Multi-scale time sequence prediction method and device

Abstract

The invention discloses a multi-scale time sequence prediction method and device, and belongs to the technical field of time sequence prediction analysis. According to the method, the fOU process, the LSTM model and the attention mechanism are combined, the long-term trend and the short-term fluctuation of the time sequence can be captured at the same time, and the prediction precision and the interpretability are improved. The method comprises the following steps: firstly, acquiring multivariate multi-scale time sequence data and preprocessing; then decomposing the target time sequence into a low-frequency component and a high-frequency component by adopting a moving average method and a residual error method; carrying out modeling and prediction on the low-frequency components by adopting an fOU-LSTM hybrid model; carrying out modeling and prediction on the high-frequency components by adopting a multi-scale LSTM model and combining an attention mechanism; and finally, fusing the low-frequency result and the high-frequency result by adopting a dynamic weighting method to obtain a final target variable prediction value. The method is suitable for time sequence prediction tasks with long memorability, multi-scale features and nonlinear features in the fields of finance, meteorology, medical treatment and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of time series analysis, and particularly relates to a multi-scale time series prediction method and device. Background Art

[0002] Time series prediction is a key issue in many fields, such as interest rate prediction in the financial market, temperature prediction of meteorological data, heart rate prediction in the medical field, etc. Time series usually has a long-term memory effect, that is, the influence of historical values on the future decays according to the power law of time, rather than exponential decay. Traditional time series prediction methods usually assume that the time series is stationary or has a linear relationship, and it is difficult to capture complex non-linear patterns and long-term dependencies; in addition, time series contains components of different time scales (such as long-term trends, medium-term cycles, short-term fluctuations). Using a single model to process multi-scale data results in mutual interference between long-term trends and short-term fluctuations, and it is difficult to model multi-scale dynamics simultaneously.

[0003] In recent years, deep learning models (such as LSTM, Transformer) have made significant progress in time series prediction. Although they can capture non-linear relationships, they need to rely on a large amount of data sets, and there are still limitations in long-term trend modeling and small data sets. Deep learning models usually require a large amount of data to train better performance and are prone to overfitting on small data sets. The fractional-order Ornstein-Uhlenbeck process (fOU process) is based on a stochastic differential equation and has a clear mathematical form and physical meaning. The fOU process performs stably on small data sets, is suitable for scenarios with less data volume, can well model the long memory and mean reversion characteristics of time series, and can better depict the long-term trend of time series, but it cannot directly process multi-variable time series and complex non-linear relationships. Therefore, how to combine the fOU process and deep learning models to achieve multi-scale time series prediction is an urgent problem to be solved.

[0004] By combining the fOU process with the LSTM model, the present invention decomposes the time series into low-frequency components (long-term trends) and high-frequency components (short-term fluctuations), models the low-frequency components using a fOU-LSTM hybrid model, models the high-frequency components using a multi-scale LSTM model combined with an attention mechanism, and uses a dynamic weighting method to fuse the prediction results, which can significantly improve the prediction accuracy and the interpretability of the results. Summary of the Invention

[0005] The purpose of the present invention is to provide a multi-scale time series prediction method and device for the deficiencies of the existing technology, aiming to realize the joint modeling of the long memory and non-linear dynamics of time series, capture the long-term trends and short-term fluctuations of time series, and effectively capture the cross-scale features of time series by designing a multi-scale decomposition and fusion mechanism, so as to improve the prediction accuracy and interpretability.

[0006] The object of the present invention is achieved by the following technical solutions: A multi-scale time series prediction method, comprising: obtaining multi-source multi-scale time series data and performing preprocessing; the multi-source multi-scale time series data includes target variable time series data and high- and low-frequency multi-source time series data related to the target variable;

[0007] Separating the high- and low-frequency components of the target variable time series data;

[0008] Based on the separated low-frequency component and the preprocessed low-frequency multi-source time series data, using the fOU-LSTM hybrid model to model and predict the low-frequency component, generating a target variable low-frequency component prediction time series;

[0009] Based on the separated high-frequency component and the preprocessed high-frequency multi-source time series data, using the multi-scale LSTM model and combining the attention mechanism to model and predict the high-frequency component, generating a target variable high-frequency component prediction time series;

[0010] Based on the target variable low- and high-frequency component prediction time series, adopting the dynamic weighting method for fusion to obtain the final target variable prediction value.

[0011] Further, the preprocessing of the multi-source multi-scale time series data includes:

[0012] Performing time alignment, frequency alignment, missing data value processing, and data standardization processing on the multi-source multi-scale time series data.

[0013] Further, the separation of the high- and low-frequency components of the target variable time series data includes:

[0014] Based on the target variable time series data, using the moving average method to calculate the low-frequency component, and the moving average time window size is adjusted according to the volatility, and the window size formula is:

[0015]

[0016] wherein, is the time window, is the volatility in the past several days, is the adjustment coefficient;

[0017] Based on the target variable time series and the low-frequency component, directly calculating the high-frequency component through the residual.

[0018] Further, the effectiveness of the separated components is verified using the Fourier transform method; if the energy proportion of the low-frequency component in the low-frequency region is higher than the preset threshold and / or the energy proportion of the high-frequency component in the high-frequency region is higher than the preset threshold, it proves that the separation effect is good, otherwise, the moving average time window is adjusted to re-extract the low-frequency and high-frequency components.

[0019] Further, based on the separated low-frequency components and the preprocessed low-frequency multivariate time series data, a fOU-LSTM hybrid model is used to model and predict the low-frequency components, generating a predicted time series of the target variable low-frequency components, including:

[0020] Use the fractional-order Ornstein-Uhlenbeck process to fit the low-frequency components, capture their long-term trends and mean reversion characteristics, and generate the benchmark values of the low-frequency trends;

[0021] Use the LSTM model to predict the residuals of the low-frequency components, capture the non-linear relationships of the low-frequency components, input the historical residual terms of the low-frequency sequence and the preprocessed low-frequency multivariate time series data, predict the residuals of the low-frequency components, and perform non-linear correction on the low-frequency components;

[0022] Fuse the benchmark values and the residual predictions to generate a predicted time series of the target variable low-frequency components.

[0023] Further, use the fOU model to fit the low-frequency components, including:

[0024] Use the fOU process to fit the low-frequency components, model the long memory by introducing fractional Brownian motion, and its differential form is , where, is the low-frequency component of the time series, is the mean reversion rate, is the long-term mean, is the volatility, is the fractional Brownian motion with the Hurst exponent of ;

[0025] Use the wavelet analysis method to estimate ;

[0026] After obtaining estimated, estimate other parameters using the maximum likelihood estimation method;

[0027] Predict the benchmark values of the low-frequency trends based on the analytical solution of the fOU process .

[0028] Further, based on the separated high-frequency components and the preprocessed high-frequency multivariate time series data, a multi-scale LSTM model is used in combination with an attention mechanism to model and predict the high-frequency components, generating a predicted time series of the high-frequency components of the target variable, including:

[0029] Using the high-frequency multivariate time series data, extract the time dependence through the LSTM model and retain the outputs of all time steps;

[0030] Use the attention mechanism to dynamically assign the importance of each time step;

[0031] Take the mean of the features of each time step to generate a high-frequency aggregated feature vector;

[0032] Using the high-frequency components, extract the long-term trend through LSTM and only output the hidden state of the last time step to generate a high-frequency component feature vector;

[0033] Concatenate the high-frequency aggregated feature vector and the high-frequency component feature vector, perform weighted fusion through feature-level attention, and use a fully connected layer to map the weighted features to the output to generate a predicted time series of the high-frequency components of the target variable.

[0034] Further, based on the predicted time series of the low-frequency components of the target variable and the predicted time series of the high-frequency components of the target variable, a dynamic weighting method is used for fusion to obtain the final predicted value of the target variable, including:

[0035] Adopt dynamic weight allocation, where the weight value is calculated based on the historical prediction errors of the low-frequency and high-frequency components, and the fusion form is:

[0036]

[0037] Among them, is the dynamic weight, is the mean absolute error of the high-frequency model in the past time window N, is the mean absolute error of the low-frequency model in the same time window.

[0038] The present invention also provides a multi-scale time series prediction device, including a data acquisition and preprocessing unit, a high-low frequency component separation unit, a low-frequency component prediction unit, a high-frequency component prediction unit, and a result fusion and prediction unit;

[0039] The data acquisition and preprocessing unit is used to: acquire multivariate multi-scale time series data and perform preprocessing; the multivariate multi-scale time series data includes target variable time series data and high-low frequency multivariate time series data related to the target variable;

[0040] The high-low frequency component separation unit is used to: separate the high-frequency and low-frequency components from the target variable time series data;

[0041] The low-frequency component prediction unit is used to: based on the separated low-frequency components and the pre-processed low-frequency multivariate time series data, use the fOU-LSTM hybrid model to model and predict the low-frequency components, and generate a target variable low-frequency component prediction time series;

[0042] The high-frequency component prediction unit is used to: based on the separated high-frequency components and the preprocessed high-frequency multivariate time series data, use a multi-scale LSTM model in combination with an attention mechanism to model and predict the high-frequency components, and generate a target variable high-frequency component prediction time series;

[0043] The result fusion and prediction unit is used to predict the time series of the low-frequency and high-frequency components of the target variable, and fuse them using a dynamic weighting method to obtain a final predicted value of the target variable.

[0044] The present invention also provides a computer program product, including a computer program, which implements the above-mentioned multi-scale time series prediction method when executed by a processor.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. The multi-scale time series prediction method proposed in this invention combines the fractional-order fOU process with the LSTM deep learning model to achieve joint modeling of the long-term memory and nonlinear relationship of the time series. The fOU process describes the long memory of the time series by introducing the Hurst index (H), and the LSTM model describes the nonlinear relationship and dependency of the time series through the gating mechanism and attention mechanism. The combination of the two can capture both the long-term trend of the time series and the nonlinear dynamic relationship of multivariate features.

[0047] 2. The high- and low-frequency component separation and fusion method proposed in the present invention decomposes the original time series into high- and low-frequency sub-components of different scales through the moving average method and models them separately, effectively capturing the cross-scale characteristics of the time series and improving the prediction accuracy and interpretability. The low-frequency components are modeled using the fOU+LSTM hybrid method to capture the long-term trend of the low-frequency components of the time series; the high-frequency components are modeled using a multi-scale LSTM model combined with an attention mechanism to capture the short-term fluctuations and nonlinear relationships of the time series. Finally, the prediction results of the two are dynamically weighted and fused, which can effectively improve the prediction accuracy of the time series. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0049] Figure 1 It is a flowchart of a multi-scale time series prediction method provided by an embodiment of the present invention.

[0050] Figure 2 It is a flowchart of the acquisition and preprocessing of multi-source multi-scale time series data provided by an embodiment of the present invention.

[0051] Figure 3 It is a flowchart of the separation of high-frequency and low-frequency components provided by an embodiment of the present invention.

[0052] Figure 4 It is a flowchart of the modeling and prediction of the low-frequency component fOU-LSTM provided by an embodiment of the present invention.

[0053] Figure 5 It is a flowchart of the multi-scale LSTM modeling and prediction of the high-frequency component provided by an embodiment of the present invention.

[0054] Figure 6 It is a flowchart of the fusion and prediction of high-frequency and low-frequency components provided by an embodiment of the present invention.

[0055] Figure 7 It is a schematic structural diagram of a multi-scale time series prediction device provided by an embodiment of the present invention. Detailed implementation manners

[0056] To make the purpose, technical solutions and advantages of the present invention clearer and more understandable, the following further details the present invention in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation manners described herein are only used to explain the present invention and do not limit the protection scope of the present invention.

[0057] Glossary:

[0058] High frequency: Refers to time series data with a short time interval, usually including data at the daily level and below (such as hourly, minute-level, etc.), with relatively large noise and strong volatility;

[0059] Low frequency: Refers to time series data with a long time interval, usually including data at the daily level and above (such as weekly, monthly, annual levels, etc.), with relatively small noise and obvious trends and seasonality.

[0060] Figure 1 It is a multi-scale time series prediction method provided by an embodiment, including the following steps:

[0061] 1) Obtain multivariate multi-scale time series data, including target variable time series data and high- and low-frequency multivariate time series data related to the target variable , and preprocess the multivariate multi-scale time series data;

[0062] 2) Based on the target variable time series data , perform high- and low-frequency component separation;

[0063] 3) Based on the separated low-frequency components and the preprocessed low-frequency multivariate time series data, use the fOU-LSTM hybrid model to model and predict the low-frequency components, generating a predicted time series of the low-frequency components of the target variable ;

[0064] 4) Based on the separated high-frequency components and the preprocessed high-frequency multivariate time series data, use the multi-scale LSTM model combined with the attention mechanism to model and predict the high-frequency components, generating a predicted time series of the high-frequency components of the target variable ;

[0065] 5) Based on the predicted time series of the low-frequency and high-frequency components of the target variable, use the dynamic weighting method for fusion to obtain the final predicted value of the target variable .

[0066] Figure 2 is the flow chart for obtaining and preprocessing multivariate multi-scale time series data provided by the embodiment, which is used to obtain multivariate multi-scale time series data and preprocess the obtained data, including the following steps:

[0067] First, obtain the target variable time series data , such as daily frequency interest rate data;

[0068] Secondly, obtain high- and low-frequency multivariate time series data related to the target variable , including weekly or monthly frequency economic data (such as CPI, GDP, MLF interest rate, etc.) and daily or hourly frequency market transaction data (such as market trading volume, hourly interest rate, market sentiment index, etc.);

[0069] Then, strictly align the obtained multivariate multi-scale time series data according to the time stamp, and take the data in the same time interval. For example, take the sequence data from January 1, 2023 to December 31, 2024, with a time span of two years;

[0070] For low-frequency multivariate time series data (such as weekly or monthly economic data), frequency alignment is performed using linear interpolation or forward filling methods according to the indicator type and data continuity to ensure consistency with the target variable frequency (daily interest rate data); for example, monthly CPI data is converted to daily data through linear interpolation. If the CPI in January is 2.1% and in February is 2.0%, the daily CPI data for each day in January after interpolation is 2.05%; the weekly MLF interest rate is converted to daily data through forward filling. If the MLF interest rate on January 10th is 2.75%, then it is 2.75% from January 11th to January 17th.

[0071] For the missing values in the obtained multivariate multi-scale time series, interpolation and filling are performed using the above linear interpolation or forward filling method.

[0072] The obtained multivariate multi-scale time series data is standardized to avoid dimension differences. Grouped by time period, different time series are standardized separately; the rolling window standardization method is used to calculate the mean and standard deviation based on a sliding window (such as the past 30 days). , where and are the mean and standard deviation within the time window respectively, is the original time series, is the standardized time series.

[0073] Finally, the preprocessed multivariate multi-scale time series data for training and prediction is obtained .

[0074] Figure 3 is the flowchart of high-low frequency component separation provided by the embodiment. The preprocessed target variable time series is decomposed into low-frequency components and high-frequency components using the moving average method and the residual method, and includes the following steps:

[0075] First, based on the target variable time series data , the low-frequency component is calculated using the moving average method. The moving average of N time periods is extracted through the moving average window as the low-frequency component, and the calculation formula is , for example days. The size of the moving average window can be adjusted according to the volatility , where is the volatility in the past 30 days, is the adjustment coefficient. For example , , then the size of the time window is adjusted to days.

[0076] Secondly, based on the target variable time series and low-frequency components The high-frequency components are directly calculated through the residuals .

[0077] In one embodiment, the separated components are verified through the Discrete Fourier Transform (DFT). The Fast Fourier Transform (FFT) is performed to calculate the DFT, and the spectral information f of each component is obtained. The form of the DFT is , where is a discrete-time sequence of length T, is the corresponding frequency-domain representation, is the imaginary unit. Ensure that the spectral energy of the low-frequency components is concentrated in the low-frequency region ( ), and the spectral energy of the high-frequency components should be concentrated in the high-frequency region ( ), where is the cut-off frequency, the demarcation point between the low-frequency and high-frequency components. The energy proportion of the low-frequency components in the low-frequency region (or the energy proportion of the high-frequency components in the high-frequency region) is higher than the threshold (such as 90%), indicating that the separation effect is good, and the final low-frequency components and high-frequency components are obtained.

[0078] Finally, if the energy proportion of the low-frequency components in the low-frequency region is lower than the threshold (such as 90%), it indicates that the separation effect is poor, and the moving average window size needs to be adjusted to re-extract the low-frequency and high-frequency components.

[0079] Figure 4 is the flow chart of the low-frequency component modeling and prediction provided by the embodiment. Based on the separated low-frequency components and the preprocessed low-frequency multivariate time series data, the fOU-LSTM hybrid model is used to model and predict the low-frequency components, generating the target variable low-frequency component prediction time series , including the following steps:

[0080] First, use the fOU process to fit the basic trend of the low-frequency components.

[0081] First, use the fOU process to model the low-frequency components to capture their long-term trends and mean-reverting characteristics. For the low-frequency components , the fOU process equation is used for fitting . is the mean reversion rate, is the long-term mean, is the volatility, is the fractional Brownian motion with a Hurst exponent of ​ Indicates long-term memory.

[0082] Secondly, the wavelet analysis method is used to estimate the Hurst exponent (H), which is the core parameter for measuring long memory. The wavelet basis function Daubechies (dbN) is selected to decompose the low-frequency components to obtain the wavelet coefficients at each scale , where is the scale, is the position, is usually selected as , is the sequence length; calculate the wavelet energy at each scale , is the scale of the number of coefficients; fit the decay relationship of energy with scale, satisfying , is a constant; through and linear regression to obtain , where is the regression coefficient; for example, obtain , indicating that the sequence has significant long memory;

[0083] Then, under the condition of obtaining the H estimate, other parameters are jointly estimated by the maximum likelihood estimation (MLE) method. The fOU process is discretized into a state space model

[0084] , ,

[0085] where, is the time interval, is the noise term, following a normal distribution with a mean of 0 and a variance of .

[0086] Construct the log-likelihood function :

[0087] ,

[0088] where, is the length of the time series, , .

[0089] Use the gradient descent method to maximize the log-likelihood function to obtain the parameter estimate value .

[0090] Finally, based on the analytical solution of the fOU process, predict the benchmark value of the low-frequency component of the future state, given the current state , and the estimated parameter values , solve the discrete form of the fOU model to generate the benchmark value of the low-frequency component of the future state :

[0091] .

[0092] Second, use the LSTM model to predict the residuals of the low-frequency component, capture the non-linear relationship of the low-frequency component, input the historical residual terms of the low-frequency sequence and the low-frequency multivariate time series data, predict the residuals of the low-frequency component, and perform non-linear correction on the low-frequency component.

[0093] First, calculate the fitting error of the fOU model, based on the low-frequency component sequence and the benchmark value of the low-frequency component , calculate the historical low-frequency residual terms ;

[0094] Then, train and predict the low-frequency component residual terms through the LSTM model. Input the historical residual terms of the low-frequency sequence (such as the past 30 days) and the preprocessed low-frequency multivariate time series data , such as macroeconomic indicators (CPI, GDP growth rate), policy interest rates (MLF, reverse repurchase rate), etc., predict the residuals of the low-frequency component (non-linear correction terms) ; Use the root mean square error (RMSE) as the loss function ; Use the Adaptive Moment Estimation (Adam) optimization algorithm as the optimizer, set the learning rate to 0.001, and obtain the low-frequency prediction residual term result .

[0095] Third, fuse the benchmark value and the residual prediction to generate the predicted time series of the low-frequency component of the target variable .

[0096] For example, through the fOU process fitting, the estimated value of the fOU parameter is , , , the actual value of the low-frequency component on a certain day is , the fOU predicted value is , then the residual value is . Calculate the residual sequence of the past 30 days in turn , and input the historical macroeconomic indicator CPI data. After being corrected by LSTM, the residual correction term is obtained; then the final fused low-frequency prediction result is , and the error value is -0.01%.

[0097] Figure 5 This is the flowchart of high-frequency component multi-scale LSTM modeling and prediction provided by the embodiments of the present invention. Combining the multi-time scale LSTM model and the attention mechanism to model high-frequency components, capture their short-term fluctuations and non-linear relationships, and based on the trained multi-scale LSTM model, predict the high-frequency components of the target time series.

[0098] First, prepare multi-scale high-frequency data, including high-frequency components and related high-frequency multivariate time series data . For example, prepare the daily-level target time series data (closing interest rate) of the past week and the high-frequency hourly interest rate and trading volume data of the past 24 hours, and predict the target time series data of the next day to ensure the time alignment of the sequence data, as shown in the following table;

[0099] Prediction target (T+1) Hourly data (T) Daily data (T-6 to T) 2024-06-08 24-hour data on 2024-06-07 Daily data from 2024-06-01 to 2024-06-07 2024-06-09 24-hour data on 2024-06-08 Daily data from 2024-06-02 to 2024-06-08 … … …

[0100] Secondly, combine LSTM and the attention mechanism to use the high-frequency hourly interest rate and trading volume data of the past 24 hours to model and predict the short-term fluctuations of the target time series data of the next day. Input the 24-hour data, extract the time dependence through LSTM, set return_sequences = True, and retain the output of all time steps , is the feature vector at time t, is the hidden state of the previous time step; use the attention mechanism (Attention layer) to dynamically assign the importance of each time step (time step attention), and calculate the attention score , where , is the dimension of the key vector, used to scale the dot product result;

[0101] Aggregate the features of 24 hours into a daily-level feature vector, take the mean of the features of 24 time steps, and generate a daily-level representation (i.e., the high-frequency aggregation feature vector);

[0102] Then, use the LSTM model to model the daily-level target time series data (closing interest rate) of the past week, extract the long-term trend through LSTM , set return_sequences = False, and only output the hidden state of the last time step (i.e., the high-frequency component feature vector);

[0103] Finally, concatenate the high-frequency aggregation feature vector and the high-frequency component feature vector, ; through feature-level attention weighted fusion, calculate the attention weight , , is a trainable parameter for weighted fusion ; map the weighted features to the output through a fully connected layer to generate the predicted time series of the high-frequency component of the target variable, , , is a trainable parameter. The root mean square error (RMSE) is used as the loss function, ; the Adaptive Moment Estimation (Adam) optimization algorithm is used as the optimizer, and the learning rate is set to 0.001.

[0104] Figure 6 This is the flowchart of high-low frequency component fusion and prediction provided by the embodiment of the present invention. The prediction capabilities of the low-frequency and high-frequency models may change over time due to short-term emotions or long-term policies. By adjusting the dynamic weights according to recent errors, the contributions of the low-frequency and high-frequency components are adaptively balanced, and the multi-scale results are fused to predict the target time series data.

[0105] First, calculate the dynamic weights based on historical prediction errors ;

[0106] is the mean absolute error of the high-frequency model in the past time window N, is the mean absolute error of the low-frequency model in the same time window. Taking the prediction result of the low-frequency component as the benchmark, no weight adjustment is performed. The smaller the prediction error of the high-frequency component, the higher the weight;

[0107] is the mean absolute error, , .

[0108] Secondly, the prediction results of the low-frequency and high-frequency components are dynamically weighted and fused as follows:

[0109]

[0110] where, is the predicted value of the low-frequency component, is the predicted value of the high-frequency component, is the final predicted value of the target time series;

[0111] For example, when the window size is , the error of the low-frequency model = 2 bps, and the error of the high-frequency model = 3 bps, then ; the prediction error of the high-frequency model is relatively large, and the influence of the result weight is reduced; the expression of the final predicted value is ; the predicted value of the low-frequency component , the predicted value of the high-frequency component , then the predicted value after final weighted fusion .

[0112] Figure 7 is a schematic structural diagram of a multi-scale time series prediction device provided by an embodiment of the present invention, including a data acquisition and preprocessing unit, a high-low frequency component separation unit, a low-frequency component prediction unit, a high-frequency component prediction unit, and a result fusion and prediction unit. The data acquisition and preprocessing unit is used to: acquire multi-source multi-scale time series data and perform preprocessing; the multi-source multi-scale time series data includes target variable time series data and high-low frequency multi-source time series data related to the target variable;

[0113] The high-low frequency component separation unit is used to: separate the high and low frequency components from the target variable time series data;

[0114] The low-frequency component prediction unit is used to: based on the separated low-frequency component and the preprocessed low-frequency multi-source time series data, use the fOU-LSTM hybrid model to model and predict the low-frequency component, and generate a target variable low-frequency component prediction time series;

[0115] The high-frequency component prediction unit is used to: based on the separated high-frequency component and the preprocessed high-frequency multi-source time series data, use the multi-scale LSTM model and combine the attention mechanism to model and predict the high-frequency component, and generate a target variable high-frequency component prediction time series;

[0116] The result fusion and prediction unit is used to: fuse the target variable low and high frequency component prediction time series by using the dynamic weighting method to obtain the final target variable predicted value.

[0117] It should be noted that the device embodiment shown in this embodiment matches the content of the above method embodiment, and the content of the above method embodiment can be referred to and will not be elaborated here.

[0118] The present invention also provides a computer program product, including a computer program, which when executed by a processor implements the above-mentioned multi-scale time series prediction method.

[0119] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0120] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing device produce means for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0121] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including instruction means that implement the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0122] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0123] The above embodiments are only used to illustrate the design concept and features of the present invention, and the purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made according to the principles and design concepts disclosed by the present invention are within the protection scope of the present invention.

Claims

1. A multi-scale time series prediction method, characterized in that, Including: Obtain multivariate multi-scale time series data and perform preprocessing; the multivariate multi-scale time series data includes target variable time series data and high- and low-frequency multivariate time series data related to the target variable; Separate the high- and low-frequency components of the target variable time series data; Based on the separated low-frequency component and the preprocessed low-frequency multivariate time series data, use the fOU-LSTM hybrid model to model and predict the low-frequency component, and generate the target variable low-frequency component prediction time series; Based on the separated high-frequency component and the preprocessed high-frequency multivariate time series data, use the multi-scale LSTM model and combine the attention mechanism to model and predict the high-frequency component, and generate the target variable high-frequency component prediction time series; Based on the target variable low-frequency and high-frequency component prediction time series, use the dynamic weighting method for fusion to obtain the final target variable prediction value.

2. The method according to claim 1, wherein Preprocess the multivariate multi-scale time series data, including: Perform time alignment, frequency alignment, data missing value processing, and data standardization processing on the multivariate multi-scale time series data.

3. The method according to claim 1, wherein Separate the high- and low-frequency components of the target variable time series data, including: Based on the target variable time series data, use the moving average method to calculate the low-frequency component, and the moving average time window size is adjusted according to the volatility. The window size formula is: Among them, is the time window, is the volatility in the past several days, is the adjustment coefficient; Based on the target variable time series and the low-frequency component, directly calculate the high-frequency component through the residual.

4. The method according to claim 3, wherein Verify the effectiveness of the separated components using the Fourier transform method; if the energy ratio of the low-frequency component in the low-frequency region is higher than the preset threshold and / or the energy ratio of the high-frequency component in the high-frequency region is higher than the preset threshold, it proves that the separation effect is good, otherwise adjust the moving average time window to re-extract the low-frequency and high-frequency components.

5. The method according to claim 1, wherein Based on the separated low-frequency component and the preprocessed low-frequency multivariate time series data, use the fOU-LSTM hybrid model to model and predict the low-frequency component, and generate the target variable low-frequency component prediction time series, including: Use the fractional-order Ornstein-Uhlenbeck process to fit the low-frequency component, capture its long-term trend and mean reversion characteristics, and generate the benchmark value of the low-frequency trend; Use the LSTM model to predict the residual of the low-frequency component, capture the non-linear relationship of the low-frequency component, input the historical residual terms of the low-frequency sequence and the preprocessed low-frequency multivariate time series data, predict the residual of the low-frequency component, and perform non-linear correction on the low-frequency component; Fuse the benchmark value and the residual prediction to generate the target variable low-frequency component prediction time series.

6. The method according to claim 5, wherein Use the fOU model to fit the low-frequency component, including: The fOU process is used to fit the low-frequency component, and the long memory is modeled by introducing fractional Brownian motion, whose differential form is , where is the low-frequency component of the time series, is the mean reversion rate, is the long-term mean, is the volatility, is the fractional Brownian motion with Hurst exponent ; Estimation using wavelet analysis ; Under the condition of obtaining the estimated value, other parameters are estimated by the maximum likelihood estimation method ; Benchmark value for predicting low-frequency trends based on the analytical solution of the fOU process .

7. The method according to claim 1, wherein Based on the separated high-frequency component and the preprocessed high-frequency multivariate time series data, use the multi-scale LSTM model and combine the attention mechanism to model and predict the high-frequency component, and generate the target variable high-frequency component prediction time series, including: Use the high-frequency multivariate time series data, extract the time dependence through the LSTM model, and retain the outputs of all time steps; Use the attention mechanism to dynamically assign the importance of each time step; Take the mean of the features of each time step to generate the high-frequency aggregation feature vector; Use high-frequency components to extract long-term trends through LSTM, only output the hidden state of the last time step, and generate high-frequency component feature vectors; Concatenate the high-frequency aggregated feature vectors and the high-frequency component feature vectors, perform weighted fusion through feature-level attention, and use a fully connected layer to map the weighted features to the output to generate the predicted time series of the high-frequency components of the target variable.

8. The method according to claim 1, wherein Based on the predicted time series of the low-frequency components of the target variable and the predicted time series of the high-frequency components of the target variable, use the dynamic weighting method for fusion to obtain the final predicted value of the target variable, including: Adopt dynamic weight allocation, and the weight values are calculated based on the historical prediction errors of the low-frequency and high-frequency components. The fusion form is: Among them, is the dynamic weight, is the mean absolute error of the high-frequency model in the past time window N, is the mean absolute error of the low-frequency model in the same time window.

9. A multi-scale time series prediction device, characterized in that, It includes a data acquisition and preprocessing unit, a high-low frequency component separation unit, a low-frequency component prediction unit, a high-frequency component prediction unit, and a result fusion and prediction unit; The data acquisition and preprocessing unit is used to: acquire multi-source multi-scale time series data and perform preprocessing; the multi-source multi-scale time series data includes the time series data of the target variable and the high-low frequency multi-source time series data related to the target variable; The high-low frequency component separation unit is used to: separate the high and low frequency components of the target variable time series data; The low-frequency component prediction unit is used to: based on the separated low-frequency components and the preprocessed low-frequency multi-source time series data, use the fOU-LSTM hybrid model to model and predict the low-frequency components, and generate the predicted time series of the low-frequency components of the target variable; The high-frequency component prediction unit is used to: based on the separated high-frequency components and the preprocessed high-frequency multi-source time series data, use the multi-scale LSTM model and combine the attention mechanism to model and predict the high-frequency components, and generate the predicted time series of the high-frequency components of the target variable; The result fusion and prediction unit is used to: perform fusion on the predicted time series of the low-frequency and high-frequency components of the target variable by using the dynamic weighting method to obtain the final predicted value of the target variable.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the multi-scale time series prediction method according to any one of claims 1-8.

Images