PMDformer Improves Long-Term Time Series Forecasting Accuracy.

Long-term time series forecasting is vital across various sectors, from energy management to finance. While Transformer-based models have adopted patch-based strategies to capture dependencies, accurately modeling shape similarities across different scales and variables remains a significant hurdle. This research introduces PMDformer, a new model designed to overcome these challenges. PMDformer employs a technique called patch-mean decoupling (PMD), which separates the trend and residual shape information within each data patch by subtracting its mean. This process preserves the original data structure, allowing the attention mechanism to more accurately capture true shape similarities without being distorted by scale differences. Additionally, the model incorporates two novel attention mechanisms: Trend Restoration Attention (TRA), which reintegrates the decoupled trend during attention calculation, and Proximal Variable Attention (PVA), which focuses cross-variable attention on the most relevant recent time segments. Extensive experiments demonstrate that PMDformer surpasses existing state-of-the-art methods in both stability and accuracy across multiple long-term time series forecasting benchmarks.