Machine learning benchmark for short-term operational wind speed forecasting: A multi-regime performance analysis at Ramakkalmedu, Kerala
Proceedings of the Institution of Mechanical Engineers, Part A: Journal of Power and Energy · 2026-11 · 240 (7) · 1415-1450
DOI: 10.1177/09576509261444820 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the Institution of Mechanical Engineers, Part A: Journal of Power and Energy, online erschienen am 25.05.2026. Verfasst von V V Muhammed Anees und 3 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 89 Literaturangaben.
Abstract (Originalsprache)
Precise short-term wind prediction is essential for incorporate increasing wind energy in Kerala grid operated by Kerala State Electricity Board (KSEB). The Ramakkalmedu wind farm (1100 m elevation in the Western Ghats) is a strategically important site were inherent wind variability challenges grid stability. This study compares five machine learning models Bayesian Ridge, AdaBoost, XGBoost, LightGBM, and LSTM against three baselines: Persistence, Climatology, and ARIMA for 1-h ahead wind speed forecasting. Models were trained using 35,064 hours of NASA POWER data (2021–2024). ML models were optimized via 5-fold expanding-window cross-validation, while final performance for all models was evaluated on a strict 20% temporal holdout test set, ensuring a true 1-h ahead forecasting. The evaluation used seven performance metrics RMSE, Theil’s U, MASE, MAE, MAPE, R 2 , and forecast bias alongside training and inference time. LightGBM was the best, getting RMSE 0.2261 m/s, MAE 0.1592 m/s, MAPE 6.06%, R 2 0.9872, and Theil’s U 0.6169. This performance represents a 38.3% improvement in RMSE over the Persistence (0.3665), 21.1% over ARIMA (0.2864) with 100,000× faster inference (0.228 s vs 26,400 s), and 88.7% over Climatology (2.0133), making it uniquely viable for real-time deployment. This accuracy enables ₹9−31 million annual savings through reduced spinning reserves (38%), lower imbalance penalties, and optimized market bidding. XGBoost delivered comparable accuracy (RMSE: 0.228) while LSTM underperformed dramatically (0.467) despite 185× longer training. Failure mode analysis showed that all the models failed during ramp events, extreme winds, directional changes and morning transitions, though LightGBM exhibited greatest stability. LightGBM is recommended for KSEB deployment based on proven accuracy, efficiency, and robustness, with projected ₹9−31 million annual savings through enhanced grid management and reduced operational costs.
Zitieren
APA
Anees, V. V. M., Nazar, K. P. A., R, S., & Maniyath, S. (2026). Machine learning benchmark for short-term operational wind speed forecasting: A multi-regime performance analysis at Ramakkalmedu, Kerala. Proceedings of the Institution of Mechanical Engineers, Part A: Journal of Power and Energy, 240(7), 1415-1450. https://doi.org/10.1177/09576509261444820
Vancouver
Anees VVM, Nazar KPA, R S, Maniyath S. Machine learning benchmark for short-term operational wind speed forecasting: A multi-regime performance analysis at Ramakkalmedu, Kerala. Proceedings of the Institution of Mechanical Engineers, Part A: Journal of Power and Energy. 2026;240(7):1415-1450. doi:10.1177/09576509261444820
BibTeX
@article{Anees2026,
author = {Anees, V V Muhammed and Nazar, K P Abdul and R, Sridharan and Maniyath, Shijin},
title = {Machine learning benchmark for short-term operational wind speed forecasting: A multi-regime performance analysis at Ramakkalmedu, Kerala},
journal = {Proceedings of the Institution of Mechanical Engineers, Part A: Journal of Power and Energy},
year = {2026},
volume = {240},
number = {7},
pages = {1415--1450},
doi = {10.1177/09576509261444820}
}
EUVIMED