Scholarlink

<p>Short-term wind power forecasting is critical for effective energy production planning, load balancing, and grid stability. This study develops a turbine-level forecasting framework with a 10-minute temporal resolution, integrating unified multi-turbine data, lag- and rolling-based feature engineering, trigonometric time encoding, and machine learning models including Ridge, Lasso, K-Nearest Neighbors, Random Forest, Extra Trees, and XGBoost. The pipeline also incorporates hyperparameter optimization, data scaling, and chronological data splitting to ensure robust model training and prevent overfitting. Model performance was assessed at both individual turbine and aggregated total power levels. All models demonstrated high predictive accuracy, with XGBoost exhibiting superior generalization, stability, and the lowest prediction errors. Linear models such as Ridge, Lasso, and Random Forest achieved comparable performance on the test set, indicating the presence of strong linear relationships within the dataset. Despite its superior accuracy, XGBoost required longer training time compared to linear models, highlighting the trade-off between computational efficiency and predictive performance. Lag and rolling statistical features provided consistent incremental improvements in forecast accuracy by capturing short-term temporal dependencies, with the lag + rolling configuration yielding the best performance in ablation experiments. Although wind speed exhibits clear diurnal and seasonal patterns — as evidenced by the site’s temporal wind profiles — explicit trigonometric time encodings contributed marginally at the 10-minute horizon, since the recent multivariate lag sequence already carries implicit temporal context; longer forecast horizons would likely benefit more from such encodings. Aggregating turbine-level predictions reduced normalised plant-level error to 5.27% (NRMSE) and enhanced forecast stability. Overall, the proposed framework demonstrates high accuracy, robust generalisation, and operational efficiency, offering a reproducible solution for real-world short-term wind power forecasting.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Short-term wind power forecasting with lag-based feature engineering: a turbine-level machine learning comparison

  • 期刊论文

  • 发表日期：

    2026年8月3日

查看全文
  • Emin Demir,
  • Sibel Akkaya Oy

摘要

Short-term wind power forecasting is critical for effective energy production planning, load balancing, and grid stability. This study develops a turbine-level forecasting framework with a 10-minute temporal resolution, integrating unified multi-turbine data, lag- and rolling-based feature engineering, trigonometric time encoding, and machine learning models including Ridge, Lasso, K-Nearest Neighbors, Random Forest, Extra Trees, and XGBoost. The pipeline also incorporates hyperparameter optimization, data scaling, and chronological data splitting to ensure robust model training and prevent overfitting. Model performance was assessed at both individual turbine and aggregated total power levels. All models demonstrated high predictive accuracy, with XGBoost exhibiting superior generalization, stability, and the lowest prediction errors. Linear models such as Ridge, Lasso, and Random Forest achieved comparable performance on the test set, indicating the presence of strong linear relationships within the dataset. Despite its superior accuracy, XGBoost required longer training time compared to linear models, highlighting the trade-off between computational efficiency and predictive performance. Lag and rolling statistical features provided consistent incremental improvements in forecast accuracy by capturing short-term temporal dependencies, with the lag + rolling configuration yielding the best performance in ablation experiments. Although wind speed exhibits clear diurnal and seasonal patterns — as evidenced by the site’s temporal wind profiles — explicit trigonometric time encodings contributed marginally at the 10-minute horizon, since the recent multivariate lag sequence already carries implicit temporal context; longer forecast horizons would likely benefit more from such encodings. Aggregating turbine-level predictions reduced normalised plant-level error to 5.27% (NRMSE) and enhanced forecast stability. Overall, the proposed framework demonstrates high accuracy, robust generalisation, and operational efficiency, offering a reproducible solution for real-world short-term wind power forecasting.