PESD-TSF: Physics-Inspired Framework for Long-Term Time Series Forecasting

A new deep learning framework, PESD-TSF, addresses long-term time series forecasting by tackling periodic perception attenuation and entangled trend-noise representations. The framework introduces a Multiplicative Periodic Gating mechanism with continuous-time priors to preserve periodic structures across deep layers. It also employs a multi-scale structured encoder with detrended attention and hierarchical sampling to decouple long-term trends from high-frequency components. Additionally, PESD-TSF overcomes the limitations of channel-independent paradigms by modeling cross-variable consistency in multivariate time series. The approach emphasizes both interpretability and predictive accuracy, offering a physics-inspired structured decomposition for forecasting.