A Multi-Scale Uncertainty Quantification Method and System for Time Series Prediction Models

By using a conditional diffusion probability framework and multi-timescale decomposition technology, a multi-scale uncertainty quantification method is generated, which solves the problems of existing time series prediction models being unable to distinguish between long-term and short-term risks, having low quantification accuracy, and poor real-time performance, and achieves efficient multi-scale risk decoupling and real-time response.

CN123069844APending Publication Date: 2026-11-10NAT UNIV OF DEFENSE TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611075355.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-20
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing time series forecasting models cannot distinguish between long-term trends and short-term volatility risks, resulting in low quantification accuracy and poor real-time performance.

Method used

By constructing a conditional diffusion probability framework and multi-timescale decomposition technology, multiple prediction samples are generated using a score function estimator and iterative sampling. Projective decomposition is then performed to calculate multi-scale uncertainty indices, and prediction results and risk assessments are output.

Benefits of technology

It enables fine-grained reliability assessment of time series prediction results in a single inference process, accurately fits complex non-Gaussian distributions without increasing the number of additional model parameters, distinguishes between long-term and short-term risks, and ensures millisecond-level real-time response capability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069844A_ABST
    Figure CN123069844A_ABST
Patent Text Reader

Abstract

This invention provides a multi-scale uncertainty quantification method and system for time-series prediction models. The method includes the following steps: acquiring time-series input data and extracting features based on a conditional generative coding model to obtain conditional variables; iteratively sampling the conditional variables using a pre-trained score function estimator and a preset sampling strategy to obtain multiple prediction samples; performing multi-time-scale projection decomposition on the multiple prediction samples to obtain sample components at different time scales; and calculating statistical features based on the multiple sample components at each different time scale to obtain a multi-scale uncertainty index. This invention achieves complex non-Gaussian distribution fitting and decoupling of long-term and short-term risks, improving the accuracy and real-time performance of reliability assessment when dealing with time-series data input.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis technology, specifically to a multi-scale uncertainty quantification method and system for time series prediction models. Background Technology

[0002] Time series forecasting models are widely used in critical decision-making scenarios such as power load dispatching, traffic flow control, financial risk control, and equipment health monitoring. However, most existing time series forecasting models are deterministic point estimates. Even if future evolution involves high ambiguity or abrupt changes, the model still only outputs a single predicted value and cannot provide information on reliability. In actual control and early warning tasks, this deviation between the predicted probability and the actual risk can lead to serious misjudgments. Therefore, quantifying the uncertainty of time series forecasting results has become a key requirement for improving system security and reliability.

[0003] Regarding the quantification of uncertainty in time series forecasting, existing technologies mainly fall into the following categories and their limitations: The first category consists of methods based on parametric distribution assumptions. In industry, this typically involves quantile regression or adding a Gaussian regression head to the model output layer to obtain the overall variance or quantile intervals through a single forward propagation. These methods are computationally efficient, but their core assumption is that the predicted distribution follows a predefined specific distribution form, such as a Gaussian or Laplace distribution. However, the noise distribution of real time-series data often exhibits complex characteristics such as non-Gaussian, multimodal, or heavy-tailed patterns. Fixed distribution assumptions are insufficient to accurately characterize the true shape of the conditional prediction distribution, resulting in overly coarse uncertainty estimates and inadequate quantization accuracy in dynamically changing time-series scenarios.

[0004] The second category comprises methods based on stochastic regularization, represented by Monte Carlo Dropout (MC Dropout). Its core mechanism involves maintaining the dropout operation during the inference phase, acquiring the statistical characteristics of the predicted distribution through multiple forward propagations. However, MC Dropout suffers from a long-standing unresolved drawback: it uses a fixed dropout rate for all input samples, failing to consider the difficulty or noise level of different samples. Subsequent improvements, such as Concrete Dropout, learn a global dropout rate during training, but once the parameters are determined, they cannot be dynamically adjusted for specific input samples during testing, still failing to achieve sample-level adaptive uncertainty estimation.

[0005] The third category comprises methods based on multi-model ensembles, represented by deep ensembles. This method estimates uncertainty by training multiple independent models and aggregating their prediction distributions, and is widely regarded as a benchmark scheme for uncertainty estimation. However, the significantly increased computational and memory overhead makes it difficult to implement in edge devices or real-time systems. For example, the same time-series model requires repeated computations and result ensemble processes, making it difficult to meet the real-time response requirements of the system in terms of computational efficiency.

[0006] The fourth category comprises methods based on Bayesian inference. Bayesian neural networks attempt to introduce a prior distribution into the parameter space and approximate the posterior through variational inference. However, their parameter dimension is much higher than the output dimension, resulting in high inference complexity and slow convergence, which is detrimental to the stable deployment of large-scale time-series models. Some improved schemes dynamically adjust parameters through iterative optimization during inference, which enhances adaptability but introduces a high-latency mechanism for online training, violating the millisecond-level response requirement of real-time prediction.

[0007] More importantly, all the aforementioned methods output single-scale overall uncertainty indicators, failing to distinguish between risks originating from different sources: long-term trend deviations and short-term local fluctuations. In time series forecasting, long-term trend uncertainty may stem from changes in the macroeconomic environment, such as seasonal migration or equipment aging; short-term fluctuation uncertainty is related to instantaneous disturbances, such as sudden load shocks or sensor noise. These two factors have drastically different implications for downstream control decisions: long-term trend uncertainty suggests the need to adjust operating strategies or plan maintenance cycles, while short-term fluctuation uncertainty indicates a decline in the reliability of the current forecast, requiring enhanced real-time monitoring or alarm triggering. Existing methods lack both a unified mathematical framework for decomposing the forecast distribution into global trend components and local detail components, and a sampling mechanism capable of simultaneously guiding the generation of both types of uncertainty while maintaining the efficiency of a single forward inference.

[0008] In summary, the existing technology has the following problems: it cannot distinguish between long-term trends and short-term volatility risks, has low quantification accuracy, and poor real-time performance. Summary of the Invention

[0009] The purpose of this invention is to overcome the problems of existing time series forecasting, such as its inability to distinguish between long-term trends and short-term volatility risks, low quantification accuracy, and poor real-time performance.

[0010] To this end, on the one hand, embodiments of the present invention provide a multi-scale uncertainty quantification method for time series prediction models, the method comprising the following steps: Acquire time-series input data, extract features from the time-series input data, and obtain condition variables; For conditional variables, iterative sampling is performed based on a pre-trained score function estimator and a preset sampling strategy to obtain multiple predicted samples; The predicted samples are decomposed by projection to obtain the sample components of each predicted sample at different time scales. Statistical characteristics are calculated based on sample components to obtain multi-scale uncertainty indicators; The system outputs prediction results and risk assessments based on a multi-scale uncertainty index. It learns the score function of the conditional prediction distribution using a conditional diffusion probability framework and generates multiple prediction samples through iterative sampling to approximate the true posterior distribution. The samples are then projectively decomposed, and prediction results and risk assessments are output based on the multi-scale uncertainty index. This achieves a unified approach to fitting complex non-Gaussian distributions and decoupling long-term and short-term risks without increasing the number of additional model parameters.

[0011] Furthermore, for the conditional variable, iterative sampling is performed based on a pre-trained score function estimator and a preset sampling strategy to obtain multiple predicted samples, including: Initialize the predicted samples; In each iteration, the condition variables and the current predicted sample are input into the score function estimator to obtain the log gradient approximation. Based on the logarithmic gradient approximation and the preset sampling strategy, the current predicted sample is updated until the iteration termination condition is met, resulting in multiple predicted samples.

[0012] Furthermore, the score function estimator is pre-trained, and the pre-training includes: A forward noise addition / denoising process is constructed. In the forward noise addition / denoising process, the input time series information is used as a conditional variable constraint to gradually add Gaussian noise of preset intensity to the historical target output samples. After denoising, the features of the input time series information are obtained through feature extraction. By adding / denoising to the samples, a generative model is used to process the samples to obtain diverse samples, and the temporal features of the samples are extracted from them. Based on the negative correlation between the sample time-series features and the preset noise intensity, the training objective of the score function estimator is constructed; The training objective is minimized to update the parameters of the conditional generative model, resulting in a trained score function estimator.

[0013] Furthermore, the preset sampling strategies include temperature annealing strategy and Langevin dynamics update rule; For the conditional variable, iterative sampling is performed based on a pre-trained score function estimator and a preset sampling strategy to obtain multiple predicted samples, including: Determine the temperature parameters corresponding to the current iteration step based on the temperature annealing strategy; Based on the logarithmic gradient approximation of the score function estimator output, the temperature parameter, and the Langevin dynamics update rule, the state update step size and random perturbation term are calculated. The state of the current predicted sample is updated using the state update step size and random perturbation term, so that the updated predicted sample distribution approximates the conditional prediction posterior distribution.

[0014] Furthermore, the temperature parameters corresponding to the current iteration step are determined based on the temperature annealing strategy, including: Obtain the initial temperature, the final temperature, and the total number of iterations; Calculate the temperature parameter corresponding to the current iteration step based on the proportion of the current iteration step to the total number of iteration steps.

[0015] Furthermore, the predicted samples are projectively decomposed to obtain the sample components of each predicted sample at different time scales, including: Construct the first time-scale projection operator and the second time-scale projection operator; The first time-scale projection operator is used to project each predicted sample and extract the first time-scale component that represents the long-term trend characteristics. The second time-scale projection operator is used to project each predicted sample and extract the second time-scale component that represents the short-term fluctuation characteristics. The first time scale component and the second time scale component are used as sample components at different time scales.

[0016] Furthermore, the first time-scale projection operator and the second time-scale projection operator respectively employ wavelet transform operator, empirical mode decomposition operator, variational mode decomposition operator, or moving average filter operator; When using wavelet transform operators, each predicted sample is projected using the first time-scale projection operator to extract the first time-scale components representing long-term trend characteristics, including: Multi-level wavelet decomposition is performed on the predicted samples to obtain approximation coefficients and detail coefficients; The approximation coefficients are reconstructed into first-time-scale components, and the detail coefficients are reconstructed into second-time-scale components.

[0017] Furthermore, statistical characteristics are calculated based on the sample components to obtain multi-scale uncertainty indices, including: For each target time scale, obtain the target sample components of multiple predicted samples at the target time scale; Calculate the mean and covariance matrix of multiple target sample components; The expected value of the prediction at the target time scale is determined based on the mean of multiple target sample components, and the confidence interval and variance index at the target time scale are determined based on the covariance matrix. The expected value, confidence interval, and variance index are used as uncertainty indicators corresponding to the target time scale.

[0018] Furthermore, multi-scale uncertainty quantification methods for time series forecasting models also include: Construct conditional scoring functions for sample components at different time scales; Dynamic weights are used to combine conditional score functions into multi-scale score functions; The state update of the predicted samples is completed during the iterative sampling process using a multi-scale scoring function.

[0019] On the other hand, embodiments of the present invention provide a multi-scale uncertainty quantification system for time series prediction models, including: The conditional coding unit is used to acquire time-series input data, extract features from the time-series input data, and obtain conditional variables. The iterative sampling unit is used to perform iterative sampling on conditional variables based on a pre-trained score function estimator and a preset sampling strategy to obtain multiple predicted samples. Multi-scale decomposition unit is used to perform projection decomposition on the predicted sample to obtain the sample components of each predicted sample at different time scales. Uncertainty quantification unit is used to calculate statistical characteristics based on sample components to obtain multi-scale uncertainty index; The output unit is used to output prediction results and risk assessments based on multi-scale uncertainty indicators.

[0020] The above technical solution has the following beneficial effects: This invention constructs a generative model conditioned on time-series inputs, utilizing a score function estimator and an iterative sampling strategy to generate multiple prediction samples. This model accurately fits the non-Gaussian posterior distribution of complex, non-stationary time-series data, overcoming the shortcomings of traditional methods that rely on fixed distribution assumptions or costly multi-model ensembles. Furthermore, by performing multi-time-scale projection decomposition on the prediction samples, the prediction results are separated into different components representing long-term trends and short-term fluctuations, and their statistical characteristics are calculated separately, thus mathematically decoupling the sources of uncertainty. This mechanism enables the system to simultaneously output refined risk indicators for different time dimensions within a single inference process, ensuring millisecond-level real-time response capabilities and providing reliable evidence for downstream control decisions to distinguish between macro-trend shifts and micro-instantaneous disturbances. This significantly improves the safety and interpretability of time-series systems in complex environments. Attached Figure Description

[0021] Figure 1 This is a flowchart of a multi-scale uncertainty quantification method for time series prediction models provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a multi-scale uncertainty quantification system for time series prediction models provided in an embodiment of the present invention; Figure 3 This is a flowchart of the conditional diffusion model training process provided in an embodiment of the present invention; Figure 4 This is a flowchart of multi-scale uncertainty quantification and reliability output provided in an embodiment of the present invention; Figure 5 This is a diagram showing the running results of the multi-scale uncertainty quantification system for time-series prediction models provided in this embodiment of the invention. Detailed Implementation

[0022] 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.

[0023] In embodiments of the present invention, such as Figure 1 This paper presents a multi-scale uncertainty quantification method for time series prediction models. This method, through the synergy of a conditional diffusion probability framework and multi-timescale decomposition techniques, enables fine-grained reliability assessment of time series prediction results within a single inference process. Specifically, the method includes the following steps: S101: Obtain time-series input data, extract features from the time-series input data, and obtain condition variables.

[0024] The time-series input data can be time-dependent sequential data such as power load, traffic flow, and equipment sensor readings. The conditional coding model captures contextual information within the historical window, mapping it to a compact representation of conditional variables. These conditional variables not only contain the statistical characteristics of historical data but also implicitly represent the current system state and evolutionary trends, serving as strong constraints for subsequent generation processes. For example, the conditional coding model can employ structures such as recurrent neural networks, Transformer encoders, or one-dimensional convolutional networks to convert variable-length historical time-series segments into fixed-dimensional vectors or tensors. In this way, the conditional variables provide explicit semantic anchors for subsequent diffusion sampling, ensuring that the generated predicted samples always remain consistent with the current input context, avoiding the pattern collapse or semantic drift problems that easily occur in unconditional generation models in time-series tasks.

[0025] S102: For conditional variables, iterative sampling is performed based on a pre-trained score function estimator and a preset sampling strategy to obtain multiple predicted samples; wherein, the score function estimator is trained based on a preset conditional diffusion process.

[0026] The score function estimator essentially learns the gradient field of the probability density function of the target output data given condition variables. During iterative sampling, a pre-defined sampling strategy guides the initial samples to move gradually along this gradient field, eventually converging to a high probability density region. It should be understood that although this embodiment mentions the score function estimator and iterative sampling, the specific dynamic update rules, temperature annealing strategies, and training details of the score function will be further elaborated in subsequent embodiments. At the architectural level of this embodiment, the key lies in establishing a "conditional" sampling mechanism, where each state update is modulated by condition variables, thereby ensuring that the generated multiple prediction samples not only reflect the diversity of uncertainty but also strictly conform to the physical constraints of the current temporal input. This mechanism enables the model to efficiently approximate the true non-Gaussian posterior distribution without relying on multi-model ensembles.

[0027] S103: Perform projection decomposition on the predicted samples to obtain the sample components of each predicted sample at different time scales; Specifically, this is the key step that distinguishes this invention from spatial multi-scale processing in the general image generation field. In the context of time series analysis, multi-timescale projection decomposition aims to decouple each predicted sample in the time dimension into components with different physical meanings. For example, a predicted sample can be separated into a first timescale component representing long-term evolution (such as seasonal trends or macroeconomic growth trends) and a second timescale component representing instantaneous disturbances or high-frequency noise (such as sudden load fluctuations or measurement noise). This decomposition can be achieved by constructing specific projection operators to map the original predicted signal into different frequency bands or modal subspaces. Through this action, the mixed uncertainties originally coupled in a single prediction curve are explicitly decomposed, laying the data foundation for subsequent independent quantification of different risk sources.

[0028] S104: Calculate statistical characteristics based on sample components to obtain a multi-scale uncertainty index. After completing the multi-timescale decomposition, for each specific timescale, statistical characteristics are calculated using the set of components corresponding to all predicted samples at that scale. For example, the mean of the sample components at that scale can be calculated as the expected value, the variance or covariance matrix can be calculated as a measure of dispersion, or confidence intervals can be constructed based on quantiles. Since these statistical characteristics are calculated independently based on the separated components, the resulting multi-scale uncertainty index can accurately distinguish which uncertainties originate from the ambiguity of long-term trends and which originate from the severity of short-term fluctuations. This fine-grained quantification result is crucial for downstream decision-making: long-term trend uncertainty may indicate the need to adjust macro-scheduling strategies or recalibrate the model baseline, while short-term fluctuation uncertainty may trigger real-time anomaly alarms or robust control mechanisms. Through the above process, this embodiment establishes a complete technical architecture from conditional sample generation to multi-scale risk decoupling without increasing the number of additional model parameters and inference latency, effectively solving the problem that existing time-series uncertainty quantification methods struggle to balance the efficiency of fitting complex distributions with multiple timescales.

[0029] S105: Output prediction results and risk assessment based on multi-scale uncertainty indicators.

[0030] The prediction results include the expected trajectory and confidence interval at each time scale, providing decision-makers with point estimates of future time series evolution and their reliable range. Risk assessment is based on a comprehensive evaluation of multi-scale uncertainty indicators, generating quantified risk scores or levels through weighted fusion or rule mapping. For example, in a power load forecasting scenario, when short-term fluctuation variance abnormally increases, the system can output an alarm signal indicating instantaneous overload risk; when long-term trend variance continues to increase and the upper bound of the confidence interval exceeds the safety boundary, a medium- to long-term warning of systemic supply-demand imbalance is output. By distinguishing the risk semantics at different time scales, this step enables downstream control systems to clearly distinguish between short-term disturbances requiring immediate response and long-term drifts requiring strategic adjustments, significantly improving the safety and operational resilience of the time series system in complex environments. Through the above process, this embodiment establishes a complete technical architecture from conditional sample generation to multi-scale risk decoupling without increasing the number of additional model parameters and inference latency, effectively solving the problem that existing time series uncertainty quantification methods struggle to balance complex distribution fitting efficiency and multi-time scale risk differentiation.

[0031] As one implementation method, for the condition variable, iterative sampling is performed based on a pre-trained score function estimator and a preset sampling strategy to obtain multiple predicted samples, including: The prediction samples are initialized; these initial samples usually follow a simple prior distribution, such as the standard normal distribution, which represents the maximum entropy state without any observation information.

[0032] In each iteration, the conditional variable and the current predicted sample are input into the score function estimator to obtain an approximate logarithmic gradient. During this process, the method of injecting the conditional variable is crucial for ensuring temporal consistency. For example, contextual features extracted by the conditional coding model can be fused into the network layers of the score function estimator through channel concatenation, additive modulation, or cross-attention mechanisms, ensuring that each denoising or sampling process is strictly controlled within the current temporal input context.

[0033] Based on the logarithmic gradient approximation and a pre-defined sampling strategy, the current predicted sample is updated until the iteration termination condition is met, resulting in multiple predicted samples. Through this multi-particle parallel iteration method, the final multiple predicted samples constitute a discretized representation of the conditional posterior distribution. As one implementation method, the score function estimator is pre-trained, and the pre-training process includes: A forward noise addition / denoising process is constructed. In the forward noise addition / denoising process, the input time series information is used as a conditional variable constraint to gradually add Gaussian noise of preset intensity to the historical target output samples. After denoising, the features of the input time series information are obtained through feature extraction. By adding / denoising to the samples and using a generative model to process them, diverse samples are obtained, and the temporal features of the samples are extracted. Specifically, the forward process is a fixed Markov chain, in which Gaussian noise is added to the original time-series data at each time step, gradually destroying the data structure until it becomes pure noise. Crucially, this adding / denoising process is always constrained by conditional variables, meaning that the model learns the conditional rule of "diversifying the temporal information within the samples using generative methods within a reasonable range, under the condition of the input sample temporal information, rather than learning sample features through a fixed preset noise distribution or temporal information distribution." Next, the model is used to predict the noise in the added / denoised samples to obtain the predicted noise.

[0034] Based on the negative correlation between sample temporal features and preset noise intensity, a training objective for the score function estimator is constructed. This training objective is then minimized to update the parameters of the conditional generative model, resulting in the trained score function estimator. This conditional generative model is typically implemented using a deep neural network, with input temporal information samples, the current noise addition / denoising time step, and conditional variables of the input temporal information. Furthermore, based on the negative correlation between sample temporal features and preset noise intensity, the training objective for constructing the score function estimator is obtained through feature extraction. Minimizing this training objective updates the parameters of the conditional generative model, resulting in the trained score function estimator.

[0035] The preset sampling strategies include temperature annealing strategy and Langevin dynamics update rule; For the conditional variable, iterative sampling is performed based on a pre-trained score function estimator and a preset sampling strategy to obtain multiple predicted samples, including: The temperature parameter corresponding to the current iteration step is determined based on the temperature annealing strategy; the temperature parameter plays a key role in controlling the randomness and exploration capability of sampling. In the early stage of sampling, a higher temperature allows the sample to explore a wide range in the state space, avoiding getting trapped in local extrema; as the iteration progresses, the temperature gradually decreases, the sampling process tends to be deterministic, and the sample accurately converges to the high probability density region. Based on the logarithmic gradient approximation output by the score function estimator, the temperature parameter, and the Langevin dynamics update rule, the state update step size and random perturbation term are calculated. Langevin dynamics provides a physical framework for simulating particle motion under the overdamped limit, and its update equation combines a deterministic gradient drift term and a random thermal fluctuation term. The state update step size and random perturbation term are used to update the state of the current predicted sample, so that the updated predicted sample distribution approximates the conditional prediction posterior distribution.

[0036] Based on the logarithmic gradient approximation of the score function estimator output, the temperature parameter, and the Langevin dynamics update rule, the state update step size and random perturbation term are calculated. The state of the current predicted sample is updated using the state update step size and random perturbation term, so that the updated predicted sample distribution approximates the conditional prediction posterior distribution.

[0037] As one implementation method, the temperature parameters corresponding to the current iteration step are determined based on the temperature annealing strategy, including: Obtain the initial temperature, the final temperature, and the total number of iterations. Calculate the temperature parameter corresponding to the current iteration step based on its proportion of the total iterations. For example, if the initial temperature is set to 1.0, the final temperature to 0.01, and the total number of iterations to 1000, an exponential decay strategy can be used to smoothly decrease the temperature as the number of steps increases. The choice of this decay rule directly affects the sampling quality: exponential decay typically maintains a longer period of high-temperature exploration in the early stages, suitable for complex multimodal distributions; while linear decay provides a more uniform cooling process. Furthermore, based on the logarithmic gradient approximation output by the score function estimator, the temperature parameter, and the Langevin dynamics update rule, the state update step size and random perturbation term are calculated. This includes: multiplying the logarithmic gradient approximation by the step size coefficient to obtain the state update step size; and determining the current noise intensity based on the mapping relationship between the temperature parameter and noise intensity. In the specific calculation, the state update step size determines the magnitude of sample movement along the gradient direction, while the variance of the random perturbation term is proportional to the current temperature, ensuring that the sampling process satisfies the detailed balance condition, thereby guaranteeing that the training results approximate the posterior distribution of the actual target. By reasonably setting the step size coefficient and the temperature decay curve, the optimal balance between sampling accuracy and computational efficiency can be achieved within a limited number of iterations, providing a high-quality sample foundation for subsequent multi-scale uncertainty quantification.

[0038] As one implementation method, the predicted samples are projectively decomposed to obtain the sample components of each predicted sample at different time scales, including: A first-time-scale projection operator and a second-time-scale projection operator are constructed. In this embodiment, the projection operator is not a general data dimensionality reduction tool, but a band separation or mode decomposition operator specifically designed for time-series signals. The first-time-scale projection operator is configured to extract a first-time-scale component representing long-term trend characteristics, which reflects the evolution pattern, seasonal cycle, or macroscopic drift of time-series data over a longer time span. The second-time-scale projection operator is configured to extract a second-time-scale component representing short-term fluctuation characteristics, which captures instantaneous disturbances, high-frequency noise, or local anomalies.

[0039] The first time-scale projection operator is used to project each predicted sample to extract the first time-scale component representing the long-term trend characteristics; the second time-scale projection operator is used to project each predicted sample to extract the second time-scale component representing the short-term fluctuation characteristics; the first time-scale component and the second time-scale component are used as sample components under different time scales.

[0040] As a preferred implementation, the first time-scale projection operator and the second time-scale projection operator respectively adopt wavelet transform operator, empirical mode decomposition operator, variational mode decomposition operator or moving average filter operator; Specifically, when using wavelet transform operators, each predicted sample is projected using the first time-scale projection operator to extract the first time-scale component representing the long-term trend characteristics. This includes performing multi-level wavelet decomposition on the predicted samples to obtain approximation coefficients and detail coefficients. During wavelet decomposition, the original time-series signal is filtered step by step, with each level generating a set of approximation coefficients reflecting the low-frequency profile and a set of detail coefficients reflecting high-frequency changes. The number of decomposition levels can be adaptively determined based on the sampling frequency and prediction step size of the time-series data. For example, for hourly sampled power load data, 3 to 5 levels of decomposition are typically chosen to effectively separate the daily cycle trend from the minute-level fluctuations. Subsequently, the approximation coefficients are reconstructed into the first time-scale component, and the detail coefficients are reconstructed into the second time-scale component. The reconstruction process is the inverse operation of the decomposition. By performing inverse wavelet transform on the approximation coefficients of specific levels individually, a smooth curve containing only long-term trend information can be restored. Similarly, by superimposing the detail coefficients of each level and performing an inverse transform, a residual sequence containing all short-term fluctuation information is obtained. This wavelet transform-based decomposition method has strict mathematical orthogonality and invertibility, which can ensure that the sum of the decomposed components is exactly equal to the original prediction sample, avoiding information loss or redundancy, and providing a lossless data foundation for subsequent independent quantification of uncertainties at each scale.

[0041] Wavelet transform is merely an exemplary implementation, and the present invention is not limited thereto. In other embodiments, the first time-scale projection operator and the second time-scale projection operator may also include at least one of empirical mode decomposition (EMD), variational mode decomposition (VMD), or moving average filtering operator. These operators are also applicable to the band separation task of time-series signals, but each has different applicable scenarios and technical advantages. For example, the empirical mode decomposition (EMD) operator is a completely data-driven adaptive decomposition method that does not require preset basis functions. It is particularly suitable for processing complex time-series signals that are nonlinear and non-stationary, and can automatically decompose the predicted sample into several intrinsic mode functions, where low-order modes correspond to short-term fluctuations and high-order modes correspond to long-term trends. The variational mode decomposition (VMD) operator extracts modal components by solving variational optimization problems. Compared with empirical mode decomposition, it has stronger noise resistance and mode separation accuracy, and can effectively alleviate mode aliasing, making it suitable for industrial sensor data with low signal-to-noise ratios. Moving average filtering operators, with their simple calculations and strong real-time performance, can serve as a lightweight alternative in latency-sensitive online monitoring scenarios. They allow for flexible control of the separation boundary between trends and fluctuations by adjusting the window size. This parallel arrangement of multiple operators not only enriches the technical implementation paths of this invention but also provides sufficient defensive depth for time series forecasting tasks with different industries and data characteristics. It ensures that regardless of changes in the downstream application environment, suitable multi-scale decomposition methods can be found to support the core function of uncertainty quantification.

[0042] As a preferred implementation, statistical characteristics are calculated based on sample components to obtain multi-scale uncertainty indices, including: For each target time scale, target sample components of multiple predicted samples are obtained at that target time scale. Specifically, after the multi-time scale projection decomposition described above, each predicted sample has been separated into a first time scale component corresponding to the long-term trend and a second time scale component corresponding to short-term fluctuations. In this step, the system aggregates the components of all predicted samples at the same time scale into a set, which serves as the data basis for statistical analysis at that scale. For example, if N predicted samples are generated, N first time scale component sequences will be obtained at the first time scale. These sequences collectively represent the set of possible evolutions of the future long-term trend under given historical conditions.

[0043] The method calculates the mean and covariance matrix of multiple target sample components. Specifically, the mean vector reflects the expected trajectory of the prediction result at that time scale, i.e., the most probable evolution path; while the covariance matrix not only contains variance information at each time step but also captures the correlation structure between different time steps. This is particularly crucial for time series forecasting because the uncertainty of time series data is often not independent and identically distributed; the deviation at a certain moment may propagate or decay along the time axis. By calculating the complete covariance matrix, this invention can preserve the temporal dynamic characteristics of uncertainty rather than simplifying it to scalar variance, thus providing a more accurate mathematical basis for subsequent confidence interval construction.

[0044] The expected value of the prediction at the target time scale is determined based on the mean of multiple target sample components, and the confidence interval and variance index at the target time scale are determined based on the covariance matrix. Specifically, the expected value of the prediction is directly given by the mean vector, serving as the point prediction result at that scale. The construction of the confidence interval depends on the distribution shape described by the covariance matrix. In a preferred implementation, assuming that the sample components approximately follow a multivariate Gaussian distribution, the standard deviation of each time step can be calculated using the diagonal elements of the covariance matrix, and then combined with a preset confidence level (e.g., 95%) to generate point-by-point confidence upper and lower bounds. In another more robust implementation, when the number of samples is sufficiently large, the confidence interval can be directly determined based on the empirical quantiles of the sample components. This method does not rely on specific distribution assumptions and is more adaptable to non-Gaussian complex distributions generated by diffusion models. The variance index can be the mean, maximum, or weighted integral of the diagonal elements, used to quantify the strength of the overall uncertainty at that scale. The expected value of the prediction, the confidence interval, and the variance index are used as uncertainty indicators corresponding to the target time scale.

[0045] The expected value, confidence interval, and variance index are used as uncertainty indicators corresponding to the target time scale. Through the above calculation process, the abstract sample set is transformed into quantitative indicators with clear physical meaning and engineering applicability, enabling the ambiguity of long-term trends and the severity of short-term fluctuations to be independently measured and monitored.

[0046] The multi-scale uncertainty quantification method for time-series prediction models in this embodiment also includes: Conditional scoring functions are constructed for sample components at different time scales; dynamic weights are used to combine the conditional scoring functions into multi-scale scoring functions; and the multi-scale scoring functions are used to update the state of the predicted samples during the iterative sampling process.

[0047] Specifically, this approach feeds back the results of uncertainty quantification to the sampling process itself, forming an adaptive optimization loop. For example, if high uncertainty in short-term fluctuations is found in the early stages of sampling, indicating that the model's fit to high-frequency noise is not yet stable, the weight of the second time-scale score function can be dynamically increased, forcing the sampled particles to focus more on the convergence of local details. Conversely, if uncertainty in long-term trends dominates, the weight of the first time-scale score function is increased to prioritize stabilizing the macroscopic evolution direction. The dynamic weight adjustment can be based on the real-time ratio of variance indices at each scale or a preset annealing strategy. This mechanism makes the sampling process no longer a blind, random exploration, but intelligently guided by multi-scale risk perception, thereby significantly improving the efficiency of generating high-quality samples while ensuring diversity.

[0048] The aforementioned method also includes: generating a risk score for the time-series system based on multi-scale uncertainty indicators and risk assessment rules; and outputting an alarm signal when the risk score exceeds a risk threshold or the confidence interval exceeds the safety boundary. Specifically, this approach directly maps the quantification results to decision-making criteria at the business level. The risk score can be calculated using a weighted summation.

[0049] Furthermore, in embodiments of the present invention, such as Figure 2 Furthermore, a multi-scale uncertainty quantification system for time series prediction models is provided, including: The conditional encoding unit 21 is used to acquire time-series input data, extract features from the time-series input data, and obtain conditional variables. The conditional encoding unit serves as the system's data access and semantic understanding interface, and integrates a pre-trained deep neural network model to map the original, high-dimensional time-series observation sequence into a compact conditional vector representation. This unit is not only responsible for data reading and preprocessing, but more importantly, it compresses contextual information such as dynamic evolution patterns, periodic patterns, and abnormal states within the historical window into the conditional variables through feature extraction. These conditional variables constitute strong constraint anchors for the subsequent generation process, ensuring that the system's output prediction results always remain consistent with the current physical scene, avoiding semantic drift problems that may result from unconditional generation. In practical deployments, the conditional encoding unit can be implemented as an independent feature extraction service or an embedded inference engine to adapt to different computing resource limitations.

[0050] The iterative sampling unit 22 is used to iteratively sample conditional variables based on a pre-trained score function estimator and a preset sampling strategy to obtain multiple predicted samples. This iterative sampling unit is the core generation engine of the system, its function being to approximate the true conditional posterior distribution using a conditional generative model. Internally, this unit maintains the model parameters of the score function estimator and the configuration information of the sampling strategy. Upon receiving the conditional variables, it drives a set of initial samples to evolve along the gradient field of the probability density in the state space through parallel or serial iterative computation mechanisms. Unlike traditional single-point prediction modules, the output of the iterative sampling unit is not a single deterministic value, but a set of samples that can characterize possible future evolution patterns. This sampling mechanism based on a diffusion probability framework enables the system to efficiently capture complex non-Gaussian and multimodal distributions in time-series data without relying on expensive multi-model ensembles, providing sufficient data support for subsequent uncertainty analysis.

[0051] The multi-scale decomposition unit 23 is used to perform projection decomposition on the predicted samples, obtaining sample components of each predicted sample at different time scales. This unit acts as a bridge between the original prediction space and the risk semantic space. It incorporates projection operators specifically designed for time-series signals (such as wavelet transform and empirical mode decomposition algorithms), enabling orthogonal or approximately orthogonal separation of each predicted sample output by the iterative sampling unit in the time dimension. Through this function, long-term trend information and short-term fluctuation information mixed in a single prediction curve are explicitly decoupled into independent component sequences. This processing method is strictly anchored to time attributes, fundamentally different from spatial multi-scale analysis in image processing, ensuring that the system can accurately identify the physical meaning corresponding to different frequency components when processing non-stationary time-series data, laying a structured foundation for fine-grained risk assessment.

[0052] Uncertainty quantification unit 24 is used to calculate statistical characteristics based on sample components to obtain multi-scale uncertainty indicators. This unit is the intermediate processing unit of the system, and its function is to transform the discrete set of sample components into statistical quantities with clear engineering meaning. This unit independently performs statistical analysis operations for each time scale, calculating indicators such as mean, variance, covariance matrix, or confidence interval, thereby quantifying the degree of ambiguity in long-term trends and the severity of short-term fluctuations, and then transmitting these multi-scale uncertainty indicators to the output unit.

[0053] Output unit 25 is used to output prediction results and risk assessments based on multi-scale uncertainty indicators. The output unit is the final interface of the system for downstream applications, and its function is to transform the multi-scale indicators generated by the uncertainty quantification unit into information that can be directly used for decision-making. Prediction results include the expected trajectory and confidence interval at each time scale, while risk assessment generates a quantified risk score or level through weighted fusion or rule mapping. For example, in a power load prediction scenario, when short-term fluctuation variance abnormally increases, the output unit can generate an alarm signal indicating instantaneous overload risk; when long-term trend variance continues to increase and the upper bound of the confidence interval exceeds the safety boundary, it outputs a medium- to long-term warning of systemic supply-demand imbalance. By distinguishing the risk semantics of different time scales, the output unit enables downstream control systems to clearly distinguish between short-term disturbances requiring immediate response and long-term drifts requiring strategic adjustments, significantly improving the safety and operational resilience of the time-series system in complex environments. The evaluation results are as follows: Figure 5 As shown.

[0054] In specific implementation, the aforementioned units can be pure software modules, dedicated hardware circuits (such as FPGAs, ASICs), or a combination of software and hardware. The connections between units can be direct data bus transmission, shared memory access, or network communication interfaces, as long as the data flow of the aforementioned functional logic can be achieved. Furthermore, although this embodiment divides the system into five independent functional units, in actual engineering, multiple units can be merged into a single processing module, or a unit can be further divided into multiple sub-modules, depending on performance optimization or integration requirements. These variations should not depart from the scope of protection of this invention.

[0055] Among them, a multi-scale uncertainty quantification system for time series prediction models adopts the above-mentioned multi-scale uncertainty quantification method for time series prediction models. Its principle and process are the same as those of the multi-scale uncertainty quantification method for time series prediction models, and will not be repeated here. Example

[0056] In this embodiment of the invention, a multi-scale uncertainty quantification method for time series prediction models is provided, which mainly includes the following steps: This method learns the score function of the conditional prediction distribution within the conditional diffusion probability framework, performs multi-particle iterative sampling using annealing Langevin dynamics, and then performs global-local multi-scale projection decomposition on the prediction particles to quantify long-term trend uncertainty and short-term fluctuation uncertainty respectively. This achieves fine-grained and highly expressive time-series prediction uncertainty estimation without adding additional parameters.

[0057] The method of this invention is implemented using computer equipment (including memory, processor, and computer program stored in memory and executable on the processor). The specific technical solution includes the following steps: Step S1: Timing Input Acquisition and Conditional Encoding The time-series input data to be processed is input into the conditional encoder in the deep neural network, and the contextual representation within the history window is extracted to obtain the conditional variables.

[0058] Step S2: Conditional generative model construction and score function learning, such as Figure 3 As shown: Based on the target output sample, a conditional diffusion noise addition / denoising process is constructed, a conditional generative model is trained, and a score function estimator is constructed through the relationship between the conditional generative model and the noise intensity, which serves as an approximation of the logarithmic gradient of the conditional prediction distribution with respect to the output variable.

[0059] Step S3: Multi-particle posterior sampling based on annealing Langevin dynamics For the same condition variable, multiple output particles are initialized. Based on the score function estimator and temperature annealing strategy, the state of each particle is iteratively updated so that the particle distribution approximates the conditional prediction posterior, resulting in multiple prediction particles corresponding to the same input condition.

[0060] Step S4: Global-Local Multiscale Decomposition Construct a global projection operator and a local projection operator, and extract a global component (long-term trend component) and a local component (short-term fluctuation component) for each predicted particle. Optionally, construct corresponding conditional score functions for the global component and the local component, and use dynamic weights to combine them into a multi-scale score function to guide subsequent sampling.

[0061] Step S5: Multi-scale uncertainty quantification and output like Figure 4 As shown: The global mean and global covariance are calculated based on the global components of multiple particles to obtain the global uncertainty index; the local mean and local covariance are calculated based on the local components to obtain the local uncertainty index; the total uncertainty index is calculated by combining the differences between particles; and finally, the prediction results, confidence intervals, risk scores or alarm signals are output.

[0062] This invention proposes a multi-scale uncertainty quantification method and system for time series prediction models, which significantly improves the performance of models in performing reliability assessments of time series systems compared to existing technologies (SWAG, SNGP, etc.). Currently, in the field of uncertainty quantification, uncertainty calibration performance mainly includes NLL and CWC. The index reflects (the lower the value, the stronger the calibration performance). The method in this application achieves significant advantages on the traditional Lorenz 96 dataset (the method in this application uses four different conditional generative models; in the table, Ours (DDPM) uses the DDPM model, Ours (DDIM) uses the DDIM model, Ours (CM) uses the consistency model, Ours (CFM) uses the flow matching model, and the time series model uses PatchTsT, corresponding to...). Figure 5 In the baseline, the table shows the results of five random seedings under different noise conditions.

[0063] Table 1. Data from the Lorenz 96 dataset under low-noise conditions.

[0064] Table 2. Data under noisy conditions in the Lorenz 96 dataset.

[0065] Table 3. Data from the Lorenz 96 dataset under high noise conditions.

[0066] As shown in Table 1, under low-noise conditions, the multi-scale uncertainty quantification method proposed in this invention (including four derived models: Ours (DDPM), Ours (DDIM), Ours (CFM), and Ours (CM)) significantly outperforms existing comparative methods (MC Dropout, SWAG, Laplace, Deep ensemble, and SNGP) in terms of uncertainty calibration metrics (CWC and NLL). Among them, the consistency model-based method (Ours (CM)) performs best, with its CWC reduced to 3.0734 and NLL reduced to 0.5031, achieving a further performance breakthrough compared to the best-performing traditional benchmark method (Deep ensemble). Simultaneously, in terms of deterministic prediction accuracy metrics (MSE, MAE, and RMSE), the method of this invention maintains a level comparable to or better than the best benchmark method, demonstrating that this invention possesses excellent uncertainty quantification capabilities and fine posterior distribution fitting accuracy in low-noise environments.

[0067] As shown in Table 2, under moderate noise conditions, the performance indicators of all methods degraded to some extent with the increase of environmental disturbances and data ambiguity. However, the method of this invention still exhibits strong robustness. The four models of this invention consistently and comprehensively outperformed all comparative methods in CWC and NLL calibration indicators. In particular, Ours (CM) achieved the lowest CWC (3.4934) and NLL (0.6367) scores across the board. Furthermore, in deterministic prediction error indicators such as MSE, MAE, and RMSE, the methods of this invention (e.g., Ours (CM) MSE of 1.2713 and Ours (DDPM) MSE of 1.2757) even surpassed the traditionally well-performing Deep ensemble (1.2928) and MC Dropout (1.2935). This indicates that when dealing with moderate noise, this invention not only provides more accurate risk assessment but also effectively improves the absolute accuracy of time series prediction itself.

[0068] As shown in Table 3, under extremely harsh conditions with high noise (strong interference), the advantages of the method of this invention compared with traditional methods are further expanded. Regarding uncertainty calibration, Ours (CM) achieves a CWC index of 4.1699, significantly lower than all other comparative methods (all of which are greater than 4.7). In terms of NLL, the four derived models of this invention are stable at around 0.90, while all comparative methods degenerate to above 0.95. This fully demonstrates that this invention, through the conditional diffusion probability framework and annealing sampling mechanism, can achieve a more accurate posterior approximation of complex non-Gaussian, multimodal distributions under strong noise. Furthermore, in terms of prediction error, the MSE of Ours (DDPM / DDIM) and Ours (CM) (approximately 2.170) is significantly lower than that of the comparative methods (e.g., Deep ensemble is 2.2556, SNGP is 2.2981).

[0069] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-scale uncertainty quantification method for time series forecasting models, characterized in that, The method includes the following steps: Acquire time-series input data, and perform feature extraction on the time-series input data to obtain condition variables; For the conditional variable, iterative sampling is performed based on a pre-trained score function estimator and a preset sampling strategy to obtain multiple predicted samples; The predicted samples are decomposed by projection to obtain the sample components of each predicted sample at different time scales. Statistical characteristics are calculated based on the sample components to obtain a multi-scale uncertainty index; The prediction results and risk assessment are output based on the multi-scale uncertainty index.

2. The multi-scale uncertainty quantification method for time series prediction models according to claim 1, characterized in that, For the condition variable, iterative sampling is performed based on a pre-trained score function estimator and a preset sampling strategy to obtain multiple predicted samples, including: The predicted samples are initialized; In each iteration, the condition variable and the current predicted sample are input into the score function estimator to obtain the log gradient approximation. Based on the logarithmic gradient approximation and the preset sampling strategy, the current predicted sample is updated until the iteration termination condition is met, resulting in multiple predicted samples.

3. The multi-scale uncertainty quantification method for time series prediction models according to claim 2, characterized in that, The score function estimator is pre-trained, and the pre-training process includes: A forward noise addition / denoising process is constructed. In the forward noise addition / denoising process, the input time series information is used as a conditional variable constraint to gradually add Gaussian noise of preset intensity to the historical target output samples. After denoising, the features of the input time series information are obtained through feature extraction. By adding / denoising to the samples, a generative model is used to process the samples to obtain diverse samples, and the temporal features of the samples are extracted from them. Based on the negative correlation between the sample time-series features and the preset noise intensity, the training objective of the score function estimator is constructed; The training objective is minimized to update the parameters of the conditional generative model, resulting in a trained score function estimator.

4. The multi-scale uncertainty quantification method for time series prediction models according to claim 1, characterized in that, The preset sampling strategy includes a temperature annealing strategy and a Langevin dynamics update rule; For the condition variable, iterative sampling is performed based on a pre-trained score function estimator and a preset sampling strategy to obtain multiple predicted samples, including: The temperature parameters corresponding to the current iteration step are determined based on the temperature annealing strategy described above. Based on the logarithmic gradient approximation output by the scoring function estimator, the temperature parameter, and the Langevin dynamics update rule, the state update step size and random perturbation term are calculated. The state of the current predicted sample is updated using the state update step size and the random perturbation term, so that the updated predicted sample distribution approximates the conditional prediction posterior distribution.

5. A multi-scale uncertainty quantification method for time series prediction models according to claim 4, characterized in that, The temperature parameters corresponding to the current iteration step are determined based on the temperature annealing strategy, including: Obtain the initial temperature, the final temperature, and the total number of iterations; The temperature parameter corresponding to the current iteration step is calculated based on the proportion of the current iteration step in the total number of iteration steps.

6. The multi-scale uncertainty quantification method for time series prediction models according to claim 1, characterized in that, The step of projecting and decomposing the predicted samples to obtain sample components of each predicted sample at different time scales includes: Construct the first time-scale projection operator and the second time-scale projection operator; The first time-scale projection operator is used to project each of the predicted samples to extract the first time-scale component that represents the long-term trend characteristics. The second time-scale projection operator is used to project each of the predicted samples to extract the second time-scale component that represents the short-term fluctuation characteristics. The first time scale component and the second time scale component are used as sample components at different time scales.

7. The multi-scale uncertainty quantification method for time series prediction models according to claim 6, characterized in that, The first time-scale projection operator and the second time-scale projection operator respectively employ wavelet transform operator, empirical mode decomposition operator, variational mode decomposition operator, or moving average filter operator; When the wavelet transform operator is used, the step of projecting each predicted sample using the first time-scale projection operator to extract the first time-scale component representing the long-term trend characteristics includes: The predicted samples are subjected to multi-level wavelet decomposition to obtain approximation coefficients and detail coefficients; The approximation coefficients are reconstructed into the first time scale component, and the detail coefficients are reconstructed into the second time scale component.

8. The multi-scale uncertainty quantification method for time series prediction models according to claim 1, characterized in that, The step of calculating statistical characteristics based on the sample components to obtain a multi-scale uncertainty index includes: For each target time scale, obtain the target sample components of multiple predicted samples at the target time scale; Calculate the mean and covariance matrix of multiple target sample components; The expected value of the prediction at the target time scale is determined based on the mean of the multiple target sample components, and the confidence interval and variance index at the target time scale are determined based on the covariance matrix. The predicted expected value, the confidence interval, and the variance index are used as uncertainty indicators corresponding to the target time scale.

9. A multi-scale uncertainty quantification method for time series prediction models according to claim 8, characterized in that, The method further includes: Construct conditional scoring functions for sample components at different time scales; The conditional scoring function is combined into a multi-scale scoring function using dynamic weights; The multi-scale scoring function is used to update the state of the predicted samples during the iterative sampling process.

10. A multi-scale uncertainty quantification system for time series prediction models, characterized in that, include: A conditional coding unit is used to acquire time-series input data, extract features from the time-series input data, and obtain conditional variables. An iterative sampling unit is used to perform iterative sampling on the condition variable based on a pre-trained score function estimator and a preset sampling strategy to obtain multiple predicted samples. A multi-scale decomposition unit is used to perform projection decomposition on the predicted sample to obtain the sample components of each predicted sample at different time scales. An uncertainty quantification unit is used to calculate statistical characteristics based on the sample components to obtain a multi-scale uncertainty index. The output unit is used to output prediction results and risk assessments based on the multi-scale uncertainty index.