Long-term time series forecasting research has primarily relied on Transformer and MLP models, but the potential of convolutional networks has not been fully explored. To address this issue, this paper proposes a novel multi-scale time series reconstruction module that effectively captures cross-period patch interactions and variable dependencies. Based on this, we develop MS-DFTVNet, a multi-scale 3D deformable convolutional framework tailored for long-term forecasting. Furthermore, to address the inherently uneven distribution of temporal features, we introduce a context-aware dynamic deformable convolutional mechanism to better capture complex temporal patterns. Through extensive experiments, we demonstrate that MS-DFTVNet significantly outperforms powerful baseline models, achieving an average performance improvement of 7.5% on six public datasets, setting a new state-of-the-art result.
