A Probabilistic Forecast Approach for Daily Precipitation Totals

Abstract

Commonly, postprocessing techniques are employed to calibrate a model forecast. Here, a probabilistic postprocessor is presented that provides calibrated probability and quantile forecasts of precipitation on the local scale. The forecasts are based on large-scale circulation patterns of the 12-h forecast from the NCEP high-resolution Global Forecast System (GFS). The censored quantile regression is used to estimate selected quantiles of the precipitation amount and the probability of the occurrence of precipitation. The approach accounts for the mixed discrete-continuous character of daily precipitation totals. The forecasts are verified using a new verification score for quantile forecasts, namely the censored quantile verification (CQV) score. The forecast approach is as follows: first, a canonical correlation is employed to correct systematic deviations in the GFS large-scale patterns compared with the NCEP?NCAR reanalysis or the 40-yr ECMWF Re-Analysis (ERA-40). Second, the statistical quantile model between the large-scale circulation and the local precipitation quantile is derived using NCEP and ERA-40 reanalysis data. Then, the statistical quantile model is applied to 12-h forecasts provided by the GFS forecast system. The probabilistic forecasts are reliable and the relative gain in performance of the quantile as well as the probability forecasts compared to the climatological forecasts range between 20% and 50%. The importance of the various parts of the postprocessing is assessed, and the performance is compared to forecasts based on the direct precipitation output from the ECMWF forecast system.