DCFA-iTimeNet: Dynamic cross-fusion attention network for interpretable time series prediction

Abstract

Although time series prediction research among engineering and technology has made breakthrough progress in performance, challenges remain in modeling complex dynamic interactions between variables and interpretability. To address these two problems, a novel two-stage strategy framework called DCFA-iTimeNet is introduced. In the first stage, this paper innovatively proposes a dynamic cross-fusion attention mechanism (DCFA) . This module facilitates the model to exchange information between different patches of the time series, thereby capturing the complex interactions between variables across time. In the second stage, we exploit a decomposition-based linear explainable Bidirectional Gated Recurrent Unit (DeLEBiGRU), which consists mainly of standard BiGRU and tensorized BiGRU. It is proposed to analyze each variable’s historical long-term, instantaneous, and future impacts. Such design is crucial for understanding how each variable impacts the overall prediction over time. Extensive experimental results demonstrate that the proposed model can effectively model and interpret complex dynamic relationships of multivariate time series and understand the model’s decision-making process. Moreover, the performance outperforms the state-of-the-art methods.