Scholarlink

<p>Data-driven approaches have emerged as a complementary strategy to physics-based models for learning unresolved processes and parameterization biases in reanalysis products. Most data-driven efforts have focused on improving reanalysis wind speeds at heights where observational data are routinely collected. In this study, data from networks of near-surface meteorological stations in Canada and the United States are combined with a smaller network to derive a theoretical wind-shear exponent. This combined dataset was used to train a machine learning model that incorporates the power-law into its loss function, enabling it to simultaneously learn a correction factor for ERA5 wind speeds at 100 m and predict the wind shear exponent. When evaluated using multi-height wind speed data from the Tall Tower Dataset (TALLDB), the new framework demonstrated significant improvements. It reduced the median absolute error and root mean squared error and improved the temporal variability of ERA5 wind speeds between 10 m and 100 m at test locations not used during training. At several TALLDB stations, the predicted shear exponent outperformed ERA5-derived values. Overall, this framework offers a practical solution for enhancing hub-height wind speed estimates from reanalysis data over large regions, relying solely on near-surface meteorological variables that are more widely available.</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

Improving reanalysis hub-height wind speeds and wind shear across large spatial domains using near-surface observational networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月4日

查看全文
  • Freddy Houndekindo,
  • Taha B.M.J. Ouarda

摘要

Data-driven approaches have emerged as a complementary strategy to physics-based models for learning unresolved processes and parameterization biases in reanalysis products. Most data-driven efforts have focused on improving reanalysis wind speeds at heights where observational data are routinely collected. In this study, data from networks of near-surface meteorological stations in Canada and the United States are combined with a smaller network to derive a theoretical wind-shear exponent. This combined dataset was used to train a machine learning model that incorporates the power-law into its loss function, enabling it to simultaneously learn a correction factor for ERA5 wind speeds at 100 m and predict the wind shear exponent. When evaluated using multi-height wind speed data from the Tall Tower Dataset (TALLDB), the new framework demonstrated significant improvements. It reduced the median absolute error and root mean squared error and improved the temporal variability of ERA5 wind speeds between 10 m and 100 m at test locations not used during training. At several TALLDB stations, the predicted shear exponent outperformed ERA5-derived values. Overall, this framework offers a practical solution for enhancing hub-height wind speed estimates from reanalysis data over large regions, relying solely on near-surface meteorological variables that are more widely available.