DAEM: A Decomposed Attention-Enhanced Mamba Model for Multivariate Time Series Forecasting

Wen Li, Jiaxin Yuan, Heming Du, Jiyanglin Li, Shouguo Du, Yiming Tang

Abstract: Multivariate time series forecasting (MTSF) is a critical task with applications spanning diverse domains. Effective MTSF requires capturing both inter-series dependencies, which represent the intricate relationships among multiple time series, and intra-series dynamics, which reflect the temporal patterns and variations within individual series. However, the complexity of real-world data, marked by nonlinearity, noise, and dynamic fluctuations, presents significant challenges. In this paper, we propose the Decomposed Attention-Enhanced Mamba Model (DAEM), which combines a learnable decomposition strategy with an attention-based module and the Mamba architecture to effectively capture dynamic trends, seasonal patterns, and the intricate relationships inherent in multivariate time series. By addressing both inter-series dependencies and intra-series dynamics, this innovative design enables DAEM to achieve more accurate and robust time series forecasting. To evaluate the effectiveness of DAEM, we conducted extensive experiments on eight widely used benchmark datasets. The results demonstrate the superiority of our approach, achieving 35 first-place rankings for MSE and 28 first-place rankings for MAE.

External IDs:dblp:conf/ijcnn/LiYDLDT25