Enhanced Transformer Architecture for Multistep Wind Power Forecasting at Sub-hourly Resolution

Accurate sub-hourly wind power forecasting is essential for ensuring the stable operation of renewable energy grids. However, high-frequency components in fine-grained temporal data, combined with measurement noise and instantaneous fluctuations, pose significant challenges for ultra-short-term prediction. This paper proposes Filter-LSTM-iTransformer (FILT), a novel architecture that integrates an improved iTransformer to capture global temporal dependencies and long short-term memory (LSTM) networks to extract features from exogenous variables. A specially designed Filter module is incorporated to suppress high-frequency noise while preserving critical signal information, thereby enhancing the robustness of the model under high-frequency conditions. Evaluated on a real-world wind farm dataset with an installed capacity of 36 MW, FILT demonstrates superior multi-step forecasting performance at a 15-min resolution over a 10-day horizon, enabling more reliable and efficient utilization of wind energy.