Improving probabilistic forecasts of extreme wind and precipitation by statistical post-processing methods (Speakers: Maurice Schmeits and Kirien Whan, KNMI)

Speakers: Maurice Schmeits and Kirien Whan, KNMI

In statistical post-processing of numerical weather prediction (NWP) output observations of the predictand (i.e. the variable to be predicted) are linked to predictors from NWP output. In this way systematic forecast errors can be corrected and hence forecast skill is increased. We use two case studies to illustrate how statistical post-processing of the high-resolution (2.5 km) Harmonie (ensemble) output can improve probabilistic forecasts of (extreme) weather. Both studies compare more traditional parametric statistical post-processing methods with relatively new machine learning techniques. We highlight positive aspects and pitfalls of each approach. First, we show that statistical post-processing of wind speed forecasts from the Harmonie ensemble increases forecast skill up to thresholds of at least 20 m/s in winter. This can be used to improve KNMI’s wind warnings. Second, we generate areal probabilistic forecasts of the maximum hourly precipitation in 11 regions of the Netherlands. We use deterministic potential predictor variables (including Harmonie precipitation, other direct model output and numerous indices of atmospheric instability) to create skilful forecasts that can be used by KNMI’s forecasters, water boards and municipalities in their decision making processes. A maximum local hourly precipitation threshold of 30 mm/hour (a criterion in the code yellow warning for severe thunderstorms) is skilfully forecast in the afternoon period at short lead times.