Veliformer: A periodicity-preserving model for short-term tidal energy forecasting in optimal power flow

With the increasing integration of renewable energy, tidal energy stands out for its high predictability, making it a valuable asset for stable power grid operation. However, accurate forecasting remains a critical challenge. Conventional deep learning models, despite their success in general time-series analysis, often struggle to preserve the inherent periodic features of tidal data, leading to reduced prediction accuracy and suboptimal grid scheduling. To address this gap, we propose Veliformer, a novel periodicity-preserving forecasting model. At its core, Veliformer introduces an innovative mask modeling technique. Unlike conventional methods that predict masked data points, our approach reconstructs the complete original sequence by learning to aggregate information from multiple, differently masked versions of the series. This unique reconstruction process is specifically designed to maintain the integrity of the underlying periodic structure of tidal energy, enabling the model to accurately capture both deterministic cycles and stochastic fluctuations. When applied to the optimal power flow (OPF) of tidal energy systems, Veliformer reduces power generation costs. Our theoretical analysis shows that the model preserves periodicity through masked sequence reconstruction. Numerical experiments demonstrate Veliformer’s superior performance in optimizing power systems and reducing prediction errors compared to other popular models. The mask modeling mechanism enhances Veliformer’s prediction accuracy by an average of 4.91%, further highlighting its effectiveness in handling tidal energy forecasting. • A novel model, Veliformer, preserves periodicity in tidal energy forecasting. • A unique masking method reconstructs the original series to retain periodic features. • Improves forecasting accuracy by an average of 4.91% over baseline models. • Reduces power generation costs in Optimal Power Flow (OPF) applications. • Theoretical analysis confirms the model’s periodicity-preserving capability.