A multi-element time series prediction method and device based on reverse attention and dynamic frequency domain modulation

CN123031179APending Publication Date: 2026-11-03UNIV OF ELECTRONICS SCI & TECH OF CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611329584.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-31
Publication Date
2026-11-03

AI Technical Summary

Technical Problem

Existing multivariate time series prediction methods suffer from low-pass bias in the graph domain of self-attention weighted aggregation, which leads to excessive attenuation of high-frequency information. Furthermore, the fixed frequency domain weights lack sample adaptability, resulting in insufficient accuracy in non-stationary multivariate time series prediction.

Method used

A method based on inverted attention and dynamic frequency domain modulation is adopted. By sharing feature mapping, inverted attention high- and low-frequency decomposition, independent gating fusion, dynamic spectrum modulation and residual fusion, dynamic spectrum modulation weights are generated to recover high-frequency information and adapt to multi-scale non-stationary changes of different samples, variables and time segments.

Benefits of technology

While retaining the self-attention long-range modeling capability, it recovers smoothed high-frequency information, improves the accuracy and stability of multivariate time series prediction, and adapts to multi-scale non-stationary changes of different samples, variables and time segments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123031179A_ABST
    Figure CN123031179A_ABST
Patent Text Reader

Abstract

The application relates to the field of artificial intelligence and time series data processing, and discloses a multi-element time series prediction method and device based on reverse attention and dynamic frequency domain modulation. The method is aimed at the problems of low-pass bias in the self-attention graph domain, lack of sample adaptability of fixed frequency domain weight, excessive attenuation of high-frequency information, and insufficient prediction accuracy of non-stationary sequences. The scheme comprises the following steps: obtaining multi-element time series and mapping the multi-element time series to a feature space through reversible instance normalization; using a normalized attention matrix and the difference between the normalized attention matrix and a unit matrix to respectively aggregate low-frequency and high-frequency features; performing weighted fusion through an independent gate branch; performing discrete Fourier transform along the time dimension to obtain a complex frequency spectrum; generating input conditioning coefficients according to a global context to combine and deform a learnable frequency spectrum prototype, so as to obtain dynamic frequency spectrum modulation weight; synchronously modulating the real part and the imaginary part of the frequency spectrum by using the weight, performing inverse transform and residual fusion, and outputting a predicted sequence; and finally performing reverse normalization to output a future time step prediction result.
Need to check novelty before this filing date? Find Prior Art