Self-adaptive time sequence prediction method based on risk perception

Abstract

The invention discloses a self-adaptive time sequence prediction method based on risk perception, which dynamically adjusts a prediction step length by perceiving the change of a prediction error, avoids the problem of high error caused by multi-step prediction, and ensures the accuracy of a prediction result. The method comprises three key modules: a feature extraction and fusion module based on multi-scale fragments, used for capturing periodic, seasonal and trending features with intricate service demands, and extracting long-term dependency and local semantic information therein; the time sequence probability prediction module based on the diffusion model is used for carrying out probability modeling on the prediction result and describing a credible value of the prediction result by probability distribution of a prediction moment; and the prediction optimization module for adaptive risk management is used for evaluating and predicting the risk and carrying out adaptive truncation. According to the invention, a new solution is provided in the aspects of accuracy and reliability of time sequence prediction.

Description

Technical Field

[0001] The present invention relates to the cross-technical field of artificial intelligence and time series data processing, and in particular to an adaptive time series prediction method based on risk perception. Background Art

[0002] The time series forecasting problem refers to predicting the values ​​of future time points based on historical data within a certain time range. This problem has a wide range of applications in many fields, such as energy management, traffic flow forecasting, financial market analysis, etc. The purpose of time series forecasting is to help decision makers arrange resources and adjust strategies reasonably to improve efficiency when facing dynamic environments and uncertainties. However, most of the current time series forecasting models are based on a fixed forecast step size, that is, forecasting is performed according to a fixed time length (such as one week, one month, etc.). Although this method is simple, it often cannot adapt to the dynamic changes of time series data and has obvious shortcomings.

[0003] Specifically, the practice of fixing the prediction step size brings two major defects. First, a prediction step size that is too short cannot provide sufficient forward-looking support for decision-making, making it difficult for the model to effectively capture long-term changes and trends. Second, a prediction step size that is too long can easily lead to the accumulation of errors. As the prediction step size increases, the error may gradually amplify in subsequent predictions, eventually leading to distortion of the prediction results, thereby affecting the effectiveness of actual decision-making. Especially when faced with complex and time-varying time series data, the impact of the error will gradually increase, further affecting the accuracy and reliability of the prediction. Summary of the invention

[0004] The purpose of the present invention is to provide an adaptive time series prediction method based on risk perception, which aims to capture the periodicity, seasonality and trend changes of the time series more accurately through the feature extraction and fusion module of multi-scale fragments according to the input time series data, and then through the time series probability prediction module based on the diffusion model, the extracted high-dimensional features are used as prior knowledge, integrated into the forward and reverse processes of the diffusion model, and the correlation of covariates in the two-way process is deeply captured, so as to train the time series probability prediction model. Finally, in the prediction optimization module of adaptive risk management, the inherent uncertainty risk of the training model and the additional risk brought by the model when predicting on new data are calculated, and the two risks are integrated into a risk value, and then adaptive truncation is implemented for the high-risk prediction moment to achieve the adaptive prediction effect of the dynamic step size.

[0005] The technical solution of the present invention is as follows: An adaptive time series forecasting method based on risk perception, comprising: Step 1) Preprocessing: Input time series data Preprocessing is performed to identify the top-k periodic patterns in the time series data through fast Fourier transform (FFT), and the original sequence is segmented using different periodic patterns as segment lengths; The original sequence is split, and the specific process is as follows: ; in, and Indicates the calculation of Fast Fourier Transform FFT and amplitude value, represents the calculated amplitude of each frequency; Represents the calculation of frequency, is the length of the time series, Represents a time interval; each channel is operated on first, then For multiple channels C To average; The top-k fragment lengths are obtained by the following formula: ; in, Indicates the frequencies of the top-k magnitudes corresponding to the amplitude values, Indicates the length of the corresponding periodic segment; Indicates selecting the top-k index values ​​with the largest amplitude values ​​from A; For a given patch length and sliding step length , and then get the patched sequence ,in ; R represents the real number field; The above explains how to process the periodic features obtained above and form a patch sequence.

[0006] Step 2: Extract and fuse the features of multi-scale fragments, and input the fragments obtained in step 1 into the local feature extraction layer and the global relationship analysis layer; the local feature extraction layer uses the multi-layer perceptron MLP to capture the temporal dependency within the fragment and the temporal correlation of different variables; the global relationship analysis layer uses multi-scale seasonal-trend decomposition to analyze the temporal changes between fragments, and then uses multiple moving average kernels to smooth periodic fluctuations to highlight long-term trends; finally, the feature extraction results are obtained through linear weighted integration ; (Model the time series between different segments to make full use of the periodic trend characteristics between adjacent time periods; since time series data usually contains patterns and characteristics on multiple time scales, a multi-scale series decomposition method is adopted to consider seasonality and trend on different time scales at the same time to more comprehensively capture the diversity of the series; decompose the time series based on different moving average kernels, and then integrate all decomposition results through averaging operations).

[0007] For each single moving average kernel kernel, the series decomposition series_decomp process is: ; in Indicates the number of patches N , the number of channels is The input sequence is Represent the seasonal component and trend component respectively, using Perform moving average, Keep the sequence length unchanged; The function represents the moving average of the time series, averaging the values ​​in the series; The function represents the padding operation to keep the length of the sequence unchanged; Represents each sliding average kernel, and the size of the kernel is recorded as ; Then a multi-scale average kernel is used to summarize the above sequence decomposition process: ; in is the final seasonal component and trend component, a total of The average moving kernels are of different sizes.

[0008] Step 3: Perform time series probability prediction based on the diffusion model. Incorporate the high-dimensional features extracted in step 2 as prior knowledge into the forward denoising and reverse denoising process of the diffusion model. Use a unified optimization objective to train the model and generate probability prediction results by sampling the input historical sequence multiple times. The entire diffusion process can be modeled as: ; in Representation feature extraction module The prediction results obtained are regarded as the estimated conditional mean Prior knowledge of represents the noised sequence in the diffusion forward process, represents the identity matrix; The conditional distribution of the forward process for all other step sizes is defined as: ; , represents the noised sequence in the forward diffusion process at time t and time t-1; represents a noise control parameter in the diffusion process; in, ; At any time step t, directly from Sampling The forward process conditional distribution of is defined as: ; in, ; The reverse process is defined as: ; in, ; represents the noised sequence at time 0; During the training process, the conditional generation model and the diffusion model are integrated into a unified optimization objective, namely: ; in, Represents the training loss of the conditional generative model, which is used to measure the conditional generative model's conditional mean Ability to make accurate estimates; represents the training loss of the diffusion model, which is used to measure the uncertainty of the denoised prediction of the diffusion model; represents the result of feature extraction, which serves as a guiding factor for approximating the true posterior distribution of z; represents the KL divergence from distribution p to distribution q; is the symbol of expectation, Latent variables for conditional generative models, used to guide generation and prediction.

[0009] Step 4: Perform adaptive risk management forecast optimization, calculate the inherent risk of the model and the risk of variation in the forecast results, screen high-risk moments based on the sigma principle in the normal distribution, and then conduct in-depth mining of high-risk moments, calculate advance and continuity, linearly weighted integrate these two indicators as a comprehensive risk score, and finally select the moment with the largest value for the stage, thereby achieving adaptive prediction of dynamic step length; Risk is divided into two parts: inherent risk and mutation risk. Inherent risk refers to the uncertainty risk inherent in the training model, which is obtained by conformal prediction theory. Mutation risk refers to the additional risk brought by the model when making predictions on new data, which is measured by the coefficient of variation. The core goal of inherent risk is to evaluate the prediction sequence generated by the feature extraction module. The confidence and reliability of the model enable the model to assign a risk indicator to each predicted time point that accurately reflects its inherent uncertainty. To this end, the present invention performs calculations on the validation set and sets a constant significance level for each time point. ε This level essentially defines the maximum acceptable error rate, ensuring that the true observations have The probability of falling within a certain interval centered on the model prediction value is: ; in, Indicated in i The forecast uncertainty bound at each point in time is the risk inherent in the model we define.

[0010] The coefficient of variation is used as a statistical indicator to quantify the volatility risk of the forecast results at each time point. Specifically, by sampling the forecast results of the diffusion model multiple times, a set of n Independent prediction samples, calculate the mean at each time point and standard deviation , then i Coefficient of variation at each time point It can be defined as: ; So far, i The risk value at a time point can be expressed as: ; in Indicates i Inherent risk of the point-in-time model, represents the dynamic risk predicted by the model at the i-th time point, represents a hyperparameter.

[0011] In the decision-making process, it is crucial to determine a reasonable cutoff point. We need to effectively resist the potential misleading caused by premature or late cutoffs and high-risk noise. In view of this, we have established an evaluation process that comprehensively considers risk intensity, time advance and continuity. First, focus on screening out those time points with higher risk values: Then, we conduct in-depth research on these high-risk time points and calculate their “advancement”, that is, the degree of advance of the high-risk points: ( Indicates high-risk locations. represents the total length of the forecast), and “continuity” – a measure of the density of consecutive high-risk time points. Finally, the risk of advance τ Peak is used to integrate the advanceness and continuity of high-risk time points.

[0012] Risk of early τ The peak is decisively truncated, and the previous predictions are integrated to form a preliminary sequence. Then, additional data is actively requested to re-forecast the truncated part until the full sequence prediction is completed to ensure the completeness and accuracy of the prediction. Through iterative prediction and data update, the overall quality of the prediction sequence can be effectively improved, the prediction length requirements can be met, and more powerful support can be provided for decision-making.

[0013] This paper proposes an adaptive time series prediction method based on risk perception, which aims to avoid the accumulation and amplification of errors by dynamically evaluating the reliability of the prediction results and adjusting the prediction step size accordingly. Specifically, the intrinsic characteristics of time series data are deeply mined through multi-scale slicing technology, and the uncertainty risk in the prediction process is quantified and evaluated by combining diffusion model and conformal prediction technology, and finally the dynamic optimization of adaptive prediction step size is achieved. This method can adaptively adjust the prediction step size according to the dynamic changes of data, thereby improving the accuracy and reliability of the prediction.

[0014] Compared with the prior art, the present invention has the following advantages: The risk-aware adaptive time series prediction method of the present invention breaks through the fixed step length limitation of the traditional time series prediction method, realizes the dynamic optimization of the prediction step length, and can overcome the decline in prediction accuracy caused by error accumulation. The diffusion model is pioneered to quantify the risk of variation of the predicted value, and the common prediction technology is combined to evaluate the inherent risk of the model, forming a comprehensive risk perception system. After comprehensively considering the adaptive prediction strategy of prediction uncertainty, foresight and continuity, the accuracy and robustness of the prediction are successfully improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A technical roadmap of an adaptive time series prediction method based on risk perception of the present invention; Figure 2 It is a flow chart of the feature extraction and fusion module based on multi-scale segments of the present invention; Figure 3 It is a flow chart of the time series probability prediction module based on the diffusion model of the present invention; Figure 4 This is a flow chart of the prediction optimization module of the adaptive risk management of the present invention; Figure 5 It is a curve diagram of mutation risk data; Figure 6 It is a graph of inherent risk data; Figure 7 This is the p-value curve. DETAILED DESCRIPTION

[0016] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of the present invention.

[0017] The objective of the present invention is achieved through the following technical solutions: like Figures 1 to 4As shown, a risk-aware adaptive time series prediction method specifically includes: Step 1: Obtain time series data and preprocess the data: Table 1: Time series table ;

[0018] In the exchange rate forecasting scenario. First, for the time series As shown in Table 1, this module uses Fast Fourier Transform (FFT) to detect the potential Top-k periodic patterns in the time series, and uses different periodic patterns as the patch length to segment the original time series. The specific process is as follows: ; in, and Indicates the calculation of Fast Fourier Transform FFT and amplitude value, represents the calculated amplitude of each frequency; Represents the calculation of frequency, is the length of the time series, Represents a time interval; each channel is operated on first, then For multiple channels C To average; The top-k fragment lengths are obtained by the following formula: ; in, Indicates the frequencies of the top-k magnitudes corresponding to the amplitude values, Indicates the length of the corresponding period segment; in the exchange rate prediction scenario, taking k=5 as an example, the obtained ,at this time .

[0019] For a given patch length and sliding step length , and then get the patched sequence ,in .

[0020] Step 2: Feature extraction and fusion based on multi-scale segments: Design intra-fragment linear layers and inter-fragment linear layers, and use fully connected neural networks to extract local deep dependencies of time series data as well as global seasonality and trend changes. The intra-fragment linear layer uses a multi-layer perceptron (MLP) to capture the features of temporal dependencies and channel relationships within the fragments; the inter-fragment linear layer uses multi-scale seasonal-trend decomposition to analyze the time series between fragments, and then uses multiple moving average kernels to smooth periodic fluctuations and highlight long-term trends. For each single moving average kernel, the series decomposition series_decomp process is: ; in Indicates the number of patches N , the number of channels is The input sequence is Represent the seasonal component and trend component respectively, using Perform moving average, Keep the sequence length unchanged; then use the multi-scale average kernel to summarize the above sequence decomposition process: ; in is the final seasonal component and trend component, a total of The average moving kernels are of different sizes.

[0021] Step 3: Time series probability prediction based on diffusion model: The time series probability prediction model based on the diffusion model uses the conditional information output by the feature extraction module as prior knowledge to guide the forward denoising and reverse denoising process of the diffusion model. By integrating this conditional information in the diffusion process, the model can make full use of richer contextual information in the denoising process, and thus more accurately grasp the intrinsic time series properties between the input sequence and the generated target. The entire diffusion process can be modeled as: ; in Representation feature extraction module The prediction results obtained are regarded as the estimated conditional mean Prior knowledge of represents the noised sequence in the diffusion forward process, Denotes the identity matrix; The conditional distribution of the forward process for all other step sizes is defined as: ; , represents the noised sequence in the forward diffusion process at time t and time t-1; represents a noise control parameter in the diffusion process; in, ; At any time step t, directly from Sampling The forward process conditional distribution of is defined as: ; in, ; The reverse process is defined as: ; in, ; represents the noised sequence at time 0; During the training process, the present invention integrates the conditional generation model and the diffusion model into a unified optimization objective, namely: ; in, Represents the training loss of the conditional generative model, which is used to measure the conditional generative model's conditional mean Ability to make accurate estimates; represents the training loss of the diffusion model, which is used to measure the uncertainty of the denoised prediction of the diffusion model; represents the result of feature extraction, which serves as a guiding factor for approximating the true posterior distribution of z; represents the KL divergence from distribution p to distribution q; is the symbol of expectation, The latent variables of the conditional generation model are used to guide generation and prediction; in the inference stage, the model can sample the input historical sequence multiple times to produce different prediction results, which will be used in the risk assessment module to quantify the uncertainty of the prediction results.

[0022] Step 4: Perform adaptive risk management forecast optimization: Risk is divided into inherent risk and mutation risk. Inherent risk refers to the uncertainty risk inherent in the training model, which is obtained by conformal prediction theory. Mutation risk refers to the additional risk brought by the model when making predictions on new data, which is measured by the coefficient of variation. The core goal of inherent risk is to evaluate the prediction sequence generated by the feature extraction module. The confidence and reliability of the model enable the model to assign a risk indicator to each predicted time point that accurately reflects its inherent uncertainty. To this end, the present invention performs calculations on the validation set and sets a constant significance level for each time point. ε , this level essentially defines the maximum acceptable error rate, ensuring that in the entire time series, the true observation value has a high probability of falling within a certain interval centered on the model prediction value, that is: ; in, Indicated in i The forecast uncertainty bound at each point in time is the risk inherent in the model we define.

[0023] The present invention uses the coefficient of variation as a statistical indicator to quantify the volatility risk of the prediction results at each time point. Specifically, by sampling the prediction results of the diffusion model multiple times, a set of n Independent prediction samples, calculate the mean at each time point and standard deviation , then i Coefficient of variation at each time point It can be defined as: ; So far, i The risk value at a time point can be expressed as: ; in Indicates i Inherent risk of the point-in-time model, represents the dynamic risk predicted by the model at the i-th time point, represents the hyperparameter; the data in Table 1 are used as input, and after data preprocessing, periodic pattern recognition, multi-scale segment feature extraction and fusion, and time series probability prediction based on the diffusion model, the inherent risk, variation risk and comprehensive risk value of the model are calculated. Get inherent risk, variant risk and Value Figures 5 to 7 As shown in the figure, the points above the set threshold are finally screened out as The corresponding comprehensive final risk value is .

[0024] In the decision-making process, it is crucial to determine a reasonable cutoff point. It is necessary to effectively resist the potential misleading caused by premature or late cutoff and high-risk noise. In view of this, we have established an evaluation process that comprehensively considers risk intensity, time advance and continuity. First, focus on screening out those time points with higher risk values: Then, we conduct in-depth research on these high-risk time points and calculate their “advancement”, that is, the degree of advance of the high-risk points: ( Indicates high-risk locations. represents the total length of the forecast), and “continuity” – measures the density of consecutive high-risk time points. Finally, the advance risk τ peak is introduced to integrate the advance and continuity of high-risk time points.

[0025] Risk of early τ The peak is decisively truncated, and the previous predictions are integrated to form a preliminary sequence. Then, additional data is actively requested to re-forecast the truncated part until the full sequence prediction is completed to ensure the completeness and accuracy of the prediction. Through iterative prediction and data update, the overall quality of the prediction sequence can be effectively improved, the prediction length requirements can be met, and more powerful support can be provided for decision-making.

[0026] As shown in Table 2, a risk-aware adaptive time series prediction method is compared with other methods (such as TimesNet, PatchTST, etc.) on multiple datasets (such as Electricity, Weather, Exchange, etc.); Table 2: Comparison of experimental results of the present invention and other methods ;

[0027] Experimental results show that the mean square error (MSE) and mean absolute error (MAE) of the model using the method of the present invention at different time slices (such as 96, 192, 336, 720, etc.) are significantly better than the existing methods, indicating that the prediction accuracy of the present invention on multiple data sets has higher stability and lower error.

Claims

1. A risk-aware adaptive time series prediction method, characterized in that: include: Step 1) Preprocessing: Input time series data Preprocessing is performed to identify the top-k periodic patterns in the time series data through fast Fourier transform (FFT), and the original sequence is segmented using different periodic patterns as segment lengths; Step 2) Extract and fuse the features of multi-scale segments, and input the segments obtained in step 1) into the local feature extraction layer and the global relationship analysis layer; the local feature extraction layer uses the multi-layer perceptron MLP to capture the temporal dependency within the segment and the temporal correlation of different variables; the global relationship analysis layer uses multi-scale seasonal-trend decomposition to analyze the temporal changes between segments, and then uses multiple moving average kernels to smooth periodic fluctuations to highlight long-term trends; Finally, the feature extraction result is obtained by linear weighted integration ; Step 3) Perform time series probability prediction based on the diffusion model and transform the high-dimensional features extracted in step 2) In the forward denoising and reverse denoising process of the diffusion model as prior knowledge, a unified optimization objective is used for model training, and probabilistic prediction results are generated by sampling the input historical sequence multiple times; Step 4) Perform adaptive risk management forecast optimization, calculate the inherent risk of the model and the risk of variation in the forecast results, screen high-risk moments based on the sigma principle in the normal distribution, and then conduct in-depth mining of high-risk moments, calculate the advance and continuity, and linearly weighted integrate these two indicators as a comprehensive risk score. Finally, select the moment with the largest value for the stage, thereby achieving adaptive prediction with dynamic step size.

2. The risk-aware adaptive time series prediction method according to claim 1, characterized in that: In step 1), the original sequence is segmented. The specific process is as follows: ； in, and Indicates the calculation of Fast Fourier Transform FFT and amplitude value, represents the calculated amplitude of each frequency; Represents an array of frequency values ​​calculated by fast Fourier transform FFT, Represents the calculation of frequency, is the length of the time series, Represents a time interval; each channel is operated on first, then For multiple channels C To average; The top-k fragment lengths are obtained by the following formula: ； in, Indicates the frequencies of the top-k magnitudes corresponding to the amplitude values, Indicates the length of the corresponding periodic segment; Indicates selecting the top-k index values ​​with the largest amplitude values ​​from A; For a given patch length and sliding step length , and then get the patched sequence ,in ; R represents the real number field.

3. The risk-aware adaptive time series prediction method according to claim 1, characterized in that: In step 2), multiple moving average kernels are used to smooth periodic fluctuations to highlight long-term trends. For each single moving average kernel, the series decomposition process series_decomp is: ； in Indicates the number of patches N , the number of channels is The input sequence is Represent the seasonal component and trend component respectively, using Perform moving average, Keep the sequence length unchanged; The function represents the moving average of the time series, averaging the values ​​in the series; The function represents the padding operation to keep the length of the sequence unchanged; Represents each sliding average kernel, and the size of the kernel is recorded as ; Then a multi-scale average kernel is used to summarize the above sequence decomposition process: ； in is the final seasonal component and trend component, a total of The average moving kernels are of different sizes.

4. The risk-aware adaptive time series prediction method according to claim 1, characterized in that: Step 3) The specific process is as follows: The diffusion model is: ； in Representation feature extraction module The prediction results obtained are regarded as the estimated conditional mean Prior knowledge of represents a normal distribution, represents the noised sequence in the diffusion forward process, represents the identity matrix; The conditional distribution of the forward process for all other step sizes is defined as: ； , represents the noised sequence in the forward diffusion process at time t and time t-1; represents a noise control parameter in the diffusion process; in, ; At any time step t, directly from Sampling The forward process conditional distribution of is defined as: ； in, ; represents the cumulative effect of the noise control parameters, and T represents the total time step of the diffusion model; The reverse process is defined as: ； in, ; represents the noised sequence at time 0; During the training process, the conditional generation model and the diffusion model are integrated into a unified optimization objective, namely: ； in, represents the unified optimization objective of the diffusion model; Represents the training loss of the conditional generative model, which is used to measure the conditional generative model's conditional mean Ability to make accurate estimates; represents the training loss of the diffusion model, which is used to measure the uncertainty of the denoised prediction of the diffusion model; represents the result of feature extraction, which serves as a guiding factor for approximating the true posterior distribution of z; represents the KL divergence from distribution p to distribution q; is the symbol of expectation, Latent variables for conditional generative models, used to guide generation and prediction.

5. The risk-aware adaptive time series prediction method according to claim 1, characterized in that: In step 4), inherent risk refers to the uncertainty risk inherent in the training model, which is obtained from conformal prediction theory; The core goal of inherent risk is to evaluate the prediction sequence generated by the feature extraction module The confidence and reliability of the model enable the model to assign a risk indicator to each prediction time point that accurately reflects its inherent uncertainty; the calculation is performed on the validation set, and a constant significance level is set for each time point ε , to define the maximum acceptable error rate, ensuring that in the entire time series, the true observations have The probability falls within a certain interval centered on the model prediction value, that is: ； in, represents the true value at time point i, represents the predicted value at time point i, Indicated in i The forecast uncertainty boundary at a point in time, i.e. the risk value inherent in the defined model; Mutation risk refers to the additional risk that the model brings when making predictions on new data, measured by the coefficient of variation; The coefficient of variation is used to quantify the volatility risk of the prediction results at each time point. By sampling the prediction results of the diffusion model multiple times, a set of n Independent prediction samples, calculate the mean at each time point and standard deviation , then the coefficient of variation at the i-th time point is Defined as: ； So far, i The risk value at each time point is expressed as: ； in Indicates i Inherent risk of the point-in-time model, Indicates i The dynamic risk predicted by the model at each time point, represents a hyperparameter; In step 4), establish an assessment process that comprehensively considers risk intensity, time advance and continuity; first, select the time point according to the following formula: ； The screened time points were recorded as high-risk moments; represents the mean of the forecast series; The overall standard deviation of performance; Then, the advance degree of high-risk moments is calculated according to the following formula: ； in, Indicates high-risk locations. Indicates the total length of prediction; Calculate the continuity of high-risk moments. Continuity refers to the number of consecutive high-risk moments at the current time point, which measures the density of consecutive high-risk moments. Finally, introducing early risk τ Peaks to integrate the advance and continuity of high-risk moments; Advance Risk τ The data is truncated at the peak and the previous predictions are integrated to form a preliminary sequence. Then, additional data is actively requested and the truncated part is re-predicted until the full sequence prediction is completed to ensure the completeness and accuracy of the prediction.

Images