ABSTRACT The effectiveness of Transformers in time series forecasting largely depends on their ability to extract and utilise key features. The model not only needs to capture temporal patterns but also focus its attention on the most informative inputs, thereby constructing sharp and discriminative attention distributions to highlight critical information and improve forecasting accuracy. However, despite their strengths in modelling global dependencies, conventional Transformers remain limited in capturing predictive key patterns and forming highly distinctive attention weights. To address this limitation, we propose FIRFormer, a framework that enhances input representation and reconstructs the attention mechanism to better capture important temporal dynamics. Specifically, we introduce the Dynamic Frequency Selector module to adaptively filter frequency components most relevant to the forecasting objective, reducing redundancy and improving input efficiency. Subsequently, we introduce the Biased Integration Window module, which integrates a temporal decay mechanism into sliding aggregation to strengthen the correlations within local contexts, achieves precise fusion of temporal dynamic patterns and further emphasises the modelling value of recent critical information. Finally, we propose the Attention Recalibration module, which employs a dual‐branch attention refinement strategy. It enhances attention selectivity from both the local feature perspective and the global subspace perspective, promoting sharper and more focused attention distributions. FIRFormer enhances the perception of critical dynamic patterns and achieves performance improvements across multiple real‐world forecasting tasks.
Chen et al. (Sun,) studied this question.