Statistical postprocessing of different variables for airports in Spain using machine learning

Abstract

The results of a deterministic calibration for the nonhydrostatic convection-permitting LAM-EPS AEMET-γSREPS are shown. LAM-EPS AEMET-γSREPS is a multiboundary condition, multimodel ensemble forecast system developed for Spain. Machine learning tools are used to calibrate the members of the ensemble. Machine learning (hereafter ML) has been considerably successful in many problems, and recent research suggests that meteorology and climatology are not an exception. These machine learning tools range from classical statistical methods to contemporary successful and powerful methods such as kernels and neural networks. The calibration has been done for airports located in many regions of Spain, representing different climatic conditions. The variables to be calibrated are the 2-meter temperature, the 10-meter wind speed, and the precipitation in 24 hours. Classical statistical methods perform very well with the temperature and the wind speed; the precipitation is a subtler case: it seems there is not a general rule, and for each point, a decision has to be taken of what method (if any) improves the direct output of the model, but even recognizing this, a slight improvement can be shown with ML methods for the precipitation.