Scholarlink

Short term wind speed forecasting is vital for efficient wind farm operations, grid stability, and renewable energy integration. Yet, many supervisory control and data (SCADA) based studies risk overstating accuracy due to temporal leakage and weak baseline comparisons. This study develops a leakage safe machine learning pipeline using SCADA data from a utility scale wind turbine in Turkey. Inputs were restricted to information available at time t, with engineered features including lagged wind speed, trigonometric encodings of wind direction, and rolling statistics. Three ensemble models; Extra Trees, Random Forest, and XGBoost, were benchmarked against persistence using blocked time series cross validation and an out of time test holdout. Results show modest but consistent gains: at 10 min, Random Forest slightly performed better than persistence (Mean Absolute Error (MAE) = 0.555 m/s vs. 0.556 m/s), while at 30 and 60 min, Extra Trees achieved the best performance, reducing MAE by 1.7 and 1.9%, respectively. These findings underscore persistence as a strong short-term predictor but highlight the value of leakage safe ensemble methods for longer horizons.

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

Leakage Safe Machine Learning for Short Term Wind Speed Forecasting from SCADA

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Muhammad Arslan Ul Haq,
  • Kashif Ali,
  • Israf Javed,
  • Murat Fahrioglu

摘要

Short term wind speed forecasting is vital for efficient wind farm operations, grid stability, and renewable energy integration. Yet, many supervisory control and data (SCADA) based studies risk overstating accuracy due to temporal leakage and weak baseline comparisons. This study develops a leakage safe machine learning pipeline using SCADA data from a utility scale wind turbine in Turkey. Inputs were restricted to information available at time t, with engineered features including lagged wind speed, trigonometric encodings of wind direction, and rolling statistics. Three ensemble models; Extra Trees, Random Forest, and XGBoost, were benchmarked against persistence using blocked time series cross validation and an out of time test holdout. Results show modest but consistent gains: at 10 min, Random Forest slightly performed better than persistence (Mean Absolute Error (MAE) = 0.555 m/s vs. 0.556 m/s), while at 30 and 60 min, Extra Trees achieved the best performance, reducing MAE by 1.7 and 1.9%, respectively. These findings underscore persistence as a strong short-term predictor but highlight the value of leakage safe ensemble methods for longer horizons.