SSEA- Statistical and machine learning based post-processing for high-resolution subseasonal ensemble predictions

Subseasonal predictions are gaining more and more attention and importance in many applications, e.g. agriculture or energy&consumption predictions. To bridge the gap between those two temporal horizons and their drivers is, however, a challenge. Several attempts have been made in recent years to improve the numerical weather predictions but they to come at a high computation cost resulting in coarse spatial resolutions.  In the past decade, significant advances were made in improving the S2S and seasonal prediction using mainly numerical weather prediction models (NWP) and in some cases climate models for generating the predictions. Recently, the application of these models in real time forecasting through the S2S Real-Time Pilot Initiative (Robbins et al., 2020) was evaluated and is ongoing. There are, however, drawbacks. Computational costs for performing one forecast cycle are high (RAM, storage, ensemble for uncertainty) and limit the spatial, and to some extent temporal, resolution which are currently roughly 1.5° in spatial and at most 6-hourly in temporal resolution. Both resolutions are not sufficient for small scale renewable production sites. To overcome this, post-processing can be applied using statistical and machine leraning methods.

In this study, statistical (EPISODES, GMOS, SAMOS) and machine learning methods (U-net, random forest) are used to downscale and post-process the coarse subseasonal ensemble predictions for temperature and precipitation. The domain in centred on Austria with a spatial resolution of 1 km  using the INCA analysis as target fields. Evaluation against INCA and point observations show the skills of all methods and highlight the need for additional downscaling.