Offline models for statistical post-processing of surface weather variables

Statistical post-processing based on machine learning (ML) methods aims to capture systematic forecasts errors, relying on information from various predictors. We explore the exclusive use of “offline” predictors for the bias correction and uncertainty estimation of 2m temperature and 10 m wind speed forecasts. Offline predictors are defined as predictors available before the start of the forecast-of-the-day. Offline predictors encompass model characteristics such as the model orography and the model vegetation cover as well as spatio-temporal markers such as the day of the year, the time of the day and the latitude. The resulting offline models are particularly simple to implement as no time-critical operations are involved. The benefits of offline models and performance compared with more complex approaches will be discussed.