Operational postprocessing at MeteoSwiss: lessons learned

Local forecasts are the most viewed product on the MeteoSwiss mobile app and website. In the past year, MeteoSwiss has integrated statistical postprocessing into its automated forecasting system for medium-range, local forecasts. The postprocessing takes advantage of limited area and global NWP ensemble systems with different forecast horizons to provide seamless probabilistic hourly predictions up to ten days ahead. Postprocessing allows us to objectively optimize forecast quality in a targeted manner, resulting in local probabilistic forecasts that are more skilful and better calibrated than the legacy system. In the process of introducing statistical postprocessing into operations, we have identified a few key challenges that we would like to share. One key challenge relates to predictability. We aimed at providing high resolution local forecasts; to achieve interpretable and useful forecasts for end users, however, postprocessing of lower-resolution features with higher signal-to-noise ratio is preferrable. This can be achieved for example by reducing the spatio-temporal granularity of the forecasts with increasing forecast lead time. Another aspect that proved challenging is the diversity of requirements from our end users. Prior to introducing new data-driven forecasts we evaluate the quality of the forecasts with a range of metrics to ensure that key requirements are met. In the future, we envisage using multi-objective optimization to target different aspects of forecast quality already during training. Generating realistic scenarios (ensemble members) of postprocessed forecasts is currently performed post-hoc using ensemble copula coupling. Using postprocessing methods that are ensemble-native would allow us to jointly optimize quality for different spatio-temporal aggregations while still providing high-resolution local forecasts as an end product. In addition to the scientific challenges outlined above, we will also address the operational challenges related to the robustness of the system, to the monitoring, and to the communication that were identified.