P2D-former for time-series forecasting with period-aware 2D representations and exogenous covariates

Time series forecasting finds extensive application across numerous real-world scenarios, yet faces challenges in modelling multi-period variability and exogenous variables. To address this, this paper proposes P2D-Former, a patch-based Transformer model integrating period-aware 2D structures with exogenous information fusion. This approach incorporates MSUnit and P2D-FFN to model complex periodic structures, while explicitly integrating exogenous variable information through an Intersect-Attention mechanism. Experiments demonstrate that P2D-Former achieves superior performance across multiple benchmark datasets for long-term forecasting, short-term forecasting, and high-missing-value imputation tasks, exhibiting significant improvements in both accuracy and robustness compared to existing approaches.