Enhanced Wind Turbine Power Forecasting via Hyperparameter-Optimized XGBoost

Dimas Ramadhani (Universitas Negeri Semarang)
Yahya Nur Ifriza (Universitas Negeri Semarang)

Information Technology Education Journal
Vol. 5, No. 2, May (2026)

Purpose – This study aims to evaluate the accuracy and computational efficiency of XGBoost in forecasting wind turbine output using a dataset aligned at hourly timestamps. This topic is important because wind turbine output exhibits fluctuating and non-linear patterns, requiring a model capable of capturing the relationship between meteorological conditions, historical turbine patterns, and the generated Energy values. Design – This study uses the 2016 Sotavento Galicia data, which consist of hourly Numerical Weather Prediction (NWP) data and historical turbine operational data originally recorded every 10 minutes. Temporal alignment was performed by retaining only turbine operational records located exactly at hourly timestamps and then merging them with the NWP data at the same timestamps. The final dataset was modeled as an hourly aligned time series dataset. The Energy variable was used as the prediction target. Since the dataset does not explicitly state the unit of Energy, RMSE and MAE were reported in the original scale of the Energy variable and cautiously interpreted as kWh per retained 10-minute record based on the variable label, recording resolution, and value range. Three model scenarios were compared, namely the XGBoost baseline, XGBoost with GridSearchCV, and XGBoost with RandomizedSearchCV. Internal validation was performed using TimeSeriesSplit, while final testing was conducted using monthly holdout on months 10, 11, and 12. Findings – The results show that XGBoost with RandomizedSearchCV produced the lowest average prediction error, with an RMSE of 135.591, MAE of 87.710, and R² of 0.907. This model reduced RMSE by 5.86% compared to the XGBoost baseline and reduced computation time by 69.51% compared to GridSearchCV. Research implications – These findings are limited to a single wind farm dataset, one observation period, and a constrained hyperparameter search space. Originality – This study demonstrates that RandomizedSearchCV can serve as an efficient tuning strategy for XGBoost-based wind power forecasting.