Reconstruction of Near‐Surface Hourly Wind Speed Data Sets During the Period 1959–2021 Over China

Abstract

Abstract The reconstruction models of hourly 10‐m wind speeds were developed for each of 2384 stations over China using stepwise regression, random forest and XGBoost machine learning approaches based on hourly observed and ERA5 reanalysis data from 2005 to 2021. The reconstruction procedures applied observed hourly data to reduce the systematic biases of reanalysis data sets. Furthermore, the procedures employed the past dynamically consistent states of the atmosphere simulated by ERA5 reanalysis techniques to reduce/remove wind‐observed data impacts of long‐term non‐meteorological condition changes over time (e.g., urbanization around weather stations), which provide homogenous hourly wind speed data sets from 1959 to 2021. The systematic errors of the models' simulations derived from three approaches are similarly small, with almost two orders of magnitude smaller than ERA5 original data sets. The systematic errors of reconstructed data sets derived from stepwise regression are similar to its simulation; however, the biases from two machine learning methods are even much greater than ERA5 original data sets. This result implies that machine learning methods are not suitable for such typical time‐series predictions using the previous‐hour wind speed as a predictor to reconstruct wind speed data for the next hour. Therefore, stepwise regression was selected to reconstruct hourly wind speed data sets, which have much better quality than ERA5 reanalysis data with the median correctness increased by >50% and the median rRMSE decreased by 25%–50%. Consequently, the reconstructed wind speed data sets have great potential to be useful for more precisely assessing the characteristics/trends of wind energy resources in the past 60 years over China.

Keywords

• hourly wind speed reconstruction
• machine learning
• stepwise regression
• homogenous wind speed data
• China