Four-hour ahead probabilistic wind power forecasting based on NWP correction

Reliable short-Term wind power forecasting is essential for secure and economical system operation. This study presents a residual-centric framework that first corrects numerical weather prediction outputs with an attention-based sequence-Tosequence LSTM and then produces probabilistic forecasts via joint kernel density estimation. Data are aggregated to a 15-minute resolution using sliding-window averaging, and multi-step predictions are generated over a four-hour horizon. The attention mechanism highlights horizon-relevant context and improves tracking of ramps, while the nonparametric density stage yields calibrated prediction intervals suitable for risk-Aware dispatch. Experiments against strong baselines show consistent reductions in point error across all lead times and improved stability under volatile inflow. Interval evaluation at common nominal levels demonstrates close adherence to coverage targets with narrow widths and limited degradation at longer horizons. Case analyses illustrate that corrected means align with rapid changes and that the derived intervals capture fluctuations without excessive conservatism. The integrated design therefore enhances deterministic accuracy and delivers reliable uncertainty information that can be consumed directly by scheduling and trading workflows.