A method for temporal modeling that fuses conditional diffusion generation and depth prediction

Abstract

The present application relates to a kind of fusion condition diffusion generation and time series modeling method of depth prediction, comprising the following steps: dynamic evolution and uncertainty characterization are carried out to multiple-source data, and multiple targeted depth prediction models are constructed;Condition diffusion generation framework is built, condition diffusion generation framework and time series depth prediction module are fused, and a depth time series prediction model based on condition diffusion generation is built;Global, local and continuous features output by time series depth prediction model are used as input information of diffusion model, and diffusion model generates prediction data by feature fusion guidance;Reliability evaluation is carried out in two ways, one is to ensure the diversity of data distribution, feature dimension and application scene;Second, noise disturbance experiment;The contribution of each feature to the prediction result is calculated using SHAP value, and whether the feature with high SHAP value conforms to the change mechanism of actual system is analyzed combined with domain expert knowledge;If the importance of certain feature does not conform to the expectation, the output process of model is traced back.

Description

Technical Field

[0001] This invention relates to the field of multi-source data processing models, and in particular to a time-series modeling method that integrates conditional diffusion generation and depth prediction. Background Technology

[0002] Traditional time series prediction models perform well on training sets, but their generalization ability declines significantly in cross-distribution applications (such as data from different regions and time periods). They lack a systematic robust assessment of noise disturbances such as random errors and missing information in data collection, making it difficult to attribute prediction results to causes. They cannot explain to domain experts which features played a key role, which is not conducive to model debugging and trust building. Consequently, they are unable to assess risks and provide early warning in high-risk decision-making scenarios. Summary of the Invention

[0003] The purpose of this invention is to provide a time series modeling method that integrates conditional diffusion generation and depth prediction, overcomes the shortcomings of existing technologies, eliminates noise disturbances in data acquisition, completes robustness assessment, uses SHAP values ​​to calculate the contribution of each feature to the prediction results, and combines domain expert knowledge to analyze whether features with high SHAP values ​​conform to the change mechanism of the actual system, thereby making the prediction results more reliable.

[0004] To achieve the above objectives, the technical solution adopted by this invention is: a temporal modeling method that integrates conditional diffusion generation and depth prediction, comprising the following steps:

[0005] Step 1: Constructing a time-series deep prediction module, which dynamically evolves and characterizes uncertainties in multi-source data, and builds multiple targeted deep prediction models;

[0006] Step 2: Integrate modeling and collaborative optimization to build a conditional diffusion generation framework. Integrate the conditional diffusion generation framework with the temporal deep prediction module to build a deep temporal prediction model based on conditional diffusion generation. Use the global, local and continuous features output by the temporal deep prediction model as input information for the diffusion model, and guide the diffusion model to generate prediction data through feature fusion.

[0007] Step 3: Model Reliability Assessment and Interpretation Analysis. Reliability assessment employs two methods: first, selecting datasets relevant to the target prediction task but not used in model training to ensure diversity in data distribution, feature dimensions, and application scenarios; second, conducting noise perturbation experiments by adding Gaussian noise and occlusion noise at different noise levels to evaluate robustness. The SHAP value is used to calculate the contribution of each feature to the prediction results, quantifying the marginal contribution of features to the model output through conditional intervention, primarily including global and local interpretations. Combining domain expert knowledge, the analysis examines whether features with high SHAP values ​​conform to the actual system's change mechanism. If the importance of certain features does not match expectations, the potential problems are identified by backtracking the model output process.

[0008] Preferably, in step one, the multi-source data is used to dynamically evolve and characterize uncertainty, including time-series dynamic feature analysis, analysis of endogenous evolution and external disturbance mechanisms, and construction of the mathematical framework of conditional diffusion model. By using stochastic differential equations and diffusion process theory, combined with conditional control variables, dynamic equations of system change are established to reveal the influence of various factors on the evolution of system state under different conditions.

[0009] Preferably, the deep prediction model in step one includes a Transformer network for learning long-term dependencies, a graph convolutional network for capturing spatiotemporal correlation patterns, and a neural differential equation characterizing the differential structure of data evolution. Specifically, it involves designing a Transformer encoder based on a multi-head self-attention mechanism to encode fused features and extract global information through multi-layer stacking; outputting global trend features; constructing a module based on a graph convolutional network (GCN), treating each time step and variable as a node, and designing an adjacency matrix to reflect the correlation between nodes; extracting local dynamic features by transmitting and aggregating information through graph convolutional layers; fusing the GCN output with the Transformer module results to form a hybrid prediction input; constructing a temporal model based on neural differential equations, using the hybrid feature input as the initial state, and using an ODE solver to continuously simulate the evolution of the data state; designing a continuous-time loss function to optimize the ODE model parameters; outputting a continuous state curve, and fusing it with the discrete prediction results to form a complete prediction.

[0010] The preferred approach includes: Time-series dynamic characteristic analysis: Based on statistical thinking and data-driven modeling theory, statistical analysis and data transformation are performed on time-series data to reveal its trend evolution, statistical characteristics, multi-scale fluctuations, mutation phenomena, and noise distribution. Endogenous evolution and external disturbance mechanism analysis: Various external disturbance variables are introduced. Utilizing domain expert knowledge and data-driven evolutionary characteristic analysis results, key external variables are selected as conditional control variables. An indicator system for external factors is constructed using data mining techniques, and these conditional variables are numerically introduced into the subsequent model to characterize and quantify different influencing factors. Granger causality tests are used to explore the causal relationship between endogenous evolution and external disturbances, identifying variables that significantly affect system changes. Simultaneously, variable selection techniques are introduced to screen representative and explanatory variables from high-dimensional data, ensuring the simplicity and robustness of the model structure. Statistical tests and cross-validation methods are used to verify the stability and sensitivity of the causal model, and the construction of conditional control variables is corrected and optimized.

[0011] Preferably, the conditional diffusion model decomposes the evolution of the system state into two parts: a drift term, describing the long-term trend of the data; and a diffusion term, describing the random fluctuations in the data. By introducing a conditional control variable 'c', external influencing factors are numerically embedded into the model, enabling the mathematical model to not only capture the endogenous dynamics of the data but also respond to changes in external conditions. The mathematical expression is as follows:

[0012]

[0013] in, This represents the system state at time t. For drift term, For diffusion term, This represents a standard Wiener procedure. These are the model parameters.

[0014] Preferably, in step two, the conditional diffusion generation framework is constructed, and the specific steps are as follows:

[0015] S1. Data Feature Enhancement and Robust Feature Extraction: A self-supervised model based on SimCLR contrastive learning is adopted. By constructing positive and negative sample pairs, the similarity between data is compared using the InfoNCE loss function to extract robust feature representations. A variational autoencoder is used to model the probability distribution of the denoised data. By minimizing the reconstruction loss (KL divergence), the reconstructed data retains as much key structural information as possible from the original data while eliminating random interference. Finally, the latent variables extracted by the variational autoencoder are used as robust features of the data and fused with the features obtained from the contrastive learning module to form the final feature representation.

[0016] S2. Dynamic evolution conditional diffusion generation modeling: Based on the stochastic differential equations of system evolution, and using the denoised diffusion probability model as the basic technology, a dynamic evolution modeling framework that can adapt to different external environmental interferences is constructed to achieve accurate characterization of the system state at different time scales.

[0017] S3, Conditional Diffusion Generation Training Optimization: Based on S1 and S2, this involves training and optimizing the conditional diffusion generation technique, specifically including conditional control strategies, multi-scale generation structure design, and training optimization strategies.

[0018] Preferably, in step two of the fusion process, the conditional features and the diffusion model are fused by encoding the features output by the deep prediction model into a latent space vector through a conditional encoder; in each prediction step of the diffusion model, the conditional features are fused with the intermediate states; the discretization of continuous-time features is matched with the diffusion model, and the features are discretized using a sampling method to match the time steps of the diffusion model; in the inversion process, the prediction results of the ODE model are used as a reference for intermediate steps to guide the diffusion process to converge to a more accurate prediction value; the model training method uses a joint loss function. To balance the loss of the diffusion model and the deep prediction model; to adopt end-to-end training and select appropriate training strategies based on model complexity and computing resources; to enhance data augmentation and robustness by introducing noise perturbation during the training phase of the deep prediction model to improve the model's generalization ability.

[0019] Preferably, the dynamic evolution modeling framework in S2 achieves data generation and trend prediction through a forward diffusion process and a reverse denoising process. Specifically, the mathematical equations of the diffusion model are discretized to obtain the forward diffusion process.

[0020]

[0021] Among them, noise scheduling parameters Control the noise injection at each step; and calculate cumulatively. The study describes the overall noise evolution of the data, analyzes the injection patterns of noise at different times during the forward diffusion process, determines noise attenuation strategies, ensures that the data is gradually transformed into Gaussian noise, and completes a quantitative description of the data in terms of randomness and external noise by combining the diffusion term σ determined by statistical characteristics.

[0022] Preferably, a conditional control variable c is introduced during the inverse denoising process. A conditional control module is designed to embed external factors into the input or intermediate representation of the neural network, and a conditional generation formula is constructed, the expression of which is:

[0023]

[0024] in, To predict the mean for the neural network, For the predicted covariance matrix, the parameter φ is optimized by minimizing the reconstruction error;

[0025] The conditional control module design includes the analysis of the impact of external conditions on data evolution and weight quantification, the integration of conditional variables into the forward diffusion and reverse denoising processes, correction of drift and diffusion terms, and the formation of a conditional mapping formula.

[0026]

[0027] in, The estimation is achieved through neural networks, reflecting the influence of external conditions on noise estimation.

[0028] Preferably, in S3, the conditional control strategy involves fusing the conditional variable c through a cross-attention layer and embedding control variables in the skip connections of the U-Net network. This leads to the construction of a dynamic conditionally coupled network module, where cosine annealing is used for dynamic weight scheduling. A multi-scale generative structure is designed, employing a hierarchical generative architecture that integrates global diffusion processes and local autoregressive constraints, integrating variable convolutions in high-frequency layers to enhance dynamic pattern capture. The training optimization method employs two-stage training and utilizes the GradNorm algorithm to balance generative quality and conditional consistency.

[0029] Unconditional diffusion loss function pre-training

[0030]

[0031] Fine-tuning the conditional diffusion loss function by adding a KL regularization term

[0032]

[0033] Through iterative training, the uncertainty of the data or trend predictions generated by conditional diffusion is quantified using the following formula.

[0034]

[0035] The above formula shows that the model generates the final predicted data from the initial noise through multiple inverse steps. Meanwhile, by sampling and making multiple predictions, a complete probability distribution of the data can be constructed, which can be used to quantify the uncertainty of prediction.

[0036] The technical effects of this invention are as follows:

[0037] To eliminate noise disturbances during data acquisition and complete robustness assessment, the contribution of each feature to the prediction results is calculated using the SHAP value. Combined with domain expert knowledge, the study analyzes whether features with high SHAP values ​​conform to the actual system's change mechanisms, thereby increasing the reliability of the prediction results.

[0038] Using the global, local, and continuous features output by the time-series deep prediction model as input information for the diffusion model, and guiding the diffusion model to generate prediction data through feature fusion, can make the input features of the diffusion model more comprehensive, thereby generating more accurate prediction data.

[0039] Intelligent preprocessing for multi-source data that can simultaneously handle noise, redundancy, and decouple internal and external factors.

[0040] Data feature enhancement and robust feature extraction can provide a high-quality data foundation for subsequent models.

[0041] Dynamic evolutionary conditional diffusion generative modeling can capture the inherent randomness of data and provide sufficient samples for the model to learn the data distribution.

[0042] By learning noise estimation and data reconstruction under different conditions, high-fidelity data samples are generated, enabling conditional modeling of global data distribution, namely the transformation and implementation of conditional diffusion generation network of mathematical equations for the dynamic evolution of multi-source data. Attached Figure Description

[0043] Figure 1 This is a technical roadmap for a time-series modeling method that integrates conditional diffusion generation and depth prediction.

[0044] Figure 2 This is a logic diagram of the implementation of a time series modeling method that integrates conditional diffusion generation and depth prediction. Detailed Implementation

[0045] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be described in detail below with reference to the accompanying drawings. The description in this part is only exemplary and explanatory, and should not be used to limit the scope of protection of the present invention in any way.

[0046] like Figure 1-2 As shown, this application presents a temporal modeling method that integrates conditional diffusion generation and depth prediction, from... Figure 1 As can be seen from the above, this application is mainly divided into three modules: 1. Dynamic evolution and uncertainty characterization of multi-source data, with the aim of providing an intelligent preprocessing scheme that can simultaneously handle noise, redundancy, and decoupling of internal and external factors.

[0047] 2. The construction of a conditional diffusion generation framework driven by multi-source data aims to build a collaborative modeling framework that integrates mechanism-driven and data-driven approaches and has both generation and discrimination capabilities. This framework overcomes the problems that existing machine learning models are mostly black-box structures, lack interpretable mathematical frameworks that can characterize the evolution mechanism of the system, and make it difficult to trace the prediction process.

[0048] 3. The purpose of integrating conditional diffusion generation and deep prediction in time series modeling is to overcome the shortcomings of existing traditional time series prediction models. Specifically, these shortcomings include: a significant decrease in generalization ability in cross-distribution applications (such as data from different regions and time periods); a lack of systematic robustness assessment of noise disturbances such as random errors and missing information in data collection; difficulty in attributing prediction results to causes; inability to explain to domain experts which features played a key role; and difficulties in model debugging and trust building. Consequently, it is difficult to assess risks and provide early warning in high-risk decision-making scenarios.

[0049] The multi-source data-driven conditional diffusion generation framework is based on the dynamic evolution and uncertainty characterization of multi-source data. It constructs a dynamically evolving conditional diffusion model and trains and optimizes it. The time series modeling that integrates conditional diffusion generation and deep prediction mainly includes three parts: 3.1, the construction of the time series deep prediction module, which is mainly based on the dynamic evolution and uncertainty characterization of multi-source data, outputting global trend features, local dynamic features, and quantifying long-term trends and dynamic uncertainties; 3.2, fusion modeling and collaborative optimization, which integrates the features output in 3.1 into the constructed and optimized conditional diffusion model for prediction; and 3.3, model reliability assessment and interpretability analysis, which improves the accuracy and efficiency of the model, while also enhancing its adaptability and generalization ability in different scenarios.

[0050] The following details the three modules of this application. Figure 2 It also shows the specific logical relationship between the three modules.

[0051] 1. Dynamic evolution and uncertainty characterization of multi-source data

[0052] 1.1 Time-Series Dynamic Feature Analysis: Any application scenario is a complex open system, influenced by the coupling of multiple variables. The data fluctuations of the target under examination exhibit diverse characteristics. For example, in the time dimension, such patterns are difficult to fully understand under scale-invariant rules, especially data fluctuations in non-stationary states. In the attribute dimension, high-dimensional variables, while containing rich process information, also contain a lot of redundant information, interfering with data modeling. Therefore, based on statistical thinking and data-driven modeling theory, statistical analysis and data transformation are performed on time-series data to reveal its trend evolution (time-series decomposition method), statistical characteristics, multi-scale fluctuations (time-frequency-modal co-analysis), mutation phenomena (mutation detection), and noise distribution (filtering techniques).

[0053] 1.2 Analysis of Endogenous Evolution and External Interference Mechanisms: Based on the dynamic evolution of multi-source data (1.1), various external interference variables of the target (such as terrain, climate, human intervention, etc.) are introduced. Using domain expert knowledge and data-driven evolutionary characteristic analysis results, key external variables are selected as conditional control variables. An indicator system of external factors is constructed through data mining techniques, and these conditional variables are introduced into the subsequent model in a numerical manner to characterize and quantify different influencing factors. Granger causality test is used to explore the causal relationship between the endogenous evolution process and external interference, and to identify variables that have a significant impact on system changes. At the same time, variable selection techniques (such as random forest and Bayesian network) are introduced to screen out representative and explanatory variables from high-dimensional data to ensure the simplicity and robustness of the model structure. Through statistical testing and cross-validation methods, the stability and sensitivity analysis of the causal model are performed, and the construction of conditional control variables is further corrected and optimized.

[0054] 1.3 Mathematical Framework Construction of the Conditional Diffusion Model: Based on the aforementioned steps, using stochastic differential equations and diffusion process theory, combined with conditional control variables, a dynamic equation for system changes is established to reveal the influence of various factors on the evolution of the system state under different circumstances. Specifically, this mathematical model decomposes the evolution of the system state into two parts: a drift term, describing the long-term trend of the data; and a diffusion term, describing the random fluctuations in the data. In particular, a conditional control variable c (determined in 1.2) is introduced to embed external influencing factors numerically into the model, enabling the mathematical model to not only capture the endogenous dynamics of the data but also respond to changes in external conditions. The mathematical expression is as follows:

[0055]

[0056] in, It represents the system state at time t (statistical, trend, multi-scale and other system characteristics). For drift term, For diffusion term, This represents a standard Wiener procedure. These are the model parameters.

[0057] 2. Construction of a conditional diffusion generation framework driven by multi-source data

[0058] 2.1 Data Feature Enhancement and Robust Feature Extraction: The acquisition of multidimensional time-series data often involves noise, outliers, and random interference. Directly using the original data to build a model often leads to unstable training results and poor generalization ability, thus affecting prediction accuracy. Simultaneously, the acquired data contains a large number of unlabeled samples (the information content of unlabeled samples is not negligible, but not all unlabeled samples contribute to modeling; therefore, it is unnecessary to invest significant expert knowledge to label all unlabeled samples). Therefore, this process employs a self-supervised model based on SimCLR contrastive learning. By constructing positive and negative sample pairs, the InfoNCE loss function is used to compare the similarity between data (labeled and unlabeled samples, regular and outliers, etc.), thereby extracting robust feature representations (samples with discriminative power). Simultaneously, a variational autoencoder is used to model the probability distribution of the denoised data. By minimizing the reconstruction loss (KL divergence), the reconstructed data retains as much key structural information as possible from the original data while eliminating random interference. Finally, the latent variables extracted by the variational autoencoder are used as robust features of the data and fused with the features obtained from the contrastive learning module to form the final feature representation. This step provides a high-quality data foundation for subsequent models.

[0059] 2.2 Dynamic Evolution Conditional Diffusion Generation Modeling: Based on the study of stochastic differential equations in system evolution (Content 1), and using the denoised diffusion probability model as the fundamental technology, a dynamic evolution modeling framework capable of adapting to different external environmental interferences is constructed to achieve accurate characterization of the system state at different time scales. Combined with the mathematical framework built in Content 1, data generation and trend prediction are achieved through forward diffusion and reverse denoising processes. Specifically, the mathematical equations of the diffusion model are discretized to obtain the forward diffusion process.

[0060]

[0061] Among them, noise scheduling parameters Control the noise injection at each step; and calculate cumulatively. This process describes the overall noise evolution of the data, analyzes the injection patterns of noise at different times during the forward diffusion process, determines noise attenuation strategies, ensures that the data is gradually transformed into Gaussian noise, and combines the diffusion term σ determined by statistical characteristics to complete a quantitative description of the data's randomness and external noise. This process can capture the inherent randomness of the data and provide sufficient samples for the model to learn the data distribution.

[0062] In the inverse denoising process, a conditional control variable c is introduced, and a conditional control module is designed to embed external factors into the input or intermediate representation of the neural network. A conditional generation formula is constructed, the expression of which is:

[0063]

[0064] in, To predict the mean for the neural network, is the predicted covariance matrix; the parameter is optimized by minimizing the reconstruction error (ensuring the model can accurately reconstruct the data distribution).

[0065] The conditional control module design includes the analysis of the impact of external conditions on data evolution and weight quantification (determined by Content 1). Conditional variables are incorporated into the forward diffusion and reverse denoising processes to correct drift and diffusion terms, forming a conditional mapping formula.

[0066]

[0067] in, The estimation is achieved through neural networks, reflecting the influence of external conditions on noise estimation.

[0068] This step generates high-fidelity data samples by learning noise estimation and data reconstruction under different conditions, realizing conditional modeling of the global data distribution, namely the transformation and implementation of the conditional diffusion generation network of the mathematical equation for the dynamic evolution of multi-source data (Content 1). In this process, different conditional diffusion generation networks will be constructed for different features (such as using attention mechanism networks for trend features and convolutional neural networks for multi-scale features).

[0069] 2.3 Conditional Diffusion Generation Training and Optimization: Based on the above research (2.1 constructing a multimodal dataset including system evolution information and conditional variables; 2.2 constructing a conditional diffusion generation model adapted to different environments using the correspondence between stochastic differential equations and diffusion processes), the training and optimization of conditional diffusion generation technology focuses on key implementation aspects including conditional control strategies, multi-scale generation structure design, and training optimization strategies. Regarding conditional control strategies, this involves fusing conditional variable c (global condition) through cross-attention layers and embedding control variables in the jump connections of the U-Net network. (Local control) is used to construct a dynamically conditionally coupled network module, where cosine annealing is employed for dynamic weight scheduling. In terms of multi-scale generative structure design, a hierarchical generative architecture is adopted, integrating global diffusion processes and local autoregressive constraints, with variable convolutions integrated in high-frequency layers to enhance dynamic pattern capture. Regarding training optimization methods, a two-stage training approach is used, employing the GradNorm algorithm to balance generative quality and conditional consistency.

[0070] Unconditional diffusion loss function pre-training

[0071]

[0072] Fine-tuning the conditional diffusion loss function by adding a KL regularization term

[0073]

[0074] Through iterative training, the uncertainty of the data or trend predictions generated by conditional diffusion is quantified using the following formula.

[0075]

[0076] The above formula shows that the model generates the final predicted data from the initial noise through multiple inverse steps. Meanwhile, by sampling and making multiple predictions, a complete probability distribution of the data can be constructed, which can be used to quantify prediction uncertainty (including the parameters of the prediction distribution output at each diffusion step and the parallel prediction of multiple quantile trajectories).

[0077] 3. A temporal modeling scheme that integrates conditional diffusion generation and depth prediction

[0078] 3.1 Construction of Time Series Deep Prediction Module: In view of the complex evolution characteristics of multi-source time series data (Content 1), several targeted deep prediction models are constructed, namely, Transformer network for learning long-term dependencies (global trend), graph convolutional network for capturing spatiotemporal correlation patterns (local dynamics), and neural differential equation for characterizing the differential structure of data evolution (quantifying long-term trends and dynamic uncertainties).

[0079] A Transformer encoder based on a multi-head self-attention mechanism is designed to encode fused features and extract global information through multi-layer stacking; outputting global trend features. A Graph Convolutional Network (GCN) module is constructed, treating each time step and variable as a node, and an adjacency matrix is ​​designed to reflect the relationships between nodes; information is transmitted and aggregated through graph convolutional layers to extract local dynamic features; the GCN output is fused with the Transformer module results to form a hybrid prediction input. A temporal model based on Neural Differential Equations (NODE) ​​is constructed, using the hybrid feature input as the initial state, and an ODE solver (Runge-Kutta method) is used to continuously simulate the evolution of the data state; a continuous-time loss function is designed to optimize the ODE model parameters; the output continuous-state curve is fused with the discrete prediction results to form a complete prediction.

[0080] 3.2 Fusion Modeling and Collaborative Optimization: This section integrates the conditional diffusion generation technique (Content 2) algorithm with the temporal deep prediction module (Content 3.1) to build a deep temporal prediction model based on conditional diffusion generation. The basic idea of ​​the fusion model is to use the global, local, and continuous features output by the temporal deep prediction model as input information for the diffusion model, and guide the diffusion model to generate prediction data through feature fusion (cross-attention method).

[0081] The key technical implementation plan for the fusion process is as follows: Fusion method of conditional features and diffusion model (the feature output of the deep prediction model is encoded into a latent space vector through a conditional encoder; in each prediction step of the diffusion model, the conditional features are fused with intermediate states (such as noise); Discretization of continuous-time features and matching with the diffusion model (the feature is discretized using a sampling method to match the time steps of the diffusion model; in the inversion process, the prediction results of the ODE model are used as a reference for intermediate steps to guide the diffusion process to converge to a more accurate prediction value); Model training method (a joint loss function is proposed). To balance the loss of the diffusion model and the deep prediction model; to adopt end-to-end training and select appropriate training strategies based on model complexity and computing resources; and to enhance data augmentation and robustness (the diffusion model naturally has data augmentation characteristics, and the focus is on introducing noise perturbation during the training stage of the deep prediction model to improve the model's generalization ability).

[0082] 3.3 Model Reliability Assessment and Interpretive Analysis: In evaluating the performance of intelligent prediction models, it is crucial to consider not only accuracy and efficiency but also adaptability and generalization capabilities across different scenarios. This stage employs two methods for reliability assessment: First, selecting datasets relevant to the target prediction task but not used in model training ensures diversity in data distribution, feature dimensions, and application scenarios (e.g., water resource data from different regions). Second, conducting noise perturbation experiments (diffusion models possess excellent denoising capabilities). By introducing Gaussian noise (simulating random errors in data acquisition) and occlusion noise (simulating incomplete information), different noise levels (noise variance) are set to evaluate robustness (baseline comparison, noise resistance analysis, etc.). Furthermore, the SHAP value is used to calculate the contribution of each feature to the prediction results. Conditional intervention quantifies the marginal contribution of features to the model output, primarily including global and local interpretations. Combining domain expert knowledge, it is analyzed whether features with high SHAP values ​​conform to the actual system's change mechanism. If the importance of certain features does not match expectations, potential problems (e.g., model overfitting, feature redundancy) are identified by tracing back the model output process.

[0083] The proposed solution can be applied to the intelligent management of various environments, such as air quality and river hydrology.

[0084] When applied to a specific water area, historical data on hydrology, meteorology, geology, and water conservancy projects of the Xijiang River system are obtained by combining previous industry-academia-research cooperation channels, public datasets, and field surveys. Theoretical, methodological, and model research results such as dynamic evolution and uncertainty characterization of multi-source time series data (content 1), data-driven conditional diffusion generation technology (content 2), and time series model integrating conditional diffusion generation and depth prediction (content 3) are used to establish a depth prediction system based on conditional diffusion. This system can achieve high-timeliness and high-precision prediction of water quality and quantity, provide real-time feedback on dynamic changes in water resources, provide risk assessment and early warning, and provide data support for intelligent hydrological management decisions.

[0085] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0086] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only for the purpose of helping to understand the method and core ideas of the present invention. The above descriptions are only preferred embodiments of the present invention. It should be noted that due to the limitations of textual expression, and the objective existence of infinite specific structures, those skilled in the art can make several improvements, modifications, or changes without departing from the principles of the present invention, and can also combine the above technical features in an appropriate manner. These improvements, modifications, changes, or combinations, or the direct application of the inventive concept and technical solution to other situations without modification, should all be considered within the scope of protection of the present invention.

Claims

1. A temporal modeling method integrating conditional diffusion generation and depth prediction, characterized in that, Includes the following steps: Step 1: Constructing a time-series deep prediction module, which dynamically evolves and characterizes uncertainties in multi-source data, and builds multiple targeted deep prediction models; Step 2: Integrate modeling and collaborative optimization to build a conditional diffusion generation framework. Integrate the conditional diffusion generation framework with the temporal deep prediction module to build a deep temporal prediction model based on conditional diffusion generation. Use the global, local and continuous features output by the temporal deep prediction model as input information for the diffusion model, and guide the diffusion model to generate prediction data through feature fusion. Step 3: Model reliability assessment and interpretability analysis. Two methods are used for reliability assessment: First, select datasets that are related to the target prediction task but were not used in model training to ensure the diversity of data distribution, feature dimensions and application scenarios; Second, conduct noise perturbation experiments by adding Gaussian noise and occlusion noise to set different levels of noise to evaluate its robustness. The contribution of each feature to the prediction result is calculated using the SHAP value. The marginal contribution of features to the model output is quantified through conditional intervention, mainly including global and local explanations. Combined with domain expert knowledge, it is analyzed whether features with high SHAP values ​​conform to the change mechanism of the actual system. If the importance of some features does not match the expectations, potential problems are located by tracing back the model output process.

2. The temporal modeling method for fusing conditional diffusion generation and depth prediction according to claim 1, characterized in that, In step one, the dynamic evolution and uncertainty of multi-source data are characterized, including time-series dynamic feature analysis, analysis of endogenous evolution and external disturbance mechanisms, and construction of the mathematical framework of conditional diffusion model. Using stochastic differential equations and diffusion process theory, combined with conditional control variables, dynamic equations of system changes are established to reveal the influence of various factors on the evolution of system state under different conditions.

3. The temporal modeling method for fusing conditional diffusion generation and depth prediction according to claim 2, characterized in that, The deep prediction model in step one includes a Transformer network for learning long-term dependencies, a graph convolutional network for capturing spatiotemporal correlation patterns, and a neural differential equation for characterizing the differential structure of data evolution. Specifically, it designs a Transformer encoder based on a multi-head self-attention mechanism to encode fused features and extract global information through multi-layer stacking. Output global trend characteristics; A module based on graph convolutional network (GCN) is constructed, treating each time step and variable as a node, and an adjacency matrix is ​​designed to reflect the relationship between nodes; information is transmitted and aggregated through graph convolutional layers to extract local dynamic features; The GCN output is fused with the Transformer module results to form a hybrid prediction input. A time series model based on neural differential equations is constructed. The hybrid feature input is used as the initial state, and the ODE solver is used to simulate the continuous evolution of the data state. A continuous-time loss function is designed to optimize the ODE model parameters. The continuous state curve is output and fused with the discrete prediction results to form a complete prediction.

4. The temporal modeling method for fusing conditional diffusion generation and depth prediction according to claim 3, characterized in that, Time-series dynamic characteristic analysis: Based on statistical thinking and data-driven modeling theory, statistical analysis and data transformation are performed on time-series data to reveal its trend evolution, statistical characteristics, multi-scale fluctuations, mutation phenomena, and noise distribution. Endogenous evolution and external disturbance mechanism analysis: Various external disturbance variables are introduced. Utilizing domain expert knowledge and data-driven evolutionary characteristic analysis results, key external variables are selected as conditional control variables. An indicator system for external factors is constructed using data mining techniques, and these conditional variables are numerically introduced into the subsequent model to characterize and quantify different influencing factors. Granger causality tests are used to explore the causal relationship between endogenous evolution and external disturbances, identifying variables that significantly affect system changes. Simultaneously, variable selection techniques are introduced to screen representative and explanatory variables from high-dimensional data, ensuring the simplicity and robustness of the model structure. Statistical tests and cross-validation methods are used to verify the stability and sensitivity of the causal model, and the construction of conditional control variables is corrected and optimized.

5. The temporal modeling method for fusing conditional diffusion generation and depth prediction according to claim 4, characterized in that, The conditional diffusion model decomposes the evolution of the system state into two parts: a drift term, describing the long-term trend of the data, and a diffusion term, describing the random fluctuations in the data. By introducing a conditional control variable 'c', external influencing factors are numerically embedded into the model, enabling the mathematical model to not only capture the endogenous dynamics of the data but also respond to changes in external conditions. The mathematical expression is as follows: in, This represents the system state at time t. For drift term, For diffusion term, This represents a standard Wiener procedure. These are the model parameters.

6. The temporal modeling method for fusing conditional diffusion generation and depth prediction according to claim 5, characterized in that, Step two involves building a conditional diffusion generation framework, the specific steps of which are as follows: S1. Data Feature Enhancement and Robust Feature Extraction: A self-supervised model based on SimCLR contrastive learning is adopted. By constructing positive and negative sample pairs, the similarity between data is compared using the InfoNCE loss function to extract robust feature representations. A variational autoencoder is used to model the probability distribution of the denoised data. By minimizing the reconstruction loss (KL divergence), the reconstructed data retains as much key structural information as possible from the original data while eliminating random interference. Finally, the latent variables extracted by the variational autoencoder are used as robust features of the data and fused with the features obtained from the contrastive learning module to form the final feature representation. S2. Dynamic evolution conditional diffusion generation modeling: Based on the stochastic differential equations of system evolution, and using the denoised diffusion probability model as the basic technology, a dynamic evolution modeling framework that can adapt to different external environmental interferences is constructed to achieve accurate characterization of the system state at different time scales. S3, Conditional Diffusion Generation Training Optimization: Based on S1 and S2, this involves training and optimizing the conditional diffusion generation technique, specifically including conditional control strategies, multi-scale generation structure design, and training optimization strategies.

7. The temporal modeling method for fusing conditional diffusion generation and depth prediction according to claim 6, characterized in that, In step two, during the fusion process, the conditional features and the diffusion model are fused. The conditional encoder encodes the features output by the deep prediction model into a latent space vector. In each prediction step of the diffusion model, the conditional features are fused with the intermediate states. The discretization of continuous-time features is matched with the diffusion model; discretization is achieved through sampling methods to match the time steps of the features with those of the diffusion model. During the inversion process, the prediction results of the ODE model are used as a reference for intermediate steps, guiding the diffusion process to converge to more accurate predictions. The model training method employs a joint loss function. To balance the loss of the diffusion model and the deep prediction model; to adopt end-to-end training and select appropriate training strategies based on model complexity and computing resources; to enhance data augmentation and robustness by introducing noise perturbation during the training phase of the deep prediction model to improve the model's generalization ability.

8. The temporal modeling method for fusing conditional diffusion generation and depth prediction according to claim 6, characterized in that, The dynamic evolution modeling framework in S2 achieves data generation and trend prediction through a forward diffusion process and a reverse denoising process. Specifically, the mathematical equations of the diffusion model are discretized to obtain the forward diffusion process. Among them, noise scheduling parameters Control the noise injection at each step; and calculate cumulatively. The study describes the overall noise evolution of the data, analyzes the injection patterns of noise at different times during the forward diffusion process, determines noise attenuation strategies, ensures that the data is gradually transformed into Gaussian noise, and completes a quantitative description of the data in terms of randomness and external noise by combining the diffusion term σ determined by statistical characteristics.

9. The time-series modeling method for fusing conditional diffusion generation and depth prediction according to claim 8, characterized in that, In the inverse denoising process, a conditional control variable c is introduced, and a conditional control module is designed to embed external factors into the input or intermediate representation of the neural network. A conditional generation formula is constructed, the expression of which is: in, To predict the mean for the neural network, For the predicted covariance matrix, the parameter φ is optimized by minimizing the reconstruction error; The conditional control module design includes the analysis of the impact of external conditions on data evolution and weight quantification, the integration of conditional variables into the forward diffusion and reverse denoising processes, correction of drift and diffusion terms, and the formation of a conditional mapping formula. in, The estimation is achieved through neural networks, reflecting the influence of external conditions on noise estimation.

10. The temporal modeling method for fusing conditional diffusion generation and depth prediction according to claim 9, characterized in that, In S3, the conditional control strategy involves fusing the conditional variable c through a cross-attention layer and embedding control variables in the jump connections of the U-Net network. This leads to the construction of a dynamic conditionally coupled network module, where cosine annealing is used for dynamic weight scheduling. A multi-scale generative structure is designed, employing a hierarchical generative architecture that integrates global diffusion processes and local autoregressive constraints, integrating variable convolutions in high-frequency layers to enhance dynamic pattern capture. The training optimization method employs two-stage training and utilizes the GradNorm algorithm to balance generative quality and conditional consistency. Unconditional diffusion loss function pre-training Fine-tuning the conditional diffusion loss function by adding a KL regularization term Through iterative training, the uncertainty of the data or trend predictions generated by conditional diffusion is quantified using the following formula. The above formula shows that the model generates the final predicted data from the initial noise through multiple inverse steps. Meanwhile, by sampling and making multiple predictions, a complete probability distribution of the data can be constructed, which can be used to quantify the uncertainty of prediction.

Images