Recalibrating DWD’s operational climate predictions: towards a user-oriented seamless climate service

The German Meteorological Service's (Deutscher Wetterdienst DWD) climate predictions website  (www.dwd.de/climatepredictions) offers a centralized platform for accessing post-processed climate predictions, including subseasonal forecasts from ECMWF's IFS and seasonal and decadal predictions from the German climate prediction system. The website design was developed in collaboration with various sectors to ensure uniformity across all time frames, and users can view maps, tables, and time series of ensemble mean and probabilistic predictions in combination with their skill. The available data covers weekly, 3-month, 1-year, and 5-year temperature means, precipitation sums and soil moisture for the world, Europe, Germany, and particular German regions. To achieve high spatial resolution, the DWD used the statistical downscaling method EPISODES. Moreover, within the BMBF project KIMoDIs (AI-based monitoring, data management and information system for coupled forecasting and early warning of low groundwater levels and salinisation) the DWD provides climate prediction data of further hydrological variables (e.g. relative humidity) with corresponding prediction skill on a regional scale.

However, all predictions on these time scales can suffer from inherent systematic errors, which can impact their usefulness. To address these issues, the recalibration method DeFoReSt was applied to decadal predictions, using a combination of 3^rd order polynomials in lead and start time, along with a boosting model selection approach. This approach addresses lead-time dependent systematic errors, such as drift, as well as inaccuracies in representing long-term changes and variability.

This study highlights the improved accuracy of the recalibration approach on decadal predictions due to an increased polynomial order compared to the original approach, and its different impact on global and regional scales. It also explores the feasibility of transferring this approach to predictions with shorter time horizons of the provided variables.