.png){kind=icon}
.png){kind=icon}
Return
MAFNet: Multi-scale active fusion network for long-term time series forecasting
.png){kind=icon}
PRE
AI
Q
X
S
J
DOI：10.1016/j.neucom.2026.134709.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Long-term time series forecasting is a challenging task because real-world series often contain non-stationary dynamics, slowly evolving trends, seasonal patterns, and localized short-term residual details. A major limitation of many existing architectures is that they either model the whole sequence with a monolithic backbone or oversimplify residual/high-frequency components after decomposition. To address this issue, we propose MAFNet, a hierarchical forecasting framework that adapts the symmetric U-Net architecture to sequence-to-sequence forecasting. The encoder recursively decomposes the input into progressively smoother trend representations and residual detail components. MAFNet further redefines conventional skip connections as active forecasting pathways: a lightweight Multi-Layer Perceptron (MLP) extrapolates the deepest latent trend, while scale-specific Temporal Convolutional Networks (TCNs) estimate the future evolution of residual details. The symmetric decoder then reconstructs the final prediction in a coarse-to-fine manner through a trend-aware gating mechanism that adaptively fuses the extrapolated trend and the predicted residual dynamics. Experiments on eight widely used benchmark datasets show that MAFNet achieves competitive performance against recent strong baselines, including the 2025 WPMixer, while maintaining a favorable accuracy–efficiency trade-off. Ablation studies further validate the contributions of the hierarchical structure, latent trend extrapolation, active residual forecasting, and gated fusion. The code is available at https://github.com/hit636/MAFNet .
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W
Organization
Cited Papers