TA-SparseMG Improves Long-Term Time Series Forecasting

A new research paper introduces TA-SparseMG, a lightweight and efficient model designed for long-term time series forecasting. This model addresses several inherent challenges in predicting dynamically varying long-term data, including statistical nonstationarity, localized high-frequency noise, and complex cross-period dependencies. Traditional lightweight models often struggle to balance parameter efficiency with robust forecasting performance in such scenarios. TA-SparseMG integrates three key modules to overcome these hurdles. First, a trend-aware reversible instance normalization module captures input-window statistics and calibrates forecast-window distributions, effectively mitigating distribution shifts. Second, a scale-adaptive gated denoising module smooths features and suppresses residuals before period rearrangement, reducing interference from high-frequency perturbations. Finally, a multiscale gated-attention MLP forecasting module enhances the prediction head's adaptive representational capacity through conditional gating and feature modulation. Extensive experiments on multiple long-term time series forecasting benchmarks demonstrate that TA-SparseMG consistently achieves superior and stable performance, with ablation studies confirming the independent contribution of each module to its overall effectiveness.