Filterformer : enhancing time series forecasting through filter

Abstract

Engineering time series signals exhibit complex spectral compositions, comprising both informative temporal patterns and non-informative high-frequency components. The high-frequency noise can distort intrinsic dependencies and undermine long-term forecasting reliability. To enhance the robustness of transformer-based forecasting under noisy conditions, we propose Filterformer, a unified framework that effectively suppresses high-frequency noise while preserving essential temporal characteristics. It integrates a padding-based sliding average filter for frequency-domain noise suppression, an SNR block for denoising control, and channel-independent patch encoding for efficient temporal representation. This design effectively suppresses non-informative high-frequency components while preserving meaningful temporal dynamics. Extensive experiments on five real-world datasets demonstrate that Filterformer consistently outperforms state-of-the-art baselines in long-term forecasting and maintains stable performance across varying signal-to-noise ratios. Moreover, its successful application to a newly collected metro subgrade-settlement dataset with 10-minute sampling intervals confirms its potential for reliable long-term monitoring and structural health assessment in real-world engineering systems.