism. The SEIR dynamic equations are as follows:
[0037] Where S, E, I, and R represent the number of susceptible individuals, latent individuals, infected individuals, and recovered individuals, respectively; SI represents the contact opportunities of susceptible individuals and infected individuals; and N is the total population. To enhance the model's expressive power, additional terms, including isolation rate, are introduced here for scalability. Reinfection rate The extended SEIR kinetic equations are as follows:
[0038] The ordinary differential equation solver employs a fourth-order Runge-Kutta numerical integration algorithm to integrate the SEIR model parameters. Substituting the equations into the system of equations and solving them, a continuous basic propagation trajectory is generated. .
[0039] Furthermore, the update formula for the fourth-order Runge-Kutta numerical integration algorithm is as follows:
[0040] , , , This is an intermediate gradient estimate;
[0041]
[0042]
[0043]
[0044] in, Denotes the right-hand function of the SEIR dynamic equations; Based on the predicted time-varying parameters, the SEIR dynamic equations are rigorously solved using the fourth-order Runge-Kutta numerical integration method, generating a basic trajectory that conforms to the fundamental laws of infectious disease transmission. This step ensures that the main trend of the model output has clear epidemiological interpretability and provides deterministic priors for residual correction in subsequent steps.
[0045] In this embodiment S4, using the basic propagation trajectory as input, a U-Net-based diffusion generation model is used to model the residual distribution between real disease propagation data and the basic propagation trajectory, generating multi-scenario propagation prediction trajectories. The diffusion generation model structure includes a symmetric encoder and decoder. The encoder uses multi-layer one-dimensional convolution and downsampling operations to extract multi-scale temporal features. The decoder uses upsampling operations to gradually restore the temporal resolution and fuses the features of the corresponding layers of the encoder through skip connections; At each layer of the encoder and decoder, SEIR model parameters and their time derivative features are injected through a conditional embedding mechanism as physical mechanism constraints. In this conditional embedding mechanism, the SEIR parameters at each time step and its first derivative with respect to time , , Linear mapping and sinusoidal position encoding are performed separately to form physical feature vectors. Then, this vector is concatenated with the feature maps of the corresponding layers of the U-Net encoder and decoder in the channel dimension. After each convolutional block, feature modulation is performed by adaptive layer normalization. This makes the generation process subject to explicit constraints of dynamic laws at multiple scales, ensuring that the generation trajectory conforms to the epidemiological mechanism described by the SEIR model in terms of propagation trend, change rate and stage transition. Furthermore, the training process for the diffusion generative model includes: Forward diffusion process: data on the actual spread of the disease Gaussian noise is added step by step to obtain a noisy sequence. , … T is the number of diffusion steps; Inverse denoising process: Training the U-Net network To recover the original data from the noisy sequence, the loss function is:
[0046] in, This represents the noisy data at step t. This represents the added Gaussian noise, and c represents the conditional information, specifically including the basic propagation trajectory, the intensity of policy intervention, and meteorological characteristics. By introducing a diffusion generation model conditioned on mechanistic trajectories, the residual distribution between real data and mechanistic predictions is learned. The conditional embedding mechanism in the U-Net structure ensures that the generation process is constrained by multi-level physical features. This module enables the model to generate multiple possible future propagation trajectories from random noise through reverse denoising sampling, realizing endogenous quantification of uncertainty and multi-scenario extrapolation capabilities.
[0047] In this embodiment, a two-stage training strategy is specifically adopted: In the first stage, the bidirectional long short-term memory network and the ordinary differential equation solver are trained independently to minimize the observation error between the basic propagation trajectory and the real data. In the second stage, the model parameters from the first stage are fixed, and the diffusion generation model is trained to learn the residual distribution.
[0048] Total loss function used for:
[0049] in, To measure the error loss in observation, the root mean square error (MSE) is used to measure the deviation between the baseline propagation trajectory and the actual data. To smooth the loss of parameters and suppress abnormal fluctuations in the parameter sequence; To denoise and reconstruct the loss of the diffusion model; , , These are the weighting coefficients used to balance the various loss terms; After training, the process of generating multi-scenario propagation prediction trajectories is as follows: from random noise We begin by using the trained U-Net network. T-step denoising is performed step by step to obtain a denoised data sequence as a predicted trajectory. Multiple predicted trajectories can be generated by sampling different random noises multiple times.
[0050] In specific application scenarios, such as when a city faces the risk of spreading a new infectious disease, the model can generate a basic transmission trajectory that conforms to the SEIR mechanism based on multi-source information such as recent case data, population flow, and prevention and control policies, reflecting the theoretical evolution trend of the disease under the established mechanism. Subsequently, through the trained diffusion generation model, starting from random noise, multiple steps of denoising are performed under the constraint of the basic trajectory to generate multiple possible transmission paths that reflect the influence of different random factors. Each path represents a potential development scenario, which together constitute a set of probabilistic trajectories used to describe future possibilities. In actual simulations, the model can generate hundreds of transmission trajectories for the next 30 days based on the current intervention intensity. The peak number of infections in most of these trajectories will be concentrated in a certain range, while the remaining trajectories reflect low-probability but high-impact transmission risks. Decision-making departments can analyze the overall distribution of these trajectories to assess the uncertainty of the transmission process, identify key stages and scale risks, and thus provide a quantitative basis for adjusting the allocation of public health resources and the intensity of intervention measures in advance.
[0051] Furthermore, the model possesses conditionally controllable generation capabilities. By adjusting input condition parameters such as population mobility intensity, isolation ratio, and detection frequency, it can proactively guide the generation of scenario predictions with different policy implications. For example, when setting conditions of high mobility and low intervention intensity, the model will generate trajectories with faster spread and higher peak values; conversely, it will generate trajectories with controlled spread and lower peak values. This allows the scheme to not only generate random scenarios driven by historical data but also serve as a policy simulation tool for systematically comparing the potential impact of different intervention strategies on the spread trend, thereby providing reliable and flexible extrapolation support in complex real-world decision-making.
[0052] In this embodiment, S5, based on the multi-scenario propagation prediction trajectory, a probabilistic prediction result is output. The probabilistic prediction results here specifically include: the predicted mean number of infections at each time point within a certain prediction period and the confidence interval at a specified confidence level, forming a complete time series probability band; and the posterior probability distribution of key transmission indicators (time-varying effective reproduction number, peak time, peak size); these outputs are presented in the form of structured data and visual charts, providing quantitative basis for public health risk assessment and prevention and control decisions.
[0053] The infectious disease transmission prediction method based on SEIR parameter prediction and diffusion generation provided in this embodiment combines a physically interpretable SEIR dynamics model with a powerful deep generative model. While ensuring that the prediction results conform to the basic laws of infectious disease transmission, it enhances the adaptability to complex real-world factors and realizes multi-scenario probabilistic prediction of future trends. It provides a scientific and reliable quantitative decision-making tool for risk assessment and strategy formulation in the field of public health.
[0054] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.
[0055] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for predicting the spread of infectious diseases based on SEIR parameter prediction and diffusion generation, characterized in that, Includes the following steps: S1. Acquire historical multi-source time series data and preprocess it to obtain standardized time series feature sequences; S2. Using the standardized temporal feature sequence, a bidirectional long short-term memory network is used to predict the SEIR model parameters that change over time; S3. Based on the SEIR model parameters, solve the SEIR dynamic equations using an ordinary differential equation solver to generate a basic transmission trajectory that conforms to the infectious disease transmission mechanism. S4. Using the basic propagation trajectory as input, a U-Net-based diffusion generation model is used to model the residual distribution between real disease propagation data and the basic propagation trajectory, generating multi-scenario propagation prediction trajectories. S5. Based on the multi-scenario propagation prediction trajectory, output the probabilistic prediction result.
2. The infectious disease transmission prediction method based on SEIR parameter prediction and diffusion generation according to claim 1, characterized in that, S1 includes: Collect multi-source time-series data including daily new confirmed cases, cumulative number of cured cases, number of deaths, population mobility intensity, policy intervention intensity, meteorological conditions and social behavior indicators; The multi-source time-series data are time-aligned and missing value imputed to eliminate data inconsistencies and missing information. The processed data is normalized to obtain a standardized time series feature sequence. L represents the length of the historical time window. Let be the multidimensional eigenvector at time t.
3. The infectious disease transmission prediction method based on SEIR parameter prediction and diffusion generation according to claim 1, characterized in that, In S2, the SEIR model parameters include propagation rate. Latency rate and recovery rate ; Bidirectional Long Short-Term Memory (LSTM) networks encode the input sequence in the forward and backward directions using forward LSTM layers, respectively, to obtain the forward hidden state sequence. With backward hidden state sequence ; The two are concatenated into a comprehensive timing representation H=[ ; The SEIR parameter sequence is output through a fully connected layer and the Softplus activation function: Where W and b are the network weights and biases, and the Softplus activation function ensures that the parameters are non-negative.
4. The infectious disease transmission prediction method based on SEIR parameter prediction and diffusion generation according to claim 3, characterized in that, The SEIR parameter sequence also includes temporal difference regularization, which involves adding a regularization term to the model training loss. accomplish: Where T is the prediction time step and λ is the regularization coefficient.
5. The infectious disease transmission prediction method based on SEIR parameter prediction and diffusion generation according to claim 1, characterized in that, In S3, the SEIR dynamic equations are: Where S, E, I, and R represent the number of susceptible individuals, latent individuals, infected individuals, and recovered individuals, respectively; SI represents the contact opportunities of susceptible individuals and infected individuals; and N is the total population. The ordinary differential equation solver employs a fourth-order Runge-Kutta numerical integration algorithm to integrate the SEIR model parameters. Substituting the equations into the system of equations and solving them, a continuous basic propagation trajectory is generated. .
6. The infectious disease transmission prediction method based on SEIR parameter prediction and diffusion generation according to claim 5, characterized in that, The update formula for the fourth-order Runge-Kutta numerical integration algorithm is as follows: in, Let i be the current state vector, and h be the iteration step index. , , , This is the intermediate gradient estimate.
7. The infectious disease transmission prediction method based on SEIR parameter prediction and diffusion generation according to claim 1, characterized in that, In S4, the diffusion generation model structure includes a symmetric encoder and decoder; The encoder employs multi-layer one-dimensional convolution and downsampling operations to extract multi-scale temporal features; The decoder uses upsampling operations to gradually restore the temporal resolution and fuses the features of the corresponding layers of the encoder through skip connections; At each layer of the encoder and decoder, SEIR model parameters and their time derivative features are injected as physical mechanism constraints through a conditional embedding mechanism.
8. The infectious disease transmission prediction method based on SEIR parameter prediction and diffusion generation according to claim 1, characterized in that, In S4, the training process of the diffusion generation model includes: Forward diffusion process: data on the actual spread of the disease Gaussian noise is added step by step to obtain a noisy sequence. , … T is the number of diffusion steps; Inverse denoising process: Training the U-Net network To recover the original data from the noisy sequence, the loss function is: in, This represents the noisy data at step t. This represents the added Gaussian noise, and c represents the conditional information.
9. The infectious disease transmission prediction method based on SEIR parameter prediction and diffusion generation according to claim 1, characterized in that, The method employs a two-stage training strategy: In the first stage, the bidirectional long short-term memory network and the ordinary differential equation solver are trained independently to minimize the observation error between the basic propagation trajectory and the real data. In the second stage, the model parameters from the first stage are fixed, and the diffusion generation model is trained to learn the residual distribution.
10. The infectious disease transmission prediction method based on SEIR parameter prediction and diffusion generation according to claim 1, characterized in that, In S5, the probabilistic prediction results include: the predicted mean and confidence interval of the number of infections at each future time point, and the probability distribution of key transmission indicators.