A robust ensemble learning model for short-term wind power forecasting using multi-height wind features

Accurate 15-minute-ahead wind-power forecasting is important for grid operation and renewable-energy integration. This study presents a systematic empirical evaluation of multi-height meteorological, physics-informed, and temporal variables for short-term wind-power forecasting. The dataset contains 35,040 observations recorded at 15-minute intervals over one year, including wind speeds measured at 10, 30, 50, and 70 m and at hub height, together with temperature, air pressure, humidity, and actual power generation. Wind shear, cubic hub-height wind speed, three power lags, four-step rolling statistics, and calendar variables were derived from these measurements. A three-hour historical window was used to predict power generation 15 min ahead. Extra Trees, LightGBM, XGBoost, CatBoost, and a voting ensemble were evaluated using a chronological 80:20 train–test division. Model performance was assessed using MAE, RMSE, sMAPE, and R², together with residual and non-parametric statistical analyses. An ablation analysis using a fixed LightGBM configuration demonstrated that multi-height measurements reduced MAE by 13.05% compared with hub-height and atmospheric inputs, while temporal features produced the largest improvement, reducing MAE by 61.36% relative to the preceding configuration. Physics-informed variables provided a modest, statistically non-significant improvement ( \\(\\:p=0.0587\\) ). PCA reduced the feature dimension from 228 to 50 components but degraded forecasting accuracy. The best ablation configuration, which retained multi-height, atmospheric, physics-informed and temporal variables without PCA, achieved an MAE of 5.9389 MW, RMSE of 10.0148 MW and an R² of 0.9634. These findings provide empirical evidence that multi-height and temporal information contribute to short-term wind-power forecasting, while dimensionality reduction does not necessarily improve predictive accuracy.