ABSTRACT Accurately modeling spatio‐temporal dependencies of multivariate time series (MTS) is crucial for MTS forecasting. Existing graph neural network‐based methods are limited to capturing either static or sample‐level dynamic spatial dependencies and lack the ability to model fine‐grained dynamic dependencies between variables. Furthermore, frequency domain approaches predominantly extract temporal features based on salient periods and neglect the distinct contributions of trend, seasonal, and high‐frequency detailed temporal patterns that correspond to different frequency bands to forecasting. To address these limitations, this paper proposes ADyTNet, which learns time‐aware dynamic graphs and tri‐band temporal pattern for MTS forecasting. The model consists of two branches: (i) In the temporal domain branch, MTS is partitioned into a series of patches and the transformer encoder is utilized to capture inter‐patch temporal dependencies. ADyTNet perceives the local contextual information of neighboring patches via 1D convolution while incorporating learnable node embeddings to learn time‐aware dynamic graphs, effectively capturing dynamic inter‐variable dependencies at the patch level; (ii) In the frequency domain branch, ADyTNet decomposes the MTS spectrum into low‐, mid‐, and high‐frequency components via a masking strategy. The three frequency components are element‐wise multiplied by learnable complex weights to derive fused spectral features. The Inverse Fourier Transform is applied to the fused spectral features to obtain the tri‐band temporal pattern features that reflect trends, seasonality, and high‐frequency details. Finally, the predicted values are obtained by adaptively fusing features of patch‐level dynamic spatio‐temporal dependencies and tri‐band temporal pattern features. Extensive experiments on eight real‐world datasets demonstrate that ADyTNet outperforms state‐of‐the‐art baselines in MTS forecasting.
Huo et al. (2026) studied this question.