WindCast: LightGBM Framework with Engineered Time-Series Features for Multi-Step Wind Turbine Forecasting

Abstract

The global expansion of wind power, with 117 GW of new capacity added in 2024 elevating the cumulative total to 1,136 GW, necessitates precise multi-step ahead forecasting of turbine output to mitigate intermittency and optimize grid operations. This paper advances a time series forecasting methodology for predicting wind turbine active power 10 to 60 minutes into the future. The framework is anchored on LightGBM (Light Gradient Boosting Machine) and rigorously enhanced with temporal feature engineering, including lagged sequences, rolling statistics, and STL decomposition, to model autocorrelation, seasonality, and non-stationarity inherent in wind data. Evaluated on a SCADA dataset from a Turkish wind turbine (50,530 records at 10minute intervals), the LightGBM model attains a peak performance of R² = 0.9885 and RMSE = 0.14 kW, eclipsing benchmark models such as CatBoost (R² 0.9838, RMSE 0.17), XGBoost, and Random Forest (R² 0.9750, RMSE 0.21), and demonstrating up to a 52% reduction in multi-step prediction error relative to a baseline Gradient Boosting model. The model exhibits enhanced fidelity at critical operational thresholds, such as cut-in wind speeds (3-4 m/s) and during directional shifts. This study confirms that integrating LightGBM with a structured time series approach delivers an efficient, interpretable, and highly accurate solution for short-term future power prediction, offering significant value for real-time renewable energy integration and grid stability.