Locally Extreme Precipitation in Models: Model Climatologies by Extreme Value Theory

Due to its critical importance in hydrology, civil engineering, urban planning, and other fields, great effort has been put forth towards attaining accurate estimates of precipitation thresholds corresponding to return periods from 1-year to hundreds, or in some cases, even thousands of years. These extremely rare, heavy rainfall events have enormous societal impacts, and accurate forecasts are critical for allowing appropriate preparedness to protect lives and property. However, almost all work to date uses observations to estimate these precipitation-frequency relationships in the true atmosphere. Due to inherent biases, approximations, and other discrepancies between true and simulated atmospheres, numerical weather prediction (NWP) models have inherent biases in quantitative precipitation forecasting (QPF), and these biases are often exacerbated in extreme cases. Thus, to optimally forecast for these extreme precipitation events, it is essential to be able to put available model guidance in an appropriate, model-based context.

To this end, a combination of reforecasts and longtime static or quasi-static NWP model guidance are used to diagnose climatologies of model QPF over the contiguous United States, with an emphasis on accurately quantifying the extreme (right) tail of the probability distribution with return periods of 2-100 years. This procedure is performed for three different numerical models across different spatial scales: 1) a global model, National Oceanic and Atmospheric Administration (NOAA)'s Second Generation Global Ensemble Forecast Reforecast (GEFS/R); 2) a regional, convection-parameterized model, a 12km horizontal grid spacing implementation of the Weather Research and Forecasting (WRF) Model run at Colorado State University (CSU-WRF); and 3) a convection allowing, 4km WRF model run by the National Severe Storms Laboratory (NSSL-WRF). Further, several methods are explored for distribution fitting to records of model QPF data, including the use of Annual Maximum Series (AMS), Partial Duration Series (PDS), Complete Daily Series (CDS), and fitting extreme value distributions (EVDs) after directly estimating return period thresholds based on forecast verification. The utility of fitted model climatologies for locally extreme precipitation forecasting is assessed via cross-validation within the model data record, and compared with forecast skill from directly comparing model QPF values with observationally derived return period thresholds from NOAA's Atlas 14 product. Preliminary results indicate value in this application of bias correction by quantile mapping, with the model precipitation climatologies serving as better predictors of observed return period exceedances than the use of observationally-derived thresholds.