Imputation of missing lidar wind speed data: a comparative study of statistical and machine learning approaches

Wind measurements conducted with lidar (light detection and ranging) devices often suffer from missing wind speed data, particularly at large heights. This study addresses this challenge by investigating five different gap filling models applied on 26 measurement locations in Germany, where wind speed was recorded at 10-minute intervals over the course of approximately one year. At first, a comprehensive analysis of the distribution of the missing values was conducted. Secondly, based on these statistics, artificial missing values were created and filled using five statistical and machine learning models. Two of those were developed based on the power law principle which is commonly used for extrapolation of wind speeds. Another two were machine learning methods, namely random forest and support vector regression. Additionally, multivariate linear regression was implemented and tested. For all models, the measured wind speeds at heights below the target height served as input data. This maximized the recovery of missing values up to 97.8%. The results finally showed that the multivariate linear regression approach outperformed the other techniques, achieving both highest accuracy and reliability (Mean Percentage Bias of 0.10%, RMSE of 0.42 m/s).