A diffusion model time series prediction method based on distillation acceleration

Abstract

The application discloses a diffusion model time series prediction method based on distillation acceleration, comprising the following steps: obtaining a historical input sequence of a target object before a current time to construct historical condition information; inputting the historical condition information into a pre-trained conditional prediction model to obtain a conditional prediction mean value of a future time sequence; adding random disturbance to the conditional prediction mean value to obtain multiple different disturbance inputs; inputting the historical condition information, the conditional prediction mean value and the multiple different disturbance inputs into a pre-trained single-step student generator to obtain corresponding multiple future prediction trajectories through network forward calculation; performing statistical processing on the multiple future prediction trajectories to obtain a point prediction result and a prediction interval of the future time sequence; and outputting the multiple future prediction trajectories, the point prediction result and the prediction interval of the future time sequence to realize time series prediction; and on the basis of ensuring prediction accuracy and probability output capability, realizing single-step diffusion type multi-element time series prediction.

Description

Technical Field

[0001] This invention belongs to the field of time series data prediction technology, specifically relating to a time series prediction method based on a diffusion model accelerated by distillation. Background Technology

[0002] Time series forecasting is a core technology in many fields, including industrial process control, energy dispatch management, resource optimization, and disaster early warning. With the increasing complexity of modern industrial systems and the enhanced capabilities of sensor data acquisition, accurately predicting future trends from high-dimensional, dynamic, and nonlinear multivariate historical data has become a hot topic of common concern in academia and industry. In recent years, with the rapid development of deep learning technology, prediction methods based on generative models, especially denoising diffusion probabilistic models (DDPM), have shown broad application prospects in the field of time series modeling due to their powerful distribution fitting capabilities and generative diversity.

[0003] In existing technologies, several time series forecasting schemes based on diffusion models have been proposed. For example, a Chinese patent document published by Beijing University of Science and Technology, entitled "A Method and System for Time Series Data Forecasting Based on a Conditional Diffusion Model" (Publication No.: CN117076931A), proposes a conditional diffusion model for time series data forecasting. This scheme maps dynamic time series data into high-dimensional features through a feature encoder and builds a conditional diffusion model with a conditional input Transformer as its backbone. The conditional features obtained by fusing historical high-dimensional dynamic time series data with static data are input into the diffusion model for training. However, the design of this scheme is highly dependent on the input of static data, and its conditional feature fusion mechanism is more suitable for yield forecasting scenarios containing static attributes. For tasks that only use historical multivariate time series as input and do not contain static features, this scheme lacks a targeted non-zero mean conditional diffusion implementation mechanism. In addition, its inference stage still adopts a conventional diffusion iterative denoising process and does not carry out lightweight or distillation compression design for the sampling process, resulting in significant computational overhead and latency issues in multi-step prediction inference, making it difficult to meet the requirements of real-time prediction scenarios.

[0004] A Chinese patent document published by Harbin Engineering University (Publication No.: CN120688557A) proposes a model called DiffMissing for multivariate time series prediction of missing variables based on a denoising diffusion model. This approach focuses on fusing information from both inter-variable and temporal relationships to improve the reconstruction and prediction capabilities of missing variables in multivariate time series. While this approach effectively utilizes the generative characteristics of the diffusion model, its core application is limited to the recovery of missing data. It does not employ specialized network structures and temporal modeling designs for multivariate, multi-step prediction with complete future windows, exhibiting significant deficiencies, particularly in error accumulation suppression and optimization of low-step sampling efficiency in long prediction time domains.

[0005] A Chinese patent document published by Zhejiang University, titled "A Power System Load Forecasting Method Based on Condition-Guided Diffusion Process" (Publication No.: CN119338283A), proposes a power system load forecasting scheme. This scheme designs a multi-stage progressively guided denoising process, using a similar-period load curve mining algorithm, multi-resolution trend features, and a hybrid feature conditional generation network to guide the denoising process, thereby improving the accuracy and controllability of load curve generation. However, this scheme relies heavily on rich prior knowledge and feature construction within the power load scenario, exhibiting significant scenario specificity and business dependence, resulting in poor versatility. Furthermore, this scheme also fails to propose effective distillation or fewer-step sampling acceleration solutions to address the issues of high sampling steps and long online inference time in the diffusion model.

[0006] A method and system for short-term trend prediction of multidimensional time-series data published by Hefei University of Technology (Publication No.: CN116258262A) proposes a prediction scheme based on trend and seasonal feature decomposition. It employs an interpretable time-series feature encoder for decomposition and combines a multi-head self-attention mechanism and a feature fusion module for prediction. However, this scheme is essentially a deterministic or discriminative prediction framework, not based on modeling the distribution of future sequences through diffusion processes. Therefore, it struggles to output future trajectories with diverse characteristics and cannot provide an effective prediction uncertainty range. Furthermore, this scheme relies on an auxiliary information encoder to fuse feature factors, which does not conform to the prediction setting based solely on historical multidimensional time-series input.

[0007] In summary, existing time series forecasting schemes generally suffer from the following limitations: 1. Some schemes rely on static data or prior features of specific scenarios, limiting their versatility; 2. Diffusion-based forecasting methods generally face problems such as numerous iterations, high computational overhead, and poor real-time performance during the inference phase, lacking efficiency optimization methods for model distillation or fewer-step sampling; 3. Existing technologies mostly focus on missing value reconstruction or load forecasting in specific domains, lacking a long-line, multi-step forecasting system that balances prediction accuracy and inference efficiency for general multivariate time series forecasts; 4. Deterministic forecasting schemes struggle to quantify model uncertainty, failing to provide reliable risk range assessments for industrial safety monitoring. Therefore, there is an urgent need for a multivariate time series forecasting method that can overcome the above deficiencies, possesses high versatility, low-latency inference capabilities, and uncertainty quantification capabilities. Summary of the Invention

[0008] To address the aforementioned problems in the existing technology, this invention provides a time series prediction method based on a distillation-accelerated diffusion model. The technical problem to be solved by this invention is achieved through the following technical solution: In a first aspect, the present invention provides a time series prediction method based on a distillation-accelerated diffusion model, comprising: Obtain the historical input sequence of the target object before the current moment to construct historical condition information; Historical conditional information is input into a pre-trained conditional prediction model to obtain the mean conditional prediction of future time series. By centering on the conditional prediction mean, random perturbations are added to obtain multiple different perturbation inputs; The historical condition information, the conditional prediction mean, and multiple different perturbations are input into a pre-trained single-step student generator, and multiple corresponding future predicted trajectories are obtained through forward computation of the network. Statistical processing is performed on multiple future predicted trajectories to obtain point prediction results and prediction intervals for future time series; multiple future predicted trajectories, point prediction results and prediction intervals for future time series are output to achieve time series prediction. Specifically, the total loss of the single-step student generator is constructed using a pre-trained teacher diffusion model and an auxiliary denoising network to constrain the training of the single-step student generator.

[0009] In one embodiment of the present invention, obtaining the historical input sequence of the target object before the current moment to construct historical condition information includes: Obtain the historical multivariate time series of the target object up to the current moment, and extract a length of [length missing]. L The historical window is used to obtain the historical input sequence; Historical condition information is constructed based on the historical input sequence and its timestamp.

[0010] In one embodiment of the present invention, the historical condition information includes: One or more of the following: historical input sequence, historical time stamp, future time stamp, and decoding input.

[0011] In one embodiment of the present invention, the training loss of the conditional prediction model is as follows: ; in, The loss function for the conditional prediction model. For mathematical expectation operators, for The square of the norm, For the true sequence, For conditional forecasting of the mean, These are the weighting coefficients corresponding to the regularization loss. This is the regularized loss in the conditional prediction model.

[0012] In one embodiment of the present invention, the first r The expression for the predicted future trajectory is as follows: ; in, For the first r A future predicted trajectory, This is the residual jump coefficient. For the first r One disturbance input, For conditional forecasting of the mean, For the first r One residual.

[0013] In one embodiment of the present invention, the first r The expression for each residual is as follows: ; in, Let the objective function be the single-step student generator. For historical conditions information, For the first r One disturbance input, The diffusion time step used by the student generator.

[0014] In one embodiment of the present invention, the total loss function of the single-step student generator is as follows: ; in, The total loss function for a single-step student generator. For non-zero mean SID generator loss, Weights are used to predict the mean loss of the window. This is for predicting the mean loss of the window.

[0015] In a second aspect, the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the steps of the distillation-accelerated diffusion model time series prediction method provided in the first aspect of the present invention.

[0016] Thirdly, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of the time series prediction method based on distillation-accelerated diffusion model provided in the first aspect of the present invention.

[0017] The beneficial effects of this invention are: The solution provided by this invention can generate multiple future predicted trajectories and output prediction intervals and risk indicators based on historical input sequences, exhibiting stronger uncertainty expression capabilities. Compared to the original multi-step diffusion time series prediction method, this invention compresses multi-step teacher sampling into single-step student generation through non-zero mean SID distillation, significantly reducing inference latency and the number of repeated network forward computations. This invention trains the single-step student generator using a teacher diffusion model, an auxiliary denoising network, and a non-zero mean SID objective, giving the single-step generation process explicit teacher distribution constraints. Furthermore, compared to general diffusion distillation methods, this invention introduces historical sequences, time stamps, conditional prediction means, and prediction window constraints specifically for time series prediction tasks, better preserving the single-step student generator's ability to model the conditional future path distribution. Attached Figure Description

[0018] Figure 1 A schematic diagram illustrating the steps of a time series prediction method based on a distillation-accelerated diffusion model provided in an embodiment of the present invention; Figure 2 The prediction graph of a diffusion model time series prediction method based on distillation acceleration provided in an embodiment of the present invention on the ETTh1 dataset; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0020] Multivariate time series multi-step forecasting is widely used in scenarios such as power load forecasting, traffic flow forecasting, financial risk forecasting, equipment status forecasting, weather forecasting, and industrial production scheduling. Multivariate time series in these scenarios typically exhibit characteristics such as non-stationarity, strong noise, multiple periods, multivariate coupling, time-lag correlations between variables, and high uncertainty in future values. Traditional point forecasting models usually output a single predicted value, making it difficult to represent the multi-peak distribution and prediction range of future trends. While traditional probabilistic forecasting models can output distribution information, they are prone to insufficient distribution characterization, weak extreme value prediction, and error accumulation under conditions of long prediction time domains, complex multivariate coupling, and sudden fluctuations.

[0021] Diffusion models are a type of progressive denoising generation model that can gradually recover target samples from random noise, making them suitable for modeling complex probability distributions. When using diffusion models for multivariate time series multi-step prediction, multiple future time window samples can be generated, thus simultaneously obtaining point prediction results and uncertainty intervals. However, existing diffusion time series prediction methods typically require tens, hundreds, or even more denoising iterations to complete a single prediction, resulting in high inference latency and high computational resource consumption, making them difficult to deploy on edge devices, online business systems, and industrial control systems with high real-time requirements.

[0022] Many diffusion model acceleration methods were initially designed for generation tasks such as images and audio, with their core objective typically being to reduce generation costs by decreasing the number of backsampling steps, improving sampling solvers, or distilling sampling trajectories. When applying these methods to multivariate time series multi-step prediction, if only the number of backsampling steps is compressed without preserving historical windows, time stamps, conditional prediction means, and prediction window constraints during distillation training, the few-step student model may struggle to fully inherit the teacher model's ability to model conditional future distributions and prediction uncertainties.

[0023] The present invention aims to solve the problems of existing diffusion models having a large number of sampling steps, slow inference speed, and high deployment cost when used for multivariate time series multi-step prediction; and to achieve single-step diffusion-based multivariate time series prediction while ensuring prediction accuracy and probability output capability.

[0024] This invention provides a time series prediction method, electronic device, and storage medium based on a distillation-accelerated diffusion model.

[0025] Below, we will first introduce a time series prediction method for a diffusion model based on distillation acceleration provided by an embodiment of the present invention.

[0026] like Figure 1 As shown in the embodiments of the present invention, a time series prediction method based on a diffusion model accelerated by distillation may include the following steps: S1, Obtain the historical input sequence of the target object before the current time to construct historical conditional information, which may include: Obtain the historical multivariate time series of the target object up to the current moment, and extract a length of [length missing]. L The historical window is used to obtain the historical input sequence; Historical condition information is constructed based on the historical input sequence and its timestamp. C .

[0027] Historical input sequence The expression is as follows: ; in, For the first t-L+2 Each time point Dimensional observations, Indicates the number of variables.

[0028] Historical conditions information C It can include: One or more of the following: historical input sequence, historical time stamp, future time stamp, and decoding input.

[0029] This invention constructs training, validation, and test sets by collecting timestamped multivariate time series data. Each sample may include a historical input sequence. Real sequence The length of the historical input sequence is denoted as The future prediction window length is denoted as .

[0030] Future forecast window as follows: .

[0031] S2 inputs historical conditional information into a pre-trained conditional prediction model to obtain the mean conditional prediction of future time series.

[0032] Historical conditions information Input the pre-trained conditional prediction model to obtain the future. The conditional forecast mean at each time step: ; Conditional forecast mean It represents the future benchmark trend obtained based on historical multivariate time series.

[0033] The training loss of the conditional prediction model is as follows: ; in, The loss function for the conditional prediction model. For mathematical expectation operators, for The square of the norm, For the true sequence, For conditional forecasting of the mean, These are the weighting coefficients corresponding to the regularization loss. This is the regularized loss in the conditional prediction model.

[0034] S3, with the conditional prediction mean as the center, adds random perturbations to obtain multiple different perturbation inputs.

[0035] In this embodiment of the invention, the total loss of the single-step student generator is constructed by using a pre-trained teacher diffusion model and an auxiliary denoising network to constrain the training of the single-step student generator.

[0036] The teacher diffusion model can use a non-zero mean forward diffusion kernel, whose condition center is the output of the conditional prediction model. The forward diffusion process is represented as: ; in, To spread to the first Noisy future sequence at each time step For the future real sequence, For diffusion time step, This is the cumulative noise scheduling coefficient. The diffusion kernel causes high-noise states to hover around the conditional prediction mean. The perturbation, rather than the perturbation around the zero mean.

[0037] The teacher denoising network is denoted as The noise prediction process is as follows: ; in, Encoding the time features corresponding to the historical input sequence.

[0038] Teacher noise prediction loss as follows: ; Based on the noise prediction results, the diffusion state can be restored to a clean sequence estimate: ; in, A clean sequence for the future reconstruction.

[0039] Total loss of the teacher diffusion model as follows: ; in, The loss weights are used for the conditional prediction model.

[0040] The single-step student generator provided in this embodiment constructs a prior perturbation around the conditional predicted mean. Specifically, it can be achieved by using the conditional predicted mean... Centered on the random perturbation, we obtain the first... The first disturbance input, the second... The expression for the perturbation input is as follows: ; in: , The input is the disturbance strength.

[0041] S4 takes historical conditional information, conditional prediction mean, and multiple different perturbation inputs as pre-trained single-step student generators, and obtains multiple corresponding future prediction trajectories through network forward computation.

[0042] The single-step student generator uses a denoising network backbone to output residuals, the first... r The expression for each residual is as follows: ; in, Let the objective function be the single-step student generator. For historical conditions information, For the first r One disturbance input, The diffusion time step used by the student generator.

[0043] Historical conditions information Conditional prediction mean and disturbance input The common input is a pre-trained single-step student generator, which is used to obtain the first step through a single network forward computation. One future predicted trajectory: ; in, For the first r A future predicted trajectory, This is the residual jump coefficient. For the first r One disturbance input, For conditional forecasting of the mean, For the first r One residual.

[0044] By repeating the single-step generation process using different random perturbations on the same historical conditions, multiple future predicted trajectories can be obtained. : ; It is important to emphasize that each future predicted trajectory is obtained through a single forward computation of the single-step student generator, without the need for multi-step backward denoising. The output of this step is: multiple possible future predicted trajectories. .

[0045] This invention embodiment constructs an auxiliary denoising network. It is used to learn noise predictions on the current student generation distribution.

[0046] Given student-generated samples The following is obtained through a non-zero mean diffusion kernel: ; in, To spread to the first A noisy future sequence at each time step.

[0047] Auxiliary denoising network predicts noise The expression is as follows: ; Its training loss for: ; During the update phase of the single-step student generator, the parameters of the teacher diffusion model and the auxiliary denoising network are fixed. The teacher diffusion model... Clean sample estimation for: ; Clean sample estimation for auxiliary denoising networks for: ; Define the difference between the two only within the prediction window: ; Weighting factors for: ; in, For numerically stable terms, Upper limit for weight pruning. Loss of the non-zero mean SID generator. for: ; in, This is the difference adjustment coefficient in the SID loss.

[0048] This invention incorporates supervision into the prediction window mean of multiple student samples. Sampling is performed under the same input conditions. This time received ,but: ; Forecast window mean loss for: ; The total loss function for the single-step student generator is as follows: ; in, The total loss function for a single-step student generator. For non-zero mean SID generator loss, Weights are used to predict the mean loss of the window. This is for predicting the mean loss of the window.

[0049] This invention proposes a non-zero mean conditional diffusion distillation framework for multivariate time series prediction tasks. It compresses a multi-step teacher diffusion model into a single-step student generator, reducing inference latency while maintaining probabilistic prediction capabilities. A conditional prediction model is introduced during the diffusion distillation process, generating a conditional prediction mean based on historical sequences, historical time stamps, decoder input, and decoder time stamps. This ensures that both the teacher diffusion model and the student generator model around this conditional prediction mean. Furthermore, a non-zero mean diffusion kernel and prediction window constraints are introduced during the student distillation process, enabling the student model to learn the future residual distribution around the conditional prediction mean and concentrating the SID loss, auxiliary denoising loss, and prediction mean loss within the prediction window.

[0050] Understandably, after training is complete, the parameters of the conditional prediction model and the single-step student generator are saved for subsequent actual predictions. The teacher diffusion model and the auxiliary denoising network are used to constrain the training of the single-step student generator and do not participate in online computation during the actual prediction phase.

[0051] S5 performs statistical processing on multiple future prediction trajectories to obtain point prediction results and prediction intervals for future time series; it outputs multiple future prediction trajectories, point prediction results and prediction intervals for future time series to achieve time series prediction. Specifically, the total loss of the single-step student generator is constructed through a pre-trained teacher diffusion model and an auxiliary denoising network to constrain the training of the single-step student generator; a non-zero mean diffusion kernel and a prediction window constraint are introduced during the student distillation process, so that the single-step student generator learns the future residual distribution around the conditional prediction mean, and the corresponding loss is concentrated in the prediction window.

[0052] Specifically, in this embodiment of the invention, the point prediction result of the future time series is obtained by averaging the predicted trajectories of each future time series. : ; Based on the quantiles of multiple predicted trajectories, the prediction interval for future time series is obtained. : ; in, The lower quantile, It represents the upper quantile.

[0053] Ultimately, the future will be output. A point prediction sequence, prediction interval, and one or more candidate prediction trajectories at each time step.

[0054] In this embodiment of the invention, during the model pre-training phase, the conditional prediction model and the single-step student generator include training sample construction, conditional prediction model training, non-zero mean teacher diffusion model training, auxiliary denoising network training, SID distillation, and prediction window mean supervision. During the actual prediction phase, the trained conditional prediction model and single-step student generator are invoked to process the multivariate time series to be predicted, outputting the point prediction results, prediction intervals, and candidate prediction trajectories for future time windows. The model pre-training phase is completed before the actual prediction; during actual prediction, the model parameters remain fixed, eliminating the need to retrain the teacher diffusion model, auxiliary denoising network, or student generator.

[0055] The prediction graph of this invention on the ETTh1 dataset, as shown in the embodiment. Figure 2 As shown, Figure 2 The horizontal axis, time index, represents the time step, and the vertical axis, feat, represents each dimension. It can be seen that the embodiment of the present invention ensures prediction accuracy while generating prediction results in one step.

[0056] This invention can generate multiple future predicted trajectories and output prediction intervals and risk indicators based on historical input sequences, demonstrating stronger uncertainty representation capabilities. Compared to traditional multi-step diffusion time series prediction methods, this invention compresses multi-step teacher sampling into single-step student generation through non-zero-mean SID distillation, significantly reducing inference latency and the number of repeated network forward computations. This invention trains the single-step student generator using a teacher diffusion model, an auxiliary denoising network, and a non-zero-mean SID objective, ensuring a clear teacher distribution constraint in the single-step generation process. Furthermore, compared to general diffusion distillation methods, this invention introduces historical sequences, time stamps, conditional prediction means, and prediction window constraints specifically for time series prediction tasks, better preserving the single-step student generator's ability to model the conditional future path distribution.

[0057] Secondly, embodiments of the present invention also provide an electronic device, such as... Figure 3 As shown, it includes a processor 001, a communication interface 002, a memory 003, and a communication bus 004, wherein the processor 001, the communication interface 002, and the memory 003 communicate with each other through the communication bus 004. The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the steps of any of the distillation-accelerated diffusion model time series prediction methods provided in the first aspect of the present invention.

[0058] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0059] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0060] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0061] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0062] The method provided in this invention can be applied to electronic devices. Specifically, the electronic device can be a desktop computer, a portable computer, a smart mobile terminal, a server, etc. No limitation is made herein; any electronic device that can implement this invention falls within the protection scope of this invention.

[0063] Thirdly, corresponding to the distillation-accelerated diffusion model time series prediction method provided in the first aspect, this embodiment of the invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the distillation-accelerated diffusion model time series prediction methods provided in the first aspect of this invention.

[0064] For the electronic device / storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and relevant details can be found in the description of the method embodiments.

[0065] It should be noted that the electronic device and storage medium in the embodiments of the present invention are respectively the device, electronic device and storage medium for applying the above-mentioned diffusion model time series prediction method based on distillation acceleration. Therefore, all embodiments of the above-mentioned diffusion model time series prediction method based on distillation acceleration are applicable to the electronic device and storage medium, and can achieve the same or similar beneficial effects.

[0066] It should be noted that, in the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0067] The above description is merely a preferred 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 are included within the scope of protection of the present invention.

Claims

1. A time series forecasting method based on a diffusion model accelerated by distillation, characterized in that, include: Obtain the historical input sequence of the target object before the current moment to construct historical condition information; Historical conditional information is input into a pre-trained conditional prediction model to obtain the mean conditional prediction of future time series. By centering on the conditional prediction mean, random perturbations are added to obtain multiple different perturbation inputs; The historical condition information, the conditional prediction mean, and multiple different perturbations are input into a pre-trained single-step student generator, and multiple corresponding future predicted trajectories are obtained through forward computation of the network. Statistical processing is performed on multiple future predicted trajectories to obtain point prediction results and prediction intervals for future time series; Output multiple future predicted trajectories, point prediction results for future time series, and prediction intervals to achieve time series prediction; Specifically, the total loss of the single-step student generator is constructed through a pre-trained teacher diffusion model and an auxiliary denoising network to constrain the training of the single-step student generator; a non-zero mean diffusion kernel and a prediction window constraint are introduced during the student distillation process, so that the single-step student generator learns the future residual distribution around the conditional prediction mean, and the corresponding loss is concentrated in the prediction window.

2. The time series prediction method based on distillation-accelerated diffusion model according to claim 1, characterized in that, The step of obtaining the target object's historical input sequence before the current moment to construct historical condition information includes: Obtain the historical multivariate time series of the target object up to the current moment, and extract a length of [length missing]. L The historical window is used to obtain the historical input sequence; Historical condition information is constructed based on the historical input sequence and its timestamp.

3. The time series prediction method based on distillation-accelerated diffusion model according to claim 1, characterized in that, The historical condition information includes: One or more of the following: historical input sequence, historical time stamp, future time stamp, and decoding input.

4. The time series prediction method based on distillation-accelerated diffusion model according to claim 1, characterized in that, The training loss of the conditional prediction model is as follows: ； in, The loss function for the conditional prediction model. For mathematical expectation operators, for The square of the norm, For the true sequence, For conditional forecasting of the mean, These are the weighting coefficients corresponding to the regularization loss. This is the regularized loss in the conditional prediction model.

5. The time series prediction method based on distillation-accelerated diffusion model according to claim 1, characterized in that, No. r The expression for the predicted future trajectory is as follows: ； in, For the first r A future predicted trajectory, This is the residual jump coefficient. For the first r One disturbance input, For conditional forecasting of the mean, For the first r One residual.

6. The time series prediction method based on distillation-accelerated diffusion model according to claim 5, characterized in that, No. r The expression for each residual is as follows: ； in, Let the objective function be the single-step student generator. For historical conditions information, For the first r One disturbance input, The diffusion time step used by the student generator.

7. The time series prediction method based on distillation-accelerated diffusion model according to claim 1, characterized in that, The total loss function of the single-step student generator is as follows: ； in, The total loss function for a single-step student generator. For non-zero mean SID generator loss, Weights are used to predict the mean loss of the window. This is for predicting the mean loss of the window.

8. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the steps of the distillation-accelerated diffusion model time series prediction method according to any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the distillation-accelerated diffusion model time series prediction method according to any one of claims 1-7.

Images