Many real-world forecasting tasks involve heterogeneous temporal sources, where high-frequency dynamic observations coexist with low-frequency semantic or contextual information. Existing methods often treat these inputs through simple broadcasting or direct concatenation, which weakens cross-scale dependency modeling and makes the prediction process vulnerable to unstable contextual correlations. To address this issue, this paper proposes a structure-guided dual-timescale learning framework for heterogeneous short-horizon time-series forecasting. A fast-scale temporal encoder is used to extract fine-grained dynamic patterns, while a slow-scale semantic encoder is introduced to characterize slowly varying contextual information. On this basis, a structure-guided semantic refinement module is designed to suppress unstable semantic components before fusion, and a structure-aware cross-scale attention mechanism is developed to adaptively align fast dynamics with the most relevant slow-varying context. In addition, a coarse-to-fine prediction structure is employed to separate background tendency modeling from short-term fluctuation correction, and intervention-based consistency regularization is incorporated to improve robustness under changing conditions. Feature contribution analysis further confirms the importance of selecting effective contextual variables to form a compact slow-timescale representation. Experiments on a real-world ultra-short-term aggregate load forecasting task demonstrate that the proposed framework achieves superior accuracy and robustness, indicating its potential as a general solution for forecasting problems with heterogeneous temporal resolutions.
Peng et al. (Wed,) studied this question.