Scholarlink

<p>Wind turbine hub-height wind speed forecast is a key factor affecting the stable operation of power grids, but its high uncertainty and chaotic fluctuations make accurate wind speed forecast challenging. In this study, we design a three-level architecture of “base learner-feature selection-meta learner” and propose a stacking ensemble (Stacking) model to forecast 100&#xa0;m wind speeds at 216 wind farms across China. The model employs Categorical Boosting (CatBoost), Light Gradient Boosting Machine (LightGBM) and eXtreme Gradient Boosting (XGBoost) as base learners. CatBoost is used for feature selection and a fivefold cross-validation approach is implemented to train the base learner models. A weighted average is then used as the meta learner to integrate the outputs of the base models. To evaluate the performance of the proposed method, we compare the Stacking model with the base models and ECMWF-IFS. The results show that the Stacking model consistently outperforms both ECMWF-IFS and the three basic models across different regions and time periods. Specifically, compared to ECMWF-IFS, the Stacking model achieves reductions in root mean square error and mean absolute error of 18.3% and 19.4%, respectively. Further error decomposition analysis indicates that these improvements are generally contributed to reductions in both the bias and sequence components, with bias component reduction playing a more dominant role. Moreover, the stacking technique is particularly effective in correcting wind speeds in the 3 ~ 12&#xa0;m/s range, which is critical for improving the power efficiency of wind power generation.</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

Enhancing 100 m wind speed forecasts in China based on CatBoost feature selection and stacking ensemble learning

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • Chenxi Jin,
  • Ting Lei,
  • Yang Yang,
  • Chen Qi,
  • Cao Weihua

摘要

Wind turbine hub-height wind speed forecast is a key factor affecting the stable operation of power grids, but its high uncertainty and chaotic fluctuations make accurate wind speed forecast challenging. In this study, we design a three-level architecture of “base learner-feature selection-meta learner” and propose a stacking ensemble (Stacking) model to forecast 100 m wind speeds at 216 wind farms across China. The model employs Categorical Boosting (CatBoost), Light Gradient Boosting Machine (LightGBM) and eXtreme Gradient Boosting (XGBoost) as base learners. CatBoost is used for feature selection and a fivefold cross-validation approach is implemented to train the base learner models. A weighted average is then used as the meta learner to integrate the outputs of the base models. To evaluate the performance of the proposed method, we compare the Stacking model with the base models and ECMWF-IFS. The results show that the Stacking model consistently outperforms both ECMWF-IFS and the three basic models across different regions and time periods. Specifically, compared to ECMWF-IFS, the Stacking model achieves reductions in root mean square error and mean absolute error of 18.3% and 19.4%, respectively. Further error decomposition analysis indicates that these improvements are generally contributed to reductions in both the bias and sequence components, with bias component reduction playing a more dominant role. Moreover, the stacking technique is particularly effective in correcting wind speeds in the 3 ~ 12 m/s range, which is critical for improving the power efficiency of wind power generation.