WT-PowerCurve-Open: an open wind-turbine power-curve dataset and reproducible pipeline for renewable-generation modelling

A reproducible 10-minute dataset for modelling a wind turbine's power curve (meteorological conditions to active power), built from the openly licensed Kelmarsh wind-farm SCADA archive (Cubico Sustainable Investments, Zenodo 10.5281/zenodo.8252025, CC-BY 4.0). It ships two aligned versions of the regression task: base_v1 (naive: wind speed only, 249,898 rows) and improved_v2 (engineered meteorological features + forced-outage cleaning, 183,197 rows, 8 features), plus the pipeline scripts and a bilingual scientific description. A model family (Linear, KNN, Decision Tree, GBM, MLP) is trained on each; engineering and cleaning reduce the error of every model, the best (GBM) improving from MAE 56.77 kW / R2 0.962 to MAE 40.80 kW / R2 0.986 (about 2% of rated power). The improved set uses meteorological features only: operational-state channels strongly correlated with the target (e.g. generator RPM, corr about 0.93 with power) are deliberately excluded, so this is a genuine meteorology-to-power curve with no target leakage; the high R2 is expected physics. This deposit supplements the author's KMOSS-2025 thesis on neural-network forecasting of electricity generation/consumption indicators (Shapovalova & Titov, 2025).