Calibration of ECMWF Precipitation Forecasts in a Dual-Resolution Ensemble

Uncertainties exist in numerical weather prediction (NWP) ensemble precipitation forecasts due to a combination of errors from the initial atmospheric conditions, the chaotic nature of the atmosphere, and because of the incorrect representation of physical processes by NWP models (e.g. due to the model spatial resolution). As ensemble precipitation forecasts are then subject to biases, numerous post-processing techniques have been developed to statistically correct them. These techniques usually compare a set of past model forecasts with observations to build relationships that can be used to correct the operational forecast.

As part of ECMWF’s contribution to the EU H2020 IMPREX (Improving PRedictions and management of hydrological Extremes) project, several experiments investigating the performance of non-calibrated and calibrated daily precipitation using two different ensemble forecast resolutions have been undertaken. A quantile mapping procedure was employed to calibrate both ensemble systems using 20 years of reforecasts and EFAS (European Flood Awareness System) 5 km gridded precipitation analysis for Europe; supplemental locations to increase the sample size were chosen based on the similarity of precipitation climatology and terrain. Five different ensemble combinations combining subsets of the 50-member operational ECMWF configuration (18 km grid) and an experimental 200-member low-resolution configuration (28 km grid) were tested. Each combination would have similar computational cost to the current operational ensemble. A weighting method is also applied to increase the range of precipitation values in the ensemble forecasts. In the calibration, it is also important to maintain the spatial consistency in the precipitation fields, as this is necessary for driving a hydrological model.

The verification of the five ensemble combinations (calibrated and raw) was undertaken with daily EFAS precipitation across Europe for June, July and August in 2016 at 1-10 day lead time; a further calibration test only using the current operational ensemble was also undertaken for autumn 2016. The CRPS, ROC, reliability, Brier Score, Quantile Score, and Relative Economic Value were evaluated for different 24-hour precipitation thresholds. The verification shows that the most skilful combination is 40 ensemble members from the operational configuration and 40 from the low-resolution configuration. In particular, the calibrated forecasts have better reliability. These results suggest that this set-up combines the advantage of the high-resolution forecast system with an improved representation of the forecast distribution, which is especially useful for longer lead times. Furthermore, the calibrated precipitation fields were used to drive the EFAS hydrological model and the initial results will be presented.