Answers are grounded in this paper and cite their sources.
Multivariate time series forecasting requires modeling both variable-specific temporal patterns and inter-variable relations, including dependencies shared across samples and interactions that vary with the input. In this paper, we propose a Hierarchical Temporal-Relation Refinement Network (HTRNet) for multivariate time series forecasting. HTRNet sequentially learns variable-specific latent representations and models inter-variable relations from the resulting residuals within each refinement layer. The framework supports both single-layer and multi-layer configurations; when multiple layers are used, the updated residual representations are propagated to subsequent layers for further refinement. In addition, a Learnable Static Relation Matrix (LSRM) learns sample-independent relation weights among variables, while a Dynamic Compensation Module (DCM) uses sample-specific inter-variable interactions to enhance the final residual representation and generate a compensation prediction. The final forecast combines the layer-wise predictions with the DCM output. Extensive experiments are conducted on widely used benchmark datasets, including ETT, Weather, Electricity, Traffic, and Solar-Energy, across multiple forecasting horizons. Averaging the relative reductions against the best baseline for each dataset and metric yields improvements of 2.01% in MSE and 4.83% in MAE. These gains vary across datasets and forecasting horizons. On Traffic, HTRNet achieves the lowest average MAE but has a higher average MSE than DeepBooTS, iTransformer, and CASA.
No takes yet. Share an insight, caveat, or question.
Lv et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: