Wind-power prediction is reasonably accurate in normal weather but degrades during high winds, gusts and rapid changing conditions, and a model fitted to one turbine does not directly transfer to turbines of other capacities. This paper presents an extreme-weather-aware, capacity-normalized framework that converts 10-minute SCADA and meteorological data into turbine power with Random Forest regressor. The model predicts power as a fraction of rated capacity, uses features such as air density, density-corrected wind speed, gust factor, rapid wind changes and one-hour variability, labels statistically extreme conditions with thresholds learned from training data, and gives extreme observations five times the training weight. The framework is demonstrated on a dataset of 11,520 records from site (coastal, flat inland, hilly and plateau) with 2.0–5.0 MW turbines, using a per-site chronological 80/20 split. On the unseen test period the model reached an NMAE of 2.01%, an NRMSE of 2.55% and an R² of 0.992, with an aggregate energy error of −0.06% (414.42 MWh predicted versus 414.68 MWh actual). Errors rose under extreme conditions (NMAE 2.25% versus 1.97% in normal conditions), confirming that extreme periods should be reported separately. Sensitivity test showed that the extreme weighting did not reduce extreme-condition error on this dataset.

Amisha Nandansingh Negi, Dr. Manoj Singh, "Inaccurate Wind Power Forecasting During Extreme Weather Events Using Normalized Machine Learning" International Journal of Scientific Research and Engineering Development, V9(5): Page(1286-1291) September - October 2026. ISSN: 2581-7175. www.ijsred.com. Published by Scientific and Academic Research Publishing.