Generating Skillful and Reliable Probabilistic Forecasts for Locally Extreme Rainfall

Historical model forecasts of heavy rainfall were compared with observations for several years at a variety of return period thresholds ranging between 2 and 1000 years, as determined by NOAA Atlas 14 data. This was conducted for several different global, regional, and convection-resolving models over the contiguous United States- including the Global Ensemble Forecast System (GEFS), the National Severe Storms Laboratory's Weather Research and Forecasting Model (NSSL-WRF), Colorado State University WRF (CSU-WRF), and High Resolution Rapid Refresh (HRRR) models- in order to gain understanding of individual model regional, seasonal, and regime-dependent biases with regard to heavy rainfall model forecasts in comparison with observed local climatologies. In so doing, numerous biases were detected in the seasonality, frequency, and spatial and temporal characteristics of heavy rain events both in comparison to observations and between models. In an attempt to improve the value of model QPF output, post-processing techniques were employed to account for the identified model precipitation biases. Several different parametric and non-parametric methods were employed to fit probability density functions (PDFs) to the historical record of model quantitative precipitation forecast (QPF) data, and in so doing, generate model QPF climatologies over the contiguous United States (CONUS). It was found that applying distribution fitting techniques to fit climatological PDFs for model QPF, and using these derived PDFs to determine appropriate model specific QPF return period thresholds generally yielded significantly improved error statistics at for all models and return periods compared with observationally-derived thresholds from NOAA's Atlas 14 dataset. From preliminary findings, computing and utilizing model precipitation climatologies appears to offer considerable value as a forecast tool for heavy rainfall and flood forecasting.