Post-processing and high-resolution downscaling of subseasonal ensemble forecasts with focus on renewables using statistics and machine learning

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.

 

In renewable energy applications, these time scales are getting more important as they can adapt their resource management strategies based on predictions of possible load/heating and cooling demand via anomalies to temperature, wind, precipitation amount, effects on the markets can be better estimated for trading, and scheduling of maintenance works. Thus, at least higher spatial resolutions could help improving the management and planning of these tasks.

 

Within the SSSEA project (SubSeasonal to Seasonal Ensemble prediction and Application), in project phase I, different methods for post-processing and downscaling the S2S challenge data to 1 km resolution and actual values instead of anomalies were implemented. The statistical methods EPISODES, GMOS, and SAMOS were adapted to be able to work with different time scales compared to their initial implementations (seasonal/hourly) and machine learning based methods were developed from scratch using a feed forward neural network, a Unet-based model, and a Random Forest. Temperature, precipitation, and in the currently ongoing project phase II, the wind components of the ECMWF S2S model were downscaled to daily analysis fields based on the INCA model.

 

For wind energy applications, specific indices were developed and applied to the downscaled results.  Verification and definition of suitable metrics is crucial to assess the skills of the different methodologies considered and a wide range of aspects and metrics were considered. Results on both grid and station verification for appx. 250 sites in Austria across nearly all altitude ranges show that all post-processing models are able to improve the ECMWF ensemble forecasts for the parameters considered, though, depending on lead time and season, differences in the models’ skill are visible. Furthermore, for most of the initial times and leadtimes in the forecast/testing period of 2020 we were able to outperform also the climatology. To assess the impact on renewable energy production, different indices were derived and evaluated with focus on wind energy and hydrology in project phases I and II. Results of SSSEA show clearly the added value of the post-processed and downscaled subseasonal predictions for both parameters and specified indices.